text
stringlengths 4
2.78M
| meta
dict |
|---|---|
---
author:
- Moussa Amrani
- Levi Lúcio
- Adrien Bibal
bibliography:
- './arXiv2018.bib'
title: |
ML + FV = ?\
A Survey on the Application of Machine Learning to Formal Verification
---
<ccs2012> <concept> <concept\_id>10002944.10011123.10011675</concept\_id> <concept\_desc>General and reference Validation</concept\_desc> <concept\_significance>500</concept\_significance> </concept> <concept> <concept\_id>10002944.10011123.10011676</concept\_id> <concept\_desc>General and reference Verification</concept\_desc> <concept\_significance>500</concept\_significance> </concept> <concept> <concept\_id>10010147.10010257</concept\_id> <concept\_desc>Computing methodologies Machine learning</concept\_desc> <concept\_significance>500</concept\_significance> </concept> </ccs2012>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: '[We study the behaviour of Yang-Mills theory under the inclusion of gravity. In the weak-gravity limit, the running gauge coupling receives no contribution from the gravitational sector, if all symmetries are preserved. This holds true with and without cosmological constant. We also show that asymptotic freedom persists in general field-theory-based gravity scenarios including gravitational shielding as well as asymptotically safe gravity. ]{}'
author:
- 'Sarah Folkerts[^1]'
- 'Daniel F. Litim'
- 'Jan M. Pawlowski'
title: 'Asymptotic freedom of Yang-Mills theory with gravity'
---
Introduction
=============
Asymptotic freedom of Yang-Mills theories [@Gross:1973id; @Politzer:1973fx] – the weakening of the strong force at asymptotically short distances – is a central characteristics of the Standard Model. Within a renormalisation group language, asymptotic freedom is signaled by a trivial ultraviolet fixed point with vanishing Yang-Mills coupling. This pattern is at the root of the existence of Yang-Mills theory as an asymptotic perturbative series valid for high energies, i.e. energies much larger than the dynamical strong-coupling scale, $\Lambda_{\rm QCD}$, and makes the high energy behaviour accessible with weak coupling methods. It is well-known, however, that this feature can be destabilized once the couplings to matter degrees of freedom are taken into account, leading to constraints for the latter to ensure asymptotic freedom within the SM and its various extensions.
Less is known about the stability of asymptotic freedom under the inclusion of gravity. Recently, quantum gravitational corrections to the Yang-Mills coupling have been studied perturbatively to one-loop order in Newton’s coupling [@Robinson:2005fj; @Pietrykowski:2006xy; @Toms:2007sk; @Ebert:2007gf; @Tang:2008ah; @Toms:2010vy], by treating gravity as an effective theory amended by an ultraviolet cutoff of the order of the Planck scale [@Donoghue:1993eb] (see [@Deser:1974nb; @Deser:1974cz; @Deser:1974xq] for earlier work). The strength of this technique is that quantum gravitational fluctuations can be computed in the low energy regime without the knowledge of the underlying quantum theory for gravity. Asymptotic freedom remains intact in these studies and extensions thereof including compact extra dimensions [@Ebert:2008ux; @Wu:2008ph] and a cosmological constant [@Toms:2008dq]. These one-loop results have recently been confirmed beyond perturbation theory in [@Daum:2009dn].
The findings open up a number of farther reaching questions. The one-loop graviton contribution to the Yang-Mills $\beta$-function comes out regularisation- and gauge-fixing dependent, which raises the question about its universality. It has also been argued that the one-loop coefficient is not invariant under re-parameterisations in field space, and hence not a measurable quantity [@Ellis:2010rw]. For specific gauges or regularisations a vanishing one-loop coefficient has been found, and a computation beyond the one-loop level becomes mandatory. Also, the applicability of the effective theory is confined to the regime of weak gravity and strong coupling effects will become important and should be taken into account for energies approaching the Planck scale. Finally, at Planck-scale energies and beyond, an effective theory description is superseded by a fundamental theory for gravity, and the ultimate fate of asymptotic freedom for Yang-Mills theory then depends on the ultraviolet completion for gravity.
In this Letter, we study Yang-Mills theory coupled to gravity with the help of the functional renormalisation group [@Wetterich:1992yh; @Litim:1998nf; @Gies:2006wv; @Pawlowski:2005xe] (see [@Daum:2009dn] for a related study). This technique is based on the infinitesimal integrating-out of momentum modes from a path integral representation of quantum field theory by means of an infrared momentum cutoff. The renormalisation group flow links the fundamental theory with the corresponding quantum effective action at low energies. A particular strength of this method is its flexibility, allowing for perturbative and non-perturbative approximations. In the past, it has been applied to Yang-Mills theory [@Litim:1998nf; @Litim:2002ce; @Gies:2002af; @Gies:2006wv; @Pawlowski:2003hq; @Pawlowski:2005xe; @Reuter:1997gx; @Litim:1998qi; @Freire:2000bq; @Pawlowski:2001df; @Pawlowski:2003sk; @Litim:2002xm; @Litim:2002hj] and gravity [@Niedermaier:2006ns; @Percacci:2007sz; @Litim:2008tt; @Reuter:1996cp; @Souma:1999at; @Lauscher:2001ya; @Percacci:2002ie; @Litim:2003vp; @Fischer:2006fz; @Codello:2007bd; @Benedetti:2009gn] individually, both at weak and strong coupling, which makes it an ideal tool for the study of the coupled system [@Folkerts:Dipl2009].
In the weak gravity limit, we analyse the gravitational corrections to Yang Mills theory for general regularisations and backgrounds. Gauge and diffeomorphism invariance of the effective action is achieved using the background field method. We evaluate the background field independence of our results by extending earlier explicit results within gauge Yang Mills theory [@Litim:2002ce] to quantum gravity. In this general setting the gravitational contribution to the running gauge coupling is computed at one-loop. The result encompasses all previous studies, also including a cosmological constant. We also address asymptotic freedom in the limit where the graviton anomalous dimension can become large. Here, we are particularily interested in the asymptotic safety scenario for gravity [@Weinberg; @Niedermaier:2006ns; @Percacci:2007sz; @Litim:2008tt; @Reuter:1996cp; @Souma:1999at; @Lauscher:2001ya; @Percacci:2002ie; @Litim:2003vp; @Fischer:2006fz; @Codello:2007bd; @Benedetti:2009gn; @Niedermaier:2009zz]. We also evaluate asymptotic freedom in the scenario where quantum gravity is shielded by black hole formation [@Banks:1999gd; @Giddings:2001bu; @DvaliGomez; @Dvali:2010ue; @Percacci:2010af].
The outline of this Letter is as follows. We introduce our renormalisation group set-up (Sec. \[RG\]), and discuss gravitational corrections to Yang-Mills theory within a background field approach (Sec. \[AF\]). A detailed discussion of the background field dependence is given (Sec. \[BFD\]), amended by a complementary RG study for flat backgrounds (Sec. \[FB\]). Our results are compared with earlier findings in perturbation theory at the one-loop order (Sec. \[PT\]). Beyond one-loop, we study the leading gravitational corrections to Yang-Mills in the presence of a gravitational fixed point (Sec. \[GFP\]) and in the presence of a cosmological constant term (Sec. \[CC\]). We close with a brief discussion (Sec. \[sec:disc\]).
Renormalisation group {#RG}
=====================
The Functional Renormalisation Group is based on a momentum cutoff for the propagating degrees of freedom and describes the change of the scale-dependent effective action $\Gamma_k$ under an infinitesimal change of the cutoff scale $k$ [@Wetterich:1992yh; @Litim:1998nf; @Gies:2006wv; @Pawlowski:2005xe]. Thereby it interpolates between a microscopic action in the ultraviolet and the full quantum effective action in the infrared, where the cutoff is removed. In its modern formulation, the renormalisation group flow of $\Gamma_k$ with the logarithmic scale parameter $t=\ln k$ is given by Wetterich’s flow equation [@Wetterich:1992yh] $$\label{eq:FRG}
\partial_t\Gamma_k=
\012\Tr\0{1}{\Gamma_k^{(2)}+R_k}\,\partial_t R_k\,.$$ The trace stands for a sum over indices and a loop integration, and $R_k$ (not to be confused with the Ricci scalar $R(g_{\mu\nu})$) is an appropriately defined momentum cutoff at the momentum scale $q^2\approx k^2$. The effective action depends on the metric field $g$, the gauge field $A_\mu^a$ including possible Abelian factors, and the related ghosts $\eta,\bar \eta$ and $C,\bar C$ for metric field and gauge field respectively. In short we have $\Gamma_k=\Gamma_k
[\phi,\bar\phi]$, where the fields are put into a superfield $\phi=
(g,\eta,\bar\eta,A,C,\bar C)$. The action may also depend on specifically chosen, non-propagating background fields $\bar \phi$. In quantum gravity , the introduction of a background metric is necessary as one has to fix a gauge, or more generally, one has to choose a specific parameterisation of the configuration space. This fact is at the root of the problem of background independence of quantum gravity.
The flow equation is derived from the insertion of infrared cut-off terms in the path integral, $$\label{eq:Sk}
\Delta S_k[\phi,\bar\phi]=\012 \int d^d x\sqrt{\bar g}\,
(\phi-\bar \phi)_i\, R_k^{ij}[\bar\phi]\, (\phi-\bar \phi)_j\,,$$ where the regulator $R_k$ is a matrix in field space and may also couple different species of fields, in particular metric and gauge fields, and may depend on the background $\bar \phi$. Note that the background metric dependence in via $\sqrt{\bar g}$ is necessary to keep the one loop nature of the flow equation . Furthermore the contraction of the spin one and spin two components in $\phi$ also depend on the background metric.
The rhs. of the flow equation consists of one loop diagrams with full field-dependent propagators multiplied with the scale derivative of the regulator. This entails that for computing a $n$-loop effective action one has to insert the $n-1$-loop effective action on the rhs. If interested in the one loop effective action we therefore have to insert the classical action on the rhs. The structure of the flow equation is such that standard perturbation theory is recovered to all loop orders by iteration [@Litim:2002xm].
Gauge and diffeomorphism symmetry is encoded in the related Slavnov-Taylor identities [@Litim:1998nf; @Freire:2000bq; @Pawlowski:2005xe]. These symmetry identities are modified in the presence of the regulator term . In the vanishing cut-off limit, $k\to 0$, the modifications also vanish and we are left with the standard symmetry identities. Furthermore, there are combined transformations of both, the background $\bar \phi$ and the fluctuation $\phi$, which leaves the effective action invariant. It is for that reason that $$\label{eq:Gammainv}
\Gamma_k[\phi]=\Gamma_k[\phi,\phi]\,$$ is both, gauge and diffeomorphism invariant.
Asymptotic freedom and gravitational interactions {#AF}
=================================================
In the present work we consider an approximation to the effective action which reproduces the full one loop effective action but allows also to discuss properties beyond one loop. Such an approximation for $\Gamma_k[\phi]$ defined in is given by
$$\label{eq:EHk}
\Gamma_k[\phi,\bar\phi]=
\int d^dx \sqrt{\det g_{\mu\nu}}
\left[\0{Z_{N,k}}{16\pi G_N}\left(-R(g_{\mu\nu})+2\bar\Lambda_k\right)
+\0{Z_{A,k}}{4 \gYM^2} F^a_{\mu\nu}F_a^{\mu\nu}\right]
+S_{\rm gh}+S_{\rm gf}+\Delta\Gamma_k\,.$$
Here we have introduced the Einstein-Hilbert action and the classical Yang-Mills action as well as gauge fixing and ghost contributions $S_{\rm gh}$ and $S_{\rm gf}$. The gauge and diffeomorphism invariant action, , is defined as $\Gamma_k[\phi]=\Gamma_k[\phi,\phi]$ where the gauge fixing term and a part of $\Delta\Gamma_k$ vanishes. Due to quantum fluctuations, all couplings have become running couplings and depend on the infrared RG-scale $k$: the running Newton coupling $G_{N,k}=G_{N}/Z_{N,k}$, the running cosmological constant $\Lambda_k$, and the running Yang-Mills coupling $\gYMk=\gYM^2/Z_{A,k}$. Further quantum corrections such as higher order curvature invariants and further matter-gravity interaction terms can be generated as well and are contained in $\Delta\Gamma_k$.
In the present work we discuss the running of the fundamental dimensionless couplings of the theory, the Yang-Mills coupling and the dimensionless Newton coupling, $$\label{eq:gN}
\gNk=\frac{k^2 G_N}{Z_{N,k}}\,.$$ The running is encoded in the respective $\beta$-functions ($\beta_X\equiv k\partial_k X$) as $$\begin{aligned}
\label{eq:betaS}
\beta_{\gYM}=& \012 \eta_A\, \gYMk\,\qquad\qquad & {\rm with} \qquad
\eta_A=-\0{\partial_t Z_{A,k}}{Z_{A,k}}\,,\\
\label{eq:betaN}
\beta_{\gN}= &(2+\eta_N) \,\gNk\,\qquad
&{\rm with} \qquad \eta_N=-\0{\partial_t Z_{N,k}}{Z_{N,k}}\,. \end{aligned}$$ The above equations , reflect the fact that in a background field approach the running couplings run inversely to the wave-function renormalisations $Z$ of the respective fields. The running of the $Z$’s splits into a pure Yang-Mills part (only Yang-Mills loops/fluctuations) and a gravity part that also contains internal graviton lines, $$\label{eq:split}
\eta_A=\eta_{A,\rm YM}+\eta_{A,\rm grav}\,,\qquad \qquad \eta_N
=\eta_{N,\rm grav}+\eta_{N,\rm YM}\,.$$ In the present work we are specifically interested in the sign $\eta_{A,\rm grav}$, since $\eta_{A,\rm YM}$ is proportional to the Yang-Mills coupling and tends to zero in the UV if asymptotic freedom is present. Hence, for $\eta_{A,\rm grav}<0$ at vanishing YM-coupling, asymptotic freedom is supported by quantum gravity corrections, whereas for $\eta_{A,\rm grav}>0$ it is spoiled.
So far we have not specified the regulator $R_k$. Indeed the full effective action at $R_k\equiv 0$ (for $k=0$) cannot depend on this choice, however at finite $k\neq 0$ with $R_k$ there will be a $k$-dependence. In the following we shall exploit this freedom to disentangle gravity and gauge theory at one loop. To that end we use a specific class of regulators which respects the renormalisation group scalings of the underlying physical theory at $k=0$, $$\begin{aligned}
\label{eq:Rk}
R_k[\bar\phi] =\Gamma_k^{(2)}[\bar\phi]\, r[\bar\phi]\,.\end{aligned}$$ The dimensionless regulator functions $r^{ij}$ are demanded to scale trivially under the RG-scalings of the underlying theory, and $$\label{eq:Gamma2}
\Gamma_{k,ij}^{(2)}[\phi,\bar\phi]
= \0{\delta^2 \Gamma_k[\phi,\bar\phi]}{\delta\phi^i\delta\phi^j}$$ are the second derivatives of the effective action w.r.t. the full fields $\phi$. Regulators of the form are called RG-adjusted [@Pawlowski:2001df; @Pawlowski:2005xe] or spectrally adjusted [@Gies:2002af], and scale like (inverse) two-point functions. With this property the RG-invariance and scaling properties of the underlying theory carry over to that of the regularised theory [@Pawlowski:2005xe], and facilitate the study and interpretation of $\beta$-functions w.r.t. to the $t$-scaling. The dimensionless regulator functions $r^{ij}$ depend on covariant Laplacians in the presence of the background $\bar\phi$. We specify $$\begin{aligned}
\label{eq:r}
r^{gg}&=& r^{gg}(-\Delta_{\bar g} /k^2)
\nonumber\\
r^{\bar\eta\eta}&=&- r^{\eta\bar\eta}=r^{\bar\eta\eta}(-\Delta_{\bar g} /k^2)
\nonumber \\
r^{AA} &=&r^{AA}(-\Delta_{\bar g}(A)/k^2)\nonumber \\
r^{\bar C C}&=&- r^{C\bar C}=r^{\bar C C}(-\Delta_{\bar g}(A) /k^2)\,.
\end{aligned}$$ All other components vanish. Regulator functions $r$ with are diagonal. They only connect the metric and gauge field with themselves and gosts with anti-ghosts. Note that this is the minimal form of regulator functions that suppress all infrared (or ultraviolet) propagation of all fields. Dropping one of the components in leads to full quantum propagation of the related field. Hence, the above choice naturally disentangles the Yang-Mills and gravitational sectors, and is therefore well-suited for our analysis.
Finally we are interested in the RG-scaling (w.r.t. $k$) of the couplings of the background gauge field $\bar A$ and the background metric field $\bar g$. Inserting the regulator in the flow and setting $\phi=\bar\phi$ we arrive at [@Pawlowski:2001df; @Litim:2002xm; @Litim:2002ce; @Litim:2002hj; @Pawlowski:2003sk; @Pawlowski:2005xe; @Gies:2002af] $$\label{eq:FRGadjusted}
\partial_t \Gamma_k[\phi]= \012 \Tr\, \0{1}{1+r[\phi]}\partial_t r[\phi]+
\Tr\, {\partial_t \Gamma_k^{(2)}[\phi,\phi]}\0{1}{\Gamma_k^{(2)}[\phi,\phi]}
\0{r[\phi]}{1+r[\phi]}\,,$$ where $\Gamma_k[\phi]=\Gamma_k[\phi,\phi]$, see . The above RG flow is the key equation in the present work. The first term on the rhs. consists out of all one loop contributions to the flow. The second term has the form of an RG improvement. It consists out of all terms beyond one loop, including all non-perturbative contributions. Note that it vanishes at one loop.
We observe that for the cutoff choice with diagonal $r$ in , the one loop flow disentangles. It is the sum of the pure gravity flow driven by $r^{gg}$ and $r^{\bar\eta\eta}$, and a pure gauge flow driven by $r^{AA}$ and $r^{\bar C C}$. Cross-terms leading to gravity loops for i.e. the gauge field propagators or gauge field loops for the graviton propagators are absent. The full one-loop effective action is then given by $$\label{eq:oneloop}
\Gamma[\phi]=\Gamma_\Lambda[\phi]+\012
\int_\Lambda^0 \0{dk}{k}\Tr\, \0{1}{1+r[\phi]}\partial_t r[\phi]\,,$$ where $\Gamma[\phi]=\Gamma_{k=0} [\phi]$. It follows from that for the class of cutoff functions the full $\beta$-function of the background gauge coupling, $\beta_{\gYM}$, in the presence of gravitational fluctuations is given by the standard result without gravity $$\begin{aligned}
\label{eq:nogravity}
\beta_{\gYM}=\beta_{\gYM,\rm YM}=-\012 \eta_{A,\rm YM}\end{aligned}$$ up to corrections starting at the 2-loop level. Beyond one loop, the potential structural disentanglement no longer holds true: $\partial_t
\Gamma_k^{(2)}/\Gamma_k^{(2)}$ is not diagonal and hence entangles Yang-Mills and gravity sector.
Background field dependence {#BFD}
===========================
The functional renormalisation group approach of Sec. III allows us to classify and understand the scheme dependence of the $\beta$-functions of a general matter-gravity system. In particular this allows us to explain the scheme dependences observed so far for the Yang-Mills – gravity theory, as well as to extract the physics stored in the $\beta$-functions, $i.e.$ the UV stability of the coupled system. We note in this context that the functional RG set-up also covers the standard perturbative regularisation schemes such as momentum cutoff, dimensional, Pauli-Villars, or heat kernel regularisation.
Scheme dependences of $\beta$-functions in the background field approach have already been studied for general field theories [@Litim:2002ce; @Pawlowski:2005xe; @Pawlowski:2001df; @Pawlowski:2003sk]. In [@Litim:2002ce] it has been shown that the YM-coupling receives contributions from the background field dependence of the regularisation. These terms even can alter the one-loop $\beta$-functions. It has also been shown explicitly how to extract the standard one-loop $\beta$-function with the help of an equation which tracks the background field dependence of the regularisation. In YM theory, the background field approach is a convenient choice. In gravity, however, it becomes a necessity. There, the background field dependence of the regularisation is encoded in $$\label{eq:fullback}
\int \0{1}{\sqrt{\bar g} }\0{\delta \sqrt{\bar g}R_k}{\delta \bar\phi}
\0{\delta \Gamma_k[\phi,\bar\phi]}
{\delta R_k}= \012 \Tr\, \0{1}{\Gamma_k^{(2)}[\phi,\bar\phi]+
R_k[\bar\phi]}
\0{1}{\sqrt{\bar g} }\0{\delta R_k[\bar\phi]}{\delta \bar\phi}\,.$$ It is left to evaluate whether the background field dependence affects the one-loop result for the background coupling as suggested by the above considerations. For the given cutoff functions , , we derive the regulator-induced background field dependence , $$\label{eq:back}
\int \0{1}{\sqrt{\bar g} }\0{\delta \sqrt{\bar g}R_k}{\delta \phi}
\0{\delta \Gamma_k[\phi,\phi]}
{\delta R_k}= \012 \Tr\, \0{1}{1+r[\phi]}
\0{\delta r}{\delta \phi}+\,
\Tr\, \0{1}{\sqrt{g} }\0{\delta \sqrt{\bar g}
\Gamma_k^{(2)}}{\delta \bar\phi}[\phi,\phi]\0{1}
{\Gamma_k^{(2)}[\phi,\phi]}
\0{r[\phi]}{1+r[\phi]}\,,$$ already evaluated at $\phi=\bar\phi$. Note that the structure of resembles that of the flow equation . The first term is purely diagonal and one loop, whereas the second term contains the non-perturbative contributions and entangles gravity and Yang-Mills due to $(\delta
\Gamma_k^{(2)}/\delta \bar \phi)/\Gamma_k^{(2)}$. In contradistinction to the second term in , it also contributes at one loop. In particular, contributions proportional to $\delta
\Gamma_k^{(2)}/\delta \bar \phi$ in the gravity sector produce terms $$\label{eq:bargterms}
\propto\frac{\delta}{\delta\phi} \int d^4 x \sqrt{g}\, \tr\, F_{\mu\nu} F^{\mu\nu}\,,$$ if evaluated at $\phi=\bar\phi$. Therefore we conclude that in general graviton contributions to the regularisation-induced background dependence of the effective action are created. We have already argued that for the regularisation schemes diagonal in $r$ defined in , the graviton-contribution to the $\beta$-function of the background field defined via vanishes. Note that this does not hold for the $\beta$-functions of the fluctuations. However, the $\beta$-function of the background field also contains unphysical contributions from the field-dependence of the regulators. These terms are given by the $t$-derivative of the $F^2$-contribution of the right hand side in . While the $t$-derivative of the $F^2$-contribution of the first term vanishes, this is not so for the one-loop contribution of the second term.
In turn, regularisations which induce a mixing between gravity and Yang-Mills fields are expected to give a non-vanishing contribution already at one-loop for all $\beta$-functions. The above arguments entail that the background $\beta$-functions of the coupled YM-gravity system are scheme-dependent even at one-loop. Note, however, that the UV stability of the YM-gravity system is solely controlled by the signs of the $\beta$-functions of the fluctuation fields.
In conclusion, the following picture has emerged. UV stability of the YM sector relates to the (negative) sign of the $\beta$-function for the fluctuating gluon. This information can be extracted directly from the correlation functions of the fluctuating gluons. Alternatively, one can exploit the fact that the regularisation-independent part of the $\beta$-function of the background coupling carries the same sign. In general, the second option requires the use of .
Flat backgrounds {#FB}
================
The results of the last section entail that a computation of the graviton contribution to the Yang-Mills $\beta$-function requires the distinction between fluctating graviton and background graviton. This is possible if the computation is done in a flat background within a standard vertex expansion. Here we put forward a computation of the YM $\beta$-function, which differs from the previous studies in two aspects [@Folkerts:Dipl2009]: Firstly, the analysis is performed for trivial backgrounds $\bar A=0$ and $\bar
g_{\mu\nu}=\eta_{\mu\nu}$. Secondly, we employ classes of momentum cutoffs with a tensorial structure different from . The result complements our previous findings and allows to make close contact with all previous 1-loop studies [@Ebert:2007gf; @Toms:2008dq; @Toms:2007sk; @Robinson:2005fj] as well as the flow study [@Daum:2009dn].
This also allows us to make use of the fact that loop contributions of fluctuating gravitons to the correlation functions of the fluctuating gluons and to the background gluons agree: the vertices of the two fields are derived only from the classical YM-action and not from the ghost and gauge fixing terms. The classical YM-action only depends on $\phi= \bar\phi+(\phi-\bar\phi)$ and hence derivatives w.r.t. the background fields $\bar\phi$ and w.r.t. the fluctuation fields $(\phi-\bar\phi)$ agree. In other words, the two possibilities of how to compute the graviton contribution to the $\beta$-function of the fluctuating gluon agree. We emphasise that this does not hold for the gluon contribution to the $\beta$-function of the fluctuating graviton.
Our Ansatz for the effective action is dictated by asymptotic freedom: a good approximation for Yang-Mills theory at high energies is given by the classical $F^2$ operator plus a gauge fixing term, both equipped with a running wave function renormalisation ${Z_{A,k}}$ and a classical scale-independent ghost action. The interaction of the gauge bosons with gravity is induced by the metric appearing in the spacetime integral measure which is now a quantum field. Also adding the Einstein-Hilbert action, we arrive at $$\begin{aligned}
\label{truncation for ymwgravity}
\Gamma_k[g,A;\bar{g},\bar{A}]
&=&\frac{Z_{N}}{16\pi{G}_N}\int d^dx\sqrt{g}\left(-R+2\bar{\lambda}_k\right)
+\frac{Z_{A}}{4}\int d^dx\sqrt{g}g^{\mu\rho}g^{\nu\sigma}F_{\mu\nu}^aF_{\rho\sigma}^a
\nonumber\\
&+&\frac{Z_{N}}{2\alpha}\int d^dx\sqrt{\bar{g}}\bar{g}^{\mu\nu}L_\mu L_\nu\
+\frac{Z_{A}}{2\xi}\int d^dx\sqrt{\bar{g}}(\bar{g}^{\mu\nu}
\bar{D}^{a b}_\mu (A-\bar{A})^b_\nu)^2+S_{\mathrm{gh}}\,.\quad\end{aligned}$$ Note that the $k$-dependence of $Z_N$ and $Z_A$ is not indicated explicitly. In , we have a general linear gauge condition $L_\mu$ with gauge parameter $\alpha$ for the graviton, and a background field gauge with parameter $\xi$ for the gluon. In general, both gauge fixing parameters are $k$-dependent, except for vanishing values. The latter constitutes a fixed point of the flow, [@Litim:1998qi]. $S_\mathrm{gh}$ consists of the sum of the classical ghost action for gravity and the one for Yang-Mills theory. $S_\mathrm{gh}$ neither contributes to the graviton-induced corrections to the running of the gluon coupling, nor to the gluon-induced running of the gravitational coupling. There is a simple reason for this. The gravitational ghost term, by construction, does not contain any gauge fields. The Yang-Mills ghost term only couples to the background metric field, eg. $S_{\rm YM, gh}=\int
d^dx\sqrt{\bar{g}}\,\bar{g}^{\mu\nu}\,\bar{c}^a\,\bar{D}^{a b}_\mu\,
D^{b c}_\nu\, c^c$ in a covariant gauge. Since ${\Gamma^{(2)}}$ is obtained by differentiation with respect to the dynamical fields, the gauge fixing terms do not generate interaction terms between gravity and the gauge field ghosts.
Hence the key input in the flow equation is the full propagator, $1/({\Gamma^{(2)}}_k+R_k)$ for the gauge field $A$ and the metric field $g$. We introduce $$\label{matrices for gamma2 and rk}
{\Gamma^{(2)}}_k=
\begin{pmatrix}
{\Gamma^{(2)}}_{AA}&&{\Gamma^{(2)}}_{Ag}\\ \\
{\Gamma^{(2)}}_{gA}&&{\Gamma^{(2)}}_{gg}\end{pmatrix}
\qquad\mathrm{and}\qquad R_k=
\begin{pmatrix} R_{A}&0\\ 0&R_{g}\end{pmatrix},$$ with the convention ${\Gamma^{(2)}}_{\varphi_i\varphi_j}\equiv
\frac{\delta^2\Gamma_k}{\delta\varphi_i\delta\varphi_j}$ for $\varphi=\{A,g\}$. Note that we have chosen a regulator $R_k$ which is diagonal. An important observation is that the off-diagonal terms of ${\Gamma^{(2)}}$ in vanish for $A=\bar{A}=0$ and $g=\bar{g}=\eta$ since the gauge field part in our ansatz involves at least two gauge fields. Consequently, the inverted matrix reads $$\label{prop zero fields}
\displaystyle \frac{1}{\Gamma^{(2)}_k+R_k}{\mathop{\bigg\vert_{A=\bar{A}=0}}_{g=\bar{g}=\eta}}=\begin{pmatrix}
\displaystyle \frac{1}{{\Gamma^{(2)}}_{AA}+R_{AA}}&&0\\\\
0&&
\displaystyle \frac{1}{{\Gamma^{(2)}}_{gg}+R_{g}}\end{pmatrix}.$$ We specify the regulator functions and their tensor structures, $$\begin{aligned}
\label{general form of regulators}
R_{A}&=&Z_A\;{T}_A(p)\; p^2\; r_A(p^2/k^2)\,,\\
\label{Rgrav}
R_{g}&=&\frac{Z_N}{32 \pi G_N}\;\frac{1}{2}\,{T}_g(p)\; p^2\; r_g(p^2/k^2)\,.\end{aligned}$$ Here $r(p^2/k^2)$ denotes the shape for the scalar part of the momentum cutoff, which can range between $0\leq r(y)\leq \infty$ for $\infty\ge y\ge 0$. Both regulators are dressed with the appropriate wave function renormalisation factor to ensure that $R$ displays the same RG scaling as $\Gamma^{(2)}$. The tensor structures ${T}_A$ and ${T}_g$ are chosen as $$\begin{aligned}
\label{cutoff gluon}
T^A_{\mu\nu}\hspace{-1mm}&=&\hspace{-1mm}
\eta_{\mu\nu}-(1-\frac{1}{\xi})\frac{p_\mu p_\nu}{p^2}\,,
\\
\label{cutoff grav}
T^g_{(\alpha\beta)(\gamma\delta)}\hspace{-1mm}&=&\hspace{-1mm}
\frac{\alpha-1}{\alpha}\left[
2\left(\eta_{\gamma\delta}\frac{p_\alpha p_\beta}{p^2}
+\eta_{\alpha\beta}\frac{p_\gamma p_\delta}{p^2}\right)
-\eta_{\beta\gamma}\frac{p_\alpha p_\delta}{p^2}-\eta_{\beta\delta}
\frac{p_\alpha p_\gamma}{p^2} -\eta_{\alpha\gamma}
\frac{p_\beta p_\delta}{p^2}-\eta_{\alpha\delta}
\frac{p_\beta p_\gamma}{p^2}\right]\nonumber\\
&&
+\left[
\eta_{\alpha\gamma}\eta_{\beta\delta}+\eta_{\alpha\delta}\eta_{\beta\gamma}
+\frac{1-2\alpha}{\alpha}\eta_{\alpha\beta}\eta_{\gamma\delta}\right]\,,\end{aligned}$$ with $T^g_{(\alpha\beta)(\gamma\delta)}=T^g_{(\gamma\delta)(\alpha\beta)}$, and $T_A$ diagonal in colour space. We extract the graviton contributions to the running for the YM coupling from the graviton contributions to the running of the YM propagator. The relevant diagrams on the rhs of the flow equation are $$\begin{aligned}
\label{RHS gluon flow}
\begin{minipage}[c]{2.5cm}
\includegraphics[width=2.5cm]{tadpolefinal.eps}
\end{minipage}+
\begin{minipage}[c]{3.6cm}
\includegraphics[width=3.6cm]{gluonselfenergyRAwomom.eps}
\end{minipage}
&&\!\!\!\!
+
\begin{minipage}[c]{3.6cm}
\includegraphics[width=3.6cm]{gluonselfenergyRgwomom.eps}
\end{minipage}\end{aligned}$$ where double lines denote the full graviton propagator, curly lines the full gluon propagator, full dots the full vertices, and the crossed circle the appropriate regulator insertion $\partial_t R_k$.
Now we come to an important observation. There is a kinematical identity which links the tadpole diagram, the first term in , to the other two self energy terms in . This identity is expressed diagramatically in : It states that the contributions to the tree level gluon-gluon-graviton-graviton amplitude from the 4-point vertex and from the 3-point vertices are proportional to each other when averaged over their angular dependencies. This holds for arbitrary gauge fixing parameter $\alpha$, and in the presence of a cosmological constant. This result can be written as $$\label{eq:angularaverage}
\includegraphics[width=0.55\textwidth]{angularaverage.eps}$$ and holds for contractions of the open-ended graviton legs with any of the tensor structures $T_{\mu\nu\delta\lambda}$ of the graviton propagator itself. The brackets indicate that this equality is only valid after the integration over $\int d\Omega_p$.
The relation relates to the fact that the graviton-gluon vertices are derived from the classical kinetic term of Yang-Mills theory. This approximation to the effective action of Yang-Mills certainly holds at the Gaußian, asymptotically free fixed point of Yang-Mills. We conclude that it should also hold in the presence of gravity if the gravity contributions to Yang-Mills sustain asymptotic freedom. We shall use this selfconsistency argument later.
A general regularisation scheme may break explicitly or implicitly the relation as it modifies either the propagation or the dynamics of the theory, or both, by cut-off effects. For example, a regularisation scheme in momentum space only changes the propagator, whereas a cutoff in covariant momenta changes propagators and vertices, and might require even tree-level counter terms. In turn, a symmetry preserving regularisation maintains the symmetry relation at the quantum level.
In order to extract the flow equation for the gluon wavefunction renormalisation $\partial_t {Z}_{A}$ we project both sides of the flow equation onto the transversal part of the inverse gluon propagator with $\Pi_{\mu\nu}^T=(\eta_{\mu\nu}-\frac{q_\mu
q_\nu}{q^2})$. Equating the coefficients in front of $q^2$ we identify the gravitational contribution to the flow for the gluon wave function renormalisation as $$\label{eq:ZAlambda0}
\frac{\partial_t Z_A}{Z_A} =48 \pi^3 \frac{G_N }{Z_N}\int_0^\infty\frac{d p\:p}{(2\pi)^4}
\frac{(1+\alpha)}{(1+r_{g})}\left[\frac{
\partial_t(Z_A r_{A})}{Z_A(1+r_{A})^2}+\frac{\partial_t(Z_N r_{g})}{
Z_N(1+r_{A})(1+r_{g})}-\frac{\partial_t(Z_N r_{g})}{Z_N(1+r_{g})}\right].$$ In we have dropped terms proportional to the flow of the gravitiational gauge fixing parameter, $\partial_t\alpha$, its impact will be discussed later. Note also that there are no terms proportional to $\xi$ and $\partial_t \xi$ due to transversal projections related to the tensor structure of the graviton -gluon vertices. With the anomalous dimensions $\eta_A$ and $\eta_N$, see and respectively, we deduce from that $$\label{eq:etaA0}
\eta_{A,\rm grav}= -\frac{3\,I}{\pi}\gNk \quad{\rm with}\quad I=I_0-\frac{\eta_A}{2}\,I_1\,.$$ with the dimensionless running Newton coupling defined in . The coefficients $$\begin{aligned}
\label{eq:IetaAgrav}
I_0&=&\int_0^\infty dx\;\frac{1+\alpha}{1+r_g}\0{ r_A}{1+r_A}
\left(1+\012 \0{ \eta_N r_g}{
1+r_g}\right) \\
\label{eq:IetaAgrav1}
I_1&=&\int_0^\infty dx\;\frac{1+\alpha}{1+r_g}\0{ r_A}{(1+r_A)^2}\end{aligned}$$ originate from the radial momentum integration with $x=p^2/k^2$. As discussed before, they only depend on the gravitational gauge fixing parameter but not on the Yang-Mills gauge fixing parameter. The dropped term proportional to $\partial_t\alpha$ would effectively lower the coefficient $1/2$ in front of the $\eta_N$-term in $I_0$. To see this we remark that in leading order the gauge fixing term is $k$-independent. This entails that the gauge fixing parameter runs like the wave function renormalisation, $\partial_t\ln \alpha\propto
-\eta_N$. As the prefactor of the $\eta_N$-term comes from a summation over all tensor component, the running of $\alpha$ effectively removes the gauge fixing direction from this sum and hence lowers the prefactor. This structure is also present beyond leading order, indeed, for $\alpha=0$ the related term even vanishes identically. Thus, for the sake of simplicity we shall drop this term as it does not change the arguments presented here. Solving for $\eta_{A,\rm grav}$ we arrive at $$\label{eq:etaA}
{ \eta_A} =\eta_{A,{\rm YM}}+\eta_{A,{\rm grav}}=\0{\eta_{A,{\rm \tiny YM}}-
\s03\pi {\gNk}I_0}{1-\s0{3}{2\pi} {\gNk}I_1} \,.$$ We conclude from with the coefficients , that the gravity-induced running of Yang-Mills is not universal. Note however that the regulators in and in general, do not respect the symmetry constraint .
Perturbation theory {#PT}
===================
The one-loop perturbative results are recovered from the flow equation by using the bare two-point functions on the right-hand side of the flow. This corresponds to setting $Z_A=Z_N=1$ on the right hand side of . Then the one-loop graviton correction to the gluon anomalous dimension reads $$\label{1loop lambda0}
\eta_{A,\rm grav}=-\frac{3 \gN}{\pi}\,I_{0,\rm 1-loop}\,,$$ with $$\label{I}
I_{0,\rm 1-loop}=\int_0^\infty dx\;\frac{1+\alpha}{1+r_g(x)} \frac{r_A(x)}{1+r_A(x)}\,.$$ $I_{0,\rm 1-loop}$ encodes the gauge-fixing and regularisation dependence of the one-loop coefficient. It is crucial to observe that $$\label{eq:eta_Agrav}
\eta_{A,\rm grav}\Big|_{\rm 1-loop}\leq0\,,$$ for $\alpha\geq -1$, and all regulators. We conclude that the full one-loop running of Yang-Mills is not universal due to the gravity correction. However, it sustains asymptotic freedom for all regularisations. Furthermore it has both, a regulator as well as a gravity gauge dependence via $\alpha$. Both are not independent as the latter one can be absorved within a specific choice of the regulator. This is not surprising as the regulator can be also partially viewed as a change of the gauge fixing. In order to disentangle these effects we first consider a vanishing gravity regularisation, $r_g\equiv 0$. The gauge dependence persists, $\partial_\alpha \eta_{A,\rm grav}=\eta_{A,\rm grav}/(1+\alpha)$ at fixed Newton coupling. Hence, $\eta_A$ is only independent of $\alpha$ for $\eta_{A,\rm grav}\equiv 0$, which is equivalent to the constraint $$\label{eq:zerogen}
\0{1}{1+r_g}\frac{r_A}{1+r_A}\equiv 0\,.$$ The constraint implies that $1/r_g$ has to vanish for finite $r_A>0$. Consequently, $r_g$ is a sharp cutoff for momenta where $r_A>0$. Legitimate choices are for example $r_g=r_A=r_{\rm sharp}$, or $r_g=r_{\rm sharp}$ and $r_A=r_{\rm opt}$ [@Litim:2000ci]. The sharp and optimised regulators are defined as $$\begin{aligned}
\label{eq:sharp}
r_{\rm sharp}(x)&=&\0{1}{\theta(x)}-1\,, \\
\label{eq:opt}
r_{\rm opt} (x)&=&\left(\0{1}{x}-1\right) \theta(1-x)\,,\end{aligned}$$ and the Heaviside step function $\theta$. Note that regulators with the constraint effectively satisfy the symmetry constraint : if the graviton legs in are contracted with graviton propagators, we are lead to . In turn, without graviton propagators the constraint necessitates $r_A=0$.
We summarise the above analysis as follows: we have computed the one loop gravity contributions $\eta_{A,\rm grav}$ to the Yang-Mills $\beta$-function for general regularisation schemes. The one-loop gravitational contribution to the running of the YM coupling is regularisation-dependent, and thus non-universal. This originates from the mass dimension inherent to the gravitational coupling. However, $\beta_{\textrm{\tiny YM},\rm grav}$ is negative semi-definite and hence sustains asymptotic freedom. It vanishes in symmetry-preserving regularisation schemes based on the symmetry relation .
Asymptotic freedom and quantum gravity {#GFP}
======================================
In this section we extend the stability analysis of asymptotic freedom to include general field-theory-based gravity scenarios [@Weinberg; @Niedermaier:2006ns; @Percacci:2007sz; @Litim:2008tt; @Niedermaier:2009zz; @Reuter:1996cp; @Souma:1999at; @Lauscher:2001ya; @Percacci:2002ie; @Litim:2003vp; @Fischer:2006fz; @Codello:2007bd; @Benedetti:2009gn; @Banks:1999gd; @Giddings:2001bu; @DvaliGomez; @Dvali:2010ue]. In some of these, short distance physics is shielded by black hole formation [@Banks:1999gd; @Giddings:2001bu; @DvaliGomez; @Dvali:2010ue]. Provided this happens at scales where the anomalous dimension of the graviton $\eta_N$ is still small, the weak coupling analysis of the previous section is sufficient to sustain asymptotic freedom of Yang Mills. However, since the graviton anomalous dimension $\eta_N$ might grow large due to strong quantum gravity effects, it is important to evaluate the stability of asymptotic freedom for general $\eta_N$.
From the RG-equation for $\gN$, , we note that the value $\eta_N=-2$ is distinguished because the $\beta$-function for the dimensionless Newton coupling $\gN$ vanishes. This is the gravitational fixed point of the asymptotic safety scenario for gravity [@Weinberg; @Niedermaier:2006ns; @Percacci:2007sz; @Litim:2008tt; @Niedermaier:2009zz; @Reuter:1996cp; @Souma:1999at; @Lauscher:2001ya; @Percacci:2002ie; @Litim:2003vp; @Fischer:2006fz; @Codello:2007bd; @Benedetti:2009gn]. In its vicinity, the gravitational $\beta$-function changes sign. Therefore the three cases $\eta_N=-2$, $\eta_N<-2$ and $\eta_N>-2$ are qualitatively different and will be discussed separately.
We first discuss the impact of a gravitational fixed point $\eta_N=-2$ for asymptotic freedom of Yang-Mills by evaluating the gravitational contribution $\eta_{A, \rm grav}$, . The related coefficient $I_0$ in becomes $$\label{eq:etaAgravfull}
I_0=\int_0^\infty dx\;\frac{1+\alpha}{(1+r_g)^2} \0{r_A}{(1+r_A)}$$ and the coefficient $I_1$ is given by . With the coefficients $I_0$ and $I_1$ and the assumption of asymptotic freedom, $g_{\rm YM}\equiv 0$ at the fixed point, the $\beta$-function at the fixed point reads $$\label{eq:etafixedA}
{ \eta_A}_* =- \0{\s03{\pi} {g_N}_* I_0}{1-\s0{3 }{2\pi} {g_N}_* I_1 } \,.$$ As for the one loop running, ${\eta_A}_*$ is not universal since the coefficients $I_0,I_1$ depend on the choices for the regulators $r_g,r_A$. Asymptotic freedom only enforces ${\eta_A}_*\leq 0$ which implies $$\label{eq:constraints}
I_0\geq 0 \quad \wedge\quad {g_N}_*\leq \frac{2 \pi}{3 I_1}\, ,
\qquad {\rm or}\qquad I_0< 0 \quad \wedge \quad {g_N}_* > \frac{2 \pi}{3 I_1}\,.$$ The latter case is irrelevant here as $I_0< 0$ can only be obtained for the singular choices $\alpha<-1$. For the symmetry preserving choices we have $I_0=0$ and we arrive at a vanishing non-perturbative gravity contribution for the Yang-Mills $\beta$-function, ${\eta_A}_*=0$.
In the general case we have to satisfy the constraint in relating $I_1$ and ${g_N}_*$. For the sake of simplicity we restrict ourselves to regulators with $r_A(x\geq
1)\equiv 0$ as this limits the size of the back-reaction of the gluon on the graviton. Then, $I_1$ in satisfies $$\label{eq:etaAgravbound2}
I_1=\int_0^1 dx\;\frac{1+\alpha}{1
+r_g} \0{r_A}{(1+r_A)^2}\leq \0{1}{4}
\frac{1 +\alpha}{1+r_{g,\rm min}}\,,$$ where $r_{g,\rm min}=\min r_g(x)$ with $x\in [0,1]$. We have also used that $r_A/(1+r_A)^2\leq 1/4$. Due to the bound the first set of constraints in is satisfied for $$\label{bound}
{g_N}_*\le
\frac{8\pi}{3}\frac{1+\alpha}{1+r_{g,\rm min}}\,.$$ The bound for ${g_N}_*$ is safely satisfied for general regulators $r_g$. Indeed, the generic value for the fixed point coupling is of order one. For example, for the pure gravity system one finds ${g_N}_*=0.893$ for an optimised regulator and $\alpha=0$. Note also in this context that ${g_N}_* (1+\alpha)$ is approximately constant for all $\alpha$. If we relax the locality condition for the gluon, that is $r_A(x\geq 1)\not\equiv 0$, one enhances the gluon fluctuations. In this case one has to take into account the back-reaction of the gluon fluctuations on the fixed point value of the Newton coupling. Here we only remark that in the full system indeed one can show that $ {g_N}_*\leq \frac{2 \pi}{3 I_1}$ is always satisfied. This will be discussed in more detail elsewhere.
In summary we find that the graviton contribution $\eta_{A,\rm grav}$ sustains asymptotic freedom $$\label{lambda0 result}
{\eta_{A,\rm grav}}_*\leq 0\,.$$ Hence the ultraviolet fixed point action of Yang-Mills theory is given by the classical action and we can invoke leading to $\eta_{A,\rm grav}=0$. This leads us to the final value of the non-perturbative $\beta$-function if all symmetries are respected, $$\label{eq:finalbeta}
{\eta_{A}}_*=0\,.$$ We conclude that the gauge coupling remains asymptotically free, even in the presence of a non-perturbative gravitational fixed point.
Next we turn to $\eta_N<-2$. In this case $I_0$ might turn negative, see . However, with we have $\beta_{\gN}<0$ and $g_N$ decreases exponentially with logarithmic RG scale $t=\ln k$, and so does the gravity contribution to $\eta_A$ in . We conclude that even though the gravity contribution to $\eta_A$ might be positive for $\eta_N<-2$, its exponential decay would eventually lead to the dominance of $\eta_{A,YM}$. Thus, asymptotic freedom would persist. Note also that $\eta_N<-2$ for $t\to\infty$ is equivalent with a Gaußian UV fixed point which is not present in gravity.
Finally, for $\eta_N>-2$ the dimensionless Newton coupling $g_N$ grows exponentially with $t$, see , and the coefficient $I_0$ is positive. This case includes classical gravity with $\eta_N=0$ as well as a gravitational shielding scenario. The constraints enforce $g_N\leq 2 \pi/(3 I_1)$ to ensure asymptotic freedom for Yang-Mills. This puts an upper limit on the value of dimensionless Newton coupling which has to be satisfied by eg. perturbation theory or a gravitational shielding scenario, or else $g_N$ de-stabilizes asymptotic freedom. Note that this bound depends on the regularisation scheme as does the definition of the dimensionless Newton constant $g_N$.
In summary we have shown that asymptotic freedom persists in general field-theory-based gravity scenarios.
Cosmological constant {#CC}
=====================
We proceed by introducing a cosmological constant $\Lambda$ to the theory. For the sake of simplicity we consider the case $\alpha=0$, but the results readily extend to $\alpha\neq 0$. The coefficient functions $I_0$ and $I_1$, , read $$\begin{aligned}
\nonumber
I_0&=&\int_0^\infty dx\;\frac{1}{1- \0{2\lambda}{x}+ r_g} \0{r_A}{1+r_A}
\left(1+\left(\0{\eta_N r_g}{2}-\0{2\lambda+\partial_t \lambda}{x}\right)\0{1}{
1-\0{2\lambda}{x}+r_g}\right)
\\
I_1&=&\int_0^\infty dx\;\frac{1}{1- \0{2\lambda}{x}+ r_g} \0{r_A}{(1+r_A)^2}
\label{eq:etaALambda}\end{aligned}$$ where $\lambda=\Lambda_k/k^2$ denotes the cosmological constant in units of the RG scale $k$. This leads to the beta-function with the coefficients $I_0$ and $I_1$ in . As before we conclude that if invoking the symmetry relation leading to with $1+r_g\to 1-2\lambda/x+r_g$, the coefficient function vanishes, $I\equiv 0$, and we have no gravity contribution to the Yang-Mills running, $$\label{eq:finalbetaLambda}
{\eta_{A,\rm grav}}=0\,.$$ It is left to prove the stability of this result under changes of the regulator, that is $\eta_{A,\rm grav}\leq 0$ for all regulators.
As in the previous section we consider the qualitatively different cases $\eta_N=-2$, $\eta_N<-2$ and $\eta_N>-2$. At the gravitational fixed point with $\eta_N=-2$ and $\partial_t \lambda=0$ we find $$\label{eq:>=}
I_0=\int_0^\infty dx\; \0{r_A}{1+r_A} \,\frac{1}{(1-\frac{2\lambda}{x}+r_g)^2}
\left(1-\0{4\lambda}{x}\right) \,.$$ We have stability of the fixed point iff $I_0\ge 0$. Evidently the integrand in is not positive but is negative for small momenta $x$ and turns positive for $x >4 \lambda$. For regulators , we find stability for $\lambda\le\lambda_{\rm crit}=\frac{1}{8}$. For general regulators one has to solve the integral equation $I_0$ numerically resulting in $\lambda_{\rm crit}(r_g,r_A)$ and the constraint $\lambda\le\lambda_{\rm crit}(r_g,r_A)$. The regulator-dependence of the critical cosmological constant is not surprising as the value of the cosmological constant (and the Newton constant) at the fixed point is not a physical observable. The important result is the existence of such a constraint.
For $\eta_N<-2$ and $\eta_N>-2$ we run into the same scenarios as in the previous section with the coefficients $I_0$ and $I_1$. Note, however, that for $\eta_N>-2$ we require $I_0>0$ which puts constraints on $\lambda$ and $\partial_t \lambda$. As one can assume $\partial_t \lambda<0$ due to the canonical running in this regime, they are less severe than the one in the asymptotic safety scenario.
Discussion {#sec:disc}
==========
First we compare our results with previous findings to leading order in perturbation theory in the $U(1)$ and $SU(N)$ cases [@Deser:1974cz; @Deser:1974nb; @Robinson:2005fj; @Pietrykowski:2006xy; @Toms:2007sk; @Ebert:2007gf; @Daum:2009dn; @Toms:2010vy]. Identifying the energy scale $E$ with the cutoff scale $k$, the one-loop gravitational correction can be written as $$\label{1loop result}
\beta_{\textrm{\tiny YM},\rm grav}=-\frac{3\, I}{2\pi}
g_\textrm{\tiny YM}\,G_N\,E^2\,.$$ Our result corresponds to the coefficient $I=I_{\rm 1-loop}$ as given in . Within dimensional regularisation and using the background field method, Deser et. al. [@Deser:1974cz; @Deser:1974nb] found no gravitational corrections to the Yang-Mills and Maxwell $\beta$-functions, respectively, and hence $I=0$. In the $U(1)$ case, the same conclusion is reached by Pietrykowski [@Pietrykowski:2006xy] by evaluating the gauge fixing dependence in a generalised $R_\xi$ gauge, and by Toms [@Toms:2007sk] within the geometrical effective action method. In the $SU(N)$ case, Ebert, Plefka and Rodigast [@Ebert:2007gf] confirmed $I=0$ based on dimensional regularisation and cutoff regularisation with Feynman gauge for the gluon and de Donder gauge for the graviton. This study shows that quadratic divergences, neglected within dimensional regularisation, cancel out within a sharp cutoff regularisation.
On the other hand, Robinson and Wilzcek [@Robinson:2005fj], Toms [@Toms:2010vy] and Daum et. al. [@Daum:2009dn] find a non-vanishing contribution to leading order, $I>0$. In particular, the result of Robinson and Wilczek [@Robinson:2005fj] is based on the Feynman gauge $\xi=1$ and $\alpha=1$, leading to $I_{\rm RW}=2$ in . A positive coefficient $I_{\rm T}=\frac{2}{3}$ has also been found by Toms [@Toms:2010vy] using a modified version of the Vilkovisky de-Witt approach. The study of Daum, Harst and Reuter [@Daum:2009dn] is based on the background field method, together with a Wilsonian momentum cutoff for the propagating modes, with $$\label{eq:DHR}
I_{\rm DHR}=4\int_0^\infty dx\frac{-x^2\, r'(x)}{1+r(x)}
=8\int_0^\infty dx\,x\, \ln [1+r(x)]\,.$$ Here $r(x)$ parametrises the shape of the momentum cutoff. The coefficient is non-universal and strictly positive, $I_{\rm DHR}> 0$, as $r'\not\equiv 0$. As has been argued in Sec. \[BFD\], the background field approach used in this study does also include terms originating in the background field dependence of the regulator. For the symmetry-preserving cut-off choice $r_{\rm sym}$ with the coefficient diverges in clear contradistinction to our result $I[r_{\rm sym}]=0$. This difference can be solely attributed to the field-dependence of the cut-off function which enters .
In the presence of a cosmological constant, a recent study based on a diffeomorphism invariant expansion scheme using the geometric effective action finds a non-vanishing result to one-loop order [@Toms:2008dq]. The expansion scheme used in this approach differs from the standard one as eg. already the Yang Mills classical propagator is massive with a mass proportional to the cosmological constant. We emphasize that the symmetry preserving scheme employed here also leads to a vanishing coefficient in this set-up, $I=0$.
Based on the general considerations underlying our RG result, we can acertain that the gauge-fixing and regularisation-dependent coefficient $I_{\rm 1-loop}\ge 0$. As long as the implicit or explicit regularisation respects the symmetry relation , the result reads $I=0$. Note in this context that regularisations such as Pauli-Villars, or other gauges such as a general $R_\xi$ gauge, do not respect the symmetry constraint. We also remark that in the computation of a non-universal coefficient one cannot tell apart regularisation and gauge-fixing dependences. It is therefore not surprising that different regularisations and gauges lead to different coefficients $I\ge 0$. We conclude that all existing studies agree in that the gravitational contribution to the Yang-Mills $\beta$-function support asymptotic freedom.
The one-loop approximation is valid in the weak gravity regime. Close to the Planck scale and beyond, the dynamics within the gravitational sector becomes relevant. We have shown that asymptotic freedom persists for general anomalous dimension $\eta_N$, which entails its compatibility with general field-theory-based gravity scenarios. This includes gravitational shielding as well as asymptotically safe gravity.
In conclusion, provided that all symmetries and in particular are preserved, the graviton induced running of the Yang-Mills coupling vanishes, $$\label{final}
{\eta_{A}}|_{\rm grav}=0\,.$$ This result stays valid in the presence of a cosmological constant, and in the presence of a gravitational fixed point. Hence, it will be interesting to extend this study to the fully coupled Yang-Mills gravity system.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work is supported by Helmholtz Alliance HA216/ EMMI, and by the Science and Technology Research Council \[grant number ST/G000573/1\].
[99]{} D. J. Gross, F. Wilczek, Phys. Rev. Lett. [**30** ]{} (1973) 1343-1346.
H. D. Politzer, Phys. Rev. Lett. [**30** ]{} (1973) 1346-1349.
S. P. Robinson and F. Wilczek, Phys. Rev. Lett. [**96**]{}, 231601 (2006) \[arXiv:hep-th/0509050\].
A. R. Pietrykowski, Phys. Rev. Lett. [**98**]{}, 061801 (2007) \[arXiv:hep-th/0606208\].
D. J. Toms, Phys. Rev. D [**76**]{}, 045015 (2007) \[arXiv:0708.2990 \[hep-th\]\]. D. Ebert, J. Plefka and A. Rodigast, Phys. Lett. B [**660**]{}, 579 (2008) \[arXiv:0710.1002 \[hep-th\]\].
Y. Tang and Y. L. Wu, arXiv:0807.0331 \[hep-ph\].
D. J. Toms, Nature [**468** ]{} (2010) 56-59. \[arXiv:1010.0793 \[hep-th\]\].
J. F. Donoghue, Phys. Rev. Lett. [**72**]{} (1994) 2996 \[gr-qc/9310024\]. S. Deser, H. -S. Tsao, P. van Nieuwenhuizen, Phys. Lett. [**B50** ]{} (1974) 491.
S. Deser, P. van Nieuwenhuizen, Phys. Rev. [**D10** ]{} (1974) 401.
S. Deser, H. -S. Tsao, P. van Nieuwenhuizen, Phys. Rev. [**D10** ]{} (1974) 3337.
D. Ebert, J. Plefka and A. Rodigast, arXiv:0809.0624 \[hep-ph\]. F. Wu and M. Zhong, arXiv:0809.1913 \[hep-ph\].
D. J. Toms, Phys. Rev. Lett. [**101**]{}, 131301 (2008) \[arXiv:0809.3897 \[hep-th\]\].
J. -E. Daum, U. Harst, M. Reuter, JHEP [**1001** ]{} (2010) 084. \[arXiv:0910.4938 \[hep-th\]\].
J. Ellis, N. E. Mavromatos, \[arXiv:1012.4353 \[hep-th\]\].
C. Wetterich, Phys. Lett. B [**301**]{} (1993) 90; J. Berges, N. Tetradis and C. Wetterich, Phys. Rept. [**363**]{} (2002) 223 \[hep-ph/0005122\].
D. F. Litim and J. M. Pawlowski, in: [*The Exact Renormalization Group*]{}, Eds. Krasnitz et al, World Sci (1999) 168 \[hep-th/9901063\]. H. Gies, arXiv:hep-ph/0611146.
J. M. Pawlowski, Annals Phys. [**322**]{} (2007) 2831 \[arXiv:hep-th/0512261\]. D. F. Litim and J. M. Pawlowski, JHEP [**0209**]{}, 049 (2002) \[arXiv:hep-th/0203005\]. H. Gies, Phys. Rev. D [**66**]{} (2002) 025006 \[arXiv:hep-th/0202207\]. J. M. Pawlowski, D. F. Litim, S. Nedelko and L. von Smekal, Phys. Rev. Lett. [**93**]{} (2004) 152002 \[hep-th/0312324\];
M. Reuter and C. Wetterich, Phys. Rev. D [**56**]{} (1997) 7893 \[hep-th/9708051\]; D. F. Litim and J. M. Pawlowski, Phys. Lett. B [**435**]{} (1998) 181.
F. Freire, D. F. Litim and J. M. Pawlowski, Phys. Lett. B [**495**]{}, 256 (2000);
J. M. Pawlowski, Int. J. Mod. Phys. A [**16**]{} (2001) 2105.
J. M. Pawlowski, arXiv:hep-th/0310018.
D. F. Litim and J. M. Pawlowski, Phys. Rev. D [**66**]{}, 025030 (2002) \[arXiv:hep-th/0202188\]. D. F. Litim and J. M. Pawlowski, Phys. Lett. B [**546**]{}, 279 (2002) \[arXiv:hep-th/0208216\].
D. F. Litim, Phys. Lett. [**B486** ]{} (2000) 92-99. \[hep-th/0005245\]. Phys. Rev. [**D64** ]{} (2001) 105007. \[hep-th/0103195\].
S. Weinberg, in [*General Relativity: An Einstein centenary survey*]{}, Eds. S.W. Hawking and W. Israel, Cambridge University Press (1979), p. 790; arXiv:0903.0568 \[hep-th\]; arXiv:0908.1964 \[hep-th\]; arXiv:0911.3165 \[hep-th\]. M. Niedermaier, Class. Quant. Grav. [**24**]{} (2007) R171 \[gr-qc/0610018\]; M. Niedermaier and M. Reuter, Living Rev. Relativity [**9**]{} (2006) 5. R. Percacci, in ’Approaches to Quantum Gravity: Towards a New Understanding of Space, Time and Matter’ ed. D. Oriti, Cambridge University Press, 0709.3851 \[hep-th\]. D. F. Litim, PoS [**QG-Ph**]{} (2008) 024, 0810.3675 \[hep-th\]. M. Reuter, Phys. Rev. D [**57**]{} (1998) 971 \[hep-th/9605030\]. W. Souma, Prog. Theor. Phys. [**102**]{} (1999) 181 \[hep-th/9907027\].
O. Lauscher and M. Reuter, Phys. Rev. D [**65**]{} (2002) 025013 \[hep-th/0108040\]; Class. Quant. Grav. [**19**]{} (2002) 483 \[hep-th/0110021\]; Phys. Rev. D [**66**]{} (2002) 025026 \[hep-th/0205062\]. R. Percacci and D. Perini, Phys. Rev. D [**67**]{} (2003) 081503 \[hep-th/0207033\]; Phys. Rev. D [**68**]{} (2003) 044018 \[hep-th/0304222\];
D. F. Litim, Phys. Rev. Lett. [**92**]{} (2004) 201301 \[hep-th/0312114\]. AIP Conf. Proc. [**841**]{} (2006) 322 \[arXiv:hep-th/0606044\]. P. Fischer and D. F. Litim, Phys. Lett. B [**638**]{} (2006) 497 \[arXiv:hep-th/0602203\]; AIP Conf. Proc. [**861**]{} (2006) 336 \[arXiv:hep-th/0606135\].
A. Codello, R. Percacci and C. Rahmede, Int. J. Mod. Phys. A [**23**]{} (2008) 143 \[arXiv:0705.1769 \[hep-th\]\]; Annals Phys. [**324**]{} (2009) 414 \[arXiv:0805.2909 \[hep-th\]\]. D. Benedetti, P. F. Machado and F. Saueressig, Nucl. Phys. B [**824**]{} (2010) 168 \[arXiv:0902.4630 \[hep-th\]\].
S. Folkerts, Diploma thesis, University of Heidelberg (August 2009).
M. R. Niedermaier, Phys. Rev. Lett. [**103**]{} (2009) 101303; PoS [**CLAQG**]{} (2008) 05.
T. Banks and W. Fischler, arXiv:hep-th/9906038.
S. B. Giddings, S. D. Thomas, Phys. Rev. [**D65** ]{} (2002) 056010. \[hep-ph/0106219\].
G. Dvali and C. Gomez, arXiv:1005.3497 \[hep-th\].
G. Dvali, S. Folkerts and C. Germani, Phys. Rev. D [**84**]{} (2011) 024039 \[arXiv:1006.0984 \[hep-th\]\]. R. Percacci and G. P. Vacca, arXiv:1008.3621 \[hep-th\].
[^1]: Present address: Arnold Sommerfeld Center, Ludwigs-Maximilians-Universität, Theresienstr. 37, 80333 München, Germany
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The behavior of the timelike and null geodesics of charged E. Ay$\acute{o}$n-Beato and A. Garcia (ABG) black hole are investigated. For circular and radial geodesics, we investigate all the possible motions by plotting the effective potentials for different parameters. In conclusion, we have shown that there is no phenomenon of *superradiance* in this case.'
author:
- Mehedi Kalam
- Nur Farhad
- 'Sk. Monowar Hossein'
title: Geodesic study of a charged black hole
---
Introduction
============
A black hole can be characterized by its mass M, charge Q and angular momentum J. A rotating charged black hole is represented by Kerr-Newman metric while uncharged one can be represented as Kerr black hole. Again, if a non-rotating black hole has charged then it is called Reissner-Nordström black hole. For uncharged case, the black hole which has only mass dependency is called Schwarzschild black hole.\
Recently, E. Ay$\acute{o}$n-Beato and A. Garcia [@Beato1999] (ABG) gave a solution of the Einstein field equations with nonlinear electrodynamics, known as ABG black hole.
The static spherical symmetric space-time is described by the metric $$ds^2 = -f(r)dt^2 + \frac{1}{f(r)}dr^2 +r^2 (d\theta^2
+sin^2\theta d\phi^2), \label{eq1}$$ where $~~f(r)=1-\frac{2M}{r}+\frac{2M}{r}\tanh(\frac{Q^2}{2Mr})$\
with M and Q corresponding to Mass,electro-magnetic charge of the black hole respectively.\
This solution corresponds to a regular black hole with mass M and electro-magnetic charge Q and avoids thus the singularity problem. Also, the metric behaves asymptotically as the Reissner-Nordström (RN)black hole solution. It is clear that the singularity of the RN solution, at r = 0, has been omitted and now it simply becomes the origin of the spherical coordinates.Several [@Kalam2009; @Raychaudhuri2009] authors studied different types of black holes (like charged brane-world black holes, dyadosphere of a charged black hole etc.). I. Radinschi [@Radinschi2000] calculate the energy distribution of a charged ABG black hole by using the energy-momentum complexes of Einstein and Møller.\
In this paper, we will analyze the behaviour of the timelike and null geodesics of charged E.Ay$\acute{o}$n-Beato and A. Garcia (ABG) black hole. For circular and radial geodesics, we will discuss the possible motions by plotting the effective potentials for various parameters. We also check the phenomenon of superradiance for an incident massless scalar field for such a black hole.
The Geodesics
=============
Now, the equation for the geodesics in the metric (1) is given $$\frac{d^2x^\mu}{d\tau^{2}}+\Gamma^{\mu}_{\sigma\lambda}\frac{dx^\sigma}{d\tau}\frac{dx^\lambda}{d\tau}=0$$ From eqn.(2) we have $$\begin{aligned}
\frac{1}{f(r)}(\frac{dr}{d\tau})^2 &=& \frac{E^2}{f(r)}-\frac{J^2}{r^2}-L \nonumber \\
r^2(\frac{d\phi}{d\tau})&=& J \nonumber \\
\frac{dt}{d\tau}&=& \frac{E}{f(r)}\end{aligned}$$ where we assume the$~\theta=\frac{\pi}{2}$ plane and constants E and J are the energy per unit mass and angular momentum,respectively about an axis perpendicular to the invariant plane $\theta=\frac{\pi}{2}$. Here, the affine parameter is $\tau$ and the Lagrangian,L has values 0 and 1,respectively for massless and massive particles.\
Now, for radial geodesic (J=0) $$\dot{r}^2 \equiv \left(\frac{dr}{d\tau}\right)^2=E^2 - Lf(r)$$ Using the above equations we get $$\begin{aligned}
\left(\frac{dr}{dt}\right)^2&=&\left(1-\frac{2M}{r}+\frac{2M}{r}\tanh(\frac{Q^2}{2Mr})\right)^2\nonumber
\\&&\left[1-\frac{L}{E^2}\left(1-\frac{2M}{r}+\frac{2M}{r}\tanh(\frac{Q^2}{2Mr})\right)\right]\end{aligned}$$
Photonlike particle motion(L=0)
-------------------------------
For photonlike particle,we have $$\begin{aligned}
\left(\frac{dr}{dt}\right)^2=\left[1-\frac{2M}{r}+\frac{2M}{r}\tanh(\frac{Q^2}{2Mr})\right]^2\end{aligned}$$ i.e. $$\pm t=\int\frac{dr}{\left(1-\frac{2M}{r}+\frac{2M}{r}\tanh(\frac{Q^2}{2Mr})\right)}$$
Considering, $1-\tanh(z)~\approx~z~$, for a certain nbd. of z where $\frac{Q^2}{2Mr}=z$, we get $$\begin{aligned}
\pm~t~=~\frac{Q^2}{2Mz}~-~\frac{Q}{2}\ln\left|\frac{Q+2Mz}{Q-2Mz}\right|\nonumber\end{aligned}$$
i.e.\
$$~\pm~t~=~r-\frac{Q}{2}\ln\left|\frac{Qr+Q^2}{Qr-Q^2}\right|$$ The relation between time and distance for this particle is shown in the Fig. 1.
Now,we get from eqn.(4) as $$\dot{r}^2 \equiv \left(\frac{dr}{d\tau}\right)^2=E^2 \nonumber$$ that implies the $\tau~ - ~r$ relationship as $$\pm E \tau =r \nonumber$$ The variation of proper time ($\tau$) with respect to the radial distance.(r).is shown in Fig.2.
![ Time-Distance relationship for photonlike particle(taking Q=1).[]{data-label="fig:1"}](Charged1.eps)
Massive particle motion (L=1)
------------------------------
![ Propertime-Distance relationship for photonlike particle(taking E=4).[]{data-label="fig:2"}](Charged2.eps)
For this case we can write, $$\begin{aligned}
\left(\frac{dr}{dt}\right)^2&=&\frac{1}{E^2}\left(1-\frac{2M}{r}+\frac{2M}{r}\tanh(\frac{Q^2}{2Mr})\right)^2\nonumber
\\&&\left[E^2-1+\frac{2M}{r}~-~\frac{2M}{r}\tanh(\frac{Q^2}{2Mr})\right] \nonumber\end{aligned}$$ Assuming $\frac{Q^2}{2Mr}~=~z$ we get
$$\begin{aligned}
\pm~t= -\frac{Q^2E}{2M}\nonumber
\\\int\frac{dz}{z^2(1-\frac{4M^2}{Q^2}z(1-\tanh z))\sqrt{E^2-1+\frac{4M^2}{Q^2}z(1-\tanh z)}}\end{aligned}$$
Again considering, $1-\tanh(z)~\approx~z~$, for a certain nbd. of z where $\frac{Q^2}{2Mr}=z$, we get $$\pm~t= -\frac{Q^2E}{2M}\nonumber
\\\int\frac{dz}{z^2(1-\frac{4M^2}{Q^2}z^2)\sqrt{E^2-1+\frac{4M^2}{Q^2}z^2}}$$
Again asssuming, $u~=~\frac{2M}{Q}z$ we get
$$\pm~t~=~-~\frac{QE}{\sqrt{E^2-1}}\int\frac{du}{u^2\left(1-u^2\right)\sqrt{1+\frac{u^2}{E^2-1}}}$$
After simplification we get $$\pm~t~=~-~QE\int\frac{dq}{q^2(1-E^2q^2)}$$ where $ q = sin \left(tan^{-1}\left(\frac{u}{\sqrt{E^2-1}}\right)\right)$\
After integrating, $$\pm~t=Er\sqrt{E^2-1+\frac{Q^2}{2r^2}}-\frac{QE^3}{2}\ln\left|\frac{\sqrt{E^2-1+\frac{Q^2}{r^2}}+E\frac{Q}{r}}{\sqrt{E^2-1+\frac{Q^2}{r^2}}-E\frac{Q}{r}}\right|$$ This will give the time(t)-distance(r) relationship(Fig.3).\
![ Time-Distance relationship for massive particle(taking E=4,Q=1).[]{data-label="fig:3"}](Charged3.eps)
Again, from equation(4) we get $$\dot{r}^2=\left(\frac{dr}{d\tau}\right)^2=E^2-\left(1-\frac{2M}{r}+\frac{2M}{r}\tanh(\frac{Q^2}{2Mr})\right)$$ After simplification in a similar manner we get $$\pm\tau=~\frac{r}{(E^2-1)Q^2}\sqrt{E^2-1+\frac{Q^2}{r^2}}$$ which gives the proper time($\tau$) -distance(r) relationship(Fig.4).
![ Propertime-Distance relationship for massive particle(taking E=4,Q=1).[]{data-label="fig:4"}](Charged4.eps)
Effective Potential
===================
From the geodesic eqn.(2) and (3) we get $$\frac{1}{2}\left(\frac{dr}{d\tau}\right)^2=\frac{1}{2}\left[E^2~-~f(r)\left(\frac{J^2}{r^2}+L\right)\right]$$
If we compare eqn.(15) with $\frac{\dot{r}^2}{2}+V_{eff}=0$, one can get the effective potential, which depends on E and L as follows : $$V_{eff}=-\frac{1}{2}\left[E^2-f(r)\left(\frac{J^2}{r^2}+L\right)\right]$$
For Photonlike particle(L=0)
----------------------------
Consider, the radial geodesics where $J=0$. Then, the corresponding $V_{eff}$ is given by -\
$$V_{eff}~=~-\frac{1}{2}E^2 \nonumber$$
The particle will behaves like a “free particle” i.e its $V_{eff}=0$ when energy,$E=0$ .\
![ Behaviour of the Effective Potential,$V_{eff}$ for $J\neq0$(taking E=1,J=10,Q=M=1)[]{data-label="fig:5"}](Charged5.eps)
We can easily say that the behavior of these geodesics is independent on the mass and charge of the black hole. Now, for circular geodesics, $J~\neq~0$. The effective potential can be written as, $$V_{eff}=-\frac{E^2}{2}+\frac{J^2}{2r^2}\left(1-\frac{2M}{r}+\frac{2M}{r}~\tanh \frac{Q^2}{2Mr}\right)$$
If $r~\rightarrow0$, the effective potential , $V_{eff}(r)$ is large enough and approaches to $-~\frac{E^2}{2}~$ when $r~\rightarrow~~\infty$ . At the horizon, the effective potential, $V_{eff}~=~-\frac{E^2}{2}~.$\
We assume, the effective potential for $E~=~0$ \[in eqn.(17), put $E~=~0$\]. The roots of the potential is nothing but the horizons for this case. It is nice to see the effective potential,$V_{eff} $ is negative between the horizons. Hence, the particle would be bounded within the horizons. Again, a stable circular orbit will definitely exist between the horizons as $V_{eff}$ has a minima there(Fig.5).\
![ Behaviour of the Effective Potential,$V_{eff}$ (for radial geodesic) for massive particle(taking Q=M=1).[]{data-label="fig:6"}](Charged6.eps)
For massive particle(L=1)
-------------------------
The effective potential is given by $$V_{eff}=-\frac{E^2}{2} + \frac{1}{2}f(r)\left(\frac{J^2}{r^2}+1\right)$$ where $~~f(r)=1-\frac{2M}{r}+\frac{2M}{r}\tanh(\frac{Q^2}{2Mr})$.\
The roots of $f(r)$ are the horizons. For radial geodesic(J=0), $V_{eff}$ will vanish for some finite value of r in the region $ 0 \leq r < r_{-}$ as $ f(r) > 0$. Therefore, it is not possible for timelike geodesic to reach the singularity. The massive particle will emerge in other region avoiding the singularity and the spacetime is geodesically complete. We can studied the motion for both $E=0$ and $E\neq 0$. In first case, if we take $E=0$ then $V_{eff}$ becomes $$V_{eff}=\frac{1}{2}\left(1-\frac{2M}{r}+\frac{2M}{r}\tanh\frac{Q^2}{2Mr}\right).$$ The roots of $V_{eff}$ coincides with the horizons (Fig. 6) The shape of the potential indicates that the particle can move only inside the black hole. If we investigate the behaviour of $V_{eff}$ for $E \neq 0$, then when $ r\rightarrow 0$ the effective potential becomes $$V_{eff} \rightarrow -\frac{2M}{r}+\frac{2M}{r}\tanh \frac{Q^2}{2Mr}\nonumber.$$ Now, for large r,$V_{eff} \rightarrow \frac{1-E^2}{2}$. For a two horizons black hole, with in the two ranges $0 \leq r < r_{-} $ and $r_{+} < r$ , the function $f(r) > 0 $. Therefore, in these two region, it is possible for $V_{eff}$ to have roots.\
Now consider for non-zero angular momentum($ J \neq 0$) particle. Roots of $V_{eff}$ coincides with the two horizons and the shape of $V_{eff}$ is shown in Fig.6. Therefore, we can say that the massive particle with zero energy never escape from black hole and would be in a bound orbit. As the potential has a minima, the particle may have circular stable orbit.\
Again,for large r,$V_{eff} \rightarrow \frac{1-E^2}{2}$ when $E \neq 0$. When $ r\rightarrow 0$ then the effective potential becomes $$V_{eff} \rightarrow \frac{J^2}{2r^2}\left(1-\frac{2M}{r}+\frac{2M}{r}\tanh \frac{Q^2}{2Mr}\right)\nonumber.$$ The same conclusion can be taken for this case also that $V_{eff}$ have finite roots in between $0 \leq r < r_{-} $ and $r_{+} < r$. Both the cases the massive particle would be in a bounded orbit.
Solution of the Massless Scalar Wave Equation in the Charged E. Ay$\acute{o}$n-Beato and Garcia metric
======================================================================================================
In this section we shall analyze the scalar wave equation for charged E. Ay$\acute{o}$n-Beato and Garcia black hole geometry following Brill et al [@Brill1972]. The wave equation for a massless particle is given by $$g^{-1/2} \frac{\partial}{\partial x^\mu} \left( g^{1/2} g^{\mu\nu} \frac{\partial}{\partial x^\nu} \right) \chi =0$$ Here, $g_{\mu\nu}$ is known from equation(1). Putting all the values one can obtain the following equation as $$\begin{aligned}
-\frac{r^4 \sin\theta}{\Delta}\frac{\partial^2 \chi}{\partial t^2} + \sin \theta \frac{\partial}{\partial r}\left(\Delta \frac{\partial \chi}{\partial r}\right) \nonumber\\
+ \frac{\partial}{\partial \theta}\left( \sin\theta \frac{\partial}{\partial \theta} \right)\chi + \frac{1}{\sin\theta}\frac{\partial^2 \chi}{\partial \phi^2}=0 \nonumber\end{aligned}$$ Where $\Delta = r^2 -2Mr+2Mr ~~tanh(\frac{Q^2}{2Mr})$.
Now, by using the separation of variable this equatin can be solved as $$\chi = e^{-i\omega t} e^{im\phi} R(r) \Theta(\theta)$$ Substuting this we get $$\begin{aligned}
\frac{r^4 \sin \theta}{\Delta}\omega^2 \chi + \frac{\sin \theta}{R} \frac{\partial}{\partial r}
\left(\Delta \frac{\partial R}{\partial r}\right) \chi \nonumber \\
+ \frac{1}{\Theta} \frac{\partial}{\partial\theta}\left(\sin\theta\frac{\partial\Theta}{\partial\theta} \right) \chi- \frac{m^2}{\sin\theta} \chi =0 \nonumber\end{aligned}$$
Now,the radial equation reduces to $$\Delta \frac{\partial}{\partial r}\left( \Delta \frac{\partial R}{\partial r} \right) + (r^4 \omega^2 - \Delta \lambda)R=0$$
and the angular part reduces to $$\frac{1}{\sin\theta}\frac{\partial}{\partial \theta} \left(\sin\theta\frac{\partial \Theta}{\partial\theta}\right) - \frac{m^2}{\sin^2\theta}\Theta + \lambda \Theta =0$$ Now, if we substitute $x=\cos \theta$, the equation becomes $$\left(1-x^2\right)\frac{d^2 \Theta}{d x^2} - 2x \frac{d \Theta}{d x} - \left( \frac{m^2}{1-x^2}-\lambda \right) \Theta=0$$ If we now write $\lambda = l(l+1)$ where $l$ is an integer, then the equation $$\left(1-x^2\right)\frac{d^2 \Theta}{d x^2} - 2x \frac{d \Theta}{d x} + \left( l(l+1)-\frac{m^2}{1-x^2} \right) \Theta=0$$
is the familiar associated Legendre equation and the solution is given by the associated Legendre polynomial $ P^m_l(x)$ and is expressed as $$\Theta^m_l(\cos\theta) = P^m_l(x) = \frac{(-1)^m}{2^l l!}\left( 1-x^2 \right)^{m/2} \frac{d^{l+m}}{d x^{l+m}}(x^2-1)^l$$
The Radial Equation:Absence of Superradiance
--------------------------------------------
Kerr-Newman black hole allows energy extraction whereas Schwarzschild black hole prohibits. By an explicit process, this can be achieved (Penrose Process). Superradiance is nothing but the wave analogy of the Penrose process on black hole.When a bosonic or fermionic wave is incident on a black hole, naturally the reflected wave carries less energy than the incident one. But under specific condition, the transmitted wave,absorbed by the black hole carries negative energy into the black hole which makes the reflection coefficient greater than unity. This phenomena is called ’superradiance’[@Misner1972].According to this process the energy can be extracted from the black hole at the cost of lossing its angular momentum.The requirred condition is\
$$0 < \omega < m \omega_H$$ where $\omega_H $ is the angular velocity of the horizon [@DeWitt1975]. Basak et al discussed this phenomenon for the acoustic analogue of the Kerr black hole [@Basak2003]. Shiraishi [@Shiraishi1992] and Ali [@Ali2007] argued that superradiance phenomenon is possible if the black hole is rotating or charged.Now, we check whether the superradiance phenomenon really happens for E. Ay$\acute{o}$n-Beato and Garcia(ABG) black hole.\
The radial equation can be written as $$\Delta \frac{d}{d r}\left( \Delta \frac{d R}{d r} \right) + \left( \omega^2 r^4 - l(l+1)\Delta\right)R=0$$
Now, we introduce the familiar tortoise coordinate,$r^*$ defined as $$\frac{dr^*}{d r} = \frac{r^2}{\Delta}$$ that implies $$\Delta \frac{d }{d r} = r^2 \frac{d }{d r^*}$$
It is to be mentioned here that the variable $r^*$ is constructed in the same way as in Schwarzschild or in Kerr metric. Therefore, in this case the variable is non-integrable. Though the basic purpose is still satisfied, the coordinate spans over the real line and pushes the horizon to -$\infty$.\
Again, we introduce another function $u(r) = rR$ which reduces the radial equation in a more familiar form $$\begin{aligned}
\frac{d^2 u}{d {r^*}^2}
+[\frac{\Delta Q^2}{r^6}sech^2(\frac{Q^2}{2Mr})+\frac{2\Delta M}{r^5}(1-tanh(\frac{Q^2}{2Mr})) \nonumber\\
-\frac{2\Delta}{r^4}-\frac{l(l+1)\Delta}{r^4}+\frac{2\Delta^2}{r^6}+\omega^2 ] u = 0 \nonumber\end{aligned}$$
Therfore, a potential barrier remains where
$$\begin{aligned}
V(r) = -[\frac{\Delta Q^2}{r^6}sech^2(\frac{Q^2}{2Mr})+\frac{2\Delta M}{r^5}(1-tanh(\frac{Q^2}{2Mr})) \nonumber\\
-\frac{2\Delta}{r^4}-\frac{l(l+1)\Delta}{r^4}+\frac{2\Delta^2}{r^6}+\omega^2 ]\nonumber\end{aligned}$$
At horizon ($\Delta \rightarrow 0, r^* \rightarrow - \infty$), the radial equation comes out to be $$\frac{d^2 u_H}{dr{*2}} + \omega^2 u_H =0$$ with $V(r) = -\omega^2$.
Now, asymptotically, $r \rightarrow \infty$, imlies that $r^* \rightarrow \infty$. The equation has the same form as in the previous case $$\frac{d^2 u_\infty}{dr{*2}} + \omega^2 u_\infty =0$$
Thus $u_H = u_\infty$, where $u_H$ represents the radial solution at horizon and $u_\infty$ is the solution at $\infty$. This equality shows that for a E. Ay$\acute{o}$n-Beato and Garcia (ABG) black hole there is *no phenomenon of superradiance for an incident massless scalar field*.
Conclusion
==========
In this investigation, we have analyzed the behavior of the timelike and null geodesics of charged E. Ay$\acute{o}$n-Beato and Garcia(ABG) Black hole. Here, we shown the behaviour of time-distance and proper time-distance graph.In case of radial geodesic, the effective potential is independent of the charge and mass of the black hole, for photonlike particle. However, from the shape of the potential, it is clear that the timelike particle can move only inside the black hole. On the other hand, for circular geodesics, the roots of the effective potential coincide with the horizon. Also, from Fig. 5, we can say that as the potential has a minima between the horizons, the photonlike as well as timelike particles would be bounded in a stable circular orbit.\
Though it is familiar that the superradiance phenomenon could be seen in charged or rotating black holes[@Shiraishi1992; @Ali2007], we have found that it is absent in the charged E. Ay$\acute{o}$n-Beato and Garcia (ABG) Black hole.
Acknowledgments {#acknowledgments .unnumbered}
===============
MK gratefully acknowledge support from IUCAA, Pune, India under Visiting Associateship under which a part of this work was carried out. SMH is thankful to IUCAA also for giving him an opportunity to visit IUCAA where a part of this work was carried out.
[99]{} E. Ay$\acute{o}$n-Beato and A. Garcia, 1999, Phys.Lett. B, 464, 25. M. Kalam et al., 2009, Int. J Mod. Phys. A, 24, 719. B. Raychaudhuri et al., 2009, Mod. Phys. Lett. A, 24, 1277. I. Radinschi, 2000, arXiv:gr-qc/0011066v1. D.R.Brill, P.L. Chrzanowski,C.M.Pereira, E.D.Fackerell and J. R. Ipser,1972, Phys.Rev. D, 5, 1913.
C. Misner,1972, Phys.Rev. Lett, 28, 994.
B.S. DeWitt,1975, Phys.Rep. 19, 295. S. Basak and P. Majumdar,2003, Class. Quantum Grav. 20, 2929. K. Shiraishi,1992, Mod. Phys. Lett. A 37, 3449. M.H.Ali,2007, Gen. Rel. Grav. 37, 977.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Shripad Vidyadhar Thite
title: Spacetime Meshing for Discontinuous Galerkin Methods
---
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Bursting X-ray binaries in globular clusters are ideal sources for measuring neutron star masses and radii, and hence, for determining the equation of state of cold, ultradense matter. We use time-resolved spectroscopic data from EXO 1745$-$248 during thermonuclear bursts that show strong evidence for photospheric radius expansion to measure the Eddington flux and the apparent surface area of the neutron star. We combine this with the recent measurement of the distance to the globular cluster Terzan 5, where this source resides, to measure the neutron star mass and radius. We find tightly constrained pairs of values for the mass and radius, which are centered around $M=1.4~M_\odot$ and $R=11$ km or around $M=1.7~M_\odot$ and $R=9$ km. These values favor nucleonic equations of state with symmetry energy that is relatively low and has a weak dependence on density.'
author:
- 'Feryal Özel, Tolga Güver and Dimitrios Psaltis'
title: 'The Mass and Radius of the Neutron Star in EXO 1745$-$248'
---
Introduction
============
Accreting neutron stars that show thermonuclear X-ray bursts are optimal sources for determining the equation of state of cold, ultradense matter. They exhibit a number of spectroscopic phenomena that depend on the neutron star mass and radius, which can be used to measure these fundamental stellar properties. In particular, the apparent surface area of thermal emission during the cooling tail of a thermonuclear burst, as well as the peak flux achieved during strong, so-called photospheric radius expansion bursts, which reach the Eddington limit, provide two such observable quantities (van Paradijs 1978, 1979; see also Lewin, van Paradijs, & Taam 1993).
The large collecting area and the systematic monitoring of a number of X-ray bursters by the Rossi X-ray Timing Explorer (RXTE) has generated a large database of high quality observations that we use in this study. The peak flux achieved in numerous photospheric radius expansion bursts from several sources has been shown to remain nearly constant, within a range as narrow as a few percent in 4U 1728$-$34 (Galloway et al. 2003), and the apparent surface areas to be reproducible between bursts (Galloway et al. 2008). This provides observational justification to the theoretical expectation that, in low magnetic field neutron stars, thermonuclear bursts not only quickly engulf the entire stellar surface but also that strong bursts reach an intrinsic limit associated with the Eddington luminosity.
These two spectroscopic phenomena can be combined either with another spectroscopic measurement, such as a surface redshift (Cottam, Paerels, & Mendez 2002; see also Cottam et al. 2008) or with an accurate, independent distance to the neutron star, to break the intrinsic degeneracies in the neutron star properties and determine both the mass and the radius of the neutron star, independently (Özel 2006). X-ray bursters in globular clusters are unique in this respect, since the distances to the clusters can be independently measured. In this paper, we use the thermonuclear burst data of the source EXO 1745$-$248 located in Terzan 5 to measure the mass and radius of the neutron star.
The low mass X-ray binary EXO 1745$-$248 was discovered with Hakucho in August 1980 (Makishima et al. 1981). The source showed Type-I X-ray bursts, with intervals as short as 8 min (Inoue et al. 1984). It was again detected in 2000 during a RXTE/PCA scan of the galactic bulge as a transient X-ray burster (Markwardt & Swank 2000). EXO 1745$-$248 was also observed by the Chandra X-ray Observatory in 2000 and 2003. Heinke et al. (2003) used the Chandra and RXTE observations to suggest that the source is an ultracompact binary and also identified a possible infrared counterpart in the Hubble Space Telescope (HST) images of the cluster. No burst oscillations have been reported from this source (Galloway et al. 2008).
Terzan 5 is one of the most metal rich globular cluster in the galaxy, with a metallicity close to solar (Origlia et al. 2004). Recently, Ortolani et al. (2007) revisited the distance measurements to Terzan 5 using HST/NICMOS data. Using NICMOS instrumental magnitudes and two separate reddening laws (Schlegel et al. 1998; Lee et al.2001) to obtain the infrared extinction slope in the instrumental bands, they measured a distance of 6.3 kpc to this cluster. The two main sources of error in this measurement are related to uncertainties in the color and magnitude measurements of the HB of the cluster, as well as to metallicity uncertainties, while the slope of the reddening law introduces a much smaller error. The combined error of 0.2 mag corresponds to a distance uncertainty of about 10%, which we will adopt here. Note that Ortolani et al. (2007) also used two calibrations for the conversion of the NICMOS to JHK magnitudes, which resulted in different values for the distance. Due to the significant width and the large displacement of the NICMOS F110W filter compared to the ground-based J filter, the transformations between these bands are color dependent and hence, suffer from systematic uncertainties that are difficult to quantify. For this reason, we will only use the distance measurement obtained from NICMOS instrumental magnitudes.
In this paper, we combine the distance measurement to Terzan 5 with the observations of radius expansion bursts obtained by RXTE to determine the mass and the radius of the neutron star in EXO 1745$-$248. In Section 2, we analyze the Eddington limited bursts from this source. In Section 3, we use these observations to determine the mass and radius of the neutron star and describe the formal method for assessing the uncertainties in the measurements using this technique. In Section 4, we discuss our results and compare them to several equations of state for neutron star matter.
Spectral Analysis of X-ray Bursts
=================================
EXO 1745$-$248 has been observed with RXTE for 148 ks, during which two Type-I X-ray bursts were discovered, with clear evidence for photospheric radius expansion (Galloway et al. 2008). Note that, while Galloway et al. (2008) identified 20 more candidate non-PRE bursts that satisfied their trigger criteria, spectral analyses of those bursts revealed that the distinctive cooling associated with Type I bursts did not occur in these events, strongly suggesting that they are Type II instead (Lewin et al. 1993).
In order to analyze the PRE bursts, we extracted time resolved 2.5$-$25.0 keV X-ray spectra using the ftool [*seextrct*]{} and included the data from all the RXTE/PCA layers. We used the Science Event mode data with the E\_125$\mu$s\_64M\_0\_1s configuration, which has a nominal time resolution of 125 $\mu$s in 64 spectral channels. We binned the X-ray spectra in 27 spectral channels and over 0.25 s (for count rates above 6000 ct s$^{-1}$) and over 0.5 s (for count rates between 3000 and 6000 ct s$^{-1}$) time intervals during each burst. Following Galloway et al. (2008), we extracted a 16 s spectrum prior to the burst and used it as background. We generated separate response matrix files for each burst using the PCARSP version 10.1 and took into account the offset pointing of the PCA during the creation of the response matrix files.
We fit the extracted spectra with a blackbody function, using the hydrogen column density value of N$_{\rm H} = 1.4 \times
10^{22}$ cm$^{-2}$ determined by Wijnands et al. (2005) from Chandra observations. We used XSPEC v12 (Arnaud 1996) for our spectral analysis. For each spectrum, we calculated bolometric fluxes using equation (3) of Galloway et al. (2008). Figure 1 shows an example countrate spectrum as well as the best fit blackbody model. There are no systematic residuals in the fit, and the addition of any other spetcral components (e.g., a power-law model) is not statistically significant.
In Figure 2, we show the distribution of the $\chi^2$/d.o.f. values that we obtained by fitting the X-ray spectra of the source during the 2 PRE bursts and compare it to the expected distribution for 25 degrees of freedom. All fits with $\chi^2/{\rm d.o.f.} < 1.5$ follow the expected distribution and are, therefore, statistically acceptable. However, the five spectra with $\chi^2/{\rm d.o.f.} >
1.5$ are outliers, which are likely to be dominated by systematic uncertainties. We rejected these fits from the subsequent analyses.
We show in Figure 3 the bolometric flux, the blackbody temperature, and the normalization of the model spectra during the evolution of the two PRE bursts. The characteristic decrease of the temperature and the increase of the photospheric radius around the burst peak, as well as the cooling of the burst emission at a constant photospheric radius for both bursts can be seen in both bursts.
In PRE bursts, the Eddington limit at the surface of the neutron star is thought to correspond to the point in each burst when the normalization of the blackbody gets its lowest value while the temperature reaches its highest (Damen et al. 1990). The spectral properties of the two PRE bursts during this so-called touchdown point are consistent with each other, as demonstrated in Figure \[touchdown\]. The combined best-fit value for the touchdown flux between the two bursts is $(6.25 \pm 0.20) \times
10^{-8}$ erg cm$^{-2}$ s$^{-1}$. Note that the ratios of the peak to touchdown fluxes in these two bursts are well within the value expected from the general relativistic effects alone, and therefore, this source is not subject to the bias discussed in Galloway, Özel, & Psaltis (2008).
The second observational quantity that we determine from the spectral fits is the apparent radius of the emitting region during the cooling phase of the bursts. This is given directly by the normalization of the blackbody function, $A \equiv (R_{\rm app}/D)^2$, where $R_{\rm
app}$ is the radius corresponding to the apparent emitting surface area and $D$ is the distance to the source. We chose the intervals 4.5$-$15 s in both bursts, during which the apparent radius is constant. Fitting the cooling tails of these bursts individually resulted in values for the ratio $A = 104.0 \pm 1.0$ km$^2$ kpc$^{-2}$ and $A = 130.0 \pm 1.0$ km$^2$ kpc$^{-2}$. Similar systematic uncertainties have been observed in the Eddington fluxes from PRE bursts from, e.g., 4U 1728$-$34 (Galloway et al. 2003) and have been attributed to the variable reflection off of the accretion disk that changes at a superorbital period. Such a phenomenon can introduce systematic uncertainties in the apparent surface areas measured during the cooling tails of bursts. Because the systematic errors dominate over the statistical errors in this particular case, we will assume a boxcar probability distribution over this quantity in the range $A =
116 \pm 13$ km$^2$ kpc$^{-2}$.
Determination of the Neutron Star Mass and Radius
=================================================
In an approach similar to Özel (2006), we use the spectroscopic measurements of the touchdown flux $\ftd$ and the ratio $A$ during the cooling tails of the bursts, together with the measurement of the distance $D$ to the source in order to determine the neutron star mass $M$ and radius $R$. The observed spectroscopic quantities depend on the stellar parameters according to the relations $$\ftd=\frac{GMc}{k_{\rm es}D^2}\left(1-\frac{2GM}{Rc^2}\right)^{1/2}$$ and $$A=\frac{R^2}{D^2f_{\rm c}^4}\left(1-\frac{2GM}{Rc^2}\right)^{-1}\;,$$ where $G$ is the gravitational constant, $c$ is the speed of light, $k_{\rm es}$ is the opacity to electron scattering, and $f_{\rm c}$ is the color correction factor.
In the absence of errors in the determination of the observable quantities, the last two equations can be solved for the mass and radius of the neutron star. However, because of the particular dependences of $\ftd$ and $A$ on the neutron star mass and radius (see also Fig. 1 in Özel 2006), the loci of mass-radius points that correspond to each observable intersect, in general, at two distinct positions. Moreover, the diverse nature of uncertainties associated to each of the observables requires a formal assessment of the propagation of errors, which we present here.
We assign a probability distribution function to each of the observable quantities and denote them by $P(D)dD$, $P(\ftd)d\ftd$, and $P(A)dA$. Because the various measurements that lead to the determination of the three observables are independent of each other, the total probability density is simply given by the product $$\begin{aligned}
&&P(D,\ftd,A)dDd\ftd dA =\nonumber\\ &&\qquad \qquad
P(D)P(\ftd)P(A) dD d\ftd dA\;.
\label{eq:firstdistrib} \end{aligned}$$
Our goal is to convert this probability density into one over the neutron-star mass, $M$, and radius, $R$. We will achieve this by making a change of variables from the pair $(\ftd,A)$ to $(M,R)$ and then by marginalizing over distance. Formally, this implies that $$\begin{aligned}
&& P(D,M,R)dDdMdR =
\frac{1}{2} P(D)P[\ftd(M,R,D)]
\nonumber \\
&& \qquad P[A(M,R,D)] J\left(\frac{\ftd,A}{M,R}\right) dD dMdR\;,
\label{eq:transform}\end{aligned}$$ where $J(\ftd,A/M,R)$ is the Jacobian of the transformation. It is important to emphasize here that, given a distance D, not all pairs of the observables $(\ftd,A)$ can be obtained with real values for the neutron-star mass and radius. For this reason, the final distribution will not be normalized, even if the three distributions of equation (\[eq:firstdistrib\]) are. In addition, the factor $1/2$ appears in equation (\[eq:transform\]) because nearly all pairs of the observables $(\ftd,A)$ correspond to two distinct pairs of $(M,D)$. There is only a region of the parameter space for which the pair of observables corresponds to a single pair of values for the mass and radius. However, this region has zero volume and, therefore, will not contribute to the final probability distribution after we marginalize over distance.
We can now use the above expressions to calculate the Jacobian of the transformation $$\begin{aligned}
J\left(\frac{\ftd,A}{M,R}\right)&=&\frac{GcR}{f_{\rm c}^4 k_{\rm es}D^4}
\left[1-\frac{6GM}{Rc^2}+7\left(\frac{GM}{Rc^2}\right)\right] \nonumber \\
&& \left(1-\frac{2GM}{Rc^2}\right)^{-5/2}\;.
\label{eq:Jac}\end{aligned}$$
For the source EXO 1745$-$248, the distance measurement is dominated by systematic errors as discussed in §1. We will, therefore, use a box-car probability distribution over distance, with a mean of $D_0=6.3$ kpc and a range of $\Delta D=0.1 D_0$, i.e., $$P(D)dD=\left\{\begin{array}{ll}
\frac{1}{\Delta D} & {\rm if}\quad\vert D-D_0\vert \le \Delta D/2\\
& \\
0 & {\rm otherwise}
\end{array}
\right.
\label{eq:PD}$$
The measurements of the touchdown flux is consistent between the two radius-expansion bursts (see Fig. \[touchdown\]) and is, therefore, dominated only by statistical uncertainties. We assign a Gaussian probability distribution for this quantity with a mean and a standard deviation that we estimate by fitting a Gaussian function to the product of the probability distributions that correspond to the confidence contours shown in Figure \[touchdown\]. The result is a mean of $F_{\rm 0}=6.25\times 10^{-8}$ erg cm$^{-2}$ s$^{-1}$ and a standard deviation of $\sigma_{\rm F}=0.2\times
10^{-8}$ erg cm$^{-2}$ s$^{-1}$, i.e., $$P(\ftd)d\ftd=\frac{1}{\sqrt{2\pi \sigma_{\rm F}^2}}
\exp\left[-\frac{(\ftd-F_0)^2}{2\sigma_{\rm F}^2}\right]\;.
\label{eq:PF}$$
Finally, the measurement of the ratio $A\equiv\ratio$ between the two bursts is dominated by systematic uncertainties. We, therefore, assign to this ratio a box car probability distribution with a mean of $A_0=116$ and a range of $\Delta A=26$, i.e., $$P(A)dA=\left\{\begin{array}{ll}
\frac{1}{\Delta A} & {\rm if}\quad\vert A-A_0\vert \le \Delta A/2\\
& \\
0 & {\rm otherwise}
\end{array}
\right.
\label{eq:PA}$$
The color correction factor that enters these expressions is determined by models of burning neutron star atmospheres (e.g., Madej et al. 2004). At the observed high temperatures of the bursts, as well as in the absence of significant magnetic fields or heavy elements (as evidenced by the lack of atomic transition lines in the high resolution spectra), the Comptonized radiative equilibrium atmosphere models can be reliably calculated. As discussed in Özel (2006), when the emerging flux is substantially sub-Eddington, as in the case of the cooling tails of the bursts, the color correction factor $f_{\infty}$ asymptotes to a value of $\simeq 1.4$, which we adopt here. Finally, we use the electron scattering opacity $\kappa_{es} = 0.20 (1+X)$ cm$^{2}$ g$^{-1}$, that depends on the hydrogen mass fraction $X$.
We obtain the final distribution over neutron-star mass and radius by inserting equations (\[eq:Jac\])–(\[eq:PA\]) into equation (\[eq:transform\]) and integrating over distance. Figure \[mr\] shows the 1- and $2-\sigma$ contours for the mass and radius of the neutron star in EXO 1745$-$248, for a hydrogen mass fraction $X=0$. For larger values of the hydrogen mass fraction ($X
\gtrsim 0.1$), the masses and radii inferred individually from the Eddington limit and the apparent surface area become rapidly inconsistent with each other. This result is in line with the identification of EXO 1745$-$248 with an ultracompact binary by Heinke et al. (2003). Note that there are two distinct regions in the mass-radius plane that are consistent with the data because of the particular dependence of $\ftd$ and $A$ on the stellar mass and radius, as discussed above.
Discussion
==========
We used time-resolved spectroscopic data from EXO 1745$-$248 during thermonuclear bursts that show strong evidence for photospheric radius expansion to measure the Eddington flux and the apparent surface area of the neutron star. We combined this with the recent measurement of the distance to the globular cluster Terzan 5 (Ortolani et al. 2007), where this source resides, to measure the neutron star mass and radius. We found tightly constrained pairs of values for the mass and radius, which are centered around $M=1.4~M_\odot$ and $R=11$ km or around $M=1.7~M_\odot$ and $R=9$ km.
The confidence contours on the mass-radius plane (see Fig. 5) are in best agreement with nucleonic equations of state without the presence of condensates or strange matter. The leftmost family of mass-radius relations is based on the assumption that the absolute ground state of matter is made up of an approximately equal mixture of up, down, and strange quarks. The primary difference between the other two families of mass-radius relations is the symmetry properties of the equation of state of neutron star matter. Moreover, the mass-radius relations with deflection points are characteristic of calculations that incorporate bosons that can condense and, thus, soften the equation of state at high densities. The radius measurements presented here favor relatively low values for the bulk symmetry energy with a weak density dependence (see Lattimer & Prakash 2001).
The measurement of the mass and radius of a neutron star can significantly constrain the range of possibilities for the equation of state of ultradense matter, as discussed above. However, it cannot uniquely pinpoint to a single equation of state because of both the measurement errors and the uncertainties in the fundamental parameters that enter the nuclear physics calculations, such as the symmetry energy of nucleonic matter or the bag constant for strange stars. Further, even tighter constraints on the equation of state can be obtained by combining observations of neutron stars with different masses that will distinguish between the slopes of the predicted mass-radius relations, which are determined entirely by the physics of the neutron star interior.
A number of other constraints on neutron star radii have been obtained to date using various methods. Özel (2006) used spectroscopic measurements of the Eddington limit and apparent surface area during thermonuclear bursts, in conjuction with the detection of a redshifted atomic line from the source EXO 0748$-$676, to determine a mass of $M
\ge 2.10 \pm 0.28~M_\odot$ and a radius $R \ge 13.8 \pm 1.8~{\rm km}$. This radius measurement is consistent with the one presented in the current paper to within 2$-\sigma$, and, therefore, several nucleonic equations of state are consistent with both measurements.
Radii have also been measured from globular cluster neutron stars in binaries emitting thermally during quiescence, such as X7 in 47 Tuc and others in $\omega$ Cen, M 13, and NGC 2808 (Heinke et al. 2006; Webb & Barret 2007). (Note that we do not consider here isolated neutron stars such as RX J1856$-$3754 because of the unquantified systematic uncertainties arising from the apparent temperature anisotropies on the neutron star surfaces and their probable magnetic nature; see Walter & Lattimer 2002; Braje & Romani 2002; Tiengo & Mereghetti 2007). These measurements have carved out large allowed bands in the mass-radius plane, all of which are also consistent with equations of state that predict neutron stars with radii $R \sim
11$ km. Future tight constraints on the masses and radii of additional neutron stars with these and other methods (see, e.g., Lattimer & Prakash 2007) will resolve this long-standing question of high energy astrophysics.
We thank Rodger Thompson for his help with understanding the NICMOS calibrations, Duncan Galloway for his help with burst analyses, Adrienne Juett for bringing the source to our attention, Jim Lattimer for sharing with us the mass-radius relations, and Martin Elvis for useful conversations on constraining the neutron star equation of state. We also thank an anonymous referee for useful suggestions. F. Ö. acknowledges support from NSF grant AST 07-08640. D. P. is supported by the NSF CAREER award NSF 0746549.
Arnaud, K. A. 1996, Astronomical Data Analysis Software and Systems V, 101, 17
Braje, T. M., & Romani, R. W. 2002, , 580, 1043
Cottam, J., Paerels, F., & Mendez, M. 2002, , 420, 51
Cottam, J., Paerels, F., M[é]{}ndez, M., Boirin, L., Lewin, W. H. G., Kuulkers, E., & Miller, J. M. 2008, , 672, 504
Damen, E., Magnier, E., Lewin, W. H. G., Tan, J., Penninx, W., & van Paradijs, J. 1990, , 237, 103
Galloway, D. K., Psaltis, D., Chakrabarty, D., & Muno, M. P. 2003, , 590, 999
Galloway, D. K., Muno, M. P., Hartman, J. M., Psaltis, D., & Chakrabarty, D. 2006, ApJ, in press (arXiv:astro-ph/0608259)
Galloway, D. K., [Ö]{}zel, F., & Psaltis, D. 2008, , 387, 268
Heinke, C. O., Edmonds, P. D., Grindlay, J. E., Lloyd, D. A., Cohn, H. N., & Lugger, P. M. 2003, , 590, 809
Heinke, C. O., Rybicki, G. B., Narayan, R., & Grindlay, J. E. 2006, , 644, 1090
Inoue, H., et al. 1984, , 36, 855
Lattimer, J. M., & Prakash, M. 2001, , 550, 426
Lattimer, J. M., & Prakash, M. 2007, , 442, 109
Lee, J.-W., Carney, B. W., Fullton, L. K., & Stetson, P. B. 2001, , 122, 3136
Lewin, W. H. G., van Paradijs, J., & Taam, R. E. 1993, Space Science Reviews, 62, 223
Madej, J., Joss, P. C., Rózańska, A. 2004, , 602, 904
Makishima, K., et al. 1981, , 247, L23
Markwardt, C. B., & Swank, J. H. 2000, , 7454, 1
Origlia, L., & Rich, R. M. 2004, , 127, 3422
Ortolani, S., Barbuy, B., Bica, E., Zoccali, M., & Renzini, A. 2007, , 470, 1043
zel, F. 2006, , 441, 1115
Schlegel, D. J., Finkbeiner, D. P., & Davis, M. 1998, , 500, 525
Tiengo, A., & Mereghetti, S. 2007, , 657, L101
van Paradijs, J. 1978, , 274, 650
van Paradijs, J. 1979, , 234, 609
Walter, F. M., & Lattimer, J. M. 2002, , 576, L145
Webb, N. A., & Barret, D. 2007, , 671, 727
Wijnands, R., Heinke, C. O., Pooley, D., Edmonds, P. D., Lewin, W. H. G., Grindlay, J. E., Jonker, P. G., & Miller, J. M. 2005, , 618, 883
![The histogram shows the distribution of $\chi^2/{\rm d.o.f.}$ values obtained by fitting the spectra during the evolution of the two photospheric radius bursts of EXO 1745$-$248. the dashed line shows the expected $\chi^2/{\rm d.o.f.}$ distribution for 25 degrees of freedom. The five spectral fits with $\chi^2/{\rm d.o.f.} > 1.5$ are outliers and are excluded from the subsequent analyses. []{data-label="chisq"}](f2.ps)
 
![The 1- and 2$-\sigma$ confidence contours of the normalization and blackbody temperature obtained from fitting the two PRE bursts during touchdown. The dashed lines show contours of constant bolometric flux.[]{data-label="touchdown"}](f4.ps)
![The $1-$ and $2-\sigma$ contours for the mass and radius of the neutron star in EXO 1745$-$248, for a hydrogen mass fraction of $X=0$, based on the spectroscopic data during thermonuclear bursts combined with a distance measurement to the globular cluster. Neutron star radii larger than $\sim 13$ km are inconsistent with the data. The descriptions of the various equations of state and the corresponding labels can be found in Lattimer & Prakash (2001). []{data-label="mr"}](f5.ps)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider pairing in a dilute system of Fermions with a short-range interaction. While the theory is ill-defined for a contact interaction, the BCS equations can be solved in the leading order of low–energy effective field theory. The integrals are evaluated with the dimensional regularization technique, giving analytic formulas relating the pairing gap, the density, and the energy density to the two-particle scattering length.'
address: |
Institute for Nuclear Theory\
University of Washington, Seattle, WA 98195 USA
author:
- 'T. Papenbrock and G.F. Bertsch'
title: 'Pairing in low-density Fermi gases'
---
\#1[eq. (\[\#1\])]{}
In the theory of fermionic matter, the expansion about the low–density limit has been invaluable for understanding the structure of the theory and the role of the interaction. At low densities, the interaction needs only be characterized by its scattering length to get expansions for the energy density, excitation spectrum, etc.[@ab63]. However, to our knowledge the pairing singularity has never been incorporated into this framework. We have for example only the qualitative statement in ref. [@ab63] that the pairing singularity is logarithmic and unimportant for integrated quantities. A more quantitative statement is needed to have complete understanding of low–density fermionic matter.
Another motivation for our study is the general reexamination of nuclear physics with effective field theory which is now taking place [@Weinberg1; @KoMany; @Parka; @LMa; @GPLa; @Bvk; @KSW; @Hsu]. In the effective field theory approach, the interaction is systematically expanded in a power series in momentum with the object of getting relationships between observables such that the details of the short-distance interaction need not be parameterized. We shall show here that the BCS theory of pairing is amenable to this approach, and the low–energy theory gives finite and analytic results. Within effective field theory many results can be obtained analytically opposed to the numerical treatment of potential models. In this sense our approach complements the large body of literature of pairing in nuclear and neutron matter that is based on potential models [@Wambach89; @Baldo90; @Chen93; @Takat93; @Khodel96; @Carlson97; @Elgaroy].
We consider a Fermi gas with two-fold degeneracy interacting with a short-range attractive interaction. Examples are neutron matter or gaseous $^3$He. The Hamiltonian is idealized to be of the form \[ham\] H = V\_k (a\_[k,]{}\^a\_[k,]{} +a\_[-k,]{}\^a\_[-k,]{}) + gV\^2 a\_[k,]{}\^ a\_[-k,]{}\^a\_[-k’,]{}a\_[k’,]{} where ${\epsilon_k=k^2/2m}$ is the kinetic energy and $V$ the volume. In effective field theory the contact interaction is the leading term in a derivative expansion of the many–body system. This limits the validity of the Hamiltonian (\[ham\]) to the regime of long wave lengths or small densities. However, corrections can systematically be implemented. We have only retained terms in the contact interaction that are needed in the wave function. The BCS wave function has the form $ |\Psi\rangle = \Pi_k (U_k + V_k a_{k,\uparrow}^\dagger
a_{-k,\downarrow}^\dagger)|0\rangle$; the energy is minimized with respect to $U_k,V_k$ to get the BCS equations [@BCS]. The equation for the pairing gap $\Delta$ is \[gapeq\] 1 = -[gV2 (2)\^3]{} , where $\lambda$ is the chemical potential. The density is given in terms of these parameters by \[densdef\] [NV]{} = . Finally, the energy density of the paired state is given by \[edens\] [E V]{} = . Note that the last two integrals are finite, although each integrand is a sum of terms that are individually divergent.
The problem with Eq. (\[gapeq\]) as derived is that the contact interaction is singular in three dimensions. One often introduces a cutoff to make the integrals converge. However, in effective field theory cutoffs are not explicitly introduced. Rather, the computed observables are expressed directly in terms of other physical quantities. To leading order in a low–energy expansion of the interaction, the physical quantity is the scattering length. With the same Hamiltonian, the scattering length $a$ is given by a similar divergent integral, \[scattlen\] -[mgV4a]{}+1 = -[gV2 ]{} Let us now subtract equations (\[gapeq\]) and (\[scattlen\]) to obtain \[master\] [mg4a]{} = - [g2 (2)\^3]{} d\^3k. Notice that the integral is now convergent and so any cutoff can be taken to infinity. Furthermore, the strength of the contact interaction $g$, which is also an unphysical quantity, can be divided out. It is convenient to evaluate both terms of the integral (\[master\]) separately by dimensional regularization (DR) [@Collins]. In DR, integrals of powers are zero so the second term in the integrand in (\[master\]) can be dropped. The first term can be evaluated using [@Gradshteyn] (3.252.11), \[integral\] \_0\^dz = -(1+x\^2)\^[/2]{} P\_(-1/), where $P_\alpha$ denotes the Legendre function.
We write the final result in the form \[scattgap\] [1k\_F a]{} = (1+x\^2)\^[1/4]{} P\_[1/2]{}(-1/) where $k_F = \sqrt{2 m\lambda}$ is the Fermi momentum and $x=\Delta/\lambda$. This is our main result. Eq. (\[scattgap\]) is graphed in Fig. \[fig1\]. For small values of $k_Fa$ the gap is exponentially small as in the usual BCS theory, \[expgap\] = [8e\^2]{}. This comes about by the behavior of $P_{1/2}(z)$, which has a logarithmic singularity at $z=-1$ [@Bateman]. Eq. (\[expgap\]) agrees with the result derived in ref. [@Khodel96]. For large values of $k_Fa$, the gap is proportional to $\lambda$, approaching $\Delta\approx 1.16\lambda$.
For neutron matter the solution of eq. (\[scattgap\]) agrees with numerical results from potential models only for small values of the Fermi momentum. The large value of the scattering length ($a=-18.8$fm) clearly limits the domain of validity of the Hamiltonian (\[ham\]). In the appendix we consider pairing in the effective range approximation. This improves on the precision of the calculation in the low–density regime but does not enlarge the domain of validity.
We complete this discussion by computing the energy density (\[edens\]) and the density (\[densdef\]). The finite integrals involved are very similar to the previous one and can be evaluated using the same DR integral, eq. (\[integral\]). The density of the BCS state is given by \[dens\] [NV]{} = -[k\_F\^3 4]{} (1+x\^2)\^[1/4]{} ,
and the energy density by \[erg\] [EV]{} = -[320]{}k\_F\^3(1+x\^2)\^[1/4]{}. For fixed density and scattering length eqs. (\[scattgap\]) and (\[dens\]) can be solved for the pairing gap and the Fermi energy. Put into eq. (\[erg\]) this yields the energy of the interacting system at fixed density. Fig. \[fig2\] shows a comparison with the energy of noninteracting neutrons. For $|k_Fa|\approx 1$ (i.e. $N/V\approx 5\times 10^{-6}$ fm$^{-3}$) pairing lowers the energy about 3% confirming the qualitative statement that the effects of pairing on the binding are mild.
[*Discussion*]{} — We now discuss the domain of validity of this low–density expansion. As pointed out above, the applicability of Hamiltonian (\[ham\]) is limited to the regime of long wave length $|k_Fa|\ll 1$ or small densities. The description of neutron or nuclear matter at nuclear densities requires the inclusion of the effective range and pions. Comparing with more microscopic calculations involving phenomenological potentials it appears that deviations from the low–density behavior are set by the scattering length. Similar considerations can be made for $^3$He where the scattering length of the Aziz potential [@aziz79] is large on an atomic scale. Many-body correlation effects will become important when $k_F a \approx 1$. It might be possible to treat them by modifying the strength of the pairing and the density of states in eq. (\[ham\]). The sign would be to increase the pairing, but we have not attempted to calculate these effects.
Another consideration is whether the low–density phase exists for fermionic systems with attractive scattering lengths. In the case of $^3$He, a low–density phase could only be metastable at zero temperature, because there is a finite binding of the liquid phase. However, the metastability could be quite significant, because the minimum size for a bound drop is thought to be of the order of fifty particles. Another indication of the metastability of a low–density phase is the sound velocity in the scattering length expansion. Taking the first three terms, the sound velocity is positive at all densities, and thus small deviations from uniformity are energetically unfavorable. In the case of neutron matter, it is thought that pressure is always positive as a function of density, so the low–density state would be stable.
In summary, we have considered the pairing in low–density Fermi systems within effective field theory. This model independent approach yields analytical expressions which relate the pairing gap, the density and the ground state energy to the scattering length. The analytical derivation of these results is quite interesting.
Appendix {#appendix .unnumbered}
========
To include the effective range we add the effective range potential \[efr\] g\_2V\^2(k-k’)\^2a\_[k,]{}\^ a\_[-k,]{}\^a\_[-k’,]{}a\_[k’,]{} to the Hamiltonian (\[ham\]). The gap equation then becomes \[rnggap\] \_p = -[V2]{} \_k and is explicitly momentum dependent. We make the quadratic ansatz $\Delta_p=\Delta + p^2\delta$ for the momentum dependence and obtain two coupled equations that express $\Delta$ and $\delta$ in terms of (divergent) integrals. To deal with the divergencies we observe that the integrals’ dependence on the Fermi momentum is given by =[k\_F\^[2n+1]{}m2\^2]{} J\_[n+1/2]{}(x,y), where $x=\Delta/\lambda$, $y=\delta k_F^2/\lambda$ and $J_{n+1/2}(x,y)$ is the dimensionless function J\_(x,y)\_0\^dt [t\^]{}. In effective field theory an expansion in momenta is quite useful [@KoMany; @KSW]. In what follows we truncate each of the gap equations to its leading order in the Fermi momentum and obtain \[coupl\] 1&=&-[Vgmk\_F4\^2]{}J\_[1/2]{}(x,y),\
&=&-[Vg\_2mk\_F4\^2]{}J\_[1/2]{}(x,y). Obviously we have $\delta/\Delta=g_2/g$. To make contact with low energy scattering data we expand the scattering amplitude \[Aham\] [A]{}(p)=Vgup to quadratic order in momenta. The loop integral is I(p) =[m p4\^2]{}\_0\^dt [t\^[1/2]{}1-t+i]{} At low energies the scattering amplitude is given in terms of the scattering length $a$ and the effective range $r_0$ \[Adata\] [A]{}(p)=[4am]{}. Note that the divergence of the integral $J_{1/2}(x,y)$ appearing in the gap equations (\[coupl\]) is similar to that of the loop integral $I(p)$ appearing in the expression (\[Aham\]) for the scattering amplitude. Thus, both divergencies may be taken care off by a a renormalization of the coupling constants $g$ and $g_2$. We use dimensional regularization to compute the divergent integrals. One obtains $I(p)=-i(m/4\pi)p$ and a comparison of (\[Aham\]) and (\[Adata\]) yields $g_2/g=ar_0/2$ and $g=4\pi a/m$. Finally we have J\_(x,y)= -(1+y\^2)\^[-1/2]{} ([1+x\^21+y\^2]{})\^[/2]{}P\_(-z), where $z= (1-xy)/\sqrt{(1+x^2)(1+y^2)}$. This yields the final results \[gapfin\] [1k\_Fa]{}&=&(1+y\^2)\^[-1/2]{}([1+x\^21+y\^2]{})\^[1/4]{}P\_[1/2]{}(-z),\
y &=& [ar\_02]{}k\_F\^2 x. Note that these equations add corrections of the order $\sim k_F^2ar_0$ to the gap equation (\[scattgap\]). These corrections are small only in the low–density regime $k_Fa \ll 1$. For a description of neutron matter ($a=-18.8$fm $r_0=2.75$fm) at larger densities, at least the inclusion of pions seems to be necessary. Note also, that the gap equations (\[gapfin\]) become singular for $k_F^2\to -2/ar_0$ (i.e. $y\to -x$) due to the logarithmic singularity of the Legendre function for $z\to -1$. This behavior results from the quadratic approximation for the interaction potential and the truncations in the gap equation. It is related to the change in sign of the truncated potential at $k_F=\sqrt{-2/ar_0}$ [@Khodel96]. Again, the introduction of pions or higher potential terms seem to be necessary to alter this behavior.
Acknowledgments {#acknowledgments .unnumbered}
===============
We acknowledge discussions with P. Bedaque, A. Bulgac, H. Grie[ß]{}hammer, D. Kaplan and M. Savage. We thank J. Hormuzdiar and S.D.H. Hsu for pointing out a correction to formula (\[expgap\]). This work was supported by the Dept. of Energy under Grant DE-FG-06-90ER-40561.
A.A. Abrikosov, L.P. Gorkov, and I.E. Dzyaloshinski, Methods of Quantum Field Theory in Statistical Physics, (Prentice-Hall,1963). S. Weinberg, B [**251**]{}(1990)288; Nucl. Phys. B [**363**]{}(1991) 3; B [**295**]{} (1992) 114 C. Ordonez, L. Ray, and U. van Kolck, (1994) 1982; T.S. Park, D.P. Min and M. Rho, (1995) 4153; Nucl. Phys. A [**596**]{} (1996) 515 M. Luke and A.V. Manohar, (1997) 4129 G.P. Lepage, [nucl-th/9706029]{}, Lectures given at 9th Jorge Andre Swieca Summer School: Particles and Fields, Sao Paulo, Brazil, 16-28 Feb 1997. P.F. Bedaque and U. van Kolck, B [**428**]{} (1998) 221; P.F. Bedaque, H.-W. Hammer and U. van Kolck, (1998) R641 D.B. Kaplan, M.J. Savage and M.B. Wise, B [**424**]{} (1998) 390; [nucl-th/9802075]{}, [*to appear in Nucl. Phys. B*]{}; S.D.H. Hsu and J. Hormuzdiar, [nucl-th/9811017]{}
T.L. Ainsworth, J. Wambach, and D. Pines, B [**222**]{} (1989) 173 M. Baldo, J. Cugnon, A. Lejeune, and U. Lombardo, Nucl. Phys. A [**515**]{} (1990) 409 J.M.C. Chen, J.W. Clark, R.D. Dav[é]{}, and V.V. Khodel, Nucl. Phys. A [**555**]{} (1993) 59 T. Takatsuka and R. Tamagaki, Prog. Theor. Phys. Suppl. [**112**]{} (1993) 27 V.A. Khodel, V.V. Khodel, and J.W. Clark, Nucl. Phys. A [**598**]{} (1996) 390 B.V. Carson, T. Frederico, and F.B. Guimaraes, (1997) 3097 Ø. Elgar[ø]{}y and M. Hjorth-Jensen, (1998) 1174 J. Bardeen, L.N. Cooper, and J.R. Schrieffer, Phys. Rev. [**108**]{} (1957) 1175 J.C. Collins, [*Renormalization*]{}, Cambridge Univ. Press, Cambridge (1984) I.S. Gradshteyn and I.M. Rezhik, [*Table of Integrals, Series and Products*]{} (1980) A. Erd[é]{}lyi (Ed.), [*Higher Transcendental Functions*]{}, Vol. I, McGraw Hill, N.Y. (1953) R.A. Aziz, V.P.S. Nain, J.S. Carley, W.L. Taylor, and G.T. McConville; (1979) 4330; A.R. Janzen and R.A. Aziz, (1995) 9626
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The photon-ion merged-beams technique for the photoionization of mass/charge selected ionized atoms, molecules and clusters by x-rays from synchrotron radiation sources is introduced. Examples for photoionization of atomic ions are discussed by going from outer-shell ionization of simple few-electron systems to inner-shell ionization of complex many-electron ions. Fundamental ionization mechanisms are elucidated and the importance of the results for applications in astrophysics and plasma physics is pointed out. Finally, the unique capabilities of the photon-ion merged-beams technique for the study of photoabsorption by nanoparticles are demonstrated by the example of endohedral fullerene ions.'
author:
- |
Stefan Schippers$^{a}$$^{\ast}$[^1], A. L. David Kilcoyne$^b$, Ronald A. Phaneuf$^{\,c}$ and Alfred Müller$^{d}$\
$^{a}$[*[I. Physikalisches Institut, Justus-Liebig-Universität Gie[ß]{}en, 35392 Gie[ß]{}en, Germany]{}*]{}\
$^{b}$[*[Advanced Light Source, Lawrence Berkeley National Laboratory, Berkeley, CA 94720, USA]{}*]{}\
$^{c}$[*[Department of Physics, University of Nevada, Reno, NV 89557-0220, USA]{}*]{}\
$^{d}$[*[Institut für Atom- und Molekülphysik, Justus-Liebig-Universität Gie[ß]{}en, 35392 Gie[ß]{}en, Germany]{}*]{}
title: |
Photoionization of ions with synchrotron radiation:\
From ions in space to atoms in cages
---
photon-ion merged-beams technique, synchrotron radiation, atomic processes in plasmas, laboratory astrophysics, endohedral fullerenes
Introduction
============
Photoionization of ions occurs in many cosmic environments where matter is in the plasma state. Examples are active galactic nuclei, hydrogen clouds around hot stars, and the interior of the Sun. Among other atomic quantities, accurate photoionization cross sections of atomic ions are required for being able to infer the physical conditions of these objects from astronomical observations [@Kallman2010]. An entire field of research, Laboratory Astrophysics [@Savin2012], is devoted to providing the atomic data that are required for the understanding of cosmic nonequilibrium plasmas. The Opacity Project, for example, aims at calculating accurate cross sections for the photoionization of ions which are needed for estimating the opacity of stellar matter (see [@Badnell2005a] and references therein). In general, the atomic data needs for the modeling of nonequilibrium plasmas are vast. Most of the body of compiled data comes from theory which has to use approximations to make calculations tractable even with modern computer technology at hand. Therefore, benchmarking by experiment is vitally needed.
A problem from solar physics illustrates this situation. The Sun produces energy by nuclear fusion reactions which proceed most efficiently at the high temperatures and pressures that prevail in the core of the Sun. The mechanism of energy transport to the Sun’s surface depends on the radial distance from the core. At smaller radii, where the solar matter is largely ionized, energy is transported by radiation. At larger distances from the core, where the solar matter is less dense, energy is transported by convective mass flow of mainly neutral hydrogen. Figure \[fig:TheSun\] visualizes the shell structure of the Sun. The radius $R_\mathrm{CZ}$ of the boundary between the radiation and the convection layers can be inferred from helioseismology [@Chaplin2005] and from solar models [@Bahcall1988] that quantitatively treat energy generation and transport in the Sun. Ingredients of the solar models are, among others, the photoionization cross sections from the Opacity Project.
The problem is now that the value $R_\mathrm{CZ}=0.713\pm0.001 R_\odot$ from helioseismology and the values of up to $R_\mathrm{CZ}=0.72 R_\odot$ from the latest solar models significantly differ from one another [@Basu2014 and references therein]. Similar discrepancies exist also for other solar quantities of interest as, e.g., the flux of neutrinos. It has been speculated [@Serenelli2009; @Bailey2015] that part of the discrepancies might be caused by inaccurate photoionization cross sections. Theoretically derived photoionization cross sections from the Opacity Project are partly accessible via the TOPbase data base [@Cunto1993]. Figure \[fig:FeAstrid\] compares experimentally measured cross sections for the photoionization of low-charged iron ions with the corresponding quantities from the TOPbase. The very significant differences between the experimental and the theoretical cross section curves clearly illustrate the need for experimental benchmarking.
The experimental data in Figure \[fig:FeAstrid\] were obtained with the photon-ion merged-beams technique which is explained in some more detail in section \[sec:exp\] of the present paper. Section \[sec:atomic\] presents and discusses selected results for atomic ions of astrophysical interest. In recent years, the photon-ion merged-beams technique has also been used for photoionization studies with cluster ions such as (endohedral) fullerene ions [see, e.g., @Phaneuf2013]. One particulary illustrative result for Xe@C$_{60}^+$ is presented in section \[sec:full\]. The present paper is intended to serve as an introduction for the non-specialized reader into the topic of photoionization of ions. It does not provide a comprehensive overview over this field of research. For this, the reader is referred to the review articles by West [@West2001a], Kennedy et al. [@Kennedy2004a], Kjeldsen [@Kjeldsen2006a], Berrah et al. [@Berrah2011] and Müller [@Mueller2015].
The photon-ion merged-beams technique {#sec:exp}
=====================================
Ionization energies [@Kramida2014] of positively charged atomic ions range from about 10 eV for Ba$^+$ up to 132 keV for hydrogen-like U$^{91+}$. Thus, experiments on photoionization of atomic ions require UV and X-ray photons. Sufficiently intense beams of such photons only became available with the advent of 2nd generation synchrotron light sources in the late 1970s. The first experiments on the photoionization of ions were carried out in the mid 1980s by Lyon et al. [@Lyon1986] who pioneered the photon-ion merged-beams technique at the Daresbury Synchroton Radiation Source in the UK. Since then, the technique has been implemented at other synchrotron radiation sources around the world including ASTRID (Denmark) [@Kjeldsen1999b], Spring-8 (Japan) [@Yamaoka2001], ALS (USA) [@Covington2002], Soleil (France) [@Gharaibeh2011a], and PETRAIII (Germany) [@Schippers2014].
A particular difficulty associated with ionic targets is their extreme diluteness. The number of ions per unit volume is limited by the mutual electrostatic repulsion of the ions. Typical ion number densities in an ion beam are about $10^6$ cm$^{-3}$. This is 10 orders of magnitude smaller as compared to the number density of neutral atoms or molecules in a gas target with a pressure of 1 mbar. The photon-ion merged-beams technique makes up for the diluteness of the ionic target by providing a large interaction volume and a high efficiency for the detection of product ions.
Figure \[fig:MB\] sketches the salient features of the experimental arrangement and Figure \[fig:PIPE\] presents details of the layout of the PIPE facility [@Schippers2014] at PETRAIII. A beam of $q$-fold charged ions (denoted by A$^{q+}$) is produced by an ion source which is kept on an acceleration potential $U_\mathrm{acc}$ of a few kV. Thus, after acceleration, the ion kinetic energy is $E_\mathrm{ion} = \frac{1}{2}m_\mathrm{ion}v_\mathrm{ion}^2 = qeU_\mathrm{acc}$ with $e$ denoting the elementary charge. The ion source usually delivers ions in different charge states. In addition, impurity ions from the residual gas in the ion source may be present in the beam. Therefore, a dipole bending magnet in combination with beam-size defining slits (labelled mass/charge analyzer in Figure \[fig:MB\]) is used for selecting ions of the desired mass-to-charge-ratio to be transported further to the photon-ion interaction region. The resolving power of such an arrangement can be adjusted (at the expense of beam intensity) by closing the beam-size defining slits. Even different isotopes of a given ion species can be individually selected if required (Figure \[fig:Xe2mass\]).
The large interaction volume is created by deflecting the ion beam onto the photon-beam axis. The length of the merging region is typically 1–2 m depending on ion-optical requirements and on the floor space available. The high efficiency for the detection of product ions is due to the use of a beam of fast ions as a target for photoionization. The photo ions that are created in the photon-ion interaction region move in the same direction as the primary ion beam. A second magnetic or electrostatic deflector that is located downstream of the interaction region demerges ion and photon beams and at the same time separates the more highly charged product ions from the primary ions. The primary beam current is measured with a Faraday cup. The photo ions are counted one by one in a single particle detector. When carefully designed, such a detector has an efficiency of practically 100% [@Rinn1982; @Spruck2015].
The well defined beam-beam collision geometry of the merged-beams technique allows an accurate determination of absolute photoionization cross sections $\sigma$ (see, e.g., [@Schippers2014]) by normalizing the measured detector count rate $R$ on the photon flux $\phi_\mathrm{ph}$, the number of ions in the interaction region, the beam overlap, and detection efficiency $\eta$, i.e., $$\label{eq:sigma}
\sigma = R \frac{q\,e\,v_\mathrm{ion}}{\eta\,I_\mathrm{ion}\,\phi_\mathrm{ph}\,\mathcal{F}_L}$$ where $I_\mathrm{ion}$ denotes the electrical ion current. The beam overlap factor $\mathcal{F}_L$ for the length $L$ of the beam overlap region is determined by separate beam-profile measurements using scanning slits or rotating wires. The systematic uncertainty of the measured absolute cross section amounts to typically 10–15% with the largest contribution to the error budget stemming from the beam overlap measurement [see, e.g., @Covington2002].
In principle, the photon-ion merged-beams technique enables studies of the photoionization of any ion including negatively charged species [@Berrah2011] and complex molecular ions (see Section \[sec:full\]). The maximum charge state that can be investigated at a given setup depends on the capabilities of the ion source in use and on the maximum photon energy available (e.g., 340 eV at the ALS [@Covington2002], 1000 eV at SOLEIL [@Gharaibeh2011a], and 3000 eV at PETRAIII [@Schippers2014]). The most highly charged ion investigated at a photon-ion merged-beams setup so far is Ce$^{9+}$ [@Habibi2009]. Atomic ions with ionization energies below 3000 eV which can, thus, be photoionized in the PIPE setup are, e.g., H-like Si$^{13+}$, Li-like Ga$^{28+}$, Na-like Sb$^{40+}$, or Ge-like Pb$^{50+}$ [@Kramida2014]. A powerful electron-cyclotron-resonance (ECR) ion source [@Trassl2003a] would be capable of delivering such ions with sufficiently high ion currents.
It should be noted that there are competing methods for the experimental investigation of photon-ion interactions in addition to the merged-beams technique. The dual laser plasma (DLP) method [@Kennedy2004a] uses two plasmas which are generated by two consecutive laser pulses. The UV emission from the first plasma backlights the second plasma that contains the atomic ions of interest. Thus, the DLP technique allows for absorption spectroscopy of multiply charged ions. However it cannot easily provide absolute cross sections, since the target volume and density as well as the intensity of the ionizing radiation field are not sufficiently well under control. More recently, photon-ion interactions have been studied using trapped ions [@Thissen2008a; @Lau2008; @Simon2010a; @Bari2011; @Milosavljevic2012; @Rudolph2013] partly in combination with VUV and X-ray free electron lasers [@Epp2007; @Bernitt2012]. Generally, it is not easy to obtain absolute cross sections from trapping techniques since the trap inventory can be a mixture of different ions and/or the overlap between the trapped ion cloud and the photon beam cannot be determined with sufficient accuracy. In this respect, the photon-ion merged-beams technique provides better defined experimental conditions than the competing methods.
Ions in space: Photoionization of atomic ions {#sec:atomic}
=============================================
Hydrogen-like ions with just one bound electron are the conceptually simplest atomic systems that can be studied by photoionization. The theoretical cross section for this fundamental process was derived already in the early days of quantum mechanics [@Stobbe1930]. For the photoionization of a $1s$ electron it is $$\label{eq:Stobbe}
\sigma(h\nu) =\frac{512\pi^2\alpha}{3(\varepsilon\!+\!1)^4}\frac{e^{-(4\arctan\sqrt{\varepsilon})/\sqrt{\varepsilon}}}{1-e^{-2\pi/\sqrt{\varepsilon}}}\left(\frac{a_0}{Z}\right)^2$$ with the photon energy $h\nu$, the fine structure constant $\alpha \approx 1/137$, the scaled energy $\varepsilon = h\nu/(I_\mathrm{H}Z^2)-1 > 0$, the ionization energy of hydrogen $I_\mathrm{H}\approx 13.6$ eV, the nuclear charge $Z$, and the Bohr radius $a_0 \approx 0.529\times10^{-10}$ m. The cross section is zero below the ionization threshold ($I_\mathrm{H}Z^2$, i.e., $\varepsilon=0$) it jumps to a finite value at the threshold and monotonically decreases as the photon energy is further increased. Figure \[fig:Heabs\] shows the calculated cross section for photoionization of He$^+$ ions together with corresponding experimental data. The theoretical cross section agrees with the measured data points within the experimental uncertainties of $\pm 15\%$.
Photoionization becomes more complicated as soon as the target ion has more than one electron. Then, the photon energy can be shared between several electrons leading to multiple excitation or even multiple ionization. In case of (multiple) excitation the photon energy has to match the energy difference between the initial and excited atomic levels. Thus excitation is only possible at discrete photon energies. At these energies, photoionization resonances may be observed if the multiply excited state subsequently decays by autoionization. It should be noted that the excitation of a single outer-shell electron does not lead to the formation of an autoionizing state. Therefore, He$^+$($1s\to n\ell$) single-excitation resonances are not observed in the He$^{2+}$ photo-ion spectrum displayed in Figure \[fig:Heabs\], for example. Figure \[fig:B1rmat\] displays the photoionization cross section of Be-like B$^+$($1s^2\,2s^2$), a quasi two-electron system[^2]. The observed resonances are due to $2s^2\to2p\,nl$ double excitation with subsequent autoionization into the B$^{2+}$($1s^2\,2s$) ground configuration. Two distinct Rydberg series of resonances can be discerned, the $2p\,ns\;^2P^o$ and the $2p\,nd\;^2P^o$ series, both converging to the same series limit at $E_\infty \approx 31.15$ eV. To a very good approximation the positions of the $2p\,nl$ resonances can be estimated from the Rydberg formula $$\label{eq:Rydberg}
E_n = E_\infty-I_\mathrm{H}\frac{(q+1)^2}{(n-\mu)^2}$$ where $q=1$ in the case of B$^+$ and the quantum defects $\mu = 0.408$ and $\mu = -0.087$ for the $2p\,ns\;^2P^o$ and $2p\,nd\;^2P^o$ series, respectively, were obtained from fitting Equation (\[eq:Rydberg\]) to the measured resonance positions [@Schippers2003b].
In addition to the resonant ionization channels, there is nonresonant direct ionization of a $2s$ electron leading to a smooth background cross section similar to the one depicted in Figure \[fig:Heabs\]. The B$^+$(2s) ionization threshold at 25.15 eV is outside of the photon-energy range of Figure \[fig:B1rmat\]. The pathways for $2s$ photoionization of B$^+$ can be written as $$\begin{aligned}
\textrm{B}^+(1s^2\,2s^2) &\to& \textrm{B}^+(1s^2\,2p\,nl)\nonumber\\
& \searrow & \downarrow\label{eq:pathways}\\
& & \textrm{B}^{2+}(1s^2\,2s)\nonumber\end{aligned}$$ Both, the pathway via $2p\,nl$ intermediate resonance states and the nonresonant pathway of direct ionization lead from the same initial to the same final state. In such a situation the quantum-mechanical amplitudes associated with both pathways interfere. This interference phenomenon leads to the manifestly asymmetric resonance line shapes in Figure \[fig:B1rmat\]. In case of the $2p\,ns$ resonances the interference is largely destructive leading to a reduced cross section in the vicinity of the resonance energies. The (red) full line in Figure \[fig:B1rmat\] is the result of a calculation [@Schippers2003b] within the nonperturbative theoretical framework of R-matrix theory. The overall agreement with the experimental data is quite satisfactory. The asymmetric resonance line shapes are reproduced very well. The same level of agreement was also found in a comprehensive study on valence-shell photoionization of the heavier Be-like ions C$^{2+}$, N$^{3+}$ and O$^{4+}$ [@Mueller2010a].
As far as absolute cross sections are concerned, the comparison between experiment and theory is often complicated by the fact that not all ions in an ion beam are in their ground level. For example, the $^3P_J$ levels ($J=0,1,2$) in Be-like ions are sufficiently long lived such that they will reach the photon-ion interaction region if they are populated in the ion source. In a photoionization experiment the presence of ions initially in metastable states is revealed by a nonzero ionization signal at photon energies below the threshold for the ionization of the ground level. Usually, the fraction of metastables in an ion beam is unknown. For Be-like C$^{2+}$, N$^{3+}$, and O$^{4+}$ [@Mueller2010a] the metastable fraction amounted to about 50% as was inferred from the comparison with the corresponding R-matrix results.
In exceptional cases, pure ground-state ion beams can be prepared by passing the ion beam through a gas cell where metastable ions are quenched by collisions [@Kjeldsen2002a; @Aguilar2003b]. Surface ionization sources produce selected singly charged ions, such as Li$^+$ [@Scully2006a], in their ground level. A more general approach is to store the ions for a sufficiently long time before exposing them to the ionizing photon beam. In a few experimental arrangements, ion traps are used for this purpose [@Thissen2008a; @Wolf2010]. A draw-back of the trapping technique is that the statistical quality of the results suffers severely from the massive particle losses from the trap during the ion-storage time. The most versatile approach would be the implementation of the photon-ion merged-beams technique at a heavy-ion storage-ring where long ion-storage times can be achieved without excessive beam losses [see, e.g., @Grieser2012]. Its realization would require a major investment and much more floor space than is usually available at synchrotron-radiation facilities. Although heavy-ion storage-rings are presently not available for photoionization studies, the time-inverse process, i.e., photorecombination of ions, is routinely studied at these facilities [@Mueller2008a; @Schippers2015]. Again, in exceptional cases, the corresponding cross sections can be compared with photoionization cross sections using the principle of detailed balance, and metastable fractions in the photoionization experiment can be inferred from such comparisons on a purely experimental basis [@Schippers2002b; @Mueller2009a; @Mueller2014a].
Despite the experimental complication concerning metastable ions, conclusive comparisons between experimental and theoretical results (such as the one shown in Figure \[fig:B1rmat\]) could be made for a number of atomic ions. Generally, state-of-the-art theoretical methods have been found to be capable of providing cross sections for photoionization of light and few-electron ions with sufficient accuracy for astrophysical purposes. More complex ions are computationally more demanding and approximations must be made in order to keep the calculations tractable. The related uncertainties are difficult to assess and usually only uncovered when theoretical cross sections are compared with experimental data. The large discrepancies between theoretical and experimental data that are revealed in Figure \[fig:FeAstrid\] for the photoionization of low-charged iron ions are largely due to the simplifications imposed by the computer technology that was available when the Opacity Project calculations were carried out in the late 1980s.
Meanwhile, computer technology has greatly advanced and more refined large-scale calculations have become feasible [e.g. @Fivet2012; @Ballance2015] allowing for a comprehensive treatment of, e.g., K-shell photoionization of light ions. These inner-shell processes are particularly important for opacity calculations [@Badnell2003b]. As an example Figure \[fig:B1Kshell\] shows measured and calculated cross sections for K-shell ionization of B$^+$ ions [@Mueller2014a]. Experimental and theoretical resonance positions, widths and strengths agree within the experimental uncertainties with the exception of the strength of the resonance at about 194.5 eV. This kind of agreement of state-of-the-art theory and experiment can be considered as typical for K-shell ionization of light ions including the astrophysically most important C [e.g. @Mueller2009a], N [e.g. @Gharaibeh2011a], and O [e.g. @Bizau2015] isonuclear sequences of ions. Next to hydrogen and helium these elements have the largest abundances in the solar system [@Asplund2009].
On the experimental side, measurements of K-shell photoionization are challenging since the cross sections are rather small except for the few strongest resonances such that only these could be investigated hitherto. Figure \[fig:B1Kshell\] provides a typical example. Smaller resonances at higher energies yield increasingly lower signal count rates and their study becomes impractical given the limited amount of beamtime at synchrotron radiation facilities. The experimental means to increase the count rate in a given experimental situation are limited. According to Equation (\[eq:sigma\]), one can try to increase either the ion current or the photon flux. The former parameter is limited by the performance of the ion source (which strongly depends on the ion species under investigation) and the latter can only be increased at the expense of resolving power within the limits imposed by the layout of the photon source and the photon beamline.
A breakthrough has been achieved recently with the implementation of the photon-ion merged-beams technique at the world’s presently brightest 3rd generation synchrotron light source, PETRAIII in Hamburg, Germany. At the PIPE setup [@Schippers2014] (Figure \[fig:PIPE\]) record-high photon fluxes are available over an extended photon energy range [@Viefhaus2013]. At the same time, the design of the ion beamline was geared towards efficient suppression of background due to stray particles and photons and, thus, towards high signal/background ratios. By these means, the quality of the data could be increased dramatically as is apparent from the comparison of one of the first results from the PIPE setup for C$^+$ [@Mueller2015a] (Figure \[fig:C1K\]) with a typical result from previous K-shell ionization studies of B$^+$ (Figure \[fig:B1Kshell\]).
As already mentioned, in these previous studies only the contribution of the strongest resonances to the single ionization channel were investigated. In contrast, the PIPE data for K-shell ionization of C$^+$ ions (Figure \[fig:C1K\]) comprise entire Rydberg series of resonances not only in the single but also in the double ionization channel. A prominent additional contribution to the latter channel is nonresonant direct ionization of a $1s$ electron and subsequent autoionization leading to the rise of the cross section beyond the lowest threshold for such a process at about 315 eV. The $1s\,2s^2\,2p\,np$ resonances that are formed via $1s\to np$ excitation of the C$^+$ ion and subsequent autoionization contribute to both ionization channels. For the strongest members of this Rydberg series, i.e., for the $1s\,2s^2\,2p^2\;^2D$ and $1s\,2s^2\,2p^2\;^2P$ resonances at about 288 eV, even triple ionization could be measured. The relative contributions of these resonances to the single, double and triple ionization channels depends on the respective branching ratios for autoionization of the C$^+$($1s\,2s^2\,2p^2$) resonance states into the C$^{2+}$($1s^2\,2l^2$), C$^{3+}$($1s^2\,2l$), and C$^{4+}$($1s^2$) final states, respectively. In this particular case, the higher charge states can only be formed when the respective Auger processes lead to the simultaneous emission of two or even three electrons (Figure \[fig:123Auger\]). The observed ratios of triple- to double- to single-Auger decay rates are of the order of $10^{-4} : 10^{-2} : 1$. In fact, by measuring all cross sections on an absolute scale the rate for the triple Auger process — representing a genuine four-electron interaction — could be quantified for the first time [@Mueller2015a].
The relatively light ions discussed up to now have in common that their photoionization resonance structure can be rather easily interpreted in terms of Rydberg series \[cf. Equation (\[eq:Rydberg\])\]. The resonance structure of heavier ions with several open electron shells can be very complex due to the associated high density of excited electronic levels. When closely spaced neighbouring resonances overlap the photoionization cross section exhibits broad irregularly shaped features (see, e.g., Figure \[fig:FeAstrid\]) instead of individually resolved resonance peaks. Theoretical calculations of these cross sections are demanding because large wave function expansions are required for an accurate description of the resonance levels.
A rather extreme case that has been studied recently by experiment and theory is photoionization of singly charged tungsten ions [@Mueller2015c] (Figure \[fig:W1PI\]). At present, tungsten receives much attention in atomic physics because of its technological importance in nuclear fusion reactors [@Mueller2015b and references therein]. The ground level of W$^+$ with an ionization energy of 16.35 eV is designated as \[Xe\]$4f^{14}\,5d^4\,(^5D)\,6s\;^6D_{1/2}$ [@Kramida2006]. However, the $5d^4\,6s$ ground configuration is strongly mixed with the neighbouring $5d^3\,6s^2$ and $5d^5$ configurations of the same parity. The associated 13 terms with in total 119 levels are all metastable and were most likely all populated in the primary ion beam. In the photon energy range of up to 245 eV photoionization proceeds via excitation or ionization of a $4f$, $5s$, $5p$, $5d$, or $6s$ electron.
Figure \[fig:W1PI\] compares the experimental cross section for photoionization of W$^+$ with the theoretical results for the three lowest terms. Although, the overall shapes of all these cross section curves are very similar it is obvious that a conclusive comparison between theory and experiment cannot easily be made. Over much of the displayed photon energy range, the calculated theoretical cross sections are consistently larger than the experimental cross section by factors 2–3 or more (in particular, at higher energies). At present, the origin of these discrepancies is not clear. It could be conceived that the not yet calculated cross sections for the higher terms might be considerably smaller and that averaging over all terms would bring the theoretical cross section down. Another explanation could be that the wave function expansions are still too limited for an accurate description of the photoionization process. The situation may be similar to electron-ion recombination of complex tungsten ions where statistical theory could successfully be applied for a satisfying theoretical description of the recombination cross section [@Spruck2014a and references therein]. This novel theoretical approach has recently been discussed also in the context of photoionization [@Flambaum2015], but it has not yet been applied to a specific case.
As compared to outer-shell ionization, inner-shell ionization of many-electron ions tends to be less complex since the ionized or excited electron is released from a closed atomic subshell. Accordingly the number of relevant levels is moderate. For example, Figure \[fig:Xe3dhires\] shows experimental cross sections for triple photoionization of low-charged xenon ions in the vicinity of the threshold for the ionization of a $3d$ inner-shell electron [@Schippers2015a]. All spectra exhibit a continuous cross section due to direct ionization of the $3d$ electron. However, the strongest features in the spectra are resonances associated with the excitation of the $3d$ electron to an empty $nf$ subshell. There are two corresponding Rydberg series because of the $j$=3/2 – $j$=5/2 fine-structure splitting (about 13 eV) of the $3d$ core hole.
With increasing charge state increasingly higher $nf$ Rydberg resonances can be excited. This is related to the so called collapse of the $nf$ wave functions and can be explained by the peculiar shape of the potential of $f$-electrons. The potential consists of an inner and an outer well separated by a centrifugal barrier. For low nuclear charges, the inner well is too shallow to confine the $nf$ electrons, so that they are mainly localized in the outer well where there is practically no overlap with the $3d$ wave function. When the charge is increased the inner potential well becomes deeper and thus capable of supporting increasingly higher $nf$ wave functions which then have a larger overlap with the $3d$ wave function. Correspondingly, strong resonances associated with $3d\to nf$ excitations are observed for those $nf$ shells which have collapsed into the inner well. Although theoretical calculations strongly corroborate this picture [@Schippers2015a], they cannot easily predict the measured triple ionization cross section since the many possible deexcitation pathways (mainly Auger cascades leading to a distribution of final charge states [@Schippers2014]) that open up after the initial creation of the core-hole cannot easily be taken into account [@Fritzsche2012a].
Atoms in cages: Photoionization of endohedral fullerene ions {#sec:full}
============================================================
The application of the photon-ion merged-beams technique is by no means limited to atomic ions. In fact, the in-situ mass/charge-selection of ions allows also for experiments with molecular ions [e.g., @Hinojosa2005a] and mass-selected cluster ions. This is an advantage compared to experiments with neutral clusters where the size distribution of clusters often cannot be tightly controlled [@Ruehl2003b]. At the same time, sample purity is not an issue since unwanted beam components can be efficiently suppressed by the in-situ mass/charge analysis. The classes of cluster ions that have been most intensely studied so far are charged fullerenes [@Scully2005a; @Bilodeau2013] and endohedral fullerenes [@Mueller2008b; @Kilcoyne2010; @Phaneuf2013a; @Hellhund2015].
Endohedral fullerenes are fascinating objects. They consist of a cage of carbon atoms surrounding encapsulated atoms or small molecules (Figure \[fig:endohedral\]). The fact that the encapsulated atoms or molecules are rather isolated from the surroundings outside of the cage has given rise to interesting ideas for possible applications in many diverse fields such as quantum computing, superconductivity, photovoltaics, medical imaging, and tumor suppression, to name just a few [@Popov2013 and references therein].
Endohedral fullerenes are also interesting from a fundamental point of view. A particularly intriguing theoretical prediction concerns photoabsorption by encaged atoms. Figure \[fig:Xe@C60andXe\] shows the experimental photoabsorption cross section of a free Xe(\[Kr\]$\,4d^{10}\,5s^2\,5p^6$) atom (dashed line) in the photon energy range where photoabsorption by the atomic 4d shell yields the dominant contribution to the total cross section. In this energy range the cross section exhibits a giant broad resonance. The theoretical prediction [@Puska1993] for photoabsorption of a xenon atom encapsulated in a C$_{60}$ fullerene cage is shown as a full red line. Accordingly and somewhat surprisingly, the broad resonance is split into several narrower resonances. These confinement resonances [@Connerade2000] can be understood in terms of multi-path interference between electron waves that are emitted directly from the Xe@C$_{60}$ molecule and electron waves that are bouncing on the inner walls of the carbon cage before they are emitted.
Although this phenomenon was investigated theoretically at different levels of sophistication for more than a decade, an experimental verification by a photoabsorption experiment with endohedral fullerenes in the gas phase was achieved only recently [@Kilcoyne2010; @Phaneuf2013a]. This long delay is due to the challenging experimental difficulties that had to be overcome before conclusive results could be measured. The most direct approach for bringing endohedral fullerenes into the gas phase is evaporation of solid material. Unfortunately, endohedral fullerenes cannot be synthesized in sufficiently large amounts for obtaining vapour targets with a sufficiently high density for detailed photoabsorption studies. Corresponding attempts with rare-earth atoms encapsulated in a C$_{82}$ cage [@Mitsuke2005a; @Katayanagi2008] have not been very conclusive. Much less material is required for the production of an endohedral fullerene ion beam, in particular, if an ion source is used which can be operated at low vapour pressures.
Figure \[fig:XeC60mass\] shows mass spectra obtained from samples containing mainly C$_{60}$ as well as traces of larger fullerenes such as C$_{70}$ and of Xe@C$_{60}$. The latter was produced by implantation of Xe$^+$ ions into a layer of 99.95% pure C$_{60}$ [@Kilcoyne2010; @Phaneuf2013a]. The samples were evaporated and ionized in an ECR ion source. The displayed mass range comprises Xe@C$_{60}$, C$_{70}$ and their heavier fragmentation products that are created in the ion source, with (endohedral) fullerene fragmentation occurring through the loss of one or several C$_2$ units from the carbon cage. Therefore, only fragments containing even numbers of carbon atoms are observed (right panels of Figure \[fig:XeC60mass\]). The strongest peaks correspond to intact C$_{70}^+$ and Xe@C$_{60}^+$ molecular ions. The fact that the yield of Xe@C$_{60}^+$ is comparable to the yield of the impurity ion C$_{70}^+$ shows that the Xe@C$_{60}$ production process was not very efficient. Xenon atoms could be implanted into at best only 1 out of 5000 C$_{60}$ molecules. Nevertheless, an ion beam of pure Xe@C$_{60}^+$ could be produced. In order to be able to optimally separate Xe@C$_{60}^+$ from C$_{70}^+$ isotopically pure $^{136}$Xe was used in the implantation process (lower panels of Figure \[fig:XeC60mass\]). The use of xenon with a natural isotope distribution (cf. Figure \[fig:Xe2mass\]) results in a wide mass distribution of several Xe@C$_{60}$ isotopologues as revealed by a corresponding high-resolution mass scan of the Xe@C$_{60}^+$ peak (upper left panel of Figure \[fig:XeC60mass\]). The number of isotopologues is greatly reduced in the case of $^{136}$Xe@C$_{60}$ (lower left panel of Figure \[fig:XeC60mass\]). These rather involved beam preparation procedures eventually resulted in $^{136}$Xe@C$_{60}^+$ ion currents of up to 5 pA. This corresponded to a number of only $N_\mathrm{ion} = I_\mathrm{ion}L/(ev_\mathrm{ion}) \approx 40$ endohedral molecular ions at a time in the photon-ion interaction region, yielding photo-product count rates of the order of 1 s$^{-1}$ in the cross section maximum \[cf. Equation (\[eq:sigma\])\].
Figure \[fig:XeC60exp\] shows the cross section that was measured under these conditions in comparison with the cross section for photoabsorption of atomic xenon. The cross section for Xe@C$_{60}$ exhibits clear signatures of confinement resonances albeit less pronounced than theoretically predicted (Figure \[fig:Xe@C60andXe\]). There is agreement about the number of resonance maxima (four), but there are differences in resonance positions and sizes. These discrepancies arise due to the rather approximate theoretical treatment of the carbon cage as an effective potential well for the outgoing photoelectron. More refined calculations yield somewhat better but still not perfect agreement between experiment and theory [@Phaneuf2013a].
In any case, the predicted phenomenon of multi-path interference in the photoionization of an encaged atom has unambiguously been verified by experiment. It should be noted that the visibility of the confinement resonances relies on the central position of the encapsulated atom inside the cage (Figure \[fig:endohedral\]). In a less symmetric geometry where the atom is located off-center the interferences are washed out when averaged over the entire solid angle [@Korol2011; @Chen2014]. For example, no signs of confinement resonances where observed in the photoabsorption of Ce@C$_{82}^+$ [@Mueller2008b] where the cerium atom is attached to the inner wall of the carbon cage.
Summary and outlook
===================
Experimental studies of photoionization of ionized matter requires energetic photons which are readily available from 3rd generation synchrotron light sources. Here we have introduced the photon-ion merged-beams technique that offers a high sensitivity for heavy photo product particles and, thus, partly makes up for the diluteness of ionic targets. So far this technique has been used predominantly for the photoionization of atomic ions, a topic that is strongly driven by atomic data needs in astrophysics and plasma physics. The field has matured in recent years and a satisfying level of theoretical understanding has been reached for few-electron ions. For more complex systems there remain formidable challenges both on the experimental and the theoretical side. The biggest experimental challenge concerns the presence of unknown fractions of metastable ions in the ion beam. Beam storage in a heavy-ion storage ring would provide a solution, but has not been realized so far because it would require a major investment. Recent progress in this direction has been made by the construction of a low-cost electrostatic storage ring for photon-ion interaction experiments, in particular, for the photofragmentation of small molecular ions [@Pedersen2015]. However, so far the ion storage times in this device are still rather short. The experiments with endohedral fullerenes have driven the photon-ion merged-beams technique to new limits of sensitivity. Experiments have been conducted successfully with minuscule amounts of sample material. The door to future studies with other types of charged nanoparticles is thus wide open.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank all our collaborators who have participated in the experiments at the Advanced Light Source in Berkeley and at PETRAIII in Hamburg or who have contributed by providing theoretical calculations. Their names appear in the list of references below. A.M. and S.S. are particulary indebted to Ticia Buhr, Sandor Ricz, and Heinz-Jürgen Schäfer for their invaluable contributions to the realization of the PIPE setup. Financial support over the years from the US Department of Energy (DOE, grant nos. DE-FG02-03ER15424 and DE-AC03-76SF0098), from the German Ministry of Education and Research (BMBF, grant nos. 05KS7RG1 and 05K10RG1), from Deutsche Forschungsgemeinschaft (DFG, grant nos. Mu1068/10, Mu1068/20, and Mu1068/22), and through the NATO Collaborative Linkage Grants 950911 and 976362 is gratefully acknowledged.
[80]{} \[1\][\#1]{}
T. Kallman, [*Modelleing of photoionized plasmas*]{}, Space Science Reviews 157 (2010), pp. 177–191.
D.W. Savin, N.S. Brickhouse, J.J. Cowan, R.P. Drake, S.R. Federman, G.J. Ferland, A. Frank, M.S. Gudipati, W.C. Haxton, E. Herbst, S. Profumo, F. Salama, L.M. Ziurys, and E.G. Zweibel, [*The impact of recent advances in laboratory astrophysics on our understanding of the cosmos*]{}, Rep. Prog. Phys. 75 (2012), p. 036901.
N.R. [Badnell]{}, M.A. [Bautista]{}, K. [Butler]{}, F. [Delahaye]{}, C. [Mendoza]{}, P. [Palmeri]{}, C.J. [Zeippen]{}, and M.J. [Seaton]{}, [*[Updated opacities from the Opacity Project]{}*]{}, Mon. Not. R. Astron. Soc. 360 (2005), pp. 458–464.
W.J. Chaplin and I. Ballai, [*The sound of the [S]{}un*]{}, Astron. Geophys. 46 (2005), pp. 4.27–4.30.
J. Bahcall and R. Ulrich, [*Solar models, neutrino experiments, and helioseismology*]{}, Rev. Mod. Phys. 60 (1988), pp. 297–372.
S. Basu, N. Grevesse, S. Mathis, and S. Turck-Chièze, [*Understanding the internal chemical composition and physical processes of the Solar interior*]{}, Space Science Review (2014), pp. 1–29.
A.M. Serenelli, S. Basu, J.W. Ferguson, and M. Asplund, [*New Solar composition: the problem with Solar models revisited*]{}, Astrophys. J. 705 (2009), p. L123.
J.E. Bailey, T. Nagayama, G.P. Loisel, G.A. Rochau, C. Blancard, J. Colgan, P. Cosse, G. Faussurier, C.J. Fontes, F. Gilleron, I. Golovkin, S.B. Hansen, C.A. Iglesias, D.P. Kilcrease, J.J. MacFarlane, R.C. Mancini, S.N. Nahar, C. Orban, J.C. Pain, A.K. Pradhan, M. Sherrill, and B.G. Wilson, [*A higher-than-predicted measurement of iron opacity at solar interior temperatures*]{}, Nature 517 (2015), pp. 56–59.
W. Cunto, C. Mendoza, F. Ochsenbein, and C.J. Zeippen, [*Topbase at the [CDS]{}*]{}, Astron. Astrophys. 275 (1993), pp. L5–L8 http://cdsweb.u-strasbg.fr/topbase/topbase.html.
N. [El Hassan]{}, J.M. Bizau, C. Blancard, P. Cosse, D. Cubaynes, G. Faussurier, and F. Folkmann, [*Photoionization cross sections of iron isonuclear sequence ions: [Fe$^{2+}$]{} to [Fe$^{6+}$]{}*]{}, Phys. Rev. A 79 (2009), p. 033415.
R.A. Phaneuf, A.L.D. Kilcoyne, A. Müller, S. Schippers, N. Aryal, K. Baral, J. Hellhund, A. Aguilar, D.A. Esteves-Macaluso, and R. Lomsadze, [ *Cross-section measurements with interacting beams*]{}, AIP Conf. Proc. 1545 (2013), pp. 72–78.
J.B. West, [*Photoionization of atomic ions*]{}, J. Phys. B 34 (2001), pp. R45–R91.
E.T. Kennedy, J.T. Costello, J.P. Mosnier, and P. [van Kampen]{}, [ *[VUV/EUV]{} ionising radiation and atoms and ions: dual laser plasma investigations*]{}, Rad. Phys. Chem. 70 (2004), pp. 291–321.
H. Kjeldsen, [*Photoionization cross sections of atomic ions from merged-beam experiments*]{}, J. Phys. B 39 (2006), pp. R325–R377.
N. Berrah, R. Bilodeau, I. Dumitriu, D. Toffoli, and R. Lucchese, [ *Shape and Feshbach resonances in inner-shell photodetachment of negative ions*]{}, J. Elec. Spectrosc. Rel. Phenom. 183 (2011), pp. 64 – 69 Electron Spectroscopy Kai Siegbahn Memorial Volume.
A. Müller, [*Precision studies of deep-inner-shell photoabsorption by atomic ions*]{}, Phys. Scr. 90 (2015), p. 054004.
A. Kramida, Y. Ralchenko, J. Reader, and [NIST ASD Team]{}, [*NIST Atomic Spectra Database (version 5.2), \[Online\]. Available: http://physics.nist.gov/asd*]{}, , National Institute of Standards and Technology, 2014.
I.C. Lyon, B. Peart, J.B. West, and K. Dolder, [*Measurements of absolute cross sections for the photoionisation of [Ba$^{+}$]{} ions*]{}, J. Phys. B 19 (1986), pp. 4137–4147.
H. Kjeldsen, F. Folkmann, H. Knudsen, M.S. Rasmussen, J.B. West, and T. Andersen, [*Absolute photoionization cross cection of [K$^{+}$]{} ions from the 3p to the 3s threshold*]{}, J. Phys. B 32 (1999), pp. 4457–4465.
H. Yamaoka, M. Oura, K. Kawatsura, T. Hayaishi, T. Sekioka, A. Agui, A. Yoshigoe, and F. Koike, [*Photoionization of singly and doubly charged neon ions following inner-shell excitation*]{}, Phys. Rev. A 65 (2001), p. 012709.
A.M. Covington, A. Aguilar, I.R. Covington, M.F. Gharaibeh, G. Hinojosa, C.A. Shirley, R.A. Phaneuf, I. [Á]{}lvarez, C. Cisneros, I. Dominguez-Lopez, M.M. Sant’Anna, A.S. Schlachter, B.M. McLaughlin, and A. Dalgarno, [ *Photoionization of [Ne$^+$]{} using synchrotron radiation*]{}, Phys. Rev. A 66 (2002), p. 062710.
M.F. Gharaibeh, J.M. Bizau, D. Cubaynes, S. Guilbaud, N. [El Hassan]{}, M.M. [Al Shorman]{}, C. Miron, C. Nicolas, E. Robert, C. Blancard, and B.M. McLaughlin, [*[K]{}-Shell photoionization of singly ionized atomic nitrogen: experiment and theory*]{}, J. Phys. B 44 (2011), p. 175208.
S. [Schippers]{}, S. [Ricz]{}, T. [Buhr]{}, A. [Borovik]{} Jr., J. [Hellhund]{}, K. [Holste]{}, K. [Huber]{}, H.J. [Sch[ä]{}fer]{}, D. [Schury]{}, S. [Klumpp]{}, K. [Mertens]{}, M. [Martins]{}, R. [Flesch]{}, G. [Ulrich]{}, E. [R[ü]{}hl]{}, T. [Jahnke]{}, J. [Lower]{}, D. [Metz]{}, L.P.H. [Schmidt]{}, M. [Sch[ö]{}ffler]{}, J.B. [Williams]{}, L. [Glaser]{}, F. [Scholz]{}, J. [Seltmann]{}, J. [Viefhaus]{}, A. [Dorn]{}, A. [Wolf]{}, J. [Ullrich]{}, and A. [M[ü]{}ller]{}, [*Absolute cross sections for photoionization of [Xe$^{q+}$]{} ions ([$1 \le q \le 5$]{}) at the 3d ionization threshold*]{}, J. Phys. B 47 (2014), p. 115602.
J. Viefhaus, F. Scholz, S. Deinert, L. Glaser, M. Ilchen, J. Seltmann, P. Walter, and F. Siewert, [*The variable polarization [XUV]{} beamline [P04]{} at [PETRA]{} [III]{}: optics, mechanics and their performance*]{}, Nucl. Instrum. Methods A 710 (2013), pp. 151 – 154.
K. Rinn, A. M[ü]{}ller, H. Eichenauer, and E. Salzborn, [*Development of single-particle detectors for [keV]{} ions*]{}, Rev. Sci. Instrum. 53 (1982), pp. 829–837.
K. Spruck, A. Becker, F. Fellenberger, M. Grieser, R. [von Hahn]{}, V. Klinkhamer, O. Novotný, S. Schippers, S. Vogel, A. Wolf, and C. Krantz, [ *An efficient, movable single-particle detector for use in cryogenic ultra-high vacuum environments*]{}, Rev. Sci. Instrum. 86 (2015), p. 023303.
M. Habibi, D.A. Esteves, R.A. Phaneuf, A.L.D. Kilcoyne, A. Aguilar, and C. Cisneros, [*Photoionization cross sections for ions of the cerium isonuclear sequence*]{}, Phys. Rev. A 80 (2009), p. 033407.
R. Trassl, [*[ECR]{} ion sources*]{}, in [*The Physics of Multiply and Highly Charged Ions, Vol. [1]{}*]{}, Kluwer Academic Publishers, Dordrecht, 2003, pp. 3–37.
R. Thissen, J.M. Bizau, C. Blancard, M. Coreno, C. Dehon, P. Franceschi, A. Giuliani, J. Lemaire, and C. Nicolas, [*Photoionization cross section of [Xe$^+$]{} ion in the pure [$5p^5\;{^2}P_{3/2}$]{} ground level*]{}, Phys. Rev. Lett. 100 (2008), p. 223001.
J.T. Lau, J. Rittmann, V. Zamudio-Bayer, M. Vogel, K. Hirsch, P. Klar, F. Lofink, T. M[ö]{}ller, and B. [von Issendorff]{}, [*Size dependence of [L]{}$_{2,3}$ branching ratio and 2p core-hole screening in X-Ray absorption of metal clusters*]{}, Phys. Rev. Lett. 101 (2008), p. 153401.
M.C. Simon, J.R. [Crespo López-Urrutia]{}, C. Beilmann, M. Schwarz, Z. Harman, S.W. Epp, B.L. Schmitt, T.M. Baumann, E. Behar, S. Bernitt, R. Follath, R. Ginzel, C.H. Keitel, R. Klawitter, K. Kubiček, V. Mäckel, P.H. Mokler, G. Reichardt, O. Schwarzkopf, and J. Ullrich, [*Resonant and near-threshold photoionization cross sections of [Fe$^{14+}$]{}*]{}, Phys. Rev. Lett. 105 (2010), p. 183001.
S. Bari, O. Gonzalez-Magaña, G. Reitsma, J. Werner, S. Schippers, R. Hoekstra, and T. Schlath[ö]{}lter, [*Photodissociation of protonated leucine-enkephalin in the [VUV]{} range of 8–40 eV*]{}, J. Chem. Phys. 134 (2011), pp. 024314–9.
A.R. Milosavljević, F. Canon, C. Nicolas, C. Miron, L. Nahon, and A. Giuliani, [*Gas-phase protein inner-shell spectroscopy by coupling an ion trap with a soft x-ray beamline*]{}, J. Phys. Chem. Lett. 3 (2012), pp. 1191–1196.
J.K. Rudolph, S. Bernitt, S.W. Epp, R. Steinbrügge, C. Beilmann, G.V. Brown, S. Eberle, A. Graf, Z. Harman, N. Hell, M. Leutenegger, A. Müller, K. Schlage, H.C. Wille, H. Yavaş, J. Ullrich, and J.R. [Crespo López-Urrutia]{}, [*X-ray resonant photoexcitation: linewidths and energies of [K]{}$\alpha$ transitions in highly charged [F]{}e ions*]{}, Phys. Rev. Lett. 111 (2013), p. 103002.
S.W. Epp, J.R. [Crespo López-Urrutia]{}, G. Brenner, V. Mäckel, P.H. Mokler, R. Treusch, M. Kuhlmann, M.V. Yurkov, J. Feldhaus, J.R. Schneider, M. Wellh[ö]{}fer, M. Martins, W. Wurth, and J. Ullrich, [*Soft x-ray laser spectroscopy on trapped highly charged ions at [FLASH]{}*]{}, Phys. Rev. Lett. 98 (2007), p. 183001.
S. Bernitt, G.V. Brown, J.K. Rudolph, R. Steinbr[ü]{}gge, A. Graf, M. Leutenegger, S.W. Epp, S. Eberle, K. Kubiček, V. M[ä]{}ckel, M.C. Simon, E. Tr[ä]{}bert, E.W. Magee, C. Beilmann, N. Hell, S. Schippers, A. M[ü]{}ller, S.M. Kahn, A. Surzhykov, Z. Harman, C.H. Keitel, J. Clementson, F.S. Porter, W. Schlotter, J.J. Turner, J. Ullrich, P. Beiersdorfer, and J.R. [Crespo L[ó]{}pez-Urrutia]{}, [*An unexpectedly low oscillator strength as the origin of the [Fe XVII]{} emission problem*]{}, Nature 492 (2012), pp. 225–228.
A. Aguilar, [*Photoionization of positive ions: the nitrogen isoelectronic sequence*]{}, Univ. of Nevada, Reno, 2003.
M. Stobbe, [*Zur [Q]{}uantenmechanik photoelektrischer [P]{}rozesse*]{}, Ann. Phys. (Leipzig) 7 (1930), pp. 661–715.
S. Schippers, A. M[ü]{}ller, B.M. McLaughlin, A. Aguilar, C. Cisneros, E. Emmons, M.F. Gharaibeh, and R.A. Phaneuf, [*Photoionization studies of the [B$^+$]{} valence shell: experiment and theory*]{}, J. Phys. B 36 (2003), pp. 3371–3381.
A. M[ü]{}ller, S. Schippers, R.A. Phaneuf, A.L.D. Kilcoyne, H. Br[ä]{}uning, A.S. Schlachter, M. Lu, and B.M. McLaughlin, [*State-resolved valence shell photoionization of [B]{}e-like ions: experiment and theory*]{}, J. Phys. B 43 (2010), p. 225201.
H. Kjeldsen, B. Kristensen, R.L. Brooks, F. Folkmann, H. Knudsen, and T. Andersen, [*Absolute state selective measurements of the photoionization cross sections of [N$^+$]{} and [O$^+$]{} ions*]{}, Astrophys. J. Suppl. Ser. 138 (2002), pp. 219–227.
A. Aguilar, A.M. Covington, G. Hinojosa, R.A. Phaneuf, I. [Á]{}lvarez, C. Cisneros, J.D. Bozek, I. Dominguez, M.M. Sant’Anna, A.S. Schlachter, S.N. Nahar, and B.M. McLaughlin, [*Absolute photoionization cross section measurements of [O II]{} ions from 29.7 e[V]{} to 46.2 e[V]{}*]{}, Astrophys. J. Suppl. Ser. 146 (2003), pp. 467–477.
S.W.J. Scully, I. [' A]{}lvarez, C. Cisneros, E.D. Emmons, M.F. Gharaibeh, D. Leitner, M.S. Lubell, A. M[ü]{}ller, R.A. Phaneuf, R. P[ü]{}ttner, A.S. Schlachter, S. Schippers, and B.M. McLaughlin, [*Doubly excited resonances in the photoionization spectrum of [Li$^+$]{}: Experiment and theory*]{}, J. Phys. B 39 (2006), pp. 3957–3968.
A. Wolf, H.B. Pedersen, L. Lammich, B. Jordon-Thaden, S. Altevogt, C. Domesle, U. Hergenhahn, M. F[ö]{}rstel, and O. Heber, [*Soft-x-ray fragmentation studies of molecular ions*]{}, J. Phys. B 43 (2010), p. 194007.
M. Grieser, [*et al.*]{}, [*Storage ring at [HIE-ISOLDE]{}*]{}, Eur. Phys. J. ST 207 (2012), pp. 1–117.
A. M[ü]{}ller, [*Electron-ion collisions: fundamental processes in the focus of applied research*]{}, Adv. At. Mol. Opt. Phys. 55 (2008), pp. 293–417.
S. Schippers, [*Electron–ion merged-beam experiments at heavy-ion storage rings*]{}, Nucl. Instrum. Methods B 350 (2015), p. 61.
S. Schippers, A. Müller, S. Ricz, M.E. Bannister, G.H. Dunn, J.D. Bozek, A.S. Schlachter, G. Hinojosa, C. Cisneros, A. Aguilar, A.M. Covington, M.F. Gharaibeh, and R.A. Phaneuf, [*Experimental link of photoionization of [Sc$^{2+}$]{} to photorecombination of [[S]{}c$^{3+}$]{}: an application of detailed balance in a unique atomic system*]{}, Phys. Rev. Lett. 89 (2002), p. 193002.
A. M[ü]{}ller, S. Schippers, R.A. Phaneuf, S.W.J. Scully, A. Aguilar, A.M. Covington, I. [Á]{}lvarez, C. Cisneros, E.D. Emmons, M.F. Gharaibeh, G. Hinojosa, A.S. Schlachter, and B.M. McLaughlin, [*K-shell photoionization of Li-like carbon ions [\[C$^{3+}$\]]{}: Experiment, theory and comparison with time-reversed photorecombination*]{}, J. Phys. B 42 (2009), p. 235602.
A. Müller, S. Schippers, R.A. Phaneuf, S.W.J. Scully, A. Aguilar, C. Cisneros, M.F. Gharaibeh, A.S. Schlachter, and B.M. McLaughlin, [*K -shell photoionization of [B]{}e-like boron ([B$^+$]{}) ions: experiment and theory*]{}, J. Phys. B 47 (2014), p. 135201.
V. Fivet, M.A. Bautista, and C.P. Ballance, [*Fine-structure photoionization cross sections of [F]{}e [II]{}*]{}, J. Phys. B 45 (2012), p. 035201.
C.P. [Ballance]{} and B.M. [McLaughlin]{}, [*[Photoionization of the valence shells of the neutral tungsten atom]{}*]{}, J. Phys. B 48 (2015), p. 085201.
N.R. Badnell and M.J. Seaton, [*On the importance of inner-shell transitions for opacity calculations*]{}, J. Phys. B 36 (2003), pp. 4367–4385.
J.M. Bizau, D. Cubaynes, S. Guilbaud, M.M. Al Shorman, M.F. Gharaibeh, I.Q. Ababneh, C. Blancard, and B.M. McLaughlin, [*[K]{}-shell photoionization of [O$^{+}$]{} and [O$^{2+}$]{} ions: Experiment and theory*]{}, Phys. Rev. A 92 (2015), p. 023401.
M. Asplund, N. Grevesse, A.J. Sauval, and P. Scott, [*The chemical composition of the Sun*]{}, Annu. Rev. Astron. Astrophys. 47 (2009), pp. 481–522.
A. Müller, A. [Borovik, Jr.]{}, T. Buhr, J. Hellhund, K. Holste, A.L.D. Kilcoyne, S. Klumpp, M. Martins, S. Ricz, J. Viefhaus, and S. Schippers, [*Observation of a four-electron Auger process in near-K-edge photoionization of singly charged carbon ions*]{}, Phys. Rev. Lett. 114 (2015), p. 013002.
A. Müller, S. Schippers, J. Hellhund, K. Holste, A.L.D. Kilcoyne, R.A. Phaneuf, C.P. Ballance, and B.M. McLaughlin, [*Single-photon single ionization of [W$^+$]{} ions: experiment and theory*]{}, J. Phys. B 48 (2015), p. 235203.
A. Müller, [*Fusion-related ionization and recombination data for tungsten ions in low to moderately high charge states*]{}, Atoms 3 (2015), pp. 120–161.
A.E. Kramida and T. Shirai, [*Compilation of Wavelengths, Energy Levels, and Transition Probabilities for [W I]{} and [W II]{}*]{}, J. Phys. Chem. Ref. Data 35 (2006), pp. 423–683.
K. Spruck, N.R. Badnell, C. Krantz, O. Novotný, A. Becker, D. Bernhardt, M. Grieser, M. Hahn, R. Repnow, D.W. Savin, A. Wolf, A. Müller, and S. Schippers, [*Recombination of [W]{}$^{18+}$ ions with electrons: [A]{}bsolute rate coefficients from a storage-ring experiment and from theoretical calculations*]{}, Phys. Rev. A 90 (2014), p. 032715.
V.V. Flambaum, M.G. Kozlov, and G.F. Gribakin, [*Coherent and stochastic contributions of compound resonances in atomic processes: electron recombination, photoionization, and scattering*]{}, Phys. Rev. A 91 (2015), p. 052704.
S. [Schippers]{}, A. [Borovik, Jr.]{}, T. [Buhr]{}, J. [Hellhund]{}, K. [Holste]{}, A.L.D. [Kilcoyne]{}, S. [Klumpp]{}, M. [Martins]{}, A. [M[ü]{}ller]{}, S. [Ricz]{}, and S. [Fritzsche]{}, [*[Stepwise contraction of the nf Rydberg shells in the 3d photoionization of multiply-charged xenon ions]{}*]{}, J. Phys. B 48 (2015), p. 144003.
S. Fritzsche, [*The [RATIP]{} program for relativistic calculations of atomic transition, ionization and recombination properties*]{}, Comput. Phys. Commun. 183 (2012), pp. 1525 – 1559.
G. Hinojosa, M.M. Sant’Anna, A.M. Covington, R.A. Phaneuf, I.R. Covington, I. Dom[í]{}nguez, A.S. Schlachter, I. Alvarez, and C. Cisneros, [ *Photofragmentation of ionic carbon monoxide*]{}, J. Phys. B 38 (2005), pp. 2701–2709.
E. R[ü]{}hl, [*Core level excitation, ionization, relaxation, and fragmentation of free clusters*]{}, Int. J. Mass Spectrom. 229 (2003), p. 117.
S.W.J. Scully, E.D. Emmons, M.F. Gharaibeh, R.A. Phaneuf, A.L.D. Kilcoyne, A.S. Schlachter, S. Schippers, A. M[ü]{}ller, H.S. Chakraborty, M.E. Madjet, and J.M. Rost, [*Photoexcitation of a volume plasmon in [[C]{}$_{60}$]{} ions*]{}, Phys. Rev. Lett. 94 (2005), p. 065503.
R.C. Bilodeau, N.D. Gibson, C.W. Walter, D.A. Esteves-Macaluso, S. Schippers, A. Müller, R.A. Phaneuf, A. Aguilar, M. Hoener, J.M. Rost, and N. Berrah, [*Single-Photon Multiple Detachment in [F]{}ullerene Negative Ions: Absolute Ionization Cross Sections and the Role of the Extra Electron*]{}, Phys. Rev. Lett. 111 (2013), p. 043003.
A. M[ü]{}ller, S. Schippers, M. Habibi, D. Esteves, J.C. Wang, R.A. Phaneuf, A.L.D. Kilcoyne, A. Aguilar, and L. Dunsch, [*Significant redistribution of [C]{}e 4d oscillator strength observed in photoionization of endohedral [Ce@[C]{}$_{82}^+$]{} ions*]{}, Phys. Rev. Lett. 101 (2008), p. 133001.
A.L.D. Kilcoyne, A. Aguilar, A. Müller, S. Schippers, C. Cisneros, G. Alna’Washi, N.B. Aryal, K.K. Baral, D.A. Esteves, C.M. Thomas, and R.A. Phaneuf, [*Confinement Resonances in Photoionization of [[X]{}e@[C]{}$_{60}^{+}$]{}*]{}, Phys. Rev. Lett. 105 (2010), p. 213001.
R.A. Phaneuf, A.L.D. Kilcoyne, N.B. Aryal, K.K. Baral, D.A. Esteves-Macaluso, C.M. Thomas, J. Hellhund, R. Lomsadze, T.W. Gorczyca, C.P. Ballance, S.T. Manson, M.F. Hasoglu, S. Schippers, and A. Müller, [*Probing confinement resonances by photoionizing [X]{}e inside a [C]{}$_{60}^{+}$ molecular cage*]{}, Phys. Rev. A 88 (2013), p. 053402.
J. Hellhund, A. [Borovik Jr.]{}, K. Holste, S. Klumpp, M. Martins, S. Ricz, S. Schippers, and A. Müller, [*Photoionization and photofragmentation of multiply charged [Lu$_3$N@C$_{80}$]{} ions*]{}, Phys. Rev. A 92 (2015), p. 013413.
A.A. Popov, S. Yang, and L. Dunsch, [*Endohedral fullerenes*]{}, Chemical Reviews 113 (2013), pp. 5989–6113.
M.J. Puska and R.M. Nieminen, [*Photoabsorption of Atoms Inside [C]{}$_{60}$*]{}, Phys. Rev. A 47 (1993), p. 1181 erratum: Phys. Rev. A 49 (1994) p. 629.
J.P. Connerade, V.K. Dolmatov, and S.T. Manson, [*On the nature and origin of confinement resonances*]{}, J. Phys. B 33 (2000), pp. 2279–2285.
K. Mitsuke, T. Mori, J. Kou, Y. Haruyama, and Y. Kubozono, [*$4d \to
4f$ Dipole Resonance of the Metal Atom Encapsulated in a Fullerene Cage: [Ce@C]{}$_{82}$*]{}, J. Chem. Phys. 122 (2005), p. 064304.
H. Katayanagi, B.P. Kafle, J. Kou, T. Mori, K. Mitsuke, Y. Takabayashi, E. Kuwahara, and Y. Kubozono, [*The $4d-4f$ dipole resonance of the [P]{}r atom in an endohedral metallofullerene, [Pr@C$_{82}$]{}*]{}, J. Quant. Spectrosc. Radiat. Transfer 109 (2008), pp. 1590–1598.
J.B. West and J. Morton, [*Absolute photoionization cross-section tables for xenon in the [VUV]{} and the soft x-ray regions*]{}, At. Data Nucl. Data Tables 22 (1978), pp. 103–107.
A.V. Korol and A.V. Solov’yov, [*Vacancy decay in endohedral atoms: the role of an atom’s non-central position*]{}, J. Phys. B 44 (2011), p. 085001.
Z. Chen and A.Z. Msezane, [*Off-center effect on the photoabsorption spectra of encapsulated [X]{}e atoms*]{}, Phys. Rev. A 89 (2014), p. 025401.
H.B. Pedersen, A. Svendsen, L.S. Harbo, H.V. Kiefer, H. Kjeldsen, L. Lammich, Y. Toker, and L.H. Andersen, [*Characterization of a new electrostatic storage ring for photofragmentation experiments*]{}, Rev. Sci. Instrum. 86 (2015), p. 063107.
\
\
\
\
\
\
\
\
\
\
[^1]: $^\ast$Corresponding author. Email: [email protected]
[^2]: The two $1s$ core electrons can safely be considered as mere spectators, since the maximum photon energy of $\sim$ 31.2 eV is well below the B$^+$($1s\to2p$) excitation energy of about 192 eV.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
In the 1980’s, Belavin and Drinfeld classified non-unitary solutions of the classical Yang-Baxter equation (CYBE) for simple Lie algebras [@BD]. They proved that all such solutions fall into finitely many continuous families and introduced combinatorial objects to label these families, Belavin-Drinfeld triples. In 1993, Gerstenhaber, Giaquinto, and Schack attempted to quantize such solutions for Lie algebras $\mathfrak{sl}(n).$ As a result, they formulated a conjecture stating that certain explicitly given elements $R \in Mat_n(\mathbb C) \otimes Mat_n(\mathbb C)$ satisfy the quantum Yang-Baxter equation (QYBE) and the Hecke condition [@GGS]. Specifically, the conjecture assigns a family of such elements $R$ to any Belavin-Drinfeld triple of type $A_{n-1}$. Until recently, this conjecture has only been known to hold for $n \leq 4$. In 1998 Giaquinto and Hodges checked the conjecture for $n=5$ by direct computation using [*Mathematica*]{} [@GH]. Here we report a computation which allowed us to check that the conjecture holds for $n
\leq 12$. The program is included which prints an element $R$ for any triple and checks that R satisfies the QYBE and Hecke conditions.
---
[Verification of the GGS conjecture for $\mathfrak{sl}(n), n
\leq 12$.]{}
12 pt
[Travis Schedler]{}
Belavin-Drinfeld triples
========================
Let $(e_i), 1 \leq i \leq n,$ be a basis for $\mathbb C^n$. Set $\Gamma = \{e_i - e_{i+1}: 1 \leq i \leq n-1\}$. We will use the notation $\alpha_i \equiv e_i - e_{i+1}$. Let $( , )$ denote the inner product on $\mathbb C^n$ having $(e_i)$ as an orthonormal basis.
A Belavin-Drinfeld triple of type $A_{n-1}$ is a triple $(\tau, \Gamma_1, \Gamma_2)$ where $\Gamma_1, \Gamma_2 \subset \Gamma$ and $\tau: \Gamma_1 \rightarrow \Gamma_2$ is a bijection, satisfying two conditions:
\(a) $\forall \alpha, \beta
\in \Gamma_1$, $(\tau \alpha,\tau \beta) = (\alpha, \beta)$.
\(b) $\tau$ is nilpotent: $\forall \alpha \in \Gamma_1, \exists k
\in \mathbb N$ such that $\tau^k \alpha \notin \Gamma_1$.
We employ three isomorphisms of Belavin-Drinfeld triples:
a\) Any triple $(\tau, \Gamma_1, \Gamma_2)$ is isomorphic to the triple $(\tau', \Gamma'_1, \Gamma'_2)$ obtained as follows: $\Gamma'_1 =
\{\alpha_m: \alpha_{n-m} \in \Gamma_1\}$, $\tau'(\alpha_m) = \alpha_k$ where $\tau(\alpha_{n-m}) = \alpha_{n-k}$.
b\) Any triple $(\tau, \Gamma_1, \Gamma_2)$ is isomorphic to the triple $(\tau^{-1}, \Gamma_2, \Gamma_1)$.
c\) The product of isomorphisms (a), (b).
Modulo these isomorphisms, we found all Belavin-Drinfeld triples for $n \leq 13$ by computer. The number of such triples is given below: 12 pt
n \# of triples n \# of triples n \# of triples
--- --------------- --- --------------- ---- ---------------
2 1 6 41 10 10434
3 2 7 161 11 45069
4 4 8 611 12 201300
5 13 9 2490 13 919479
The GGS conjecture
==================
Let $\mathfrak g = {\mathfrak sl}(n)$ be the Lie algebra of $n \times
n$ matrices of trace zero. Set $\mathfrak h \subset \mathfrak g$ to be the subset of diagonal matrices. Elements of $\mathbb C^n$ define linear functions on $\mathfrak h$ by $\bigl( \sum_i \lambda_i e_i
\bigr) \bigl( \sum_i a_i e_{ii} \bigr)= \sum_i \lambda_i a_i$. Set $\sigma = \sum_{1 \leq i,j \leq n} e_{ij} \otimes e_{ji}$, and let $P$ be the orthogonal projection of $\sigma$ to $\mathfrak g \otimes
\mathfrak g$ with respect to the form $(X,Y) = Tr(XY)$ on $Mat_n(\mathbb C)$. Then, set $P^0$ to be the projection of $P$ to $\mathfrak h \otimes \mathfrak h$. Thus $P^0
= \sum_i \frac{n-1}{n} e_{ii} \otimes e_{ii} - \sum_{i \neq j}
\frac{1}{n} e_{ii} \otimes e_{jj}$.
For any Belavin-Drinfeld triple, consider the following equations:
$$\begin{gathered}
\label{r01}
r^0_{12} + r^0_{21} = P^0. \\ \label{r02} \forall \alpha \in \Gamma_1,
(\tau \alpha \otimes 1)r^0 + (1 \otimes \alpha) r^{0} = 0.\end{gathered}$$
Belavin and Drinfeld showed that nonunitary solutions of the CYBE correspond to solutions of these equations. Define $\tilde r^0 =
r^0-P^0/2$.
The GGS conjecture gives an explicit form of a matrix $R \in
Mat_n(\mathbb C) \otimes Mat_n(\mathbb C)$ for any given triple and any given $r^0 \in \mathfrak h \otimes \mathfrak h$ satisfying , as follows:
Set $\tilde \Gamma_1 = \{v \in \text{Span}(\Gamma_1): v = e_i - e_j, 0
\leq i < j \leq n, i \neq j\}$, and define $\tilde \Gamma_2$ similarly. Then, extend $\tau$ to a map $\tilde \Gamma_1 \rightarrow \tilde
\Gamma_2$ so that $\tau$ is additive, i.e. $\tau(a+b) = \tau(a) +
\tau(b)$ provided $a,b,(a+b) \in \tilde \Gamma_1$. Further, define $\alpha \prec \beta$ if $\alpha \in \tilde \Gamma_1$ and $\tau^k(\alpha) = \beta$, for some $k \geq 1$. It is clear from the conditions on $\tau$ that this means, given $\alpha = \alpha_i +
\ldots + \alpha_{i+p}$, that $\beta = \alpha_j + \ldots +
\alpha_{j+p}$, $0 \leq p \leq n-2, 1 \leq i,j \leq n, i \neq j$. Assume $\beta = \tau^k(\alpha), k \geq 1$. If, in this case, $\tau^k(\alpha_i) = \alpha_{j+p}$, that is, $\tau^k$ sends the left endpoint of $\alpha$ to the right endpoint of $\beta$, then define $\text{sign}(\alpha,\beta) = (-1)^p$. Otherwise, set $\text{sign}(\alpha,\beta) = 1$.
We will use the notation $x \wedge y \equiv \frac{1}{2} (x \otimes y - y
\otimes x)$. Furthermore, for all matrices $x \in Mat_n(\mathbb C)
\otimes Mat_n(\mathbb C)$ we will use the notation $x = \sum_{i,j,k,l}
x_{ik}^{jl} e_{ij} \otimes e_{kl}$. Let $q$ be indeterminate and set $\hat q \equiv q-q^{-1}$. Finally, for any $\alpha = e_i - e_j$, set $e_{\alpha} = e_{ij}$, and say $\alpha > 0$ if $i < j$, otherwise $\alpha < 0$. Now, we can define the matrix $R$ as follows:
$$\begin{gathered}
\label{ace}
a = 2 \sum_{\underset{\alpha \prec \beta}{\alpha, \beta > 0}}
\text{sign}(\alpha,\beta)\: e_{-\alpha} \wedge e_{\beta}, \quad c =
\sum_{\alpha > 0} e_{-\alpha} \wedge e_\alpha, \quad \epsilon = ac +
ca + a^2, \\ \label{tars}
\tilde a = \sum_{i,j,k,l} a_{ik}^{jl} q^{a_{ik}^{jl}
\epsilon_{ik}^{jl}}, \quad R_s = q \sum_{i} e_{ii} \otimes e_{ii} +
\sum_{i \neq j} e_{ii} \otimes e_{jj} + \hat q \sum_{i>j} e_{ij}
\otimes e_{ji}, \\ \label{r} R = q^{\tilde r^0} (R_s + \hat q \tilde a)
q^{\tilde r^0}.\end{gathered}$$
[**(GGS)**]{} The matrix $R$ satisfies the quantum Yang-Baxter equation, $R_{12} R_{13} R_{23} = R_{23} R_{13} R_{12}$, and $PR$ satisfies the Hecke relation, $(PR-q)(PR+q^{-1}) = 0$.
Checking GGS by computer
========================
We checked the GGS conjecture through a program written in C, which takes as input any list of Belavin-Drinfeld triples. For each triple, it finds a valid $\tilde r^0$, constructs the matrix $R$, and checks the QYBE and Hecke conditions. Following is a more detailed description of the procedure.
We will use the notation $\tau(\alpha) = 0$ if $\alpha \notin \tilde
\Gamma_1$. Given a triple, the first step is to find an appropriate $\tilde r^0$. We rewrite the equations , as follows:
$$\begin{gathered}
\label{tr01}
\tilde r^0_{12} + \tilde r^0_{21} = 0, \\ \label{tr02} \forall \alpha
\in \Gamma_1, ((\alpha - \tau \alpha) \otimes 1) \tilde r^0 =
\textstyle{\frac{1}{2}} ((\alpha + \tau \alpha) \otimes 1) P^0.\end{gathered}$$
As before, we view elements of $\mathbb C^n$ as linear functions on $\mathfrak h$. Then, it is easy to check $(\alpha_i)_{1 \leq i \leq
n-1}$ and $(\alpha_i - \tau \alpha_i)_{1 \leq i \leq n-1}$ are bases of $\mathfrak h^*$. Let $(g_i)$ and $(f_i)$ be dual to the bases $(\alpha_i)$ and $(\alpha_i - \tau \alpha_i)$, respectively. Then, if we view $\tilde r^0$ as an element of $Mat_{n-1}(\mathbb C)$ in the basis $(f_i)$, it is clear that $\tilde r^0 = (b_{ij})$ where $b_{ij}
= \frac{1}{2} (\alpha_i + \tau \alpha_i, \alpha_j - \tau \alpha_j), i
\in \Gamma_1,$ where the inner product is the same we defined earlier on $\mathbb C^n$, and $b_{ji} = -b_{ij}, i \notin \Gamma_1, j \in
\Gamma_1$. Then, the free components of $\tilde r^0$ are those $b_{ij}$ with $i,j \notin \Gamma_1, i < j$, which determine those $b_{ij}, i,j \notin \Gamma_1, i > j$ since $\tilde r^0$ is skew-symmetric. Thus, the dimension of the space of all valid $\tilde
r^0$ is $n-m-1 \choose 2$.
The computer program merely chooses $b_{ij} = 0$ whenever $i,j \notin
\Gamma_1$. It is known that it is sufficient to consider one element from the family of possible $\tilde r^0$ in verifying the GGS conjecture. Namely, this follows from
If $R$ of the form satisfies the QYBE and PR satisfies the Hecke relation for a given $\tilde r^0$ satisfying , , then for any other solution $\tilde r^0 + r'$ of , , $q^{r'} R q^{r'}$ also satisfies the QYBE and $P q^{r'} R q^{r'}$ satisfies the Hecke relation.
[*Proof.*]{} It is clear that $P q^{r'} R q^{r'} = q^{r'_{21}} PR
q^{r'}$. Since $r'_{21} = -r'$ by , the Hecke relation may be rewritten as $q^{-r'} (PR - q) (PR + q^{-1}) q^{r'} = 0$, which is true iff $PR$ satisfies the Hecke relation.
To see that $q^{r'} R q^{r'}$ satisfies the QYBE, we take the following steps. By , $$\label{rp1}
((\alpha - \tau \alpha) \otimes 1) r' = 0.$$ Suppose that $r' = \sum_i a_i \otimes b_i$ where the $b_i$ are linearly independent. By , we know that $\alpha(a_i) =
\beta(a_i)$ whenever $\alpha \prec \beta$. Then we consider the commutator $[a_i \otimes 1 + 1 \otimes a_i, R]$ = $[a_i \otimes 1 + 1
\otimes a_i, q^{\tilde r^0}R_s q^{\tilde r^0} + \hat q q^{\tilde r^0}
\tilde a q^{\tilde r^0}]$. First note that $[a_i, e_{\alpha}] =
\alpha(a_i) e_\alpha$ for any $a_i \in \mathfrak h$. Then, it is clear $[a_i \otimes 1 + 1 \otimes a_i, q^{\tilde r^0} R_s q^{\tilde
r^0}] = [a_i \otimes 1 + 1 \otimes a_i, \sum_{i > j} d_{ij} e_{ij}
\otimes e_{ji}] = \sum_{i > j} d_{ij} (\alpha(a_i)-\alpha(a_i)) e_{ij}
\otimes e_{ji} = 0$ for the appropriate coefficients $d_{ij}$. Now, we see that $$\begin{gathered}
[a_i \otimes 1 + 1\otimes a_i, q^{\tilde r^0} \tilde a q^{\tilde r^0}]
= [a_i \otimes 1 + 1 \otimes a_i, \sum_{\alpha, \beta > 0, \alpha
\prec \beta} (f_{\alpha,\beta} e_{-\alpha} \otimes e_{\beta} +
g_{\alpha, \beta} e_{\beta} \otimes e_{-\alpha})] \\ = \sum_{\alpha,
\beta > 0, \alpha \prec \beta} (\beta(a_i) - \alpha(a_i))
(f_{\alpha,\beta} e_{-\alpha} \otimes e_{\beta} + g_{\alpha, \beta}
e_{\beta} \otimes e_{-\alpha}) = 0.\end{gathered}$$
This implies that $r' \in \Lambda^2 K$ where $K$ is the space of symmetries of $R$, that is, $K = \{ x \in Mat_{n}(\mathbb C): [1
\otimes x + x \otimes 1, R] = 0\}$. Furthermore, it is well-known and easy to check that if $x \in \Lambda^2 K$ and $R$ satisfies the QYBE, then $e^x R e^x$ also satisfies the QYBE. Thus, in our case, we have proved that $q^{r'} R q^{r'}$ satisfies the QYBE. The proposition is proved.$\quad\square$
Now, given the chosen $\tilde r^0$ in the basis $(f_i)$, the computer program changes bases to $(g_i)$. This is accomplished via the transformation $[\tilde r^0]_{(g_i)} =
([(1-\tau)]_{(\alpha_i)}^{-1})^{T} [\tilde r^0]_{(f_i)}
[(1-\tau)]_{(\alpha_i)}^{-1}$ where $(1-\tau)$ is considered to be a linear transformation on $\mathfrak h^*$, with $(1-\tau)\alpha_i =
\alpha_i - \tau \alpha_i$. Denote this new matrix by $(b'_{ij})$.
Then, the computer program obtains the matrix $[\tilde r^0]_{(e_{ii})}
\in Mat_n(\mathbb C)$ from this matrix by two quick transformations. First it finds the intermediate matrix $(b''_{ij}) = [\tilde
r^0]_{(e_{ii}),(g_i)} \in Mat_{n \times (n-1)}(\mathbb C)$ by $b''_{i1} = \frac{1}{n} ((n-1)b'_{i1} + (n-2)b'_{i2} + \ldots +
b'_{i,n-1})$, and the other terms follow easily. The same technique on the other side finally gives $[\tilde r^0]_{(e_{ii})}$.
Once $\tilde r^0$ is obtained, the computer constructs the matrix $R
\in Mat_n(\mathbb C) \otimes Mat_n(\mathbb C)$ in the basis $e_{ij}
\otimes e_{kl}, 1 \leq i,j,k,l \leq n$. First it computes $a$, $c$, and $\epsilon$ by . Then, formulas , are implemented for each entry separately. Elements $x \in
Mat_n(\mathbb C) \otimes Mat_n(\mathbb C)$ are implemented as 3-dimensional arrays $(x_{ik}^j)$, since all matrices presented in the GGS conjecture take the form $\sum_{i,j,k} x_{ik}^j e_{ij} \otimes
e_{k,i+k-j}$. Polynomials in $q$ are implemented as structures containing two arrays of integers, one for positive and one for negative powers of $q$. The sizes of the arrays are determined in the input of the program.
The computer checks the QYBE and Hecke conditions in the following manner. For the QYBE condition, the corresponding entries of $R_{12}
R_{13} R_{23}$ and $R_{23} R_{13} R_{23}$ are computed and compared individually; both take the form $\sum_{i,j,k,l,m} d_{ikm}^{jl} e_{ij} \otimes e_{kl} \otimes
e_{m,i+k+m-j-l}$. The same method is applied to the Hecke condition with matrices $\sum_{i,j,k} d_{ik}^j e_{ij} \otimes
e_{k,i+k-j}$. Explicitly, if $R = \sum_{i,j,k} r_{ik}^j e_{ij} \otimes
e_{k,i+k-j}$, the QYBE and Hecke conditions become, respectively:
$$\begin{gathered}
\sum_p r_{ik}^{k+i-p} r_{k+i-p,m}^j r_{p,m+k+i-p-j}^l = \sum_p
r_{km}^p r_{i,m+k-p}^{j+l-p} r_{j+l-p,p}^j, \forall i,j,k,l,m. \\
\sum_l r_{ki}^l r_{k+i-l,l}^j = \delta_{ij} + \hat q r_{ki}^j, \forall
i,j,k.\end{gathered}$$
Then, the computer prints the matrices $\tilde r^0$ and $R$ and reports whether or not the conditions passed.
After generating all Belavin-Drinfeld triples for $n \leq 13$ as described in the previous section, all tests were performed on each triple where $n \leq 12$ with this procedure, all of which passed. Thus, by application of the previous proposition, we have the following result:
The GGS conjecture is true for Lie algebras $\mathfrak{sl}(n)$ with $n
\leq 10$.
The computer program is included with this paper, with instructions on usage included with the program itself.
Acknowledgements
================
I would like to thank Pavel Etingof for his generous help and advice. I would also like to thank the Harvard College Research Program for their support.
[VancVanRL]{}
Belavin, A.A. and Drinfeld, V.G.: Triangle equations and simple Lie algebras, in: S.P.Novikov (ed.), [*Mathematical Physics Reviews*]{}, Harwood, New York, 1984, 93-166.
Gerstenhaber, M., Giaquinto, A., and Schack, S.D.: Construction of quantum groups from Belavin-Drinfeld infinitesimals, in: A. Joseph and S. Shnider (eds), [*Quantum Deformations of Algebras and their Representations,*]{} Israel Math. Conf. Series 7, Bar-Ilan Univ., Ramat Gan, 1993, pp. 45-64.
Giaquinto, A. and Hodges, T.: Nonstandard solutions of the Yang-Baxter equation, [*Letters in Mathematical Physics*]{} [**44**]{} (1998), 67-75.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Three entanglement concentration protocols (ECPs) are proposed. The first ECP and a modified version of that are shown to be useful for the creation of maximally entangled cat and GHZ-like states from their non-maximally entangled counterparts. The last two ECPs are designed for the creation of maximally entangled $(n+1)$-qubit state $\frac{1}{\sqrt{2}}\left(|\Psi_{0}\rangle|0\rangle+|\Psi_{1}\rangle|1\rangle\right)$ from the partially entangled $(n+1)$-qubit normalized state $\alpha|\Psi_{0}\rangle|0\rangle+\beta|\Psi_{1}\rangle|1\rangle,$ where $\langle\Psi_{1}|\Psi_{0}\rangle=0$ and $|\alpha|\neq\frac{1}{\sqrt{2}}$. It is also shown that W, GHZ, GHZ-like, Bell and cat states and specific states from the 9 SLOCC-nonequivalent families of 4-qubit entangled states can be expressed as $\frac{1}{\sqrt{2}}\left(|\Psi_{0}\rangle|0\rangle+|\Psi_{1}\rangle|1\rangle\right)$ and consequently the last two ECPs proposed here are applicable to all these states. Quantum circuits for implementation of the proposed ECPs are provided and it is shown that the proposed ECPs can be realized using linear optics. Efficiency of the ECPs are studied using a recently introduced quantitative measure (Phys. Rev. A **85**, 012307 (2012)). Limitations of the measure are also reported.'
author:
- 'Chitra Shukla$^{a}$, Anindita Banerjee$^{b}$, Anirban Pathak$^{a,c,}$[^1]'
title: 'Protocols and quantum circuits for implementing entanglement concentration in cat state, GHZ-like state and 9 families of 4-qubit entangled states'
---
Introduction\[sec:Introduction\]
================================
Entanglement plays a crucial role in quantum computation and quantum communication [@Non; @local; @CNOT; @Manu-annalender; @Bennett-teleportation; @Bennett-densecoding; @ekert; @Ping-Pong; @DLL; @Beyond; @GV]. It is essential for realization of various protocols and algorithms [@Non; @local; @CNOT; @Manu-annalender; @Bennett-teleportation; @Bennett-densecoding; @ekert; @Ping-Pong; @DLL; @Beyond; @GV]. Specifically, maximally entangled state is required for realization of nonlocal quantum gates [@Non; @local; @CNOT], distributed quantum search algorithm [@Manu-annalender], teleportation [@Bennett-teleportation], densecoding [@Bennett-densecoding], entanglement based quantum key distribution [@ekert] and entanglement based secure direct quantum communication [@Ping-Pong; @DLL; @Beyond; @GV]. In these applications, entanglement is generally produced locally and distributed to different parties involved in the communication or computation process. During the transmission, processing and storing a maximally entangled pure state may interact with the environment and become mixed state or less entangled (i.e., non-maximally entangled) pure state. This may happen because of many different reasons. For example, the transmission channel may be noisy. In general, the amount of entanglement is usually reduced during transmission process. Unfortunately, such degradation of entanglement is unavoidable, but for the proper execution of the protocols mentioned above [@Manu-annalender; @Bennett-teleportation; @Bennett-densecoding; @ekert; @Ping-Pong; @DLL; @Beyond; @GV] we need perfect quantum channels for distribution of ebits. In absence of such a channel it would be sufficient to design a protocol that can convert the non-maximally entangled state back into maximally entangled state. Interestingly, such protocols exist and the protocols are divided into two classes (i) Entanglement concentration protocol (ECP) which can transform a partially entangled pure state (pure non-maximally entangled state) into a maximally entangled state and (ii) Entanglement purification/distillation protocols (EP) that can transform a mixed non-maximally entangled state into a maximally entangled state. In 1996, Bennet et al. [@origional; @ECP] proposed the first ECP. In the same year they proposed an EP, too [@Entaglement; @purification]. In their pioneering work Bennet et al. used collective entanglement concentration procedure (i.e., Schmidt projective method). Since the pioneering works of Bennet et al., several ECPs and EPs are proposed [@S.; @Bose; @bandyopadhyay; @zhao-proposal; @yamamoto; @Sheng-1; @entropy; @Bell-determinstic-ckt; @Bell_Deng; @Sheng-2; @opt-comm-2013-he; @Dhara; @GHZ; @state; @GHZ-zhou; @Dhara; @Cluster; @state; @T; @Xu-cluster; @Zhao-cluster; @NOON]. Initially most of the ECPs were proposed for non-maximally entangled Bell state. Specifically, Bose et al. [@S.; @Bose], Bandyopadhyay [@bandyopadhyay], Zhao et al. [@zhao-proposal], Yamamoto et al. [@yamamoto], Sheng et al. [@Sheng-1], Sheng and Zhou [@entropy], Gu et al. [@Bell-determinstic-ckt], Deng [@Bell_Deng] proposed ECPs and EPs for partially entangled Bell state. While Bose et al.’s proposal involved entanglement swapping and Bell state measurement and Gu et al.’s proposal involved projective operator valued measurement (POVM), other proposals [@zhao-proposal; @yamamoto; @Sheng-1] circumvented the use of Bell measurement and POVM and discussed the possibilities of optical implementation of proposed ECPs using polarizing beam splitter (PBS) and wave-plates. In fact, ECPs proposed by Zhao et al. and Yamamoto et al. were experimentally realized in 2003 [@experiment2-zhao; @experiment1-yamamoto]. It was shown in some of these initial works that the ECPs designed for partially entangled Bell states can be generalized to build ECPs for partially entangled cat ($N$-partite GHZ) states. Recently, several efforts have been made to extend the applicability of ECPs beyond the production of Bell states. For example, in 2012, Sheng et al. have proposed an ECP for partially entangled arbitrary W state [@Sheng-2]. In 2013, Ling-yan He [@opt-comm-2013-he] proposed a single nitrogen vacancy (N-V) center assisted ECP for a specific type of partially entangled W state. Very recently, ECPs for partially entangled GHZ states are proposed by Choudhury and Dhara [@Dhara; @GHZ; @state] and Zhou et al. [@GHZ-zhou], ECPs for 4-qubit cluster state are proposed by Choudhury and Dhara [@Dhara; @Cluster; @state], Ting-Ting Xu et al. [@T; @Xu-cluster] and Zhau et al. [@Zhao-cluster], ECP for partially entangled NOON states is proposed by Zhou et al. [@NOON]. Clearly, much attention has recently been paid to develop ECPs for partially entangled states other than Bell state. Interestingly, majority of these recent efforts are concentrated toward the construction of ECPs for partially entangled GHZ states [\[]{}[@Dhara; @GHZ; @state; @GHZ-zhou] and references therein[\]]{} and W states [\[]{}[@Sheng-2; @opt-comm-2013-he] and references therein[\]]{}. In case of 3-qubit pure states it is well-known that there are only 2 families of entangled states [@3-qubit; @SLOCC] under stochastic local quantum operations assisted by classical communication (SLOCC). These two families of 3-qubit entangled states are referred to as GHZ and W states. Thus the recent trend of developing ECPs for partially entangled GHZ and W states is reasonable. Interestingly, no ECP has yet been proposed to explicitly concentrate GHZ-like states. However, in recent past several applications of maximally entangled GHZ-like states have been reported [@GHZ-like-1; @GHZ-like2; @GHZ-like3]. Keeping this in mind, in the present work we have proposed an ECP for partially entangled cat state and have modified that to develop an ECP for partially entangled GHZ-like state.
Recent success in developing ECPs for both the families of 3-qubit entangled states (i.e., for GHZ and W class) also motivated us to ask: How to concentrate pure states of different families of 4-qubit partially entangled states? Present paper’s main objective is to answer this question. The effort is timely as to the best of our knowledge until date for 4-qubit pure states ECPs are proposed only for partially entangled 4-qubit cat state and 4-qubit cluster state. No effort has yet been made to concentrate other families of 4-qubit entangled states. Keeping this in mind, present paper is focused around construction of a general ECP for 9 families of 4-qubit entangled states. Before we discuss further detail of our idea, it would be apt to note that in 2002, Verstraete et al. [@4-qubit; @SLOCC] had shown that 4-qubit pure states can be entangled in 9 different ways under SLOCC. A rigorous proof of this classification of 4-qubit pure states was subsequently provided by Chterental and Djokovic in 2007 [@4-Qubit-SLOCC-proof]. In 2010, a similar SLOCC classification of 4-qubit pure states was obtained using string theory [@string]. However, in 2010, another interesting result was reported by Gaur and Wallach [@gilad], in which they had established the existence of uncountable number SLOCC-nonequivalent classes of 4-qubit entangled states. In what follows we will use Verstraete et al.’s [@4-qubit; @SLOCC] classification and propose two ECPs that can concentrate some states from each of the 9 families proposed by Verstraete et al. [@4-qubit; @SLOCC].
Until now different strategies have been used for developing ECPs and EPs. For example, ECPs are proposed using linear optics (specifically, using PBSs and wave plates), cross-Kerr-nonlinearities (i.e., using quantum non-demolition (QND) measurements) [@Sheng-1; @Sheng-2], entanglement swapping and Bell measurement [@S.; @Bose], unitary transformation [@zhao-proposal] and quantum electrodynamics (QED) based techniques [@QED] etc. However, most of the recent works discuss ECP with a perspective where qubits are realized using the polarization of photon. As the qubit can be realized using different systems, such as superconductivity, NMR, photon etc., in what follows we have not restricted ourselves to any specific technology and have presented our protocol in general as a quantum circuit. This makes it applicable to any specific kind of realization of qubits. Only at the end of the paper we have shown that the present work can be realized using PBSs, wave-plates and photon-detectors. As the proposed protocols do not require anything other than implementation of Bell measurement it can also be realized in other implementations of qubits. For example, it is straight forward to implement the ECPs proposed here using NMR-based approach as Bell measurement is possible in NMR [@NMR-Anil; @Kumar].
Rest of the paper is organized as follows. In Section \[sec:Entanglement-concentration-protocols\] an ECP for cat state is proposed and it is modified to develop an ECP for GHZ-like state. Quantum circuits for these ECPs are also described. In Section \[sec:A-generalized-scheme\] we propose two ECPs for $(n+1)$-qubit normalized states of the form $|\psi\rangle=\alpha|\Psi_{0}\rangle|0\rangle+\beta|\Psi_{1}\rangle|1\rangle,$ where $|\Psi_{0}\rangle$ and $|\Psi_{1}\rangle$ are mutually orthogonal $n$-qubit states and $|\alpha|\neq\frac{1}{\sqrt{2}}$. In this section we have also shown that each of the 9 families of four qubit states contain some states of the form $\alpha|\Psi_{0}\rangle|0\rangle+\beta|\Psi_{1}\rangle|1\rangle.$ In Section \[sec:Optical-implementation\] we have shown that the ECPs proposed here can be realized using linear optics. In Section \[sec:Efficiency\] efficiencies of the proposed ECPs are discussed using a quantitative measure of ECP introduced by Sheng et al. [@Sheng-1] and finally the paper is concluded in Section \[sec:Conclusion\].
Entanglement concentration protocols (ECPs) \[sec:Entanglement-concentration-protocols\]
========================================================================================
In the previous section we have already described the importance of ECPs in quantum information processing. In this section we propose new ECPs for partially entangled cat state and GHZ-like states. To begin with we first propose an ECP for a non-maximally entangled cat state.
ECP for partially entangled cat state\[sub:For-Cat-state\]
----------------------------------------------------------
A non-maximally entangled Bell-type state may be defined as
$$|\psi\rangle_{{\rm Bell}}=(\alpha|00\rangle+\beta|11\rangle)_{12},\label{eq:Bell state (2 qubit)}$$
where $|\alpha|^{2}+|\beta|^{2}=1$ and $|\alpha|\neq|\beta|.$ Similarly, we may define a non-maximally entangled $n$-qubit cat state as
$$|\psi\rangle_{{\rm cat}}=(\alpha|000\cdots0\rangle+\beta|111\cdots1\rangle)_{12\cdots n}.\label{eq:cat state}$$
We wish to devise an ECP for $|\psi\rangle_{{\rm cat}}$ with the help of non-maximally entangled Bell state $|\psi\rangle_{{\rm Bell}}.$ For this purpose we introduce the quantum circuit shown in Fig. \[fig:Quantum-circuit-for-cat\]. The circuit is composed of two parts. In the first part (See the left most box of Fig. \[fig:Quantum-circuit-for-cat\]) we produce non-maximally entangled $n$-qubit cat state $|\psi\rangle_{{\rm cat}}$ starting from a non-maximally entangled Bell state $|\psi\rangle_{{\rm Bell}}$ and $(n-2)$ auxiliary qubits each prepared in $|0\rangle$. Working of this part of the circuit may be understood as follows: Assume that we have a non-maximally entangled $n$-qubit cat state $|\psi\rangle_{{\rm cat}}$ and we add an auxiliary qubit (prepared in $|0\rangle$) with that as $(n+1)^{th}$ qubit and apply a ${\rm CNOT}$ operation with any one of the qubits of the $n$-qubit $|\psi\rangle_{{\rm cat}}$ state (in Fig. \[fig:Quantum-circuit-for-cat\] it is the second qubit of $n$-qubit $|\psi\rangle_{{\rm cat}}$) as the control qubit and the auxiliary qubit as the target qubit. This would yield an $(n+1)$-qubit $|\psi\rangle_{{\rm cat}}$ state as
$$\begin{array}{lcl}
{\rm CNOT_{2\rightarrow n+1}}(|\psi\rangle_{{\rm cat}}\otimes|0\rangle) & = & {\rm CNOT_{2\rightarrow n+1}}((\alpha|000\cdots0\rangle+\beta|111\cdots1\rangle)_{12\cdots n}\otimes|0\rangle_{n+1})\\
& = & {\rm CNOT_{2\rightarrow n+1}}(\alpha|000\cdots00\rangle+\beta|111\cdots10\rangle)_{12\cdots n+1}\\
& = & (\alpha|000\cdots00\rangle+\beta|111\cdots11\rangle)_{12\cdots n+1}.\end{array}\label{eq:CNOT CAT state}$$
This part of the circuit is not the main component of the proposed ECP and it can be ignored. However, this may be relevant in the following scenario: Assume that we have a machine for generation of maximally entangled Bell state $|\psi^{+}\rangle=\frac{|00\rangle+|11\rangle}{\sqrt{2}}$. The machine is not working properly and producing $|\psi\rangle_{{\rm Bell}}$ (with a fix but unknown value of $\alpha$) instead of $|\psi^{+}\rangle$. In such a scenario we may first use an output of that imperfect Bell state generator and $(n-2)$ auxiliary qubits prepared in $|0\rangle$ to produce $|\psi\rangle_{{\rm cat}}$. Once we obtain $|\psi\rangle_{{\rm cat}}$ (either prepared from imperfect Bell state generator or supplied) we may use an entanglement swapping operation between another output of that machine $\left(|\psi\rangle_{{\rm Bell}}\right)$ and $|\psi\rangle_{{\rm cat}}$ to obtain the desired ECP for $|\psi\rangle_{{\rm cat}}$ as shown in the right most box of the circuit shown in Fig. \[fig:Quantum-circuit-for-cat\]. This part of the circuit works as follows: As the input of the main part of the circuit for implementation of the ECP is $|\psi_{1}\rangle=|\psi\rangle_{{\rm Bell}}\otimes|\psi\rangle_{{\rm cat}}$ using (\[eq:Bell state (2 qubit)\]) and (\[eq:CNOT CAT state\]) we can write the input state as
$$\begin{array}{lcl}
|\psi_{1}\rangle & = & (\alpha|00\rangle+\beta|11\rangle)_{12}\otimes(\alpha|000\cdots0\rangle+\beta|111\cdots1\rangle)_{345\cdots n+2}\\
& = & (\alpha^{2}|00000\cdots0\rangle+\alpha\beta|00111\cdots1\rangle+\alpha\beta|11000\cdots0\rangle+\beta^{2}|11111\cdots1\rangle)_{12345\cdots n+2}.\end{array}\label{eq:Intermidiate Cat state 1}$$
Now after applying the SWAP gate shown in the circuit (i.e., after swapping the second and third qubits of $|\psi\rangle_{{\rm 1}}$ we obtain
$$\begin{array}{lcl}
|\psi_{2}\rangle & = & \frac{1}{\sqrt{2}}[(\alpha^{2}(|\psi^{+}\rangle+|\psi^{-}\rangle)_{13}|000\cdots0\rangle_{245\cdots n+2}+\alpha\beta(|\phi^{+}\rangle+|\phi^{-}\rangle)_{13}|011\cdots1\rangle_{245\cdots n+2}\\
& + & \alpha\beta(|\phi^{+}\rangle-|\phi^{-}\rangle)_{13}|100\cdots0\rangle_{245\cdots n+2}+\beta^{2}(|\psi^{+}\rangle-|\psi^{-}\rangle)_{13}|111\cdots1\rangle)_{245\cdots n+2}]\\
& = & |\psi^{+}\rangle_{13}\frac{(\alpha^{2}|000\cdots0\rangle+\beta^{2}|111\cdots1\rangle)_{245\cdots n+2}}{\sqrt{2}}+|\psi^{-}\rangle_{13}\frac{(\alpha^{2}|000\cdots0\rangle-\beta^{2}|111\cdots1\rangle)_{245\cdots n+2}}{\sqrt{2}}\\
& + & |\phi^{+}\rangle_{13}\frac{\alpha\beta(|011\cdots1\rangle+|100\cdots0\rangle)_{245\cdots n+2}}{\sqrt{2}}+|\phi^{-}\rangle_{13}\frac{\alpha\beta(|011\cdots1\rangle-|100\cdots0\rangle)_{245\cdots n+2}}{\sqrt{2}},\end{array}\label{eq: cat state final}$$
where we have used $|\psi^{\pm}\rangle=\frac{|00\rangle\pm|11\rangle}{\sqrt{2}}$ and $|\phi^{\pm}\rangle=\frac{|01\rangle\pm|10\rangle}{\sqrt{2}}$.
![\[fig:Quantum-circuit-for-cat\]Quantum circuit for ECP of non-maximally entangled cat state $|\psi\rangle_{{\rm cat}}$ ](fig2)
After the application of the SWAP gate a Bell measurement is performed on the first two qubits of $|\psi\rangle_{2}.$ As a consequence of that remaining $n$-qubits would collapse to one of the four states as shown in Column 2 of Table \[tab:tab1\]. If the outcome of the Bell measurement is $|\psi^{\pm}\rangle$ then the protocol fails, otherwise depending upon the outcome we apply a single qubit unitary operation as shown in Column 3 of Table \[tab:tab1\]. It is easy to observe that on application of the single qubit unitary operation in both the cases (i.e., if the outcome of Bell measurement is $|\phi^{+}\rangle$ or $|\phi^{-}\rangle$) we obtain a maximally entangled $n$-qubit cat state $\frac{|000\cdots0\rangle+|111\cdots1\rangle}{\sqrt{2}}.$ Thus the circuit shown in Fig. \[fig:Quantum-circuit-for-cat\] is equivalent to an ECP for non-maximally entangled cat state. Knowledge of values of $\alpha$ and $\beta$ are not required in the above scenario where we create the state $|\psi\rangle_{{\rm cat}}$ starting from a defective Bell state generator which always produces $|\psi\rangle_{{\rm Bell}}$ with same values of $\alpha$ and $\beta$. In case of any other scenario, we would require to know the value of $\alpha$ or $\beta$ and use that to produce $|\psi\rangle_{{\rm Bell}}$. Design of such a circuit is a trivial exercise as $|\psi\rangle_{{\rm Bell}}$ may be easily created using a modified EPR circuit where the Hadamard gate is replaced by a single qubit unitary gate which maps $|0\rangle\rightarrow\alpha|0\rangle+\beta|1\rangle.$ A specific example of such a circuit is shown in Fig. \[fig:fig 2\] where the single qubit gate $U_{1}$ that maps $|0\rangle\rightarrow\alpha|0\rangle+\beta|1\rangle$ is introduced as $$U_{1}=\left(\begin{array}{lc}
\alpha & -\beta^{*}\\
\beta & \alpha^{*}\end{array}\right)\label{eq:u1}$$ In general, the proposed ECP is probabilistic. The existing ECPs contain the same probabilistic nature but the feature is not explicitly mentioned. For example, ECPs proposed in recent works of Choudhury and Dhara [@Dhara; @GHZ; @state; @Dhara; @Cluster; @state] are essentially probabilistic. However, they didn’t mention the probabilistic nature. To be consistent with the conventional ECPs, we may assume that Alice prepares an $n$-qubit cat state, keeps the first qubit with herself and sends the remaining $n-1$ qubits to $n-1$ parties, say ${\rm Bob_{1},\, Bob_{2},\,\cdots,Bob_{n-1}.}$ At a later time the transmitted cat state may be reduced to partially entangled cat state $|\psi\rangle_{{\rm cat}}.$ In order to concentrate that in the above described protocol Alice prepares $|\psi\rangle_{{\rm Bell}},$ swaps her share of the $|\psi\rangle_{{\rm cat}}$ with the second qubit of $|\psi\rangle_{{\rm Bell}},$ performs a Bell measurement on the first two qubits of her possession and if the protocol succeeds then she applies appropriate unitary operation (as described in Table \[tab:tab1\]) on the third qubit to obtain a maximally entangled cat state shared between her and ${\rm Bob_{1},\, Bob_{2},\,\cdots,Bob_{n-1}.}$ ECPs proposed in the remaining part of this paper can also be illustrated in the similar fashion. However, protocols proposed in the remaining part of the paper are not described in this manner as it is a trivial exercise.
Outcome of Bell measurement State of the remaining $n$ qubits Operation applied on qubit 2 Final state
----------------------------- ----------------------------------------------------------------------------------------------- ------------------------------ ----------------------------------------------------------
$|\psi^{+}\rangle_{13}$ $\frac{(\alpha^{2}|000\cdots0\rangle+\beta^{2}|111\cdots1\rangle)_{245\cdots n+2}}{\sqrt{2}}$
$|\psi^{-}\rangle_{13}$ $\frac{(\alpha^{2}|000\cdots0\rangle-\beta^{2}|111\cdots1\rangle)_{245\cdots n+2}}{\sqrt{2}}$
$|\phi^{+}\rangle_{13}$ $\frac{\alpha\beta(|011\cdots1\rangle+|100\cdots0\rangle)_{245\cdots n+2}}{\sqrt{2}}$ $X$ $\frac{|000\cdots0\rangle+|111\cdots1\rangle}{\sqrt{2}}$
$|\phi^{-}\rangle_{13}$ $\frac{\alpha\beta(|011\cdots1\rangle-|100\cdots0\rangle)_{245\cdots n+2}}{\sqrt{2}}$ $iY$ $\frac{|000\cdots0\rangle+|111\cdots1\rangle}{\sqrt{2}}$
: \[tab:tab1\]Relation among Alice’s Bell state measurement outcome, cat state and operation applied
![\[fig:fig 2\]Quantum circuit for generation of $|\psi\rangle_{{\rm Bell}}=\alpha|00\rangle+\beta|11\rangle$ where $U_{1}=\left(\protect\begin{array}{lc}
\alpha & -\beta^{*}\protect\\
\beta & \alpha^{*}\protect\end{array}\right)$](fig1.eps)
### Special cases of cat state
Interestingly Bell state and GHZ state are special cases of an $n$-qubit cat state for $n=2$ and $n=3,$ respectively. Thus the Bell-measurement based ECP presented here also works as ECP for non-maximally entangled Bell and GHZ states. In 1999, S. Bose et al. [@S.; @Bose] proposed a Bell-measurement based scheme for obtaining the maximally entangled Bell state $|\psi^{+}\rangle$ from $|\psi\rangle_{{\rm Bell}}$. Bose et al.’s scheme can now be viewed as a special case of the above proposed ECP for cat state for $n=2$. Similarly, for $n=3$ qubits proposed ECP reduces to an ECP for GHZ state. Very recently a Bell-measurement based ECP for a non-maximally entangled 3-qubit GHZ state is proposed by Choudhury and Dhara [@Dhara; @GHZ; @state]. Their ECP can also be viewed as a special case of our ECP for cat state. The scheme proposed here is more general than Choudhury and Dhara [@Dhara; @GHZ; @state] scheme for several other reasons, too. For example, $\alpha$ and $\beta$ are unknown and complex numbers here while in work of Choudhury and Dhara $\alpha$ and $\beta$ were considered as real and known. Further, complexity of their approach is high as the ECP proposed by Choudhury and Dhara involves many steps that are not essential.
ECP for partially entangled GHZ-like state\[sub:For-GHZ-like-state\]
--------------------------------------------------------------------
A maximally entangled GHZ-like state is defined as $$\frac{|\psi_{i}0\rangle\pm|\psi_{j}1\rangle}{\sqrt{2}},\label{eq:GHZ-like-gen}$$ where $|\psi_{i}\rangle,|\psi_{j}\rangle\in\{|\psi^{\pm}\rangle,|\phi^{\pm}\rangle\}$ and $i\neq j.$ For example, we may consider a specific GHZ-like state as $$|\psi\rangle=\frac{|\psi^{+}0\rangle+|\phi^{+}1\rangle}{\sqrt{2}}.\label{eq:GHZ-like-specific}$$ Corresponding non-maximally entangled 3-qubit GHZ-like state should be defined as
$$|\psi\rangle_{{\rm GHZ-like}}=\alpha|\psi^{+}0\rangle+\beta|\phi^{+}1\rangle.\label{eq:GHZ-like state}$$
As several applications of maximally entangled GHZ-like states have been reported in recent past [@GHZ-like-1; @GHZ-like2; @GHZ-like3], successful implementation of these applications using GHZ-like state would require an ECP for GHZ-like state. Unfortunately no ECP for GHZ-like state has been proposed until now. Keeping this in mind, we wish to show that a slightly modified version of the ECP described above for cat state works for GHZ-like states. If we start with our defected Bell state generator as before, then with the help of an EPR circuit (a Hadamard gate followed by a ${\rm CNOT}$ gate) and an auxiliary qubit as shown in the left box of Fig. \[fig:Quantum-circuit-ghz-like\] we can easily produce $|\psi\rangle_{{\rm GHZ-like}}$ state. However, such construction of GHZ-like state is not an essential part of the ECP as discussed above. Now after combining the non-maximally entangled GHZ-like state produced as the output of the first block with a non-maximally entangled Bell-type state $|\psi\rangle_{{\rm Bell}}$ we obtain
$$\begin{array}{lcl}
|\psi_{3}\rangle=|\psi\rangle_{{\rm Bell}}\otimes|\psi\rangle_{{\rm GHZ-like}} & = & (\alpha|00\rangle+\beta|11\rangle)_{12}\otimes(\alpha|\psi^{+}0\rangle+\beta|\phi^{+}1\rangle)_{345}\\
& = & (\alpha^{2}|00\psi^{+}0\rangle+\alpha\beta|00\phi^{+}1\rangle+\alpha\beta|11\psi^{+}0\rangle+\beta^{2}|11\phi^{+}1\rangle)_{12345},\end{array}\label{eq:Intermidiate GHZ-like state}$$
which can be decomposed as
$$\begin{array}{lcl}
|\psi_{3}\rangle & = & \frac{1}{\sqrt{2}}\left[(\alpha^{2}(|\psi^{+}\rangle+|\psi^{-}\rangle)_{15}|0\psi^{+}\rangle_{234}+\alpha\beta(|\phi^{+}\rangle+|\phi^{-}\rangle)_{15}|0\phi^{+}\rangle_{234}\right.\\
& + & \left.\alpha\beta(|\phi^{+}\rangle-|\phi^{-}\rangle)_{15}|1\psi^{+}\rangle_{234}+\beta^{2}(|\psi^{+}\rangle-|\psi^{-}\rangle)_{15}|1\phi^{+}\rangle)_{234}\right]\\
& = & |\psi^{+}\rangle_{15}(\frac{\alpha^{2}(|000\rangle+|011\rangle)+\beta^{2}(|101\rangle+|110\rangle)_{234}}{\sqrt{2}})+|\psi^{-}\rangle_{15}(\frac{\alpha^{2}(|000\rangle+|011\rangle)-\beta^{2}(|101\rangle+|110\rangle)_{234}}{\sqrt{2}})\\
& + & |\phi^{+}\rangle{}_{15}(\frac{\alpha\beta(|001\rangle+|010\rangle+|100\rangle+|111\rangle)_{234}}{\sqrt{2}})+|\phi^{-}\rangle_{15}(\frac{\alpha\beta(|001\rangle+|010\rangle-|100\rangle-|111\rangle)_{234}}{\sqrt{2}})\\
& = & |\psi^{+}\rangle_{15}(\frac{\alpha^{2}(|000\rangle+|011\rangle)+\beta^{2}(|101\rangle+|110\rangle)_{234}}{\sqrt{2}})+|\psi^{-}\rangle_{15}(\frac{\alpha^{2}(|000\rangle+|011\rangle)-\beta^{2}(|101\rangle+|110\rangle)_{234}}{\sqrt{2}})\\
& + & |\phi^{+}\rangle{}_{15}(\frac{\alpha\beta(|\psi^{+}1\rangle+|\phi^{+}0\rangle)_{234}}{\sqrt{2}})+|\phi^{-}\rangle_{15}(\frac{\alpha\beta(|\psi^{-}1\rangle+|\phi^{-}0\rangle)_{234}}{\sqrt{2}}).\end{array}\label{eq:GHZ-like state final}$$
Thus after re-ordering the qubit sequence as $12345\rightarrow15234$[^2] and a Bell measurement on first two qubits of $|\psi_{3}\rangle$ (as shown in Fig. \[fig:Quantum-circuit-ghz-like\]) the state of the remaining 3 qubits would collapse to one of the four states as shown in Column 2 of Table \[tab:Tab2\]. If the outcome of Bell measurement is $|\psi^{\pm}\rangle$ then the protocol fails, otherwise depending upon the outcome we apply a unitary operation as shown in Column 3 of Table \[tab:Tab2\]. It is easy to observe that on application of the unitary operation in both the cases (i.e., if the outcome of Bell measurement is $|\phi^{+}\rangle$ or $|\phi^{-}\rangle$) we obtain a maximally entangled 3-qubit GHZ-like state. Thus we have an ECP for GHZ-like state and as far as the main ECP part is concerned this ECP is similar to the ECP designed for cat states with only difference in the choice of qubits to be swapped and to be modified through unitary operation. Thus a quantum circuit designed for ECP of cat states as shown above will also work for GHZ-like states if suitably modified.
![\[fig:Quantum-circuit-ghz-like\]Quantum circuit for ECP of non-maximally entangled GHZ-like state $|\psi\rangle_{{\rm GHZ-like}}.$](fig3)
[|>p[1.5in]{}|>p[2in]{}|>p[1.5in]{}|>p[1in]{}|]{} Outcome of Bell measurement on qubits $1$ and $5$ & Qubits $2,3$ and $4$ collapses to & Operation applied on qubits $2$ and $3$ & Final state of qubits $2,3$ and $4$ [\
]{}$|\psi^{+}\rangle_{15}$ & $\frac{\alpha^{2}(|000\rangle+|011\rangle)_{234}+\beta^{2}(|101\rangle+|110\rangle)_{234}}{\sqrt{2\left(\alpha^{4}+\beta^{4}\right)}}$ & [\
]{}$|\psi^{-}\rangle_{15}$ & $\frac{\alpha^{2}(|000\rangle+|011\rangle)_{234}-\beta^{2}(|101\rangle+|110\rangle)_{234}}{\sqrt{2\left(\alpha^{4}+\beta^{4}\right)}}$ & [\
]{} $|\phi^{+}\rangle_{15}$ & $\frac{\alpha\beta(|\psi^{+}1\rangle+|\phi^{+}0\rangle)_{234}}{\sqrt{2}}$ & $I\otimes X$ & $\frac{(|\psi^{+}0\rangle+|\phi^{+}1\rangle)_{234}}{\sqrt{2}}$[\
]{} $|\phi^{-}\rangle_{15}$ & $\frac{\alpha\beta(|\psi^{-}1\rangle+|\phi^{-}0\rangle)_{234}}{\sqrt{2}}$ & $iY\otimes I$ & $\frac{(|\psi^{+}0\rangle+|\phi^{+}1\rangle)_{234}}{\sqrt{2}}$[\
]{}
ECPs for quantum states of the form $\alpha|\Psi_{0}\rangle|0\rangle+\beta|\Psi_{1}\rangle|1\rangle$ \[sec:A-generalized-scheme\]
=================================================================================================================================
In this section we will propose two ECPs for non-maximally entangled $(n+1)$-qubit state of the following form $$|\psi\rangle=\alpha|\Psi_{0}\rangle|0\rangle+\beta|\Psi_{1}\rangle|1\rangle,\label{eq:channel}$$ $|\Psi_{0}\rangle$ and $|\Psi_{1}\rangle$ are arbitrary $n$-qubit states that are mutually orthogonal. Before we propose the ECP it would be apt to mention a few words about the relevance of the states of the form $\alpha|\Psi_{0}\rangle|0\rangle+\beta|\Psi_{1}\rangle|1\rangle.$ This would justify why are we interested in constructing ECPs for states of this particular form. Clearly GHZ-like, GHZ, Bell and cat states described above are of this form. Further, recently we have shown that states of this particular form are useful in bidirectional quantum teleportation [@bi-directional] and hierarchical quantum communication schemes (e.g., hierarchical quantum information splitting (HQIS), probabilistic HQIS and hierarchical quantum secret sharing (HQSS)) [@HQIS]. These recently reported applications and the fact that many well-known entangled states are of these form indicate that the states of the form (\[eq:channel\]) are of particular importance. Its relevance can be further established by showing that 9 different families of SLOCC-nonequivalent 4-qubit entangled states can be expressed in this form.
In Section \[sec:Introduction\], we have already mentioned that there exist 9 families of 4-qubit entangled states. Following Verstraete et al. [@4-qubit; @SLOCC] we may describe them as
$$\begin{array}{lcl}
G_{abcd} & = & \frac{a+d}{2}(|0000\rangle+|1111\rangle)+\frac{a-d}{2}(|0011\rangle+|1100\rangle)+\frac{b+c}{2}(|0101\rangle+|1010\rangle)\\
& + & \frac{b-c}{2}(|0110\rangle+|1001\rangle),\\
L_{abc_{2}} & = & \frac{a+b}{2}(|0000\rangle+|1111\rangle)+\frac{a-b}{2}(|0011\rangle+|1100\rangle)+c(|0101\rangle+|1010\rangle)\\
& + & |0110\rangle,\\
L_{a_{2}b_{2}} & = & a(|0000\rangle+|1111\rangle)+b(|0101\rangle+|1010\rangle)+|0110\rangle+|0011\rangle,\\
L_{ab_{3}} & = & a(|0000\rangle+|1111\rangle)+\frac{a+b}{2}(|0101\rangle+|1010\rangle)+\frac{a-b}{2}(|0110\rangle+|1001\rangle)\\
& + & \frac{i}{\sqrt{2}}(|0001\rangle+|0010\rangle+|0111\rangle+|1011\rangle),\\
L_{a_{4}} & = & a(|0000\rangle+|0101\rangle+|1010\rangle+|1111\rangle)+(i|0001\rangle+|0110\rangle-i|1011\rangle),\\
L_{a_{2}0_{3\oplus\bar{1}}} & = & a(|0000\rangle+|1111\rangle)+|0011\rangle+|0101\rangle+|0110\rangle,\\
L_{0_{5\oplus\bar{3}}} & = & |0000\rangle+|0101\rangle+|1000\rangle+|1110\rangle,\\
L_{0_{7\oplus\bar{1}}} & = & |0000\rangle+|1011\rangle+|1101\rangle+|1110\rangle,\\
L_{0_{3\oplus\bar{1}}0_{3\oplus\bar{1}}} & = & |0000\rangle+|0111\rangle.\end{array}$$
Clearly, among 9 families listed above first 6 families (i.e., $G_{abcd},L_{abc_{2}},L_{a_{2}b_{2}},L_{ab_{3}},L_{a_{4}},L_{a_{2}0_{3\oplus\bar{1}}}$) are parameter-dependent and remaining 3 families are parameter independent. Now we may note that for the specific choices of parameters $a,\, b,\, c$ and $d,$ the parameter dependent families yield different quantum states of the form (\[eq:channel\]) and all the parameter independent families are already in form (\[eq:channel\]) as shown in last three rows of Table \[tab:9 families\]. Specific examples of interesting quantum states of the form (\[eq:channel\]) obtained from the parameter dependent families are also shown in Table \[tab:9 families\]. Interestingly, each of the 9 families contains state of the form (\[eq:channel\]). Since a state of a family can be transformed to any other state of the family by SLOCC, so if we can construct an ECP for the quantum states of the form (\[eq:channel\]) in general, that would imply that ECPs can be constructed for a large class of entangled states involving 4-qubits. Here we further note that several applications of the quantum states obtained as examples in $5^{th}$ Column of Table \[tab:9 families\] are known. For example, applications of Bell state and GHZ state are well-known, recently protocol of quantum dialogue using $Q_{4}$ and $Q_{5}$ is shown by us in Ref. [@GHZ-like2]. Further, we have recently shown that sates of this form are useful for various kind of hierarchical quantum communication [@HQIS]. The general nature and applicability of quantum states of the form (\[eq:channel\]) motivated us to construct an ECP for $\alpha|\Psi_{0}\rangle|0\rangle+\beta|\Psi_{1}\rangle|1\rangle$ in general. The same is described in the following section.
[|>p[1.3cm]{}|>p[3.3cm]{}|>p[5.4cm]{}|>p[6.2cm]{}|>p[1.8cm]{}|]{} Family of states & Values of the parameters $a,b,c,d$ & Corresponding normalized states $|\psi\rangle$ & State of the form $\frac{1}{\sqrt{2}}\left(|\Psi_{0}\rangle|0\rangle+|\Psi_{1}\rangle|1\rangle\right)$ that belong to the family & Name of the state[\
]{} $G_{abcd}$ & $\begin{array}{l}
a=d=\frac{1}{\sqrt{2}},b=c=0\\
a=1,b=c=d=0\end{array}$ & $\frac{1}{\sqrt{2}}(|0000\rangle+|1111\rangle),$
$\frac{1}{2}\left(|0000\rangle+|1111\rangle+|0011\rangle+|1100\rangle\right)$
$=\frac{1}{2}[(|00\rangle+|11\rangle)\otimes(|00\rangle+|11\rangle)]$ & $\frac{1}{\sqrt{2}}(|000\rangle|0\rangle+|111\rangle|1\rangle),$
$\frac{1}{\sqrt{2}}[(|0\rangle|0\rangle+|1\rangle|1\rangle]$ & cat state,
Bell state[\
]{} $L_{abc_{2}}$ & $a=b=1,$ $c=0$ & $\frac{1}{\sqrt{3}}\left(|0000\rangle+|1111\rangle+|0110\rangle\right)$ & $\frac{1}{\sqrt{3}}[(|000\rangle+|011\rangle)|0\rangle+|111\rangle|1\rangle]$ & -[\
]{} $L_{a_{2}b_{2}}$ & $a=1,b=0$ & $\frac{1}{2}\left(|0000\rangle+|1111\rangle+|0110\rangle+|0011\rangle\right)$ & $\frac{1}{2}[(|000\rangle+|011\rangle)|0\rangle+(|111\rangle+|001\rangle)|1\rangle]$ & -[\
]{} $L_{ab_{3}}$ & $a=b=0$ & $\frac{1}{2}(|0001\rangle+|0010\rangle+|0111\rangle+|1011\rangle)$ & $\frac{1}{2}[|001\rangle|0\rangle+(|000\rangle+|011\rangle+|101\rangle)|1\rangle]$ & -[\
]{} $L_{a_{4}}$ & $a=0$ & $\frac{1}{\sqrt{3}}(|0001\rangle+|0110\rangle+|1000\rangle)$ & $\frac{1}{\sqrt{3}}[(|011\rangle+|100\rangle)|0\rangle+|000\rangle|1\rangle]$ & -[\
]{} $L_{a_{2}0_{3\oplus\bar{1}}}$ & $a=0$ & $|0011\rangle+|0101\rangle+|0110\rangle$$=|0\rangle\otimes\frac{1}{\sqrt{3}}(|011\rangle+|101\rangle+|110\rangle)$ & $\frac{1}{\sqrt{3}}[|11\rangle|0\rangle+(|01\rangle+|10\rangle)|1\rangle]$ & 3-qubit$W$ state[\
]{} $L_{0_{5\oplus\bar{3}}}$ & parameter independent & $\frac{1}{2}\left(|0000\rangle+|0101\rangle+|1000\rangle+|1110\rangle\right)$ & $\frac{1}{2}[(|000\rangle+|100\rangle+|111\rangle)|0\rangle+|010\rangle|1\rangle]$ & $Q_{4}$ state [@Pati][\
]{} $L_{0_{7\oplus\bar{1}}}$ & parameter independent & $\frac{1}{2}(|0000\rangle+|1011\rangle+|1101\rangle+|1110\rangle)$ & $\frac{1}{2}[(|000\rangle+|111\rangle)|0\rangle+(|101\rangle+|110\rangle)|1\rangle]$ & $Q_{5}$ state [@Pati][\
]{} $L_{0_{3\oplus\bar{1}}0_{3\oplus\bar{1}}}$ & parameter independent & $|0000\rangle+|0111\rangle$$=|0\rangle\otimes\frac{1}{\sqrt{2}}(|000\rangle+|111\rangle)$ & $\frac{1}{\sqrt{2}}(|00\rangle|0\rangle+|11\rangle|1\rangle)$ & $GHZ$ state[\
]{}
ECP1 for quantum states of the form
------------------------------------
The ECP is described here through the quantum circuit shown in the Fig. \[fig:General-circuit-for\]. In this quantum circuit initial $(n+1)$-qubit state $\alpha|\Psi_{0}\rangle|0\rangle+\beta|\Psi_{1}\rangle|1\rangle$ is combined with $|\psi\rangle_{{\rm Bell}}$ and we obtain the combined input state as
$$\begin{array}{lcl}
|\psi_{5}\rangle & = & \left(\alpha|\Psi_{0}\rangle|0\rangle+\beta|\Psi_{1}\rangle|1\rangle\right)_{1,2,\cdots,n+1}\otimes(\alpha|00\rangle+\beta|11\rangle)_{n+2,n+3}\\
& = & \left(\alpha^{2}|\Psi_{0}\rangle|000\rangle+\alpha\beta|\Psi_{1}\rangle|100\rangle+\alpha\beta|\Psi_{0}\rangle|011\rangle+\beta^{2}|\Psi_{1}\rangle|111\rangle\right)_{1,2,\cdots,n+1,n+2,n+3}.\end{array}$$ After swapping the $(n+1)$-th with $(n+2)$-th qubits we obtain $$\begin{array}{lcl}
|\psi_{6}\rangle & = & \left(\alpha^{2}|\Psi_{0}\rangle|000\rangle+\alpha\beta|\Psi_{1}\rangle|010\rangle+\alpha\beta|\Psi_{0}\rangle|101\rangle+\beta^{2}|\Psi_{1}\rangle|111\rangle\right)_{1,2,\cdots,n,n+2,n+1,n+3}\\
& = & \frac{1}{\sqrt{2}}\left(\left(\alpha^{2}|\Psi_{0}\rangle|0\rangle+\beta^{2}|\Psi_{1}\rangle|1\rangle\right)|\psi^{+}\rangle+\left(\alpha^{2}|\Psi_{0}\rangle|0\rangle-\beta^{2}|\Psi_{1}\rangle|1\rangle\right)|\psi^{-}\rangle\right.\\
& + & \left.\alpha\beta\left(|\Psi_{0}\rangle|1\rangle+|\Psi_{1}\rangle|0\rangle\right)|\phi^{+}\rangle+\alpha\beta\left(|\Psi_{0}\rangle|1\rangle-|\Psi_{1}\rangle|0\rangle\right)|\phi^{-}\rangle\right)_{1,2,\cdots,n,n+2,n+1,n+3}.\end{array}$$
Now a Bell measurement is performed on the last two qubits of $|\psi_{6}\rangle.$ If the Bell measurement yields $|\psi^{\pm}\rangle$ then the protocol fails, but if it yields $|\phi^{+}\rangle$ $\left(|\phi^{-}\rangle\right)$ then we can obtain the desired state (i.e., $\left(\frac{|\Psi_{0}\rangle|0\rangle+|\Psi_{1}\rangle|1\rangle}{\sqrt{2}}\right)_{1,2,\cdots,n,n+2}$) by applying $X$ $(iY)$ on the $(n+2)$-th qubit. This provides a simple, but very useful ECP schemes for quantum states of the form (\[eq:channel\]) in general. Consequently, we obtain ECP for GHZ-like state, GHZ-state, 9 families of SLOCC-nonequivalent 4-qubit entangled state, cluster state, cat state etc. In this ECP, knowledge of $\alpha,$ $\beta$ is required for the construction of $|\psi\rangle_{{\rm Bell}},$ but $\alpha,\,\beta$ can be complex. In what follows we propose another alternative quantum circuit for ECP of $\alpha|\Psi_{0}\rangle|0\rangle+\beta|\Psi_{1}\rangle|1\rangle.$
![\[fig:General-circuit-for\]Quantum circuit for ECP for partially entangled $(n+1)$-qubit state of the form $\alpha|\Psi_{0}\rangle|0\rangle+\beta|\Psi_{1}\rangle|1\rangle.$ ](fig4)
ECP2 for quantum states of the form
------------------------------------
In this subsection we propose an alternative circuit for ECP for quantum states of the form $\alpha|\Psi_{0}\rangle|0\rangle+\beta|\Psi_{1}\rangle|1\rangle$. The proposed ECP is shown in the Fig. \[fig:alternative ckt for ECP\]. In this case input state is $$\begin{array}{lcl}
|\psi_{5}\rangle & = & \left(\alpha|\Psi_{0}\rangle|0\rangle+\beta|\Psi_{1}\rangle|1\rangle\right)_{1,2,\cdots,n+1,}\otimes|0\rangle_{n+2}\\
& = & \frac{1}{\sqrt{2}}\left(\alpha|\Psi_{0}\rangle|00\rangle+\beta|\Psi_{1}\rangle|10\rangle\right)_{1,2,\cdots,n+1,n+2,}\end{array}$$ Now a ${\rm CNOT}_{\left(n+1\right)\rightarrow\left(n+2\right)}$ gate is applied on $|\psi_{5}\rangle$ using $(n+1)$-th qubit as control qubit and $(n+2)$-th qubit as target qubit to yield $$|\psi_{6}\rangle=\frac{1}{\sqrt{2}}\left(\alpha|\Psi_{0}\rangle|00\rangle+\beta|\Psi_{1}\rangle|11\rangle\right)_{1,2,\cdots,n+1,n+2}.$$ Now we may apply a unitary operator $U_{2}=\left(\begin{array}{cc}
\alpha & \beta\\
-\beta & \alpha\end{array}\right)$ on $(n+2)$-th qubit to obtain (here $\alpha,$ $\beta$ are real)[^3]
$$\begin{array}{lcl}
|\psi_{7}\rangle & = & U_{2,n+2}|\psi_{6}\rangle\\
& = & \left(\alpha^{2}|\Psi_{0}\rangle|00\rangle-\alpha\beta|\Psi_{0}\rangle|01\rangle+\beta^{2}|\Psi_{1}\rangle|10\rangle+\alpha\beta|\Psi_{1}\rangle|11\rangle\right)_{1,2,\cdots,n+1,n+2}\\
& = & \left(\left(\alpha^{2}|\Psi_{0}\rangle|0\rangle+\beta^{2}|\Psi_{1}\rangle|1\rangle\right)|0\rangle-\alpha\beta\left(|\Psi_{0}\rangle|0\rangle-|\Psi_{1}\rangle|1\rangle\right)|1\rangle\right)_{1,2,\cdots,n+1,n+2},\end{array}$$
and subsequently measure the last qubit (i.e., $\left(n+2\right)$-th qubit) in computational basis. If the measurement result yields $|0\rangle$ then the protocol fails, otherwise we apply $Z=\sigma_{z}$ gate on the $\left(n+1\right)$-th qubit to obtain the desired maximally entangled state: $$\left(\frac{|\Psi_{0}\rangle|0\rangle+|\Psi_{1}\rangle|1\rangle}{\sqrt{2}}\right)_{1,2,\cdots,n+1}.$$ Thus we have two alternative ECPs for quantum states of the form (\[eq:channel\]) in general. In the first scheme (ECP1) we don’t need $\alpha,\beta$ to be real, while the same is required in the second scheme (ECP2). This is indicative of superiority of ECP1 over ECP2. Still ECP2 is interesting for its extreme simplicity. Specifically, ECP2 neither require any Bell measurement nor any particle swapping. Further, as the amount of initial entanglement is less compared to ECP1, its efficiency would be more compared to ECP1 if we use Sheng et al.’s quantitative measure of quality of ECP [@Sheng-1]. This point is elaborated in Section \[sec:Efficiency\].
![\[fig:alternative ckt for ECP\] An alternative quantum circuit for ECP for partially entangled $(n+1)$-qubit state of the form $\alpha|\Psi_{0}\rangle|0\rangle+\beta|\Psi_{1}\rangle|1\rangle$. Here CM stands for measurement in computational basis and PF stands for protocol fails.](fig5)
Optical implementation \[sec:Optical-implementation\]
=====================================================
Bell measurement was used in one of the pioneering work on ECP [@S.; @Bose]. However, in most of the subsequent works the use of Bell measurement was criticized by noting that the Bell measurement is a complex task and using linear optics we cannot distinguish all the four Bell states [@criticized; @bell]. Here it is important to note that all the Bell-measurement based ECPs proposed here fail when outcome of the Bell measurement is $|\psi^{\pm}\rangle$. Thus for the practical implementation of the first two ECPs proposed here it would be sufficient to distinguish $|\phi^{+}\rangle$ and $|\phi^{-}\rangle.$ A simple linear optics setup shown in Fig. \[fig:Bell measurement\] can perform the task [@Lee-Bell; @measuremnt]. The working of the optical circuit shown in Fig. \[fig:Bell measurement\] is elaborately described in Ref. [@Lee-Bell; @measuremnt]. Here for the completeness of our discussion we may briefly note that when information is encoded using polarization degree of freedom then usually horizontal ($H$) and vertical ($V$) polarized states represent $|0\rangle$ and $|1\rangle,$ respectively. Thus Bell states can be expressed as $|\psi^{\pm}\rangle=\frac{1}{\sqrt{2}}\left(|HH\rangle\pm|VV\rangle\right)$ and $|\phi^{\pm}\rangle=\frac{1}{\sqrt{2}}\left(|HV\rangle\pm|VH\rangle\right)$. If $|\phi^{+}\rangle$ enters the optical circuit then the detectors click as either $H_{{\rm up}},V_{{\rm down}}$ or $V_{{\rm {\rm up}}},H_{{\rm down}}$ where the subscript up (down) denotes the outcome of top (bottom) two detectors. Similarly, when $|\phi^{-}\rangle$ enters the optical circuit then the detectors clicks as either $H_{{\rm up}},H_{{\rm down}}$ or $V_{{\rm {\rm up}}},V_{{\rm down}}$. Thus $|\phi^{+}\rangle$ can be distinguished from $|\phi^{-}\rangle$. However, we cannot distinguish $|\psi^{+}\rangle$ and $|\psi^{-}\rangle$ as in both of the cases detectors click as $H_{{\rm up}},V_{{\rm up}}$ or $H_{{\rm {\rm down}}},V_{{\rm down}}$. Thus for the Bell-measurement based ECPs proposed here if both of the upper or lower detectors click then the protocol fails, otherwise we apply appropriate unitary operations as described above. Further, the ${\rm CNOT}$ used in ECP2 can be implemented using optical circuits implemented by J. L. OBrien et al. [@CNOT-obrien]. Thus in general ECPs proposed here can be realized optically. However, the applicability of the circuits is not limited to optical realization. For example, these ECPs may be practically realized using NMR as Bell measurement is possible in NMR based technologies [@NMR-Anil; @Kumar].
![\[fig:Bell measurement\]A linear optics based scheme for Bell-state measurement for (a) single photon polarization qubit. The scheme uses polarizing Beam splitters (PBSs) that allows horizontally polarized photon to transmit and reflects the vertically polarized photon, wave plates, and on/off photo detectors [@Lee-Bell; @measuremnt].](OpticalCkt)
Efficiency\[sec:Efficiency\]
============================
Recently Sheng et al. [@Sheng-1] have introduced a quantitative measure of entanglement concentration efficiency of an ECP. They have referred to it as *entanglement transformation efficiency* $\eta$ and explicitly defined it as
$$\eta=\frac{E_{c}}{E_{0}}\label{eq:efficiency1}$$
where, $E_{0}$ is the amount of entanglement in the initial partially entangled state and $E_{c}$ is the amount of entanglement of the state after concentration. Further, they have defined $E_{c}$ as
$$E_{c}=P_{s}\times E_{m}+(1-P_{s})E^{\prime}\label{eq:efficiency2}$$
where $P_{s}$ is the success probability of obtaining the maximally entangled state on execution of the ECP and $E_{m}$ is the amount of entanglement in the maximally entangled state. Sheng et al. assumed that the measure of entanglement is chosen in such a way that the amount of entanglement in a maximally entangled state is 1, however it may be different in general. For example, we often use a definition of negativity where negativity of maximally entangled Bell state is $0.5$ and log negativity is $1.$ Thus the first (second) term of $E_{c}$ corresponds to success (failure) of the ECP. Up to this point the definition of $\eta$ seems clear and straight forward. However, there exists an ambiguity in the definition, it is not clear whether $E_{0}$ is the amount of entanglement of the state to be concentrated or that of the entire initial state. To remove this ambiguity we choose $E_{0}$ to be the total initial entanglement. This choice naturally implies higher efficiency of single photon assisted ECPs over Bell-type state assisted ECPs (such as the first 2 ECPs of the present paper). Further, a closer look into (\[eq:efficiency1\]) would reveal that it is neither unique nor easily expendable to the multipartite case. Specifically, the definition does not define which measure of entanglement is to be used for obtaining $E_{0}.$ Existence of different measures of bipartite entanglement and the fact that these measures are not monotone of each other makes the definition (\[eq:efficiency1\]) non-unique. In fact, Sheng et al. [@Sheng-1] have used von Neumann entropy as a measure of entanglement, but von Neumann entropy is a good measure of entanglement for bipartite systems only. This limitation exists for most of the well known measures of entanglement and this fact leads to an interesting question: How to find $\eta$ for an ECP that is designed for multipartite case. Interestingly, the problem is equivalent to provide a quantitative measure of multipartite entanglement. In last two decades several efforts have been made to introduce measures of multipartite entanglement [@ent-measure-1; @ent-measure-2; @ent-measure-3; @ent-measure-4; @ent-measure-5; @ent-measure-6]. We may use some of the approaches followed in [@ent-measure-1; @ent-measure-2; @ent-measure-3; @ent-measure-4; @ent-measure-5; @ent-measure-6] to obtain $\eta$ for multipartite ECP. To show the dependence of $\eta$ one choice of entanglement measure we may choose *tangle* [@ent-measure-2; @ent-measure-3; @ent-measure-5] as a measure of entanglement. In that case, entanglement of $|\psi\rangle_{{\rm Bell}}$ is $4|\alpha\beta|^{2}$ and success probability for all ECPs presented here and for Sheng et al.S ECP is $2|\alpha\beta|^{2}.$ Thus $\eta$ for Sheng protocol and our last protocol (i.e., ECP2) will be $\frac{1}{2}.$ Whereas that of our first two ECPs and ECP of Zhao [@zhao-proposal] will be $\frac{1}{4}.$ As the tangle for a partially entangled GHZ state is same as that of a partially entangled Bell state. Efficiency of our protocol would remain same ($\frac{1}{4}$) for ECP for partially entangled GHZ state. Clearly in all these three cases $\eta$ is independent of $\alpha$ which is in contrast with the result obtained by Sheng using von Neumann entropy as a measure of entanglement. Specifically, they had observed $\eta$ was a function of $\alpha$. To extend the definition of efficiency $\eta$ to the multipartite case and to elaborate it’s dependence on choice of the entanglement measure we may note that in 2004, Yu and Song established [@ent-measure-4] that any good measure $M_{A-B}$ of bi-partite entanglement can be generalized to multipartite systems, by considering bipartite partitions of the multipartite system. Yu and Song defined a simple measure of tripartite entanglement as $$M_{ABC}=\frac{1}{3}\left(M_{A-BC}+M_{B-AC}+M_{C-AB}\right),\label{eq:yu song}$$ where $M_{i-jk}$ is a measure of entanglement between subsystem $i$ and subsystem $jk$. $M_{i-jk}$ may be any good measure of bi-partite entanglement (e.g., von Neumann entropy, negativity etc.). Yu and Song’s idea was used to measure tripartite entanglement in various systems using different measures of bipartite entanglement e.g., negativity, concurrence, and von Neumanns entropy (cf. [@ent-measure-1] and references therein). However, some limitations of the above measure (\[eq:yu song\]) were found and a new measure of tripartite entanglement was introduced by Sabin and Garca-Alcaine by replacing arithmetic mean present in (\[eq:yu song\]) by geometric mean. Thus Sabin and Garca-Alcaine’s measure of tripartite entanglement is given as [@ent-measure-1]$$M_{ABC}=\left(M_{A-BC}M_{B-AC}M_{C-AB}\right)^{\frac{1}{3}}.\label{eq:sabin}$$ In what follows we have provided analytic expressions for efficiencies of ECPs proposed here using (\[eq:efficiency1\]) and (\[eq:sabin\]). To be precise, if we use negativity as a measure of bipartite entanglement then the efficiency of the first two protocols (i.e., protocols assisted by $|\psi\rangle_{{\rm Bell}}$) proposed here are as follows $$\eta_{{\rm Bell},{\rm GHZ}}^{{\rm Bell-type}}=\frac{2|\alpha\beta|^{2}}{2|\alpha\beta|}=|\alpha\beta|,\label{eq:eta1}$$ and
$$\eta_{{\rm GHZ-like}}^{{\rm Bell}{\rm -type}}=\frac{2|\alpha\beta|^{2}}{\sqrt[3]{\frac{1}{4}|\alpha\beta|}+|\alpha\beta|}.\label{eq:eta2}$$
Similarly, in case of ECP2 (i.e., for the single qubit assisted protocol) we obtain $$\eta_{{\rm Bell},{\rm GHZ}}^{{\rm 1-qubit}}=2|\alpha\beta|,\label{eq:eta3}$$ and
$$\eta_{{\rm GHZ-like}}^{{\rm 1-qubit}}=\frac{4|\alpha\beta|^{2}}{\sqrt[3]{2|\alpha\beta|}}.\label{eq:eta4}$$
From the above equations it is clear that the ECP2 is more efficient than ECP1. The same is illustrated in Fig. \[fig:Variaontion-of-eta\]. In the left (right) panel of Fig. \[fig:Variaontion-of-eta\] the variation of efficiency of ECPs proposed for partially entangled Bell and GHZ state (GHZ-like state) with $\alpha$ are shown.
 Variation of $\eta$ with $\alpha$. (a) for Bell and GHZ states and (b) for GHZ-like state. Solid line represents ECPs realized using assistance of $|\psi\rangle_{{\rm Bell}}$ and the dashed line represent single-qubit assisted ECP. ](f1)
Conclusion\[sec:Conclusion\]
============================
We have proposed three ECPs in the present paper. The first one is shown to generate a maximally entangled cat state from the corresponding partially entangled state. A modified version of this ECP is also introduced as an ECP for GHZ-like state. ECPs for cat states were proposed earlier, too. However, no ECP for GHZ-like states were proposed until now. Thus this is the first ever ECP reported precisely for GHZ-like state. The last two ECPs are designed for quantum states of the general form $\alpha|\Psi_{0}\rangle|0\rangle+\beta|\Psi_{1}\rangle|1\rangle.$ These two protocols are extremely interesting as several applications of the states of the form $\alpha|\Psi_{0}\rangle|0\rangle+\beta|\Psi_{1}\rangle|1\rangle$ are reported in recent past [@GHZ-like-1; @GHZ-like2; @GHZ-like3; @bi-directional; @HQIS]. Further, its very important as specific states from the 9 families of SLOCC-nonequivalent 4-qubit entangled states can be described in this form. Thus the proposed ECPs are valid for the 9 families of 4-qubit entangled states. Further, partially entangled cat-like, GHZ-like, GHZ, W and Bell states can also be expressed as $\alpha|\Psi_{0}\rangle|0\rangle+\beta|\Psi_{1}\rangle|1\rangle$. The ECPs are not described in the usual style, rather they are described as quantum circuit. From the quantum circuits described above one can clearly see that all the ECPs proposed here require local measurement, classical communication and post selection. According to Vedral et al. [@Vedral] these are the basic steps required by any good ECP or EP. Further, the efficiency of the proposed protocols are discussed in detail using a quantitative measure of efficiency that was recently introduced by Sheng et al. [@Sheng-1]. Apparently any Bell-type state assisted ECP (e.g., the first two ECPs of the present paper and the Zhao et al. proposal [@zhao-proposal]) will have lesser efficiency compared to linear optics-based single qubit assisted ECPs [@Sheng-1]. Again if we go beyond linear optics and use nonlinear resources then the efficiency would increase further. However, this parameter cannot be considered as a basis of choosing ECPs as the measure of efficiency introduced by Sheng et al. [@Sheng-1] is really a weak measure. Keeping these in mind and the fact that the proposed ECPs that are applicable to a large class of quantum states of practical interest are experimentally realizable using linear optical resources and NMR, we conclude the paper with an expectation that experimentalists will find it interesting to implement the ECPs proposed here.
**Acknowledgment:** AP thanks Department of Science and Technology (DST), India for support provided through the DST project No. SR/S2/LOP-0012/2010 and he also acknowledges the supports received from the projects CZ.1.05/2.1.00/03.0058 and CZ.1.07/2.3.00/20.0017 of the Ministry of Education, Youth and Sports of the Czech Republic.
[51]{} J. Eisert, K. Jacobs, P. Papadopoulos and M. B. Plenio, Phys. Rev. A **62**, 052317 (2000).
M. Gupta and A. Pathak, Annalen der Physik **16**, 791 (2007).
C. H. Bennett, G. Brassard, C. Cr�peau, R. Jozsa, A. Peres and W. K. Wootters, Phys. Rev. Lett. **70**, 1895 (1993).
C. H. Bennett and S. J. Wiesner, Phys. Rev. Lett. **69**, 2881 (1992).
A. K. Ekert, Phys. Rev. Lett. **67**, 661 (1991).
K. Bostrom and T. Felbinger, Phys. Rev. Lett. **89**, 187902 (2002).
F. G. Deng, G. L. Long and X. S. Liu, Phys. Rev. A **68**, 042317 (2003).
C. Shukla, A. Pathak and R. Srikanth, Int. J. Quant. Info. **10**, 1241009 (2012).
C. H. Bennett, H. J. Bernstein, S. Popescu and B. Schumacher, Phys. Rev. A **53**, 2046 (1996).
C. H. Bennett, G. Brassard, S. Popescu, B. Schumacher, J. A. Smolin and W. K. Wootters, Phys. Rev. Lett. **76**, 722 (1996).
S. Bose, V. Vedral and P. L. Knight, Phys. Rev. A **60**, 194 (1999).
S. Bandyopadhyay, Phys. Rev. A **62**, 032308 (2000).
Z. Zhao, J. W. Pan and M. S. Zhan, Phys. Rev. A **64,** 014301 (2001).
T. Yamamoto, M. Koashi and N. Imoto, Phys. Rev. A **64**, 012304 (2001).
Y.-B. Sheng, L. Zhou, S.-M. Zhao and B.-Y. Zheng, Phys. Rev. A **85**, 012307 (2012).
Y.-B. Sheng and L. Zhou, Entropy **15**, 1776 (2013).
Y.-J. Gu, W.-D. Li and G.-C. Guo, Phys. Rev. A **73**, 022321 (2006).
F.-G. Deng, Phys. Rev. A **85**, 022311 (2012).
Y.-B. Sheng, L. Zhou and S.-M. Zhao, Phys. Rev. A **85**, 042302 (2012).
L.-y. He, C. Cao and C. Wang, Opt. Commun. **** **** .
B. S. Choudhury and A. Dhara, Int. J. Theor. Phys. **52**, 3965 (2013).
L. Zhou, X.-F. Wang and Y.-B. Sheng Int. J. Theor. Phys. DOI 10.1007/s10773-013-1974-8 (2014).
B. S. Choudhury and A. Dhara, Quant. Inf. Process. **12**, 2577 (2013).
T.-T. Xu, W. Xiong and L. Ye, Int. J. Theor. Phys. **52**, 2981 (2013).
S.-Y. Zhao, J. Liu, L. Zhou and Y.-B. Sheng, Quant. Inf. Process. **12**, 3633 (2013).
L. Zhou, Y.-B. Sheng, W.-W. Cheng, L.-Y. Gong and S.-M. Zhao, Quant. Inf. Process. **12**, 1307 (2013).
Z. Zhao, T. Yang, Y. A. Chen, A. N. Zhang and J. W. Pan, Phys. Rev. Lett. **90**, 207901 (2003).
T. Yamamoto, M. Koashi, S. K. Ozdemir and N. Imoto, Nature **421**, 343 (2003).
W. ${\rm D\ddot{u}r}$, G. Vidal and J. I. Cirac, Phys. Rev. A **62**, 062314 (2000).
A. Banerjee, K. Patel and A. Pathak, Int. J. Theor. Phys. **50**, 507 (2011).
C. Shukla, V. Kothari, A. Banerjee and A. Pathak, Phys. Lett. A **377**, 518 (2013).
K. Yang, L. Huang, W. Yang and F. Song, Int. J. Theor. Phys. **48**, 516 (2009).
F. Verstraete, J. Dehaene, B. De Moor and H. Verschelde, Phys. Rev. A **65**, 052112 (2002).
O. Chterental and D. Z. Djokovic, arxiv:0612184 [\[]{}quant-ph[\]]{}.
L. Borsten, D. Dahanayake, M. J. Duff, A. Marrani and W. Rubens, Phys. Rev. Lett. **105**, 100507 (2010).
G. Gour and N. R. Wallach, J. Math. Phys. **51**, 112201 (2010).
J. L. Romero, L. Roa, J. C. Retamal and C. Saavedra, Phys. Rev. A **65**, 052319 (2002).
J. R. Samal, M. Gupta, P. K. Panigrahi and A. Kumar, J. Phys. B **43**, 095508 (2010).
C. Shukla, A. Banerjee and A. Pathak, Int. J. Theor. Phys. **52**, 3790 (2013).
C. Shukla and A. Pathak, Phys. Lett. A **377**, 1337 (2013).
B. Pradhan, P. Agrawal and A. K. Pati, arxiv:0705.1917v1 [\[]{}quant-ph[\]]{}.
S.-W. Lee and H. Jeong, arxiv:1304.1214 [\[]{}quant-ph[\]]{}.
S.-H. Xiang, Z.-G. Shi, W. Wen, D.-H. Lu, X.-X. Zhu, K.-H. Song, Opt. Commun. **284**, 2402 (2011).
J. L. O’Brien, G. J. Pryde, A. G. White, T. C. Ralph and D. Branning, Nature **426**, 264 (2003).
C. Sabin and G. Garca-Alcaine, Eur. Phys. J. D **48**, 435 (2008).
C. Eltschka, A. Osterloh, J. Siewert and A. Uhlmann, New J. Phys. **10**, 043014 (2008).
W. K. Wootters, arxiv:1402.2219v1 [\[]{}quant-ph[\]]{}.
C-s. Yu, H.-s. Song, Phys. Lett. A **330**, 377 (2004).
V. Coffman, J. Kundu and W. K. Wootters, Phys. Rev. A **61**, **** 052306 (2000).
P. Rungta, V. Buzek, C. M. Caves, M. Hillery and G. J. Milburn, Phys. Rev. A **64**, 042315 (2001).
V. Vedral and M. B. Plenio, Phys. Rev. A **57**, 1619 (1998).
[^1]: Email: [email protected], Phone: +91 9717066494
[^2]: This swapping operation that changes the particle sequence as $12345\rightarrow15234$ can be viewed as a sequence of three conventional SWAP gates that work as follows: $12345\rightarrow15342\rightarrow15243\rightarrow15234$.
[^3]: Here $U_{2,n+2}|\psi_{6}\rangle$ denotes that single qubit operation $U_{2}$ operates on $(n+2)$-th qubit of $|\psi_{6}\rangle$ and identity operators operate on rest of the qubits.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In neutral graphene, two prominent cusps known as Kohn anomalies are found in the phonon dispersion of the highest optical phonon at $q=\Gamma$ (LO branch) and $q=K$ (TO branch), reflecting a significant electron-phonon coupling (EPC) to undoped Dirac electrons. In this work, high-resolution electron energy loss spectroscopy is used to measure the phonon dispersion around the $\Gamma$ point in quasi-freestanding graphene epitaxially grown on Pt(111). The Kohn anomaly for the LO phonon is observed at finite momentum $q\sim2k_F$ from $\Gamma$, with a shape in excellent agreement with the theory and consistent with known values of the EPC and the Fermi level. More strikingly, we also observe a Kohn anomaly at the same momentum for the out-of-plane optical phonon (ZO) branch. This observation is the first direct evidence of the coupling of the ZO mode with Dirac electrons, which is forbidden for freestanding graphene but becomes allowed in the presence of a substrate. Moreover, we estimate the EPC to be even greater than that of the LO mode, making graphene on Pt(111) an optimal system to explore the effects of this new coupling in the electronic properties.'
author:
- Antonio Politano
- Fernando de Juan
- Gennaro Chiarello
- 'Herbert A. Fertig'
bibliography:
- 'kohn.bib'
title: 'Emergence of a ZO Kohn anomaly in quasi-freestanding epitaxial graphene'
---
*Introduction* - Kohn anomalies are kinks in the phonon dispersion of a material produced by abrupt variations in the screening of atomic vibrations by gapless electrons [@K59; @BRZ09; @AKB08]. The properties of these anomalies are completely determined by the electron-phonon coupling (EPC) and the shape of the Fermi surface, and their observation thus offers a window to study the interplay between electronic properties and phonon dynamics.
As a paradigmatic example, in graphite [@MRT04; @PLM04; @LPM06; @MMD07] and graphene [@LM06; @ZSF08; @THD08; @SYM09; @MMD09; @MKD10; @JF12a; @JF12], Kohn anomalies are realized as linear cusps in the dispersion of the highest optical phonon branches at $\Gamma$ (the $E_{2g}$ phonon) and at $K$ (the $A_1'$ phonon). The form and position of this cusp is determined by the Dirac fermion dispersion of the electronic $\pi$ bands around $K$ and the Fermi level $\mu \approx 0$. These two in-plane phonon branches are commonly accepted to be the only ones with a significant EPC in graphene [@BPF09]. Within a tight-binding picture, these are the only branches that modify the nearest neighbor hopping integrals, so other in-plane phonons require a different mechanism to couple to electrons.
For the case of out-of-plane phonons, the EPC is strongly constrained by the presence of mirror symmetry with respect to the horizontal plane which forbids a first order coupling to electrons, at least in free-standing graphene. When graphene is supported by a substrate a first order EPC becomes allowed in principle [@L11; @B08], but it has remained unknown whether its strength is large enough to produce any significant effect. In particular, for the out-of-plane optical mode (the ZO mode), it has been speculated that a finite coupling could be responsible for a Peierls instability to a spontaneous buckling of the lattice [@FL07; @ZKM11], which has not been observed so far. A coupling of the ZO mode to electrons would influence many physical properties, since the ZO has significantly lower energy than the in-plane branches with strong EPC. The observation of a Kohn anomaly for the ZO phonon [@GP09] would represent a definitive proof of the existence of the EPC for this phonon branch and would allow an estimate of it. Herein, we demonstrate the existence of precisely this Kohn anomaly by means of high-resolution electron energy loss spectroscopy [@PCB13] (HREELS) measurements on monolayer graphene grown on Pt(111). In addition, we also present evidence of the expected Kohn anomaly for the $E_{2g}$ phonon branch.
Graphene on Pt(111) is an ideal playground for investigating the possible emergence of a ZO Kohn anomaly, as it is characterized by the weakest graphene-substrate interaction [@GPH11]. The graphene-Pt distance (3.31 $\rm{\AA}$) lies close to the $c$-axis spacing in graphite. Moreover, the electronic structure of graphene on Pt(111) resembles that of isolated graphene [@SSS09], with the Dirac fermion dispersion of $\pi$ bands preserved. Angle-resolved photoemission spectroscopy (ARPES) experiments [@SSS09] do not show any significant hybridization of the graphene $\pi$ states with Pt $d$-states, which simply superpose in energy with minimal interaction between them. This is in contrast to the case of, for example, graphene on Ni(111), where the hybridization of the graphene $\pi$-states with the Ni $d$ bands has a very strong effect on the $\pi$ bands [@AW10]. The graphene sheet grown on Pt(111) is $p$-doped with a shift by $0.30\pm 0.15$ eV of the Fermi level from the Dirac point.
![HREELS intensity plot for phonon dispersion of graphene/Pt(111). Phonon modes have been recorded in HREELS spectra acquired in off- specular geometry with the sample oriented along the $\Gamma-M$ direction. The incidence angle is $\theta_i= 80 \degree$ with respect to the surface normal and the primary electron beam energy is $E_p = 20$ eV. To put in evidence only inelastic losses due to phonons, the tail of the elastic peak is grayed out for clarity. The LO, LA, ZO and ZA branches are clearly identified, while the TO and TA are barely visible due to a selection rule [@JPC14]. Note the very low intensity of the ZO mode close to $\Gamma$. Yellow lines are fits to the dispersion obtained with higher resolution data, where Kohn anomalies can be appreciated (see text for details).[]{data-label="EELSmap"}](fig1.jpg){width="8cm"}
.
*Experimental methods* - Experiments were carried out in an ultra-high vacuum chamber operating at a base pressure of $5\cdot 10^{-9}$ Pa. The sample was a single crystal of Pt(111). The substrate was cleaned by repeated cycles of ion sputtering and annealing at 1300 K. Surface cleanliness and order were checked using Auger electron spectroscopy and low-energy electron diffraction measurements, respectively. Graphene was attained by dosing ethylene onto the clean Pt(111) substrate held at 1150 K. The completion of the first layer was reached upon an exposure of $3\cdot10^{-8}$ mbar for ten minutes (24 $L$, 1 $L=1.33\cdot10^{-6}$ mbar·s). The graphene layer was characterized by Raman spectroscopy, low-energy electron diffraction, and scanning electron microscopy experiments [@CCC13; @SM]. In particular, Raman measurements indicate the unique presence of monolayer graphene domains.
HREELS experiments were performed by using an electron energy loss spectrometer (Delta 0.5, SPECS) with an energy resolution ranging from 1 to 3 meV. The dispersion of the peaks in the energy loss $E_{\rm{loss}}$ was measured by moving the analyzer while keeping the sample and the monochromator fixed. The phonon in-plane momentum was determined from $\vec q_{\parallel} = \vec k_i\sin \theta_i-\vec k_s \sin \theta_s$ as $$q_{\parallel} = \sqrt{2m E_p} \left(\sin \theta_i-\sqrt{1-E_{\rm{loss}}/E_p} \sin \theta_s \right)$$ where $\theta_i$ and $\theta_s$ are the incident and scattering angles, and we set $\hbar=1$. The impinging energy $E_p$ and the incident angle $\theta_i$ were chosen so as to obtain the highest signal-to-noise ratio. A primary beam energy of $E_p$=20 eV provided the best compromise among surface sensitivity, the highest cross-section for mode excitation and momentum resolution. The angular acceptance of the apparatus was $\alpha=\pm0.5 \degree$, which determines the momentum resolution as $$\Delta q_{\parallel} = \sqrt{2m E_p}\left(\cos \theta_i+\sqrt{1-E_{\rm{loss}}/E_p} \cos \theta_s \right)\alpha$$ For the investigated range of $q_{\parallel}$, $\Delta q_{\parallel}$ was found to range from 0.005 near $\Gamma$ to 0.022 $\AA^{-1}$ at $K$. To obtain the energies of loss peaks, a polynomial background was subtracted from each spectrum. The resulting spectra were fitted by a Voigt line shape [@SM]. All measurements were made at room temperature.
![Selected phonon spectrum from the dataset of Fig \[EELSmap\] at $\theta_s= 48 \degree$, which corresponds to $q_\parallel = 0.56$ ${\rm \AA}^{-1}$. The TA and TO modes are not observed due to a selection rule. Inset: Phonon displacements for the four phonons observed. Crosses and circles denote displacements in and out of the plane respectively.[]{data-label="losspeaks"}](fig2.pdf){width="8cm"}
{width="18cm"}
*Experimental results* - The phonon spectrum in graphene is composed by six phonon branches. Four vibrate in-plane and are labeled transverse and longitudinal acoustic (TA and LA) and optical (TO and LO). The other two are acoustic and optical out-of-plane vibrations (ZA and ZO). Fig. \[EELSmap\] shows a map of the HREELS intensity as a function of energy and momentum, where four phonon branches (ZA,ZO,LA,LO) can be clearly observed. Their dispersion is consistent with previous works [@AW10; @PMC12; @EMW13]. The other two phonon branches (TO and TA) are practically absent because a selection rule that forbids the emission of odd phonons under reflections by the scattering plane [@JPC14]. This is also clearly observed in Fig. \[losspeaks\], where the energy loss is plotted for a selected angle, and four peaks are clearly identified. The displacement pattern for each of these four modes is shown as an inset to Fig. \[losspeaks\].
We now focus on the dispersion of the ZO and LO modes around $\Gamma$. As usual, the ZO mode is significantly softened compared to the LO mode, due to the higher freedom for atom motion perpendicular to the plane with respect to the in-plane motion. The intensity of the ZO phonon at small momenta has very low weight, as shown in the intensity plot reported in Fig. \[EELSmap\] (see also [@SM]). To enhance the signal-to-noise ratio from the ZO phonon in the nearness of $\Gamma$, each spectrum was acquired for several days. While this allowed us to resolve the ZO loss peak, the error bars for its frequency are still in general larger than for other branches. A careful fitting procedure [@SM] was used to extract the experimental frequencies for the LO and ZO phonons for small momenta, which are reported in Fig. \[fits\]. The most striking feature in these plots is that they both display a clear cusp *at the same momentum* $q \sim 0.13 \rm{\AA}^{-1}$. This strongly suggests that both cusps are Kohn anomalies at $q=2k_F$, which originate from the interaction of phonons with electrons. The Fermi wave-vector $k_F = E_F/v_F$ can be estimated from ARPES measurements of graphene/Pt(111) [@SSS09]. The reported values of the Fermi energy and Fermi velocity are $E_F \approx 0.30 \pm 0.15$ eV and $v_F \approx 6$ eV$\rm{\AA}$. Thus, $2k_F \approx 0.10 \pm 0.05 \rm{\AA}^{-1}$, in good agreement with the position of the cusp found in this work. The position of the G peak in our Raman measurements [@SM] can be used to estimate [@F07; @DPC08] $E_F \sim 0.42$ eV, also in good agreement. While a Kohn anomaly for the LO mode is known to occur, the presence of the cusp for the ZO mode is unexpected and represents the first evidence of the coupling of the ZO mode to electrons. In hindsight, by carefully inspecting the phonon dispersion recorded for graphene on Pt(111) in a previous experiment by some of the authors [@PMC12], a dip for both LO and ZO at finite momentum is in fact seen but was not noticed or discussed. In Ref. [@PMF11], the LO anomaly was incorrectly presented by shifting it to the $\Gamma$ point due to a misunderstanding, but in fact occurs at finite momentum as well. To substantiate the claim that these cusps are in fact Kohn anomalies, we now compute the self-energies of both LO and ZO phonons following the conventions of Refs. [@BA08; @BPF09; @JF12a] and compare the predictions to the experimental data.
*Theory* - The dispersion of the phonons around $\Gamma$ is modified because of their coupling to electrons, which at low energies can be modeled with a Dirac Hamiltonian $$\label{H_el}
H = \int \frac{d^2k}{(2\pi)^2} \psi^{\dagger} \left( v_F \vec \sigma \cdot \vec k -\mu \right)\psi,$$ with $\vec \sigma = (\sigma_x, \sigma_y)$ are Pauli matrices, $v_F$ is the Fermi velocity and $\mu$ is the chemical potential. This effective model is applicable up to energies $\Lambda_E \approx$ 1.5 eV or momenta $\Lambda_q = \Lambda_E / v_F =$ 0.25 $\AA ^{-1}$, beyond which the dispersion is no longer linear. Therefore, our predictions are only valid for phonon momenta within this range as well. The phonon Hamiltonian is $$\begin{aligned}
H = \sum_i \int \frac{d^2q}{(2\pi)^2} \omega_{i}(q) b^{\dagger}_{i,q} b_{i,q},\label{phononH}\end{aligned}$$ with creation and destruction operators defined by the effective displacements associated with each phonon mode $$\begin{aligned}
u_i = \sqrt{\frac{A_c}{4 \omega_i M}} \int \frac{d^2q}{(2\pi)^2} (b_{i,q}e^{i \vec q \vec r} +
b^{\dagger}_{i,q}e^{-i \vec q \vec r}),\label{phononDisp}\end{aligned}$$ where $i=LO,TO,ZO$ and $A_c = 3\sqrt{3}a^2/2$ is the unit cell area, with $a=1.42 \rm{\AA}$ the nearest neighbor distance. The dispersions of the phonons $\omega_i(q)$ are analytic in the absence of EPC. For momenta close to the $\Gamma$ point they can be expanded as $\omega_i(q) = \omega^0_i + a_i q^2 +b_i q^4$, where $\omega_i^0$, $a_i$ and $b_i$ are parameters to be fitted from experimental data. The coupling between electrons and phonons is described by $$H_{e-ph} = F_i \int d^2 r u_i \psi^{\dagger} \mathcal{M}_i \psi,\label{epv}$$ with $F_i$ the electron-phonon coupling. The matrix $\mathcal{M}_i$ for the different phonons is equal to $\mathcal{M}_{LO}= \hat q \times \vec \sigma $, $\mathcal{M}_{TO}=\hat q \cdot \vec \sigma$, $\mathcal{M}_{ZO}=\sigma_3$. Following Ref. [@BPF09], it is customary to introduce a dimensionless EPC as $$\lambda_i =\frac{F_i^2 A_c}{2 M \omega_i v_F^2}\label{dimensionless}$$ The electron-phonon coupling induces a phonon self-energy $\Pi_i(q)$ that corrects the dispersion according to $$\omega_{R,i} = \omega^0_i + a_i q^2 +b_i q^4 +\frac{\lambda_i}{2}\Pi_i(q/k_F).\label{dispersion}$$ In the static approximation, the self-energies for the different phonon branches in terms of the dimensionless variable $x=q/k_F$ are given by [@SM] $$\begin{aligned}
\Pi_{LO}(x) &=
\frac{g_{s,v} \mu}{4\pi} \left(\sqrt{1-\tfrac{4}{x^2}} +\tfrac{x}{2} \arccos(2/x)\right)\theta(2-x),
\\
\Pi_{TO}(x) &= 0,\\
\Pi_{ZO}(x)& =
\frac{g_{s,v} \mu}{4\pi} \left(2 +x \arccos(2/x)\theta(2-x)\right).\end{aligned}$$ These expressions can be used to fit the cusps in the experimental curves to obtain an estimate for the different electron-phonon couplings. To do so, we have fitted the dispersion parameters for both curves and determined the optimal Fermi level as $E_F = v_F q_c/2 \approx 0.401$ eV, which corresponds to a cusp momentum of $q_c=0.133$ $\AA^{-1}$. The fits are shown in Fig. \[fits\] with the experimental dispersion, and the result for the LO EPC is $\lambda_{LO} = 0.029$, in excellent agreement with the estimates obtained from Raman, $\lambda_{LO} = 0.027-0.034$ [@BPF09]. The fit for the ZO gives $\lambda_{ZO}=0.087$, an even greater value.
*Discussion* - Since this is the first work to observe the effects of electron-phonon coupling to the ZO mode in graphene, there are no measurements of $\lambda_{ZO}$ available to compare with. Theoretically, this coupling has only been estimated for a SiO$_2$ substrate in Ref. [@FL07], giving a maximum value of $F_{ZO} = 7$ $\rm{eV \AA}$ or $\lambda_{ZO} =0.011$ [^1]. Our significantly higher value reflects the fact that the substrate is metallic and that, while direct hybridization with graphene is small, the surface electric field induces a stronger breaking of reflection symmetry which is responsible for the EPC.
Our finding that $\lambda_{ZO}$ has a significant magnitude implies that its effects should also be observable in other experiments. For example, the EPC is visible in the electron dispersion, in the form of a kink at the phonon frequency that is observable by ARPES [@ZSF08]. The existence of $\lambda_{ZO}$ would be responsible for an extra kink at $\omega_{ZO}$ that could be resolved in future experiments. One could also expect a finite $\lambda_{ZO}$ to induce new Raman peaks. A peak at $\omega_{ZO}$ from a first order Raman process (an analog of the $G$ peak) for the ZO is not allowed because the ZO transforms as $B_2$, which is not contained in $E_1 \times E_1$. However, a second order process (analogous to the $2D'$ peak [@BPF09]) is in principle allowed by symmetry. This process would give a peak at roughly $2\omega_{ZO}$, which unfortunately may be difficult to detect because of its overlap with the $G$ peak. Therefore, either Raman or ARPES could provide an independent confirmation of electron-phonon coupling for the ZO mode.
The finding that the substrate can induce an EPC to the ZO may also have important implications for transport in the high-field regime. The phenomenon of current saturation at high bias originates from inelastic scattering with optical phonons, both intrinsic [@BLM09] or from the substrate [@MHY08]. Since the energy of the ZO ($\sim 100$ meV) is significantly lower than the known phonons with strong EPC, the effects of the ZO may already be present in current experiments [@MHY08; @BLM09; @PA10; @DZJ10], and its presence should be accounted for in theory [@BM09; @TS09; @FKA11].
A related system where the standard (LO and TO) Kohn anomalies have been observed is graphene on Ir(111) [@EMW13]. According to ARPES [@PKP09], the Fermi level in this system is $E_F\approx 0.1$ eV which corresponds to $2k_F\approx0.03$ $\rm{\AA}^{-1}$. Kohn anomalies at such small momentum are probably too difficult to resolve, which is consistent with the observed broadened dip around $\Gamma$ instead of a cusp. Similarly, if the ZO EPC in this system is significant, such a feature should be observed for the ZO around $\Gamma$. The data in Ref. [@EMW13] are inconclusive on this matter. Finite momentum Kohn anomalies will be more likely to be found in graphene on metals more similar to Pt(111), with weak hybridization but large charge transfer [@KGR09]. In metallic substrates with stronger hybridization such as as Ni(111), Kohn anomalies are not present for the LO and TO phonons [@SFA99; @FRS00; @FSR99; @SFR98; @ASI90] because the coupling of the graphene $\pi$ bands with the Ni $d$ orbitals completely rearranges the electron bands. A Kohn anomaly for the ZO is therefore not expected in this type of system.
In summary, in this work we have shown that the presence of a substrate induces a significant electron-phonon coupling to ZO phonons in graphene, which is responsible for a strong Kohn anomaly at $q=2k_F$. This finding paves the way to explore the many implications of the coupling of flexural phonons to Dirac electrons in supported graphene samples and graphene-metal contacts.
*Acknowledgments* We thank Davide Campi and Sinisa Coh for helpful discussions. A.P. and G.C. thank Fabio Vito for technical support. F. de J. acknowledges support from the “Programa Nacional de Movilidad de Recursos Humanos” (Spanish MECD). This work was supported in part by the US-Israel Binational Science Foundation.
[**Supplementary Material: Emergence of a ZO Kohn anomaly in quasi-freestanding epitaxial graphene**]{}
Review of experimental techniques
=================================
In this section we compare the different experimental techniques to measure phonons and explain why the ZO Kohn anomaly can only be observed with HREELS. There are four basic techniques to probe phonons: Neutron Inelastic Scattering, Raman scattering of X-Rays, Inelastic He Atom Scattering and Electron Energy Loss Spectroscopy (EELS).
The first two techniques are better suited to the study of bulk 3D phonons. Neutrons are particularly well suited for bulk phonon dispersion studies because of their very small cross section with matter which allows them to penetrate deeply into the crystal. X-ray photons in grazing incidence, available at synchrotron radiation sources, are also a standard probe for bulk phonons and have in fact been used to study phonon dispersion in bulk graphite [@MMD07]. But none of this methods is suitable for studying phonon modes of a single layer of graphite because of the lack of surface sensitivity.
Inelastic helium atom scattering is a more appropriate surface probe. However, in standard conditions the impinging energy is 65 meV. In order to observe modes at 100 meV it would be necessary to keep the nozzle at very high temperature, which implies a strong decrease of the signal. Inelastic helium atom scattering is more appropriate to study phonons up to 20-30 meV. The ZO phonon at around 100 meV has too high energy to be probed by inelastic helium atom scattering.
HREELS measurements, on the other hand, have no limit in the energy loss and can easily access the high frequency range. They are surface sensitive measurements and have excellent resolution in both momentum and energy. This is thus the most suitable technique to study the full surface phonon spectrum, and in particular the only one capable of detecting the presence of the ZO Kohn anomaly.
![a) Selected HREELS spectra for different values of the parallel momentum transfer $q_\parallel$, evaluated by inserting the kinematic conditions and the loss energies in Eq. 1 of the main text. The intensity of the various phonons has been normalized to that of the elastic peak. Note the poor intensity of the ZO mode near $\Gamma$. b) EELS Intensity plot for phonon dispersion of graphene/Pt(111), reproduced from Fig. 1 of the main text. Note again the very low intensity of the ZO mode close to $\Gamma$. The fits in Fig. 3 of the main text are displayed as yellow lines.[]{data-label="sup1"}](fig1sup.pdf){width="16cm"}
Fit procedures
==============
In this section we present more details of the process employed to obtain the phonon dispersions displayed in the main text, and in particular in Fig. 3 of the main text where the Kohn anomalies are demonstrated. First, the intensity plot in Fig. 1 of the main text was produced from many different spectra acquired at fixed incidence angle ($\theta_i = 80 \degree$, all angles with respect to the sample normal) and at scattering angles varying from $\theta_s = 80 \degree$ up to $15 \degree$. A representative subset of this series of spectra is shown in Fig. \[sup1\] as a waterfall plot, along with the intensity plot which is reproduced here for ease of comparison.
It is clear that finding the location of the Kohn anomalies in an intensity plot by simple inspection is not really possible due to the low intensity near $\Gamma$. To help the reader locate the Kohn anomalies in the dispersion of the LO and ZO in the intensity plot, we have also included the fits of Fig. 3 of the main text as yellow lines in Fig. \[sup1\](b). To quantify the low intensity, in Fig. \[intensity\] we also present the integrated weight of the ZO and LO peaks as a function of the off-scattering angle with incidence angle kept at $80 \degree$. The intensity of the ZO mode increases with the off-specular angle, and thus with the momentum $q_\parallel$, with a continuous behavior. Instead, the LO mode show maximum intensity for $15 \degree$-$40 \degree$ off-specular angles. Near $\Gamma$, i.e., near the specular conditions $\theta_i=\theta_s = 80 \degree$, the intensity of the LO and especially of the ZO mode is extremely poor. The spectra near $\Gamma$ actually required several days of measurement in order to reach a sufficient signal-to-noise ratio.
![Intensity of the ZO (filled squares) and the LO (empty circles) phonon modes as a function of the off-specular scattering angle, with the incidence angle fixed at $80 \degree$ with respect to the sample normal. The intensity of phonon modes has been normalized to that of the elastic peak. []{data-label="intensity"}](fig2sup.pdf){width="9cm"}
To quantify the dispersion of the LO and ZO around $\Gamma$ accurately, we extracted the energy for each peak in each spectrum with a given $q_\parallel$ with the following fitting procedure, exemplified in Fig. \[fits\]. First, an exponential background was subtracted from the raw HREELS spectrum. Then, the obtained spectrum was fitted with Voigt line-shapes. The Voigt line-shape is a convolution of Lorentzian and Gaussian line-shapes. The Lorentzian part takes into account the intrinsic line-shape of phonon modes, while the Gaussian line-shape is necessary to account for the experimental broadening due to the energy resolution of spectrometer. The extracted set of pairs $(E,q_\parallel)$ was then fitted with the model dispersion relations discussed in the text. Both the data and the fits are presented in Fig. 3 of the main text.
![a) Selected HREELS spectrum. The fit to an exponential background is reported with a dashed line. b) Fit procedures for TO+LO (left panel) and ZO modes (right panel) with the exponential background in a) substracted.[]{data-label="fits"}](fig3sup.pdf){width="14cm"}
Sample characterization
=======================
Scanning electron microscopy investigation
------------------------------------------
In Fig. \[SEMRaman\](a) we report a scanning electron microscopy (SEM) of the graphene/Pt(111) sample, also reported elsewhere [@CCC13]. The surface morphology appears to be homogeneous across all the sample (1x1 cm$^2$). The SEM image shows a full coverage of the substrate surface by the graphene which forms a network of wrinkles (darker horizontal and vertical lines in Fig. \[SEMRaman\](a)). The wrinkles network develops on a micrometric length scale. Its morphology matches that obtained by low-energy electron microscopy measurements for graphene grown on Pt(111) by carbon segregation from the Pt(111) substrate and other metallic substrates [@SSS09].
Raman investigation
-------------------
The sample has been also analyzed ex-situ by Raman spectroscopy. Such experiments indicate that the surface of the Pt(111) substrate is well coated with monolayer graphene, which is more ordered and homogeneous than for similar preparations previously reported in literature. A compressive strain is found for the graphene as the effect of the growth process. For more details on Raman investigation, see Ref. [@CCC13]. In Fig. \[SEMRaman\](b) we report a comparison between the Raman signal from the G peak in graphite and graphene/Pt(111). Graphite shows a single peak (data taken from Ref. [@F07]), which can be fitted by a single Voigt line-shape. By contrast, graphene/Pt(111) shows a clearly asymmetric G peak, located at $\omega = 1603$ cm$^{-1}$. The position of the G peak can be used to estimate the Fermi level. A peak at $1603$ cm$^{-1}$ corresponds to a hole concentration [@DPC08] of $n = 1.5 \cdot 10^{13}$ cm$^{-2}$, which corresponds to $\mu = v_F (n\pi)^{1/2} \sim 0.42$ eV, in very good agreement with the estimate obtained from the position of the cusp, and within the error of the ARPES estimate $\mu = 0.3 \pm 0.15$ eV.
It is interesting to note that the asymmetry of the Raman G peak could be originated in the emergence of a second order process with a peak around $2\omega_{ZO}$, activated by the linear electron-phonon coupling. However, since this type of second order peak can probe a range of energies because momentum is not restricted to $\vec k=0$, further theoretical modeling is required to determine unambiguously that this peak indeed originates from the ZO.
![a) SEM image of a region of the graphene layer on a Pt(111) substrate. Taken from Ref. [@CCC13]. b) Raman spectrum of graphite (data taken from Ref. [@F07]) and MLG/Pt(111) (data taken from Ref. [@CCC13])[]{data-label="SEMRaman"}](fig4sub.pdf){width="17cm"}
Theory: Phonon self-energy
==========================
All calculations are performed at $T=0$. Room temperature (300 K) corresponds to 25 meV, which is much smaller than $E_F$ or the optical phonon frequencies. With the definitions given in Eqs. \[phononH\]-\[dimensionless\] in the main text, the phonon propagator is written as $$\label{phonon}
G_{i}(\omega,q) = \frac{2\omega_i(q)}{\omega^2 - \omega_i(q)^2 -2\omega_i(q)\Sigma_i(\omega,q)},$$ where $\Sigma_i(\omega,q) = \frac{\lambda_i}{2} \Pi_i(\omega,q)$ is the self-energy, and $i=LO,TO,ZO$ as in the text. The dispersion relation for the phonon can be obtained by solving for the pole in Eq. (\[phonon\]) for small $\lambda_i$, so that the renormalized phonon dispersion relation is given by $$\omega_{R,i}(q) \approx \omega_i(q) + \frac{\lambda_i}{2}\Pi_i(\omega_i,q). \label{omega1}$$ which yields Eq. \[dispersion\] in the text. The response function $\Pi_i$ for a particular phonon is given by $$\Pi_i(E,\vec q) = g_{s,v}i\int \frac{d \omega d^2k}{(2\pi)^3} \;{\rm tr} \; \mathcal{M}_i G(\omega,\vec k)\mathcal{M}_i G(\omega+E,\vec k+\vec q),\label{resp}$$ where $g_{s,v}=4$ accounts for spin and valley degeneracy, $\mathcal{M}_i$ is the electron-phonon coupling matrix as defined in the text, and $$G(\omega,\vec k) = \frac{\omega+\mu +\vec \sigma \cdot \vec k}{(\omega(1+i \epsilon)+\mu)^2 -k^2},$$ is the electron propagator, where we set $v_F=1$. Evaluating the trace in Eq. \[resp\] we obtain $$\Pi_i(E,\vec q) = g_{s,v}i\int \frac{d \omega d^2k}{(2\pi)^3} \frac{N_i}{[(\omega(1+i \epsilon)+\mu)^2 -k^2][((\omega+E)(1+i \epsilon)+\mu)^2 -|\vec k+\vec q|^2]},$$ where the numerators $N_i$ for the different phonons are given by $$\begin{aligned}
N_{LO} =& 2[\omega(\omega+E) - \vec k \cdot (\vec k +\vec q) + 2(\vec k \cdot \vec q)^2/q^2 +2\vec k \cdot \vec q],\\
N_{TO} =& 2[\omega(\omega+E) - \vec k \cdot (\vec k +\vec q) + 2(\vec k \times \vec q)^2/q^2 ],\\
N_{ZO} =& 2[\omega(\omega+E) - \vec k \cdot (\vec k +\vec q)],\end{aligned}$$ The frequency integral is performed with the residue method, giving $$\Pi_i(E,\vec q) = g_{s,v}\int \frac{d^2k}{(2\pi)^2} \left[\frac{ \theta(k-\mu)N_i |_{\omega = k-\mu}}{2|\vec k| [(k+E(1+i \epsilon))^2 -|\vec k+\vec q|^2]}+\frac{ \theta(|\vec k+\vec q|-\mu)N_i |_{\omega = -E + |\vec k+\vec q|-\mu}}{2|\vec k+\vec q| [(|\vec k+\vec q|+E(1+i \epsilon))^2 -|\vec k|^2]}\right],$$ Since these integrals are ultraviolet divergent as power counting reveals, it is customary to split them into the contribution coming from the undoped Dirac cone plus a finite correction induced by finite chemical potential. This is done by replacing $\theta(k-\mu) = 1 - \theta(\mu-k)$ in the integrals, giving $$\Pi_i(E,\vec q) = \Pi_i^{\mu=0}(E,\vec q) -\Delta \Pi_i(E,\vec q),\label{splitpi}$$ with $$\Delta \Pi_i(E,\vec q) = g_{s,v}\int \frac{d^2k}{(2\pi)^2} \left[\frac{ \theta(\mu-k)N_i |_{\omega = k-\mu}}{2|\vec k| [(k+E(1+i \epsilon))^2 -|\vec k+\vec q|^2]}+\frac{ \theta(\mu-|\vec k+\vec q|)N_i |_{\omega = -E + |\vec k+\vec q|-\mu}}{2|\vec k+\vec q| [(|\vec k+\vec q|+E(1+i \epsilon))^2 -|\vec k|^2]}\right].$$ $\Pi_i^{\mu=0}(E,\vec q)$ can then be computed with dimensional regularization methods giving [@PLM04; @BPF09] $$\begin{aligned}
\Pi_{LO}^{\mu=0} =& \frac{g_{s,v}}{16} \sqrt{-E^2+q^2}, \\
\Pi_{TO}^{\mu=0} =& -\frac{g_{s,v}}{16} \frac{E^2}{\sqrt{-E^2+q^2}},\\
\Pi_{ZO}^{\mu=0} =& \frac{g_{s,v}}{8}\sqrt{-E^2+q^2}.\end{aligned}$$ The correction terms become complicated expressions but can be computed analytically in the static approximation ($E=0$) giving [@THD08] $$\begin{aligned}
\Delta\Pi_{LO} =& \frac{g_{s,v}}{4\pi} \left[\frac{q\pi}{4}-\left(\frac{q}{2} \arccos(\tfrac{2\mu}{q})+\frac{\mu}{q}\sqrt{q^2-4\mu^2}\right)\theta(q-2\mu)\right], \\
\Delta\Pi_{TO} =& 0, \\
\Delta\Pi_{ZO} =& \frac{g_{s,v}}{4\pi} \left[\frac{q\pi}{2}-2\mu-q \arccos(\tfrac{2\mu}{q})\theta(q-2\mu)\right] .\end{aligned}$$ Combining the two contributions, Eq. \[splitpi\], we finally get $$\begin{aligned}
\Pi_{LO}(E=0) =& \frac{g_{s,v}}{4\pi} \left[\frac{q}{2} \arccos(\tfrac{2\mu}{q})+\frac{\mu}{q}\sqrt{q^2-4\mu^2}\right]\theta(q-2\mu), \label{fin1}\\
\Pi_{TO}(E=0)=& 0, \label{fin2} \\
\Pi_{ZO}(E=0) =& \frac{g_{s,v}}{4\pi} \left[2\mu+q \arccos(\tfrac{2\mu}{q})\theta(q-2\mu)\right]. \label{fin3}\end{aligned}$$ The results in the main text are obtained by recovering $v_F$ and setting $x=q/(v_F\mu))$. Note that the functional dependence of the $\mu=0$ part for the ZO was anticipated in Ref. [@GP09], and by symmetry arguments [@JF12], it is the same as that of the TO phonon at the $K$ point [@B08].
These self-energies can be then fed into Eq. \[omega1\] to determine the phonon energies. As an example, in Fig. \[muDep\] we have plotted $\omega_{LO}$ and $\omega_{ZO}$ for different values of $\mu$, with the coeficients of the analytic part as obtained from the fits to the experiment. This plots emphasize the dependence of the cusp position on the chemical potential, or equivalently, on $k_F$.
![Left: $\omega_{LO}(q)$ for different values of $\mu$. Right: $\omega_{ZO}(q)$ for different values of $\mu$. In both cases the coeficients of the analytic part of the dispersion are those obtained from the experimental fits in the main text.[]{data-label="muDep"}](fig5asub.pdf "fig:"){width="8cm"} ![Left: $\omega_{LO}(q)$ for different values of $\mu$. Right: $\omega_{ZO}(q)$ for different values of $\mu$. In both cases the coeficients of the analytic part of the dispersion are those obtained from the experimental fits in the main text.[]{data-label="muDep"}](fig5bsub.pdf "fig:"){width="8cm"}
List of acronyms used
=====================
- [EPC - Electron-phonon coupling]{}
- [ZA - Out-of-plane acoustical]{}
- [TA - Transverse acoustical]{}
- [LA - Longitudinal acoustical]{}
- [ZO - Out-of-plane optical]{}
- [TO - Transverse optical]{}
- [LO - Longitudinal optical]{}
- [HREELS - High-resolution electron energy loss spectroscopy]{}
- [ARPES - Angle-resolved photoemission spectroscopy]{}
[^1]: In the notation of Ref. [@FL07], when the sublattice displacements are $h_i=(\eta,-\eta)$, the mass at the Dirac point is $M = D \eta$. In our case $M = F_{ZO} u_{ZO}$, and the displacement is $h_i=u_{ZO}(1,-1)/\sqrt{2}$ because we use normalized phonon eigenstates. This implies $F_{ZO} = D/\sqrt{2}$. With their estimate of $Da = 1-14$ eV and $a=1.42$ $\rm{\AA}$ we obtain $F_{ZO} = 0.5-7$ $\rm{eV \AA}$
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'I give some personal remarks on some current issues in the nucleon spin structure study. At an elementary level I propose a new angular momentum separation for the massless Dirac field in a free theory which mimics the usual free photon angular momentum separation pattern in Coulomb gauge. In connection with this construction I introduce a somewhat idiosyncratic formalism in a free massless Dirac theory which I call “dressed axial $U(1)_A$ symmetry”. I show that this new “fermion spin operator”, which is more correctly called “helicity vector operator”, can be incorporated into this new symmetry pattern in a natural way. This set of “dressed axial vector current” and its corresponding charges show an interesting internal structure and may be useful in a broader physical context. I then discuss the case of the QED model with a massless Dirac fermion. In the case of covariant quantization, I give a new and correct proof that the additional term in the total angular momentum operator stemming from the gauge-fixing term does not contribute at the level of physical matrix elements, which is discussed in the relevant literature. At the level of asymptotic fields I construct the helicity vector operator of the QED theory which actually coincides with the usual projection of the total angular momentum operator when acting on a beam of collinearly moving free particles. I then consider the QCD model with only two massless light quarks. In this context I discuss the approximate concept of “asymptotic quark and gluon fields” which is relevant to the usual parton model picture and propose to use the asymptotic “quark helicity vector” operator to describe the quark helicity contribution to an IMF proton. Finally, I give some remarks on the concept of IMF itself, which show that this very concept should be understood with some reservation from a rigorous mathematical consideration.'
address: 'School of Physics, Nanjing University, Nanjing 210093, the People’s Republic of China'
author:
- Israel Weimin Sun
title: Some remarks on the current issues in nucleon spin structure study
---
``gauge field,angular momentum,nucleon spin structure
Introduction: an angular momentum separation for a free massless fermion theory
===============================================================================
The study of momentum and angular momentum separation problem in a gauge field system has gained much progress in the last ten years in the context of nucleon spin structure research due to the work of Chen ${\it et al.}$ [@Chen] in 2008. Many new developments emerge later on, which are nicely summarized in two recent review articles [@LeaderLorce; @Wakamatsu]. However, there still remain some interesting issues which deserve further investigations. In this letter I shall present some personal remarks on the various problems in these topics. First, let me begin with the simplest case, a free massless Dirac field system.
For a free massless Dirac field, one has a standard split of the total angular momentum $$\begin{aligned}
\nonumber {\bf J}_{free}&=& \int d^3x \big(\psi^\dagger {\bf x}\times \frac{{\bf \nabla}}{i}\psi+\psi^\dagger \frac{{\bf \Sigma}}{2}\psi\big) \\
&=& {\bf L}_{free}+{\bf S}_{free},\end{aligned}$$ where both ${\bf L}_{free}$ and ${\bf S}_{free}$ satisfy the standard $SU(2)$ algebra and are usually regarded as the “standard” definition of orbital and spin angular momentum operator of a Dirac field. However, for a free photon, one also has a standard separation of the total angular momentum operator in the Coulomb gauge $$\begin{aligned}
\nonumber {\bf J}_{em}&=&\int d^3x \big(E^i_\perp {\bf x}\times{\bf \nabla} A^i_\perp+{\bf E}_\perp \times {\bf A}_\perp \big)\\
&=& {\bf L}_{\gamma free}+{\bf S}_{\gamma free},\end{aligned}$$ where the “spin operator” ${\bf S}_{\gamma free}$ satisfies an unusual commutator ${\bf S}_{\gamma free}\times{\bf S}_{\gamma free}=0$. Physically, this phenomenon stems from the massless nature of the photon. Then, a natural question arises: does the same thing also hold for a massless Dirac particle case?
In fact, one can write down a new angular momentum separation $$\begin{aligned}
\nonumber {\bf J}_{free}&=& \int d^3x \big(\psi^\dagger {\bf x}\times \frac{{\bf \nabla}}{i}\psi+\cdots+\psi^\dagger \frac{{\bf \Sigma}}{2}\cdot \frac{-i{\bf \nabla}}{\sqrt{-{\bf \nabla}^2}}\frac{-i{\bf \nabla}}{\sqrt{-{\bf \nabla}^2}}\psi\big) \\
&=& {\bf L}'_{free}+{\bf S}'_{free}.\end{aligned}$$ One can readily check that both ${\bf L}'_{free}$ and ${\bf S}'_{free}$ are conserved (whereas ${\bf L}_{free}$ and ${\bf S}_{free}$ are not separately), and the ${\bf S}'_{free}$ also satisfies a commutator ${\bf S}'_{free}\times{\bf S}'_{free}=0$, which is quite similar to the photon spin case.
This ${\bf S}'_{free}$ has an explicit expansion in terms of the free fermion creation and annihilation operators in the helicity basis $${\bf S}'_{free}=\int \frac{d^3 p}{(2\pi)^3 2 E_{{\bf p}}}\sum_{s=\pm\frac{1}{2}} \frac{1}{2}\frac{{\bf p}}{|{\bf p}|}\epsilon(s)(b^\dagger (p,s)b(p,s)- d^\dagger (p,s)d(p,s)).$$ Because of this structure ${\bf S}'_{free}$ actually measures the helicity of a massless fermion times the associated unit vector ${\bf p}/|{\bf p}|$, hence I call it the “helicity vector”. Due to the same reason, the free “photon spin operator” ${\bf S}_{\gamma free}$ should also be called the “helicity vector” of the free electromagnetic field.
To unravel the internal structure of this formalism, let us also look at the helicity operator itself which reads $$\begin{aligned}
\nonumber h &=& \int d^3x \psi^\dagger \frac{{\bf \Sigma}}{2}\cdot \frac{-i{\bf \nabla}}{\sqrt{-{\bf \nabla}^2}}\psi
= \int \frac{d^3 p}{(2\pi)^3 2 E_{{\bf p}}}\sum_{s=\pm\frac{1}{2}} \frac{1}{2} \epsilon(s)(b^\dagger (p,s)b(p,s)- d^\dagger (p,s)d(p,s)).\end{aligned}$$ Now, let me start from the classical massless Dirac equation $i\gamma^\mu \partial_\mu \psi=0$ to obtain $${\bf \Sigma}\cdot (-i {\bf \nabla})\psi=\gamma^5 i\frac{\partial}{\partial t}\psi,$$ which implies $${\bf \Sigma}\cdot \frac{{\bf P}}{|{\bf P}|}\psi_{plane~wave}=\gamma^5 \epsilon(P^0) \psi_{plane~wave}.$$ Then, by the substitution $(P^0,{\bf P})\rightarrow (i\frac{\partial}{\partial t},-i{\bf \nabla})$ for a plane wave state, one can establish a curious identity: $$\frac{{\bf \Sigma}}{2}\cdot \frac{-i{\bf \nabla}}{\sqrt{-{\bf \nabla}^2}}\psi_{clasical/operator}=\frac{1}{2}\gamma^5 \epsilon (i\frac{\partial}{\partial t})\psi_{clasical/operator},$$ which holds for both classically on-shell $\psi_{clasical}(x)$ and quantum Dirac field operator $\psi_{operator}(x)$. With this at hand, one sees immediately $$h=\int d^3x \psi^\dagger \frac{{\bf \Sigma}}{2}\cdot \frac{-i{\bf \nabla}}{\sqrt{-{\bf \nabla}^2}}\psi=\int d^3x \psi^\dagger \frac{1}{2}\gamma^5 \epsilon (i\frac{\partial}{\partial t})\psi.$$
At this point let me introduce a general type of current which I call “dressed axial vector current” $$j^{\mu 5}_f(x)={\bar \psi}\gamma^\mu \gamma^5 f(i\frac{\partial}{\partial x^\rho})\psi,$$ where $f(\cdot)$ is an arbitrary real valued function. Using the on-shell Dirac equation, it is very easy to check that it is conserved $$\partial_\mu j^{\mu 5}_f=\partial_\mu{\bar \psi}\gamma^\mu \gamma^5 f(\cdot)\psi+{\bar \psi}\gamma^\mu \gamma^5 f(\cdot)\partial_\mu\psi=0.$$ Classically, such a dressed axial vector current is actually connected with the so-called “dressed $U(1)_A$ rotation” $$\left \{
\begin{array}{c}
\psi(x)\rightarrow e^{i\theta \gamma^5 f(i\frac{\partial}{\partial x^\rho})}\psi(x) \\
{\bar \psi}(x)\rightarrow {\bar \psi}(x) e^{i\theta \gamma^5 f(\stackrel{\longleftarrow}{-i\frac{\partial}{\partial x^\rho}})}
\end{array}
\right.$$ which can be readily shown to be a symmetry of the free massless Dirac field theory. In fact, under an infinitesimal transformation $$\left \{
\begin{array}{c}
\delta \psi=i\delta \theta \gamma^5 f(i\frac{\partial}{\partial x^\rho})\psi \\
\delta {\bar \psi}=i\delta \theta f(-i\frac{\partial}{\partial x^\rho}){\bar \psi} \gamma^5
\end{array}
\right.$$ one easily verifies $$\begin{aligned}
\nonumber \delta \mathcal{L} &\sim& i \delta \theta\big( f(-i\frac{\partial}{\partial x^\rho}){\bar \psi}\gamma^5\gamma^\mu \partial_\mu \psi+
{\bar \psi}\gamma^\mu \gamma^5 f(i\frac{\partial}{\partial x^\rho})\partial_\mu \psi \big)\\
\nonumber &=& i \delta \theta\big( {\bar \psi}\gamma^5\gamma^\mu f(i\frac{\partial}{\partial x^\rho})\partial_\mu \psi+
{\bar \psi}\gamma^\mu \gamma^5 f(i\frac{\partial}{\partial x^\rho})\partial_\mu \psi \big) +\rm{total~divergence~term}\\
&=& \rm{total~divergence~term},\end{aligned}$$ where the “derivative moving pattern” is apparent, e.g., for a monomial $f(\cdot)=i\frac{\partial}{\partial x^\alpha}i\frac{\partial}{\partial x^\beta}\cdots i\frac{\partial}{\partial x^\kappa}$. The corresponding conserved charge is easily constructed $$Q^5(f)=\int d^3x \psi^\dagger(x) \gamma^5 f(i\frac{\partial}{\partial x^\rho})\psi(x),$$ which depends functionally on the real valued function $f(\cdot)$.
Now, let us come back to the helicity or helicity vector. The helicity operator $h$ is constructed using the function $f(\cdot)=\epsilon(i\frac{\partial}{\partial t})$ which is not a smooth one. One can approximate it with a family of smooth (in fact real analytic) functions. One introduces the standard Gaussian distribution $\delta_\alpha (\tau)=\frac{1}{\sqrt{\pi}\alpha}e^{-\tau^2/\alpha}$ which tends weakly to $\delta(\tau)$ and then defines $\theta_\alpha (u)=\int^{u}_{-\infty}~d\tau \delta_\alpha (\tau)$. With this one obtains $\epsilon_\alpha (u)=\theta_\alpha (u)-\theta_\alpha (-u)$ whose weak limit is $\epsilon(u)$. Then, one can define $$h=\lim_{\alpha\rightarrow 0^{+}}\int d^3x \psi^\dagger \frac{1}{2}\gamma^5 \epsilon_\alpha (i\frac{\partial}{\partial t})\psi.$$ The helicity vector can be constructed in the following way. One introduces a dressed current ${\bar \psi}\gamma^\mu \gamma^5\frac{1}{2}\epsilon (i\frac{\partial}{\partial t})
\frac{-i{\bf \nabla}}{\sqrt{-{\bf \nabla}^2}}\psi$ whose conserved charge is just the helicity vector $${\bf S}'=\int d^3x \psi^\dagger \frac{1}{2}\gamma^5 \epsilon (i\frac{\partial}{\partial t})\frac{-i{\bf \nabla}}{\sqrt{-{\bf \nabla}^2}}\psi=\int d^3x \psi^\dagger \frac{{\bf \Sigma}}{2}\cdot \frac{-i{\bf \nabla}}{\sqrt{-{\bf \nabla}^2}}\frac{-i{\bf \nabla}}{\sqrt{-{\bf \nabla}^2}}\psi.$$ If one likes, he or she could also define the helicity vector as the weak limit $${\bf S}'=\lim_{\alpha\rightarrow 0^{+}}\int d^3x \psi^\dagger \frac{1}{2}\gamma^5 \epsilon_\alpha (i\frac{\partial}{\partial t})\frac{-i{\bf \nabla}}{\sqrt{-{\bf \nabla}^2}}\psi.$$
In this context I would like to dwell on the algebraic properties of such a family of quantum conserved charges in the free field theory. First of all, a generic charge $Q^5(f)$ is an hermitian one on the free fermion Fock space. In fact, a direct hermitian operation yields $$\begin{aligned}
\nonumber (Q^{5}(f))^\dagger &=& \int d^3x f(-i\frac{\partial}{\partial x^\rho})\psi^\dagger(x)\gamma^5 \psi(x) \\
&=& \int d^3x \psi^\dagger(x)\gamma^5f(i\frac{\partial}{\partial x^\rho}) \psi(x),\end{aligned}$$ which coincides exactly with the original $Q^5(f)$. Such a reasoning depends on a formal “integration by parts”. This operation is legitimate for the spatial partial derivative case but it seems to be in jeopardy for a time derivative $\frac{\partial}{\partial t}$. Fortunately, the on-shell-ness of the quantum Dirac field operator rescues everything. To see this, one first notes that the so-called “doubly dressed axial-vector current” $$j^{\mu 5}_{(f_1,f_2)}(x)=f_1(i\frac{\partial}{\partial x^\rho}){\bar \psi}\gamma^\mu \gamma^5 f_2(i\frac{\partial}{\partial x^\rho})\psi$$ is also conserved. As a consequence, the corresponding charge operator $Q^5(f_1,f_2)$ is independent of time $$\frac{d Q^5(f_1,f_2)}{dt}=\int d^3x \frac{\partial}{\partial t}\big( f_1(\cdot)\psi^\dagger(x) \gamma^5 f_2(\cdot)\psi(x)\big)=0,$$ which shows that a formal “integration by parts” w.r.t. $\frac{\partial}{\partial t}$ is actually valid. Then, a direct computation using the fundamental anticommutation relations of a free Dirac field gives $$\label{generator}
[Q^5(f),\psi(x)]=-\gamma^5 f(i\frac{\partial}{\partial x^\rho})\psi(x),$$ which verifies the status of $Q_f^5$ as the generator of the dressed $U(1)_A$ rotation of the Dirac field. Using (\[generator\]) one easily finds $$[Q^5(f),Q^5(g)]=0.$$ Thus, one actually has a commuting family of conserved hermitian charges. Finally, one notices that the correspondence $f(\cdot) \mapsto Q^5(f)$ is an algebraic homomorphism in the sense that $$Q^5(c_1f_1+c_2 f_2)=c_1Q^5(f_1)+c_2Q^5(f_2).$$
The above analysis shows vividly that some “physical observables” such as the helicity or helicity vector can be incorporated into the general framework of dressed $U(1)_A$ symmetries and the corresponding physical outcome. In some sense this could be regarded as a type of mathematical design based on some particular purposes. Nonetheless, the clearness and flexibility of this framework proves its usefulness for handling potentially interesting physical situations.
Quantum Electrodynamics
=======================
The free fermion field case being clear, let us turn to the case of QED, e.g., the QED of just one species of massless Dirac fermion. This simple theoretical toy model is useful since some interesting structural points already show up in such a hypothetical case.
The Lagrangian takes the form $$\label{QED_Lagrangian}
\mathcal{L}=-\frac{1}{4}F^2+{\bar \psi}i\gamma^\mu\partial_\mu \psi-e{\bar \psi}i\gamma^\mu\psi A_\mu,$$ where the bare mass $m_0$ of the Dirac field vanishes. It is a standard result in the perturbative QED that to all orders of the bare electric charge a seed of the vanishing bare fermion mass $m_0=0$ will yield a vanishing pole mass $m_{pole}=0$, which implies that to all orders of perturbation theory the “physical mass” of an asymptotic free electron state also vanishes.
In this theory the total angular momentum operator ${\bf J}$ actually depends on the quantum gauge choice. In the usual Coulomb gauge choice the angular momentum operator is of the form $${\bf J}=\int d^3x \big(\psi^\dagger {\bf x}\times \frac{{\bf \nabla}}{i}\psi+\psi^\dagger \frac{{\bf \Sigma}}{2}\psi+E^i_\perp {\bf x}\times{\bf \nabla} A^i_\perp+{\bf E}_\perp \times {\bf A}_\perp \big),$$ where the contributions of the fermion and the photon are clearly separated. In the standard covariant gauge quantization, the classical Lagrangian is modified as $$\mathcal{L}_{cov}=\mathcal{L}-\frac{1}{2\alpha}(\partial\cdot A)^2$$ so that manifest Lorentz covariance is preserved but the quantization procedure yields a Hilbert-Krein structure [@Strocchi] with an indefinite metric. With this Lagrangian form, the angular momentum operator changes into $${\bf J}=\int d^3x \big(\psi^\dagger {\bf x}\times \frac{{\bf \nabla}}{i}\psi+\psi^\dagger \frac{{\bf \Sigma}}{2}\psi+E^i {\bf x}\times{\bf \nabla} A^i-\frac{1}{\alpha}(\partial\cdot A)({\bf x}\times{\bf \nabla})A^0+{\bf E} \times {\bf A} \big)$$ which includes an additional contribution stemming from the gauge-fixing term. In Ref. [@Leader] it is pointed out that such a term does not contribute at the level of the physical matrix elements. Physically, such an intuitive conclusion is undoubtedly correct, however, the proof provided in Ref. [@Leader] is not without flaws. The main idea of the proof is like this. To evaluate the physical matrix element $\langle \Phi'|\partial\cdot A~ \partial^i A^0|\Phi\rangle$ for two arbitrary physical state vectors $|\Phi'\rangle$ and $|\Phi\rangle$, one can insert “a complete set of physical states” between the two operators to obtain $\sum_n\langle \Phi'|\partial\cdot A|\Phi_n\rangle \langle \Phi_n| \partial^i A^0|\Phi\rangle$, then because $\langle phys'|\partial\cdot A(x)|phys\rangle=0$, one arrives at $\langle \Phi'|\partial\cdot A~ \partial^i A^0|\Phi\rangle=0 $. However, this proof is flawed. This is because the relevant operators act on the whole indefinite metric Hilbert space, hence one should insert a complete set of intermediate states which also includes the contribution of nonphysical states (i.e., those state vectors with a negative norm), so the original proof in Ref. [@Leader] cannot go through without any difficulties.
Here, I shall give a new proof. The idea is rather simple: one can express all the interpolating field operators using the free asymptotic fields and then consider everything on the in/out particle Fock space. For simplicity, I shall choose the $\alpha=1$ theory (the Feynman gauge case) at the unrenormalized level (a subsequent renormalization procedure will “renormalize” such a gauge parameter but this is irrelevant for our essential discussions), and the quantum equation of motion of the interpolating field reads $$\partial^2 A^\mu=j^\mu=e {\bar \psi}\gamma^\mu \psi.$$ Then, a standard formal process will establish $$\label{interpolation}
A^\mu(x)=\sqrt{Z_3}~A^\mu_{in}(x)+\int d^4y ~D_{ret}(x-y)j^\mu(y),$$ where everything has its standard meaning. With this at hand, let us consider the physical matrix elements $\langle \Phi'|\partial\cdot A~ \partial^i A^0|\Phi\rangle$. First, using (\[interpolation\]) together with the current conservation condition $\partial_\mu j^\mu=0$ gives $\partial\cdot A=\sqrt{Z_3}~ \partial\cdot A_{in}$, and one also obtains $$\label{interpolation2}
\partial^i A^0(x)=\sqrt{Z_3}~\partial^i A^0_{in}(x)+\int d^4y ~\partial^i D_{ret}(x-y)j^0(y),$$ then, because the electromagnetic current operator $j^\mu=e {\bar \psi}\gamma^\mu \psi$ is a gauge-invariant one, the action of the “interaction part” in (\[interpolation2\]) on the physical state $|\Phi\rangle$ will produce a new physical state vector, which does not contribute to the matrix element $\langle \Phi'|\partial\cdot A~ \partial^i A^0|\Phi\rangle$. Consequently, what remains is a “free part” $\langle \Phi'|\partial\cdot A~ \partial^i A^0|\Phi\rangle=Z_3 \langle \Phi'|\partial\cdot A_{in}~ \partial^i A^0_{in}|\Phi\rangle$, which could then be analyzed on the free in-state Fock space. According to the usual asymptotic completeness hypothesis, the total Hilbert space of the interacting theory is actually isomorphic to the free incoming photon/fermion Fock space, hence in a purely mathematical sense one can identify the arbitrary physical state vectors $|\Phi\rangle$ and $|\Phi'\rangle$ appearing in the relevant matrix elements as a pair of “free physical states” in the in-state Fock space. Then I will show that the matrix element $\langle \Phi'|\partial\cdot A_{in}~ \partial^i A^0_{in}|\Phi\rangle$ actually vanishes.
The reasoning is as follows. On the in-state Fock space, one has the standard operator expansion $$\label{a}
A^0_{in}(x)=\int \frac{d^3 k}{(2\pi)^3 2\omega_k} \big( a^{(0)}(k)e^{-i k\cdot x}+ a^{(0)\dagger}(k)e^{i k\cdot x} \big),$$ $$\label{b}
i\partial\cdot A_{in}(x)=\int \frac{d^3 k}{(2\pi)^3 2\omega_k}|{\bf k}|\big(L(k)e^{-i k\cdot x} - L^\dagger(k)e^{i k\cdot x}\big),$$ where the $L(k)=a^{(0)}(k)-a^{(3)}(k)$ and the physical in-state is identified by the Gupta-Bleuler condition: $\partial\cdot A_{in}^{(+)}(x)|phys\rangle=0$ (which is actually the same as the GB condition $\partial\cdot A^{(+)}(x)|phys\rangle=0$ in the interacting theory). Then, one has $$\langle \Phi'|\partial\cdot A_{in}~ \partial^i A^0_{in}|\Phi\rangle=\langle \Phi'|\partial\cdot A_{in}^{(+)}~ \partial^i A^0_{in}|\Phi\rangle=\langle \Phi'|\partial\cdot A_{in}^{(+)}~ \partial^i A^{0~(-)}_{in}|\Phi\rangle,$$ where the last step follows from $\partial\cdot A_{in}^{(+)}~ \partial^i A^{0~(+)}_{in}|\Phi\rangle= 0$. Now, one inserts the expansion (\[a\]) and (\[b\]) into the relevant matrix element and finds $$\begin{aligned}
\nonumber && \langle \Phi'|\partial\cdot A_{in}^{(+)}(x)~ \partial^i A^{0~(-)}_{in}(x)|\Phi\rangle \\
\nonumber &=& \int \frac{d^3 k~d^3 q}{(2\pi)^6 2\omega_k 2\omega_q}|{\bf k}|q^i \langle \Phi'|L(k)a^{(0)\dagger}(q)|\Phi\rangle e^{i(q-k)\cdot x} \\
\nonumber &=& \int \frac{d^3 k~d^3 q}{(2\pi)^6 2\omega_k 2\omega_q}|{\bf k}|q^i \langle \Phi'|[L(k),a^{(0)\dagger}(q)]|\Phi\rangle e^{i(q-k)\cdot x} \\
\nonumber &=& -\int \frac{d^3 k}{(2\pi)^3 }\frac{k^i}{2}\langle \Phi'|\Phi\rangle \\
&=& 0,\end{aligned}$$ which verifies $\langle \Phi'|\partial\cdot A~ \partial^i A^0|\Phi\rangle=0$. Therefore, the results discovered in Ref. [@Leader] are in fact correct.
Now, let us come back to the main line of our presentation. If one takes, for instance, the Coulomb gauge, one can similarly define the “helicity vector” for the fermion and the photon, respectively $${\bf S}'_f=\int d^3x \psi^\dagger \frac{{\bf \Sigma}}{2}\cdot \frac{-i{\bf \nabla}}{\sqrt{-{\bf \nabla}^2}}\frac{-i{\bf \nabla}}{\sqrt{-{\bf \nabla}^2}}\psi,$$ $${\bf S}'_\gamma=\int d^3x {\bf E}_\perp \times {\bf A}_\perp.$$ However, in the interacting field case, these two “helicity vectors” are not conserved. But at the level of asymptotic fields one could introduce the corresponding free helicity vector operators $$\begin{aligned}
\nonumber {\bf S}'_{in/out}&=&\int d^3x \big(\psi^\dagger_{in/out} \frac{{\bf \Sigma}}{2}\cdot \frac{-i{\bf \nabla}}{\sqrt{-{\bf \nabla}^2}}\frac{-i{\bf \nabla}}{\sqrt{-{\bf \nabla}^2}}\psi_{in/out}+{\bf E}_{\perp in/out}\times {\bf A}_{\perp in/out}\big) \\
&=& {\bf S}'_{f ~in/out}+{\bf S}'_{\gamma ~in/out},\end{aligned}$$ which describe the “helicity vector” of free asymptotic fermions and photons and are conserved. At the formal LSZ level, one has a relation $${\bf S}'_{in} =S~{\bf S}'_{out} S^{-1},$$ where the $S$ stands for the S-matrix of the underlying theory.
In the case of a beam of collinear incoming/outgoing free particles which I denote as $|\psi_{colinear},in/out\rangle$, it is obvious that the “helicity vector operator” ${\bf S}'_{in/out}$ can be identified as the usual projection of the total angular momentum operator of the full theory $${\bf S}'_{in/out}|\psi_{colinear},in/out\rangle= \frac{{\bf J}\cdot{\bf P}}{|{\bf P}|}\frac{{\bf P}}{|{\bf P}|}|\psi_{colinear},in/out\rangle,$$ which shows that in this specific subspace the ${\bf S}'_{in/out}$ operator has a “gauge-invariant” meaning.
Quantum Chromodynamics
======================
Now let us turn to the case of QCD with two massless $u$ and $d$ quarks. This is a toy model which differs from real QCD, but is appropriate for our purposes. In such a hypothetical world there should exist two bound states, the proton and the neutron, which is stable w.r.t. strong interactions. The real mass of the nucleon should be due to $\chi$SB in the underlying theory. Then, the angular momentum of the system reads (ignoring the gauge choice issues) $${\bf J}_{QCD}=\int d^3x \big(\psi^\dagger {\bf x}\times \frac{{\bf \nabla}}{i}\psi+\psi^\dagger \frac{{\bf \Sigma}}{2}\psi+E^{ai} {\bf x}\times{\bf \nabla} A^{ai}+{\bf E}^{a} \times {\bf A}^{a} \big).$$ The real problem of nucleon spin structure is of course a bound-state problem. One should remember that, from a purely axiomatic QFT point of view, the stable nucleon state is an isolated point in the mass spectrum of QCD theory, and should be regarded as a separate composite particle with its own asymptotic in/out fields. This is rather different from the case of QED where no stable bound state exists (the only such “bound state”, i.e., the positronium, is actually unstable w.r.t. electromagnetic interactions), and one is confronted with a true bound state problem. Nevertheless, if one introduces a stable proton into the QED theory, then the Hamiltonian of the system will have a stable bound state, the hydrogen atom, which is orthogonal to all asymptotic scattering states. In this case one would have a two-sided problem: an internal spin structure one for the proton (please remember in that case one is studying the spin structure of an almost static proton state, not a proton observed in the IMF), and an electromagnetic angular momentum partition problem in the hydrogen atom.
Let me come back to the QCD theory and the bound state nucleon. In the usual parton model picture, an nucleon, for instance, the proton, which is observed in the IMF, could be regarded, rather approximately, as a beam of almost free and collinearly moving partons, the colored quarks and gluons. In connection with this point I would like to give some remarks on the concept of asymptotic fields in the QCD theory. It is apparent that because of the color confinement mechanism, there are no true asymptotic quark/gluon fields which are defined in the whole $\mathbf{R}^3$ region. However, within the parton model picture, one could introduce a (frankly speaking, rather approximate and indefinite) concept of “asymptotic quark/gluon fields”. This concept is based on the intuitive idea that when an actual physical process could be clearly separated into two stages: a near distance (and short time) one in which the quark/gluon particles interact sufficient weakly so that they could be effectively taken to be free ones, and a subsequent one in which confinement effects bind the colored particles into color singlet hadrons, one could “define” some sort of “asymptotic quark/gluon fields” within such a specifically chosen finite spacetime region. This concept of “asymptotic colored field” must be an approximate and spacetime-region-dependent one which is valid in a rather limited sense. On the large-scale regions, which are effectively the whole $\mathbf{R}^3$, one can define the usual asymptotic field operator of the colorless composite hadrons which have a finite size by its very formation.
Within this intuitive picture, a fast-moving proton in the parton model will be expanded as a global colorless combination of in/out quarks/gluons which are moving collinearly in the parent hadron direction. According to this approximate idea of “asymptotic quark/gluon field”, one can introduce the asymptotic quark/gluon helicity vector operator acting on the IMF proton state, for instance, the quark helicity one $${\bf S}'_{q ~in/out}=\int d^3x \big(\psi^\dagger_{in/out} \frac{{\bf \Sigma}}{2}\cdot \frac{-i{\bf \nabla}}{\sqrt{-{\bf \nabla}^2}}\frac{-i{\bf \nabla}}{\sqrt{-{\bf \nabla}^2}}\psi_{in/out}\big),$$ which measures the quark helicity contribution to the proton. It should be noted that in such a situation either ${\bf S}'_{q~in}$ or ${\bf S}'_{q~out}$ can be used. This is because the proton is a stable bound state, therefore according to the usual LSZ formalism there is no distinction between its in-state representation and out-state representation: $|{\rm proton}, in\rangle=|{\rm proton}, out\rangle$. As a consequence, one has $$\langle{\rm proton}| {\bf S}'_{q ~in}|{\rm proton} \rangle=\langle {\rm proton}|S~ {\bf S}'_{q ~out}S^{-1}|{\rm proton}\rangle=\langle {\rm proton}| {\bf S}'_{q ~out}|{\rm proton}\rangle.$$
Finally, some remarks on the IMF itself. From a physical point of view, the IMF should be regarded as some kind of limit of a sequence of Lorentz frames whose moving velocity tends to $c=1$. Mathematically, one can imagine to obtain an “infinite momentum proton” state, e.g., that moving in the third direction, by a boost acting on a static proton state: $$|{\rm proton}, p_z=\infty\rangle=\lim_{\eta\rightarrow \infty}e^{i\eta K_z}|{\rm proton}, {\rm static}\rangle$$ I will argue that such a limit could not be reached, at least mathematically. In fact, in the usual axiomatic field theory, the quantum state vector space of the QCD theory should be a separable Hilbert space, on which one has defined a strongly continuous unitary representation of the ${\rm Poincar\acute{e}}$ group. To see why this is so, let me consider more closely the state vector: $e^{i\eta K_z}|{\rm proton}, {\rm static}\rangle$. First note that the boost operator $e^{i\eta K_z}$ is a unitary one which keeps the length of a state vector unchanged. Admittedly, a static proton state is an “improper state vector” whose length is infinite, however, one can build a normalized state vector by using, for example, a smearing process. Hence, without loss of correctness of the argument, one can assume the initial state vector $|\psi_0\rangle$ to be a normalized one. Then, could the vector $\lim_{\eta\rightarrow \infty}e^{i\eta K_z}|\psi_0\rangle$ really exist or not? Because the boost operator is indeed unitary, the whole family of the vectors $e^{i\eta K_z}|\psi_0\rangle$ all lie on the unit sphere of a separable Hilbert space which I call $\mathcal{H}_{QCD}$. However, the “limit vector” should not exist, because all state vectors in $\mathcal{H}_{QCD}$ should have a finite energy, while a $v=c=1$ boosted state has an infinite energy ! This should be compared with the situation of a pure spatial rotation operation, e.g., $U(\theta)=e^{-i \theta J_z}$ which does not change the energy. At first sight, this seems rather strange. Why there does not exist such a limit vector on this unit sphere? The reason is actually quite simple. We note that one is dealing with an infinite dimensional separable Hilbert space. Therefore, its “unit sphere” is not compact, which differs from the case of a finite-dimensional Euclidean space whose unit sphere is necessarily compact. Because of this, a sequence of vectors $\{ e^{i\eta_n K_z}|\psi_0\rangle\}$ on that unit sphere does not necessarily have an accumulation point, which is just what actual physics teaches us. Hence, our conclusion is that the IMF and the associated parton model picture is necessarily an approximate and incomplete framework which has a limited physical significance.
0.3cm [**Acknowledgments**]{}
The author thanks the financial support from the Natural Science Funds of Jiangsu Province of the People’s Republic of China under Grant No. BK20151376.
References {#references .unnumbered}
==========
X.S. Chen, X.F. L$\rm{\ddot{u}}$, W.M. Sun, F. Wang and T. Goldman, Spin and orbital angular momentum in gauge theories: nucleon spin structure and multipole radiation revisited. Phys. Rev. Lett. [**100**]{}, 232002 (2008). E. Leader and C. $\rm{Lorc\acute{e}}$, The angular momentum controversy: What’s it all about and does it matter? Phys. Rept. [**541**]{},163 (2014). M. Wakamatsu, Is gauge-invariant complete decomposition of the nucleon spin possible? Int. J. Mod. Phys. A [**29**]{},1430012 (2014). F. Strocchi, [*Selected Topics on the General Properties of Quantum Field Theory*]{}. World Scientific, 1993. E. Leader, Controversy concerning the definition of quark and gluon angular momentum. Phys. Rev. D [**83**]{}, 096012 (2011).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the probability for a random line to intersect a given plane curve, defined over a finite field, in a given number of points defined over the same field. In particular, we focus on the limits of these probabilities under successive finite field extensions. Supposing absolute irreducibility for the curve, we show how a variant of Chebotarev density theorem for function fields can be used to prove the existence of these limits, and to compute them under a mildly stronger condition, known as simple tangency. Partial results have already appeared in the literature, and we propose this work as an introduction to the use of Chebotarev theorem in the context of incidence geometry. Finally, Veronese maps allow us to compute similar probabilities of intersection between a given curve and random curves of given degree.'
address:
- 'Johann Radon Institute for Computational and Applied Mathematics (RICAM), Austrian Academy of Sciences, Linz and Research Institute for Symbolic Computation (RISC), Johannes Kepler University, Linz'
- 'Research Institute for Symbolic Computation (RISC), Johannes Kepler University, Linz'
- 'Research Institute for Symbolic Computation (RISC), Johannes Kepler University, Linz'
author:
- 'Mehdi Makhul$^{\ast}$'
- 'Josef Schicho$^{\ast}$'
- 'Matteo Gallet$^{\ast, \circ}$'
title: Probabilities of incidence between lines and a plane curve over finite fields
---
Introduction
============
What is the probability that a random line in the (affine or projective) plane intersects a curve of given degree in a given number of points? More precisely, what happens if we consider a finite field with $q$ elements as base field, and then we ask the same question for a field with $q^2, q^3, \ldots, q^N$ elements, analyzing how these probabilities behave as $N$ goes to infinity? In this work we investigate this problem by means of algebro-geometric techniques. In recent years, the interplay between combinatorial problems and algebraic techniques has become more and more common, and has been revealing to be extremely fruitful. Here we refer in particular to the area called *combinatorial geometry*, which deals with, among others, distance or intersection relations between finite sets of geometric objects such as points, lines, or circles (seee [@Tao2014]). In the last decade, algebraic geometry and algebraic topology helped solving several outstanding problems and conjectures in this area. Amongst the most prominent of such problems, one can mention the *distinct distance problem* (see [@Guth2015]), the *Kakeya problem over finite fields* (see [@Dvir2009] and later improvements in [@Saraf2008] and [@Dvir2013]) and the *Dirac-Motzkin conjecture* (see [@Green2013]). For a nice survey about these topics, see [@Tao2014].
A motivation for the problem we investigate in our paper comes from the famous *Sylvester-Gallai theorem*. Sylvester posed it as a question in [@Sylvester1893], which was raised again by Erdös in [@Erdoes1943] and later solved by Melchior (see [@Melchior1941]) and Gallai (see [@Gallai1944]). Given a set of points in the affine plane, a line is called *ordinary* if it passes through exactly two of them.
Suppose that $P$ is a finite set of points in the real plane, not all on a line. Then $P$ admits an ordinary line.
This theorem is clearly false if we work over finite fields, since in this case we can pick $P$ to be the the whole plane.
In the same circle of ideas, in the early $60$s, Erdös asked the following question: is it possible for a set of points in the real plane to contain many collinear four-tuples, but to contain no five points on a line? Here by “many” we mean a number which is quadratic in terms of the cardinality of the set of points. In [@Solymosi2013], Solymosi pointed out that, if we weaken the quadratic condition, then the problem has a affirmative solution. Given a set $P$ of points in the plane, a line is called $k$-*rich*, if it contains precisely $k$ points of $P$. For example, a $2$-rich line is an ordinary line. Then, Solymosi’s theorem reads as:
For any $k \ge 4$, there is a positive integer $n_0$ such that for $n >
n_0$ there exists $P \subseteq {\mathbb{R}}^2$ such that there are at least $n^{2-\frac{c}{\sqrt{\log n}}}$ $k$-rich lines, but no $k+1$-rich lines. Here, $c=2\log(4k+1)$.
A recent outstanding result of Green and Tao (see [@Green2013]) gives an almost complete description of the structure of sets with few ordinary lines in the real plane. In the same paper, the authors also proved the Dirac-Motzkin conjecture and a less known problem, referred in the literature as the *orchard problem*.
Our work is inspired by these results and conjectures. Here, we consider an algebraic plane curve $C$ of degree $d$ over a finite field ${\mathbb{F}}_q$ with $q$ elements, where $q$ is a prime power, namely the set of points in the projective plane ${\mathbb{P}}^2({\mathbb{F}}_q)$ that are zeros of a homogeneous trivariate polynomial of degree $d$. Given such a curve, we can define the probability for a line in ${\mathbb{P}}^2({\mathbb{F}}_q)$ to intersect it in exactly $k$ points. Notice that here we consider the mere set-theoretic intersection: no multiplicities are taken into account. We can then consider the same kind of probability, keeping the same curve $C$ — namely, the same trivariate polynomial — but changing the base field from ${\mathbb{F}}_q$ to ${\mathbb{F}}_{q^2}$, ${\mathbb{F}}_{q^3}$ and so on. In this way, for every $N
\in {\mathbb{N}}$ we define the numbers $p_k^N(C)$, namely the probability for a line in ${\mathbb{P}}^2({\mathbb{F}}_{q^N})$ to intersect $C$ in exactly $k$ points. If the limit as $N$ goes to infinity of the sequence $\bigl( p_k^N(C) \bigr)_{N \in {\mathbb{N}}}$ exists, we denote this number by $p_k(C)$. The main tool we use to compute these numbers when the curve $C$ is absolutely irreducible and with *simple tangency* is an effective version of the *Chebotarev theorem* for function fields. Here, by *absolutely irreducible* we mean that the curve is irreducible over the algebraic closure of its field of definition. By asking that the curve has *simple tangency* we require that there exists a line whose intersection with $C$ consists of simple intersections except for one, which is a double intersection. These are the main results of our paper:
\[thm:existence\] Let $C$ be an absolutely irreducible plane algebraic curve of degree $d$ over ${\mathbb{F}}_q$, where $q$ is a prime power. Then the numbers $\{ p_k(C) \}$ are well-defined, namely the corresponding limits exist.
\[thm:formula\] Let $C$ be an absolutely irreducible plane algebraic curve of degree $d$ over ${\mathbb{F}}_q$, where $q$ is a prime power. Suppose that $C$ has simple tangency. Then for every $k \in \{0, \dotsc, d\}$ we have $$p_k(C) = \sum_{s=k}^d \frac{(-1)^{k+s}}{s!} \binom{s}{k}.$$ In particular, $p_{d-1}(C) = 0$ and $p_d(C) = 1 / d!$.
A consequence of this theorem is that the question by Erdös we already mentioned — asking whether there exists a set of points in the plane containing many collinear four-tuples, but not containing any five points on a line — has a positive answer for the plane over a finite field (see Corollary \[cor:erdos\_problem\]).
We approached this problem using Galois theory techniques; during a revision of our work, we have been informed[^1] that some of the questions investigated in this paper (or similar ones) have already appeared in the literature, though expressed in a different language and with different purposes (see [@Bary-Soroker2012] and [@Diem2012]). Both the two cited paper use the Chebotarev theorem for function field as a key ingredient. After studying Chebotarev theorem, we realized that we could use it to provide a much shorter proof for our result than the one we initially used, and that our initial approach, although we were not aware of that, did not differ too much from the techniques that lead to Chebotarev theorem. However, we still think that our initial approach could be of interest for researchers in discrete and combinatorial geometry. In fact, although it provides less information than Chebotarev theorem, it can serve as an introduction to this technique because of its self-containedness and of the avoidance of technical aspects that are present in other works. Because of this, in the initial part of this paper we report our initial approach to the problem, and then we explain how to use Chebotarev theorem to obtain Theorem \[thm:formula\]. After that, we show how the same technique provides a formula for the probabilities of intersection between a given plane curve of degree $d$ and a random plane curve of degree $e$ (Proposition \[prop:higher\_formulas\]).
We briefly summarize how the problem we investigate is discussed in the aforementioned literature. In [@Bary-Soroker2012], the focus is a variant of the so-called *Bateman-Horn conjecture* for polynomial rings of finite fields. The original Bateman-Horn conjecture concerns the frequency of prime numbers among the values of a system of polynomials at integer numbers. One of its consequences is *Schinzel conjecture*, which asks whether, given polynomials $f_1, \dotsc, f_r \in {\mathbb{Z}}[x]$, then for infinitely many $n \in {\mathbb{Z}}$ we have that $f_1(n), \dotsc, f_r(n)$ are all prime. Bary-Soroker and Jarden consider the situation in which ${\mathbb{Z}}$ is replaced by ${\mathbb{F}}_q[t]$ for some prime power $q$. More precisely, given polynomials $f_1, \dotsc, f_r \in {\mathbb{F}}_q[t][x]$, they want to compute the number of polynomials $g \in {\mathbb{F}}_q[t]$ such that $f_1
\bigl( t,g(t) \bigr), \dotsc, f_r \bigl( t,g(t) \bigr)$ are irreducible. In particular, they focus on the case when $g$ is linear, namely on the computation of the pairs $(a_1, a_2) \in {\mathbb{F}}_q^2$ such that $f_1(t, a_1t + a_2), \dotsc
f_r(t, a_1t + a_2)$ are irreducible. In our language, this is the number of lines in the plane such that the polynomial obtained by restricting a plane curve on such a line is irreducible. The authors improve a result by Bender and Wittenberg (see [@Bender2005 Theorem 1.1 and Proposition 4.1]) and show that this number goes as $q^2 / d$. To prove this, they make use of an effective version of the Chebotarev density theorem (see the appendix of [@Andrade2015]). The number computed by Bary-Soroker and Jarden is similar to the quantity $p_0$ that we define, though it is not the same, since it can happen that a line does not intersect a curve at any point over ${\mathbb{F}}_q$, but the polynomial given by the restriction of the curve to the line can be reducible. Also the behaviour as $d \to \infty$ of these two quantities is different: the one by Bary-Soroker and Jarden goes to zero, while $p_0$ tends to $1/e$.
In [@Diem2012], the author focuses on the complexity of computation of the so-called *discrete logarithm* in the group of divisors of degree $0$ of a nonsingular curve. Given two elements $a$ and $b$ in a group $G$, the discrete logarithm $\log_b a$ is an integer $k$ such that $b^k = a$. On a smooth curve $C$, one can consider formal integer sums of points of $C$, and define an equivalence relation on them in order to obtain the class group of $C$. One can therefore try to compute discrete logarithms in the class group of a curve, and in particular for those formal integer sums of points whose coefficients add up to zero, namely the ones of degree $0$; this has important applications in cryptography. In [@Diem2012 Theorem 2], Diem proves that computing the discrete logarithm has an expected time of ${\widetilde}{O}(q^{2- \frac{2}{d-2}})$ for those curves defined over ${\mathbb{F}}_q$ that admit a birational plane model $D$ of degree $d$ such that there exists a line in the plane intersecting $D$ in $d$ distinct points over ${\mathbb{F}}_q$. Then the author computes the number of lines in the plane intersecting $D$ in exactly $d$ points over ${\mathbb{F}}_q$ (see [@Diem2012 Theorem 3]), namely the quantity $p_d(D)$ in our language. As in the previous paper, this is done using an effective version of Chebotarev density theorem (see [@KumarMurty1994]).
Recently, a new paper [@Entin2018] appeared dealing with the same problem we investigate in our work, but allowing the given curve to be constituted of several irreducible components.
The rest of the paper is structured as follows. Section \[preliminaries\] introduces some preliminary results, namely the Lang-Weil bound for the number of points of a variety over a finite field (Subsection \[preliminaries:lang\_weil\]) and some known facts about Galois groups of plane curves (Subsection \[preliminaries:galois\]). Sections \[galois\] and \[probabilities\] present our initial approach to the problem, which provides less information than the one obtained via Chebotarev theorem, but uses more elementary tools (namely, some basic facts about étale maps). Section \[chebotarev\] shows how to use Chebotarev density theorem in order to prove Theorems \[thm:existence\] and \[thm:formula\] and Proposition \[prop:higher\_formulas\].
Preliminaries
=============
Lang-Weil bound {#preliminaries:lang_weil}
---------------
One of the main tools we use in our work is the so-called *Lang-Weil bound* for the number of points of a variety over a finite field (see [@Lang1954 Theorem 1]). For a nice exposition of this result, see Terence Tao’s blog[^2]. Let ${\mathbb{F}}$ be a field and consider an affine algebraic variety $V$ defined over ${\mathbb{F}}$. This means that $V$ is the set of common zeros in ${\mathbb{F}}^n$ of finitely many polynomials $P_1, \dotsc, P_r \in {\mathbb{F}}[x_1, \dotsc, x_n]$. To a variety $V$ defined over ${\mathbb{F}}$ we can then associate the ideal $I(V)$ of all polynomials $P \in {\mathbb{F}}[x_1, \dotsc, x_n]$ that vanish at all points of $V$. For any extension of fields ${\mathbb{F}}\subseteq \mathbb{K}$, we denote by $V(\mathbb{K})$ the set of common zeros in $\mathbb{K}^n$ of the polynomials in the ideal $I(V)$, considered now as an ideal in $\mathbb{K}[x_1, \dotsc,
x_n]$. One says that a variety $V \subseteq {\mathbb{F}}^n$ is *irreducible* if $I(V)$ is prime in ${\mathbb{F}}[x_1, \dotsc, x_n]$. For our considerations we will need a stronger notion of irreducibility, which we introduce in the following definition.
We say that an affine variety $V$ over a field ${\mathbb{F}}$ is *absolutely irreducible* if the ideal $I(V)$ is prime in $\overline{{\mathbb{F}}}[x_1, \dotsc, x_n]$, where $\overline{{\mathbb{F}}}$ is an algebraic closure of ${\mathbb{F}}$. This is equivalent to the fact that $V(\overline{{\mathbb{F}}})$ is irreducible.
We say that an affine variety $V \subseteq {\mathbb{F}}^n$ defined by polynomials $P_1, \dotsc, P_r$ has *complexity* $M$ if $n, r \leq M$ and $\deg(P_i) \leq M$ for all $i \in \{1, \dotsc r\}$.
\[thm:lang\_weil\] Let $V$ be an absolutely irreducible variety over a finite field ${\mathbb{F}}$ of complexity at most $M$. Then $$|V(\mathbb{F})| =
\bigl(
1 + O_M(|\mathbb{F}|^{-\frac{1}{2}})
\bigr) \thinspace
|\mathbb{F}|^{\dim(V)} \,.$$
By writing $O_M(|{\mathbb{F}}|^{-\frac{1}{2}})$ we mean that there exists a nonnegative constant $\delta_M$ depending on $M$, but not on $V$, such that $$\bigl( 1-\delta_M |{\mathbb{F}}|^{-\frac{1}{2}}) \bigr) \thinspace |{\mathbb{F}}|^{\dim(V)}
\leq
|V({\mathbb{F}})|
\leq
\bigl( 1+\delta_M |{\mathbb{F}}|^{-\frac{1}{2}}) \bigr) \thinspace |{\mathbb{F}}|^{\dim(V)} \,.$$ Using an inclusion-exclusion argument, one obtains by induction on the dimension:
\[cor:lang\_weil\] Let $V$ be a variety over a finite field ${\mathbb{F}}$ of complexity at most $M$. Then $$|V(\mathbb{F})| =
\bigl(
c + O_M(|\mathbb{F}|^{-\frac{1}{2}})
\bigr) \thinspace
|\mathbb{F}|^{\dim(V)} \,,$$ where $c$ is the number of irreducible components of $V(\overline{{\mathbb{F}}})$.
All the considerations and results we stated so far hold also for projective varieties defined over finite fields. By a *projective variety* defined over a field ${\mathbb{F}}$ we mean the set of common zeros in the projective space ${\mathbb{P}}^n({\mathbb{F}})$ of finitely many *homogeneous* polynomials $P_1, \dotsc,
P_r \in {\mathbb{F}}[x_0, \dotsc, x_n]$. From now on, all the varieties we consider are projective, or are open subsets of projective varieties.
Galois group of a plane curve {#preliminaries:galois}
-----------------------------
The aim of this section is to recall a construction (see [@Rathmann1987]) which associates a Galois group to a plane algebraic curve. We will see in the following sections that this group determines the irreducibility of certain surfaces; this will be the key to derive a formula for the probabilities we are interested in.
Let $q$ be a prime power, namely $q = p^r$ for some prime number $p$. We denote by ${\mathbb{F}}_q$ the finite field with $q$ elements. Let $C$ be an absolutely irreducible algebraic curve in ${\mathbb{P}}^2({\mathbb{F}}_{q})$. Define $$X_1 :=
\bigl\{
(w,[\ell]) \in C \times \check{{\mathbb{P}}}^2({\mathbb{F}}_{q}): w \in \ell
\bigr\}
\quad \textrm{and} \quad
X_0 := \check{{\mathbb{P}}}^2({\mathbb{F}}_{q}).$$ Here $\check{{\mathbb{P}}}^2({\mathbb{F}}_{q})$ denotes the *dual* projective plane, namely the projective plane whose points are in bijection with the lines in ${\mathbb{P}}^2 \bigl( {\mathbb{F}}_q \bigr)$. For a line $\ell \subseteq {\mathbb{P}}^2({\mathbb{F}}_{q})$, we write $[\ell]$ for the corresponding point in $\check{{\mathbb{P}}}^2({\mathbb{F}}_{q})$. The correspondence is given by $$\check{{\mathbb{P}}}^2({\mathbb{F}}_{q}) \ni (a:b:c)
\quad \longleftrightarrow \quad
\bigl\{ (x:y:z) \in {\mathbb{P}}^2({\mathbb{F}}_q) \, : \, ax + by + cz = 0 \bigr\}.$$
\[def:projection\] Using the notation we have already introduced, we define the map $\pi
\colon X_1 \longrightarrow X_0$ to be the projection onto the second component.
Since $X_0$ is irreducible, we can define its *function field*, denoted $K(X_0)$. This is the field of equivalence classes of morphisms $\varphi
\colon U \longrightarrow {\mathbb{F}}_q$, where $U$ is any (Zariski) open subset of $X_0$; two morphisms are considered equivalent if they agree on a non-empty open subset. Consider the projection $\rho \colon X_1 \longrightarrow C$ on the first component: its fibers are lines in $\check{{\mathbb{P}}}^2\bigl( {\mathbb{F}}_q \bigr)$, because the elements in the fiber over a point $w \in C$ correspond to the lines in ${\mathbb{P}}^2
\bigl( {\mathbb{F}}_q \bigr)$ through $w$. Hence all these fibers are irreducible varieties of the same dimension. This implies that $X_1$ is irreducible by [@Shafarevich2013 Chapter 1, Section 6.3, Theorem 1.26]; its function field is denoted $K(X_1)$.
\[lemma:projection\_quasi\_finite\] The projection $\pi \colon X_1 \longrightarrow X_0$ is a quasi-finite dominant separable morphism of degree $d$.
Because of Lemma \[lemma:projection\_quasi\_finite\], the induced map $\pi^{*}
\colon K(X_0) \longrightarrow K(X_1)$ between fields of rational functions realizes $K(X_1)$ as a finite separable extension of $K(X_0)$ of degree $d$. By the primitive element theorem, the field $K(X_1)$ is generated over $K(X_0)$ by a single rational function $h \in K(X_1)$ satisfying $P(h)=0$ for an irreducible monic polynomial $P$ over $K(X_0)$ of degree $d$.
\[def:galois\_curve\] Using the notation just introduced, we define the *Galois group* ${\operatorname{Gal}}(C)$ of $C$ to be the Galois group of a splitting field of the polynomial $P$ over $K(X_0)$. In other words, ${\operatorname{Gal}}(C)$ is the Galois group of a Galois closure (see [@Rowen2006 Remark 4.77]) of the field extension $K(X_0)
\hookrightarrow K(X_1)$. The group ${\operatorname{Gal}}(C)$ is independent of the choice of $h$ and it can be regarded as a subgroup of the permutation group $S_d$ of the roots of $P$.
\[def:simple\_tangency\] Let $C$ be an absolutely irreducible curve of degree $d$ in ${\mathbb{P}}^2(\overline{{\mathbb{F}}_{q}})$. We say that $C$ has *simple tangency* if there exists a line $\ell \subseteq
{\mathbb{P}}^2(\overline{{\mathbb{F}}_{q}})$ intersecting $C$ in $d-1$ smooth points of $C$ such that $\ell$ intersects $C$ transversely at $d-2$ points and has intersection multiplicity $2$ at the remaining point.
A general curve $C \subseteq {\mathbb{P}}^2({\mathbb{F}}_q)$ of degree $d$ has simple tangency. In fact, notice that having simple tangency is an open condition, therefore it is enough to exhibit a single example in order to obtain the claim. To do that, consider the curve of equation $$x^2 \, P(x,y) + z \, Q(x,y,z) = 0,$$ where $P$ is a homogeneous polynomial with $d-2$ distinct roots in $\overline{{\mathbb{F}}}_a$ and $Q$ is a homogeneous polynomial of degree $d-1$.
\[prop:symmetric\_galois\] Let $C \subseteq {\mathbb{P}}^2({\mathbb{F}}_{q})$ be an absolutely irreducible plane curve of degree $d$ with simple tangency. Then the Galois group ${\operatorname{Gal}}(C)$ of $C$ is the whole symmetric group $S_d$.
Galois theory for étale maps {#galois}
============================
In this section we associate a Galois group to a morphism (satisfying certain conditions) between two irreducible smooth varieties. We show (Proposition \[prop:galois\_isomorphic\]) that this concept admits a geometric counterpart, and we use this characterization in the next section. As we pointed out in the Introduction, the results of this and the following section are subsumed by the ones of Section \[chebotarev\]. Nevertheless, we believe that the approach presented in these sections can be useful to help understanding the setting that is used also in Chebotarev theorem to solve this kind of problems.
[**Note.**]{} All varieties considered in this section are supposed to be defined over an algebraically closed field.
For technical reasons, we develop the theory for a special class of morphisms, namely the one of *étale* maps. They model, in the algebraic setting, the notion of “local isomorphism” for the analytic topology. Recall that, in differential geometry, a smooth map between two smooth manifolds is a *local diffeomorphism* if it induces an isomorphism at the level of tangent spaces. For an affine variety $X$ cut out by polynomials $P_1, \dotsc, P_r$, one defines the *tangent cone* $C_{x}(X)$ of $X$ at the origin as the variety defined by the homogeneous parts of minimal degree of each of the polynomials $P_1, \dotsc, P_r$; the tangent cone at any other point is obtained by translating it to the origin and by applying the previous definition. The tangent cone plays for étale morphisms the role played by the tangent space for local diffeomorphisms:
\[def:etale\_map\] A morphism $f \colon X \longrightarrow Y$ between irreducible varieties is *étale at a point* $x \in X$ if $f$ induces an isomorphism between the tangent cones $C_{x}(X)$ and $C_{f(x)}(Y)$. A map is called *étale* if it is étale at every point.
\[def:galois\_scheme\] Let $f \colon X \longrightarrow Y$ be a finite étale map of degree $d$ between two irreducible smooth varieties. We define the *Galois scheme* (see [@Vakil2006 Section $3$]) of $f$ as $${\operatorname{GS}}(f) :=
\bigl\{
(x_1, \dotsc, x_d) \in X^d
\, : \,
f(x_1) = \dotsb = f(x_d), \;
x_i \neq x_j \text{ for all } i \neq j
\bigr\}.$$
Notice that the Galois scheme is the fiber product of $d$ copies of the map $f$ minus the big diagonal. Because of this, and since $f$ is a finite map, we have $$\label{eq:dimension_galois_scheme}
\dim {\operatorname{GS}}(f) \, = \, \dim X \, = \, \dim Y.$$ There is an induced map $F
\colon {\operatorname{GS}}(f) \longrightarrow Y$, sending $(x_1, \dotsc, x_d)$ to $f(x_1)$, which is dominant of degree $d!$.
Because of Definition \[def:galois\_scheme\], in the following we will consider often morphisms between varieties satisfying the following condition: $$\label{eq:condition_morphism}
\tag{$\ast$}
\begin{array}{c}
\text{the morphism is a finite \'etale map of degree } d\\
\text{between smooth absolutely irreducible varieties}.
\end{array}$$
We have a natural action of the symmetric group $S_d$ on ${\operatorname{GS}}(f)$ given by $$\sigma \cdot (x_1, \dotsc, x_d)
:=
\bigl( x_{\sigma(1)}, \dotsc, x_{\sigma(d)} \bigr) \quad
\text{for evey } \sigma \in S_d.$$
\[lemma:transitivity\_components\] Let $f \colon X \longrightarrow Y$ be a morphism satisfying . Then, for any pair of irreducible components $Z$ and $Z'$ of ${\operatorname{GS}}(f)$ there exists an element $\tau \in S_d$ such that $\tau \cdot Z = Z'$. Namely, the action of $S_d$ on ${\operatorname{GS}}(f)$ is transitive on irreducible components.
Since fiber products of étale maps are étale (see [@stacks-project [Tag 03PA](http://stacks.math.columbia.edu/tag/03PA), Proposition 53.26.2]), the morphism $F$ is étale. Moreover, by construction $F$ is dominant, and it is finite, since the fiber product of finite morphisms is finite. Hence $F$ is surjective. Since $Y$ is irreducible and smooth and $F$ is étale, then ${\operatorname{GS}}(f)$ is smooth (see [@stacks-project [Tag 03PA](http://stacks.math.columbia.edu/tag/03PA), Proposition 53.26.2]) and equidimensional (namely, all of its irreducible components have the same dimension, see [@Hartshorne1977 Corollary 9.6] and [@Hartshorne1977 Theorem 10.2] taking into account that an étale morphism is smooth of relative dimension $0$). For any two irreducible components $Z$ and $Z'$, consider the restriction maps $$F_{|_{Z}} \colon Z \longrightarrow Y
\quad \text{and} \quad
F_{|_{Z'}} \colon Z' \longrightarrow Y.$$ These maps are also étale, in fact open immersions are étale, and the composition of étale maps is étale. Moreover, both restrictions are dominant. In fact, since ${\operatorname{GS}}(f)$ is equidimensional, it follows $\dim Z = \dim
{\operatorname{GS}}(f)$, and since $F$ is finite, we get $\dim F(Z) = \dim(Z) = \dim(Y)$ because of Equation .
Since $f$ is finite and étale, then for every $y \in Y$ the fiber $f^{-1}(y)$ is constituted of $d$ distinct points (see [@Goertz2010 Equation 12.6.2][^3]). Hence, for all $y \in Y$ we have $|F^{-1}(y)| = d!$. Fix $y \in Y$ and write $f^{-1}(y) = \{ x_1, \dotsc, x_d \}$. Therefore $$F^{-1}(y) =
\bigl\{
(x_{\sigma(1)}, \dotsc, x_{\sigma(d)}) \, : \, \sigma \in S_d
\bigr\} \, ,$$ where $S_d$ is the symmetric group. Suppose that $F^{-1}(y)$ intersects nontrivially both $Z$ and $Z'$. It follows that there exists $a \in Z$ and $a'
\in Z'$ and an element $\tau \in S_d$ such that $\tau \cdot a = a'$. Since the action of $S_d$ on ${\operatorname{GS}}(f)$ is algebraic, then the action of any element $\sigma
\in S_d$ determines an automorphism of ${\operatorname{GS}}(f)$; in particular, such an action sends irreducible components to irreducible components. It follows that $\tau
\cdot Z = Z'$. Hence, we are left to show that such an element $y \in Y$ exists. This is the case because of the following argument. Define $${\widetilde}{Y}_{Z} =
\bigl\{
y \in Y \, : \, F^{-1}(y) \cap Z \neq \emptyset
\bigr\}
\quad \text{and} \quad
{\widetilde}{Y}_{Z'} =
\bigl\{
y \in Y \, : \, F^{-1}(y) \cap Z' \neq \emptyset
\bigr\}.$$ These two sets are open, and since the restrictions $F_{|_{Z}}$ and $F_{|_{Z'}}$ are dominant, they are non-empty. Since $Y$ is irreducible, then ${\widetilde}{Y}_{Z} \cap {\widetilde}{Y}_{Z'}$ is open and non-empty. Any point in ${\widetilde}{Y}_{Z} \cap {\widetilde}{Y}_{Z'}$ satisfies the desired requirement.
Lemma \[lemma:transitivity\_components\] shows that the action of the symmetric group $S_d$ on the Galois scheme is transitive on irreducible components. Because of this, the stabilizers of these components are conjugate subgroups of $S_d$.
We define the *geometric Galois group* ${\operatorname{Gal}}_{g}(f)$ of a morphism $f
\colon X \longrightarrow Y$ of degree $d$ satisfying to be the stabilizer of any irreducible component of ${\operatorname{GS}}(f)$ with respect to the action of the symmetric group $S_d$. This definition is well-posed up to conjugacy in $S_d$.
It follows that the number of irreducible components of the Galois scheme ${\operatorname{GS}}(f)$ coincides with the number of cosets of the geometric Galois group ${\operatorname{Gal}}_{g}(f)$ in $S_d$.
Let $f \colon X \longrightarrow Y$ be a morphism satisfying . Since $f$ is dominant, it determines a field extension $K(Y) \hookrightarrow K(X)$. We define the *algebraic Galois group* ${\operatorname{Gal}}_{a}(f)$ of $f$ to be the Galois group of the extension $K(Y)
\hookrightarrow E$, where $E$ is a Galois closure (see [@Rowen2006 Remark 4.77]) of $K(Y) \hookrightarrow K(X)$.
\[prop:galois\_isomorphic\] For every morphism $f \colon X \longrightarrow Y$ of degree $d$ satisfying , the two groups ${\operatorname{Gal}}_{g}(f)$ and ${\operatorname{Gal}}_{a}(f)$ are isomorphic.
Let $Z$ be any component of ${\operatorname{GS}}(f)$ and realize ${\operatorname{Gal}}_{g}(f)$ as the stabilizer of $Z$. Since the restriction $F_{|_{Z}} \colon Z \longrightarrow Y$ is dominant (see Lemma \[lemma:transitivity\_components\]) we have a field inclusion $K(Y) \hookrightarrow K(Z)$.
[**Claim**]{}. ${\operatorname{Gal}}_g(f) \cong {\operatorname{Gal}}\bigl( K(Z) / K(Y) \bigr)$.\
[*Proof of the claim*]{}. Since ${\operatorname{Gal}}_{g}(f)$ is the stabilizer of $Z$, then for every $\sigma \in {\operatorname{Gal}}_{g}(f)$ we have an automorphism $\varphi_\sigma
\colon Z \longrightarrow Z$, which induces an automorphism $\psi_\sigma
\colon
K(Z) \longrightarrow K(Z)$ fixing $K(Y)$. We define a group homomorphism $$\label{eq:isomorphism_galois}
\begin{array}{rccc}
\Psi \colon & {\operatorname{Gal}}_{g}(f) & \longrightarrow & {\operatorname{Gal}}\bigl( K(Z) / K(Y) \bigr)\\
& \sigma & \mapsto & \psi_\sigma
\end{array}$$ Let $b$ be the number of components of ${\operatorname{GS}}(f)$. Notice that the field extension $K(Y) \hookrightarrow K(Z)$ has degree $d! /
b$, since $K(Y) \hookrightarrow K({\operatorname{GS}}(f))$ has degree $d!$ and all components of ${\operatorname{GS}}(f)$ differ by the action of an element of $S_d$. In particular, $|{\operatorname{Gal}}\bigl( K(Z) / K(Y) \bigr)| \leq d! / b$. The group homomorphism $\Psi$ is injective because any automorphism which is the identity on an open subset is the identity everywhere. Hence the set $\Psi \bigl( {\operatorname{Gal}}_{g}(f)
\bigr)$ has cardinality $|{\operatorname{Gal}}_{g}(f)|$. By what we noticed before, the number $|{\operatorname{Gal}}_{g}(f)|$ equals $|S_d| / b = d! / b$ because the number of components of ${\operatorname{GS}}(f)$ equals the number of cosets of ${\operatorname{Gal}}_{g}(f)$. Hence, $\Psi$ is an isomorphism and $K(Y)
\hookrightarrow K(Z)$ is a Galois extension (see [@Chambert-Loir2005 Definition 3.2.5]).
[**Claim**]{}. $K(Z)$ is a Galois closure of $K(Y) \hookrightarrow K(X)$.\
[*Proof of the claim*]{}. We know already that $K(Y) \hookrightarrow K(Z)$ is a Galois extension. Moreover, the latter factors via $K(Y) \hookrightarrow K(X)$ by considering the projection $Z \longrightarrow X$ on the first factor. The only thing left to prove is that $K(Y) \hookrightarrow K(Z)$ is minimal among Galois extensions of $K(Y) \hookrightarrow K(X)$. Namely, we have to show that if $E \subseteq K(Z)$ is a field such that the image of the inclusion $K(X)
\hookrightarrow K(Z)$ is contained in $E$, and $K(Y) \hookrightarrow E$ is Galois, then $E = K(Z)$. Define $G := {\operatorname{Gal}}\bigl( K(Z) / K(Y) \bigr)$. By the Galois correspondence, and recalling that the inclusion $K(X) \hookrightarrow
K(Z)$ is given by the projection on the first factor, we have $$\label{eq:galois_correspondence}
\begin{array}{ccccc}
G & \supset & {\operatorname{Stab}}_{G}(1) & \supset & \{\mathrm{id}\} \\
\updownarrow & & \updownarrow && \updownarrow \\
K(Y) & \hookrightarrow & K(X) & \hookrightarrow & K(Z)
\end{array}$$ where ${\operatorname{Stab}}_{G}(1) = \bigl\{\sigma \in G \, : \, \sigma(1)=1
\bigr\}$ is the stabilizer of $1$ under the action of $G$ on $\{1,
\dotsc, d\}$. This action is given by the identification of $G$ with ${\operatorname{Gal}}_g(f)$ provided by the map $\Psi$ in Equation . Under this correspondence, the field $E$ is associated to a subgroup $H \subseteq
G$, which is normal since $K(Y) \hookrightarrow E$ is Galois, and which is contained in ${\operatorname{Stab}}_{G}(1)$. To conclude, we prove that $H = \{ \mathrm{id} \}$, which implies $E = K(Z)$. We start by showing that $G$ acts transitively on $\{1, \dotsc, d\}$. If we denote by $G \cdot 1$ the orbit of $1$ under $G$, then by standard results in Galois theory and taking into account Equation , we have $$|G \cdot 1|=\left[G : {\operatorname{Stab}}_{G}(1) \right]=\left[K(X):K(Y) \right]=d.$$ This implies that $G \cdot 1 = \{ 1, \dotsc, d\}$, showing that the action is transitive. By normality, $H \subseteq \sigma \, {\operatorname{Stab}}_{G}(1) \, \sigma^{-1}$ for any $\sigma \in G$. Since $G$ is transitive on $\{1, \dotsc, d\}$, it follows $$\label{eq:intersection_stabilizers}
H \; \subseteq \;
{\operatorname{Stab}}_{G}(1) \cap {\operatorname{Stab}}_{G}(2)
\cap \dotsb \cap {\operatorname{Stab}}_{G}(d).$$ The right hand side of Equation equals $\{\mathrm{id}\}$, so the claim is proved.
Summing up, the first claim says that ${\operatorname{Gal}}_g(f) \cong {\operatorname{Gal}}\bigl( K(Z) / K(Y)
\bigr)$, and the second claim implies that the latter group is isomorphic to ${\operatorname{Gal}}_a(f)$. This concludes the proof of the proposition.
Now we cast the notions defined so far into the framework of Galois schemes of morphisms (Corollary \[cor:galois\_curve\_as\_map\]). After that, we recall the notion of simple tangency for a curve and highlight its consequences on Galois groups.
\[def:etale\_locus\] For an absolutely irreducible curve $C \subseteq {\mathbb{P}}^2({\mathbb{F}}_q)$ of degree $d$, define ${\mathcal{V}}_C$ to be the set of points in $X_0(\overline{{\mathbb{F}}}_q) =
\check{{\mathbb{P}}}^2 \bigl( \overline{{\mathbb{F}}}_q \bigr)$ such that the restriction of the map $\pi \colon X_1(\overline{{\mathbb{F}}}_q) \longrightarrow X_0(\overline{{\mathbb{F}}}_q)$ from Definition \[def:projection\] to ${\mathcal{U}}_{C} :=
\pi^{-1}\bigl({\mathcal{V}}_C\bigr)$ is étale.
\[rem:etale\] Notice that the set ${\mathcal{V}}_C$ is open and non-empty. In fact, since the map $\pi \colon X_1(\overline{{\mathbb{F}}}_q) \longrightarrow X_0(\overline{{\mathbb{F}}}_q)$ is separable, it is enough to ensure that $\pi \colon {\mathcal{U}}_C \longrightarrow
{\mathcal{V}}_C$ is flat. Now, the locus in the domain where a map is flat is open (see [@stacks-project [Tag 0398](http://stacks.math.columbia.edu/tag/0398), Theorem 36.15.1,]), and flat maps are open morphisms (see [@stacks-project [Tag 01U2](http://stacks.math.columbia.edu/tag/01U2), Lemma 28.24.9]), so this shows that ${\mathcal{V}}_C$ is open. The fact that ${\mathcal{V}}_C$ is non-empty is ensured by the generic flatness result (see [@stacks-project [Tag 0529](http://stacks.math.columbia.edu/tag/0529), Proposition 28.26.1]).
\[lemma:restriction\_etale\] Let $C$ be an absolutely irreducible curve of degree $d$ defined over ${\mathbb{F}}_q$. Then the restriction to ${\mathcal{U}}_{C} := \pi^{-1}\bigl({\mathcal{V}}_C\bigr)$ of the map $\pi \colon X_1\bigl( \overline{{\mathbb{F}}}_q \bigr) \longrightarrow X_0\bigl(
\overline{{\mathbb{F}}}_q \bigr)$ from Definition \[def:projection\] is a finite separable dominant étale morphism between smooth absolutely irreducible varieties, namely it satisfies condition .
We know from Section \[preliminaries:galois\] that both $X_0$ and $X_1$ are smooth and absolutely irreducible. Since ${\mathcal{V}}_C$ and ${\mathcal{U}}_C$ are open and non-empty, the same is true for them. Moreover, $\pi$ is a quasi-finite separable dominant morphism between projective varieties (Lemma \[lemma:projection\_quasi\_finite\]) and so it is finite. Hence, the same holds for its restriction $\pi_{|_{{\mathcal{U}}_C}}$. By Remark \[rem:etale\], the map is étale, and this concludes the proof.
By unravelling the definition, in the light of Lemma \[lemma:restriction\_etale\] we obtain:
\[cor:galois\_curve\_as\_map\] For an absolutely irreducible curve $C$ defined over ${\mathbb{F}}_q$, we have ${\operatorname{Gal}}(C) \cong {\operatorname{Gal}}_{a} \bigl( \pi_{|_{{\mathcal{U}}_C}} \bigr)$.
The interpretation of the Galois group of a curve provided by Corollary \[cor:galois\_curve\_as\_map\] allows to use Proposition \[prop:galois\_isomorphic\] and hence to deduce the irreducibility of the Galois scheme when the Galois group is the full symmetric group.
Suppose that $C$ is an absolutely irreducible curve in ${\mathbb{P}}^2({\mathbb{F}}_{q})$ of degree $d$ with simple tangency. Then, the Galois group ${\operatorname{Gal}}_{g}\bigl(
\pi_{|_{{\mathcal{U}}_C}} \bigr) \cong {\operatorname{Gal}}_{a}\bigl( \pi_{|_{{\mathcal{U}}_C}} \bigr)$ is the full symmetric group, and so the Galois scheme ${\operatorname{GS}}\bigl(
\pi_{|_{{\mathcal{U}}_C}} \bigr)$ is irreducible.
This follows from Corollary \[cor:galois\_curve\_as\_map\] and Proposition \[prop:galois\_isomorphic\].
Probabilities of incidence {#probabilities}
==========================
In this section we define probabilities of intersection between a random line and a given curve in the projective plane over a finite field (Definition \[def:density\]). We then prove the main result of our paper, namely Theorems \[thm:existence\] and \[thm:formula\], by showing that its counterpart for morphisms hold (Theorems \[thm:existence\_maps\] and \[thm:formula\_maps\]). We will re-prove these results in Section \[chebotarev\] by using Chebotarev density theorem.
\[def:density\] Let $q$ be a prime power and let $C \subseteq {\mathbb{P}}^2 \bigl( {\mathbb{F}}_q
\bigr)$ be an absolutely irreducible curve of degree $d$ defined over ${\mathbb{F}}_q$. For every $N \in {\mathbb{N}}$ and for every $k \in \{0, \dotsc, d\}$, the *$k$-th probability of intersection* $p_k^N(C)$ of lines with $C$ over ${\mathbb{F}}_{q^N}$ is $$p_k^N(C) :=
\frac{\Bigl|
\bigl\{
\text{lines } \ell \subseteq {\mathbb{P}}^2 ({\mathbb{F}}_{q^N}) \, : \,
|\ell({\mathbb{F}}_{q^N}) \cap C({\mathbb{F}}_{q^N}) | = k
\bigr\}
\Bigr|}{q^{2N}+q^N+1} \,.$$ Notice that $q^{2N}+q^N+1$ is the number of lines in ${\mathbb{P}}^2 \bigl( {\mathbb{F}}_{q^N}
\bigr)$.
The aim of this paper is to prove that the limit as $N$ goes to infinity of the quantities $p_k^N(C)$ exists for every $k$, and to give a formula for these limits, provided that some conditions on the curve $C$ are fulfilled.
The following result is a direct consequence of Definitions \[def:density\] and \[def:projection\].
Let $C \subseteq {\mathbb{P}}^2({\mathbb{F}}_q)$ be an absolutely irreducible curve of degree $d$ defined over ${\mathbb{F}}_q$. For every $k \in \{0, \dotsc, d\}$ we have $$p_k^N(C) =
\frac{\Bigl|
\bigl\{
[\ell] \in \check{{\mathbb{P}}}^2 ({\mathbb{F}}_{q^N}) \, : \,
| \pi^{-1}([\ell])({\mathbb{F}}_{q^N})| = k
\bigr\}
\Bigr|}{q^{2N}+q^N+1} \,.$$
Via Lemma \[lemma:reduction\_etale\] and Definition \[def:probabilities\_map\] we reduce the problem of computing intersection probabilities for curves to the analogous problem for morphisms.
\[lemma:reduction\_etale\] Let $C \subseteq {\mathbb{P}}^2({\mathbb{F}}_q)$ be an absolutely irreducible curve of degree $d$ defined over ${\mathbb{F}}_q$. Let ${\mathcal{V}}_C \subseteq \check{{\mathbb{P}}}^2(\overline{{\mathbb{F}}}_q)$ be as in Definition \[def:etale\_locus\]. For every $N \in {\mathbb{N}}$ and for every $k
\in \{0, \dotsc, d\}$, define $${\widetilde}{p}_k^N(C) :=
\frac{\Bigl|
\bigl\{ [\ell] \in {\mathcal{V}}_C({\mathbb{F}}_{q^N}) \, : \,
| \pi^{-1}([\ell]) \bigl( {\mathbb{F}}_{q^N} \bigr)| = k \bigr\}
\Bigr|}{|{\mathcal{V}}_C({\mathbb{F}}_{q^N})|} \,.$$ Then $\displaystyle \lim_{N \to \infty} p_k^N(C)$ exists if and only if $\displaystyle \lim_{N \to \infty} {\widetilde}{p}_k^N(C)$ exists, in which case the two numbers coincide.
It is enough to show that the probability for a point to lie in ${\mathbb{P}}^2({\mathbb{F}}_{q^N}) \setminus {\mathcal{V}}_C({\mathbb{F}}_{q^N})$ goes to zero as $N$ goes to infinity. This is a consequence of the Lang-Weil bound (Theorem \[thm:lang\_weil\]). In fact, since ${\mathbb{P}}^2({\mathbb{F}}_{q^N}) \setminus
{\mathcal{V}}_C({\mathbb{F}}_{q^N})$ has dimension at most $1$: $$\frac{\bigl| {\mathbb{P}}^2({\mathbb{F}}_{q^N}) \setminus {\mathcal{V}}_C({\mathbb{F}}_{q^N})
\bigr|}{q^{2N}+q^N+1}
\, \sim \,
\frac{a \, q^N}{q^{2N}} \to 0 \, ,$$ where the constant $a$ is the number of irreducible components of ${\mathbb{P}}^2(\overline{{\mathbb{F}}}_q) \setminus {\mathcal{V}}_C(\overline{{\mathbb{F}}}_q)$.
\[def:probabilities\_map\] Let $f \colon X \longrightarrow Y$ be a morphism of degree $d$ defined over ${\mathbb{F}}_q$, where $q$ is a prime power, satisfying . For every $N
\in {\mathbb{N}}$ and for every $k \in \{0, \dotsc, d\}$, we define the *$k$-th preimage probability* $p_k^N(f)$ to be $$p_k^N(f) :=
\frac{\Bigl|
\bigl\{
y \in Y({\mathbb{F}}_{q^N}) \, : \, |f^{-1}(y)({\mathbb{F}}_{q^N})| = k
\bigr\}
\Bigr|}{|Y({\mathbb{F}}_{q^N})|} \,.$$
Notice that if $C$ is an absolutely irreducible algebraic plane curve of degree $d$, then for every $N \in {\mathbb{N}}$ and for every $k \in \{0, \dotsc, d\}$ we have ${\widetilde}{p}_k^N(C) = p_k^N\bigl(\pi_{|_{{\mathcal{U}}_C}}\bigr)$. Hence, by Lemma \[lemma:reduction\_etale\], in order to show the existence of the limits of $k$-th probabilities of intersections for a curve, it is enough to show the existence of $k$-th preimage probabilities for morphisms over ${\mathbb{F}}_q$ satisfying .
\[thm:existence\_maps\] Let $f \colon X \longrightarrow Y$ be a morphism of degree $d$ defined over ${\mathbb{F}}_q$, where $q$ is a prime power, satisfying . Then for every $k \in \{0, \dotsc, d\}$ the limit as $N$ goes to infinity of the sequence $\bigl( p_k^N(f) \bigr)_{N \in {\mathbb{N}}}$ exists.
We generalize the construction of the Galois scheme of the morphism $f$. For every $k \in \{0, \dotsc, d\}$, define $$G_k(f) :=
\bigl\{
(x_1, \dotsc, x_k) \in X^k
\, : \,
f(x_1) = \dotsc = f(x_k), \;
x_i \neq x_j \text{ for all } i \neq j
\bigr\}.$$ In particular $G_d(f) = {\operatorname{GS}}(f)$. As we showed for the Galois scheme, see Equation , for every $k$ the variety $G_k(f)$ has the same dimension of $X$ and $Y$. There is a natural finite morphism $F_k \colon G_k(f) \longrightarrow Y$, the fiber product of $f$ with itself $k$ times. This map has degree $d(d-1)
\dotsb
(d-k+1)$. The main idea of the proof is to compute, in two different ways, the expected cardinality $\mu_k^N(f)$ of a fiber $F_k^{-1}(y)$ over ${\mathbb{F}}_{q^N}$, where $y$ is a random element in $Y$. On one hand, $$\mu_k^N(f) = \frac{\bigl| G_k(f)({\mathbb{F}}_{q^N}) \bigr|}{\bigl| Y({\mathbb{F}}_{q^N}) \bigr|}
\,.$$ On the other hand, we can express $\mu_k^N(f)$ in terms of the preimage probabilities: $$\label{eq:expectation}
\mu_k^N(f) = \sum_{s = k}^{d} s (s-1) \dotsb (s-k+1) \, p_s^N(f) \,.$$ In matrix form: $$\label{eq:matrix_form}
\begin{pmatrix}
\mu_0^N(f) \\ \vdots \\ \mu_d^N(f)
\end{pmatrix}
=
\begin{pmatrix}
1 & \ast & \cdots & & \cdots & \ast\\
0 & 1 & \ast & & & \vdots \\
\vdots & & \ddots & & & \vdots\\
0 & \cdots & 0 & k! & \ast & \ast\\
\vdots & & & & \ddots & \vdots \\
0 & \cdots & & & \cdots & d!
\end{pmatrix}
\begin{pmatrix}
p_0^N(f) \\ \vdots \\ p_d^N(f)
\end{pmatrix}.$$ Since the matrix in Equation has non-zero determinant, we can write $$\label{eq:probabilities_expectation}
p_k^N(f) = \sum_{s = 0}^d \alpha_{k,s} \, \mu_s^N$$ for some numbers $(\alpha_{k,s})_{k,s}$. Using the Lang-Weil bound on Equation , we have $$\label{eq:limit}
\mu_k^N \, \sim \, \frac{\delta_k \, q^{N \cdot \dim G_k(f)}}{q^{N \cdot \dim
Y}}
\qquad \text{as } N \to \infty \, ,$$ where $\delta_k$ is the number of irreducible components of $G_k(f)(\overline{{\mathbb{F}}}_q)$. Since $\dim G_k(f) = \dim Y$, we conclude that the limit in Equation exists, and so by Equation also $\displaystyle \lim_{N \to
\infty} p_k^N(f)$ exists.
Theorem \[thm:existence\] holds. In fact, the map $\pi_{|_{{\mathcal{U}}_C}}$ satisfies the hypotheses of Theorem \[thm:existence\_maps\], so the numbers $p_k \bigl( \pi_{|_{{\mathcal{U}}_C}} \bigr)$ exist, and we have already proved that this implies that the limits $p_k(C)$ exist.
\[thm:formula\_maps\] Let $f \colon X \longrightarrow Y$ be a morphism of degree $d$ defined over ${\mathbb{F}}_q$, where $q$ is a prime power, satisfying . Suppose that ${\operatorname{Gal}}_g(f) \cong
{\operatorname{Gal}}_a(f)$ is the full symmetric group $S_d$. Then for every $k \in \{0,
\dotsc,
d\}$ we have $$p_k(f) = \sum_{s=k}^d \frac{(-1)^{k+s}}{s!} \binom{s}{k}.$$ In particular, $p_{d-1}(f) = 0$ and $p_d(f) = 1 / d!$.
Since ${\operatorname{Gal}}_g(f)$ is the full symmetric group, the Galois scheme ${\operatorname{GS}}(f)$ is absolutely irreducible. Hence, using the notation of the proof of Theorem \[thm:existence\_maps\], for all $k \in \{0,
\dotsc, d\}$ we have $$\label{eq:expected_preimage}
\lim_{N \to \infty} \mu_k^N(f) =
\lim_{N \to \infty} \frac{q^{N \cdot \dim G_k(f)}}{q^{N \cdot \dim Y}} = 1.$$ In fact, every variety $G_k(f)$ is an image (under a projection) of ${\operatorname{GS}}(f) =
G_d(f)$, thus is absolutely irreducible and so Equation follows from Equation . Again using the notation as in Theorem \[thm:existence\_maps\], we get $$\label{eq:formula_alpha}
\lim_{N \to \infty} p_k^N(f) = \sum_{s = 0}^{d} \alpha_{k,s}.$$ Therefore, the statement is proved once we are able to explicitly compute the coefficients $(\alpha_{k,s})_{k,s}$. Recall that $\alpha_{k,s}$ is the $(k,s)$-entry of the inverse of the matrix $M_d$ appearing in Equation . A direct inspection of the matrices $M_d$ shows that they admit the following structure: $$M_d =
\left(
\begin{array}{ccc|c}
\multicolumn{3}{c|}{\multirow{3}{*}{$M_{d-1}$}} & 1 \\
&&& \vdots \\
&&& d! / 1! \\ \cline{1-3}
0 & \cdots & 0 & d! / 0!
\end{array}
\right) \,.$$ A direct computation shows that $$M_d^{-1} =
\left(
\begin{array}{ccc|c}
\multicolumn{3}{c|}{\multirow{3}{*}{$M_{d-1}^{-1}$}} & \frac{(-1)^d}{d!} \cdot
\binom{d}{0} \\
&&& \vdots \\
&&& \frac{(-1)}{d!} \cdot \binom{d}{d-1} \\ \cline{1-3}
0 & \cdots & 0 & \frac{1}{d!} \cdot \binom{d}{d}
\end{array}
\right) \,.$$ Hence $$\alpha_{k,s} = \frac{(-1)^{k+s}}{s!} \binom{s}{k}
\quad \text{ for all } k,s \in \{0, \dotsc, d\}.$$ It follows from Equation that for all $k \in \{0,
\dotsc, d\}$, $$p_k(f) = \sum_{s=0}^d \frac{(-1)^{k+s}}{s!} \binom{s}{k}
= \sum_{s=k}^d \frac{(-1)^{k+s}}{s!} \binom{s}{k}$$ and so the statement is proved.
As a consequence of Proposition \[prop:symmetric\_galois\] and Theorem \[thm:formula\_maps\], we obtain:
Theorem \[thm:formula\] holds.
As we pointed out in the Introduction, one of the consequences of Theorem \[thm:formula\] is that a question raised by Erdös concerning $4$- and $5$-rich lines has a positive answer over finite fields.
\[cor:erdos\_problem\] Let $C$ be an absolutely irreducible plane algebraic curve of degree $4$ in the plane ${\mathbb{P}}^2({\mathbb{F}}_{q})$. By Theorem \[thm:lang\_weil\], the curve $C$ has $cq+O(\sqrt{q})$ elements for some $c>0$, and by Theorem \[thm:formula\] it has $\epsilon q^2$ $4$-rich lines for some $\epsilon>0$, after possibly taking a finite extension of the base field, since $p_4(C) > 0$. Hence, if we take $P$ as the set of points of $C$, then $P$ spans a quadratic number of $4$-rich lines, but no five points of $P$ are collinear.
Probabilities of intersection via Chebotarev theorem {#chebotarev}
====================================================
In this section, we show how to use an effective version of Chebotarev density theorem for function fields as exposed in [@Andrade2015 Appendix A], and used in [@Bary-Soroker2012] and [@Diem2012] to prove the results reported in the Introduction, to prove Theorems \[thm:existence\] and \[thm:formula\]. We recall the setting of the paper and specialize Chebotarev theorem to our case; we refer to the cited appendix for the proofs of the claims we make in this section regarding the objects introduced to state Chebotarev theorem.
We start by considering an integrally closed finitely generated ${\mathbb{F}}_q$-algebra $R$ and a monic polynomial ${\mathcal{F}} \in R[T]$ such that the discriminant of ${\mathcal{F}}$ is invertible in $R$. In our case, we take $R$ to be the ${\mathbb{F}}_q$-algebra $$R := \frac{{\mathbb{F}}_q[a,b,u]}{{\operatorname{Disc}}_x \bigl(F(x, ax+b) \bigr) \cdot u - 1} \cong
{\mathbb{F}}_q[a,b]_{(f)} \quad \text{with } f := {\operatorname{Disc}}_x \bigl(F(x, ax+b) \bigr),$$ where the last ring is the localization of the polynomial ring ${\mathbb{F}}_q[a,b]$ at the element $f$. In geometric terms, $R$ is the coordinate ring of the open subset of the dual projective plane parametrizing lines in the plane that intersect the curve $\{F = 0\}$ in $d$ distinct points over the algebraic closure of ${\mathbb{F}}_q$. We then take the polynomial ${\mathcal{F}}$ to be $F(T, aT + b)$. Then by construction, its discriminant is invertible in $R$.
Starting from $R$ and ${\mathcal{F}}$, we consider $K$, the quotient field of $R$, and we define $L$ to be the splitting field of ${\mathcal{F}}$ over $K$. In other words, if $\{y_1, \dotsc, y_d\}$ are the roots of ${\mathcal{F}}$, we set $L :=
K(y_1, \dotsc, y_d)$. In our situation, we have $$L = \frac{K[t_1, \dotsc, t_d]}{\bigl( F(t_i, a t_i + b) \text{ for } i \in
\{1, \dotsc, d \} \bigr)} \,.$$ Then we define $S$ to be the integral closure of $R$ in $L$, namely $S = R[y_1,
\dotsc, y_d]$. Geometrically, $S$ is the coordinate ring of an open subset of the variety $$X_d := \bigl\{ (x_1, \dotsc, x_d, [\ell]) \in C^d \times
\check{{\mathbb{P}}}^2 \, \colon \, x_i \in \ell \bigr\} \,.$$ The strategy we adopt to compute probabilities of intersections is the following: our goal is to count the number of lines $\ell$ in ${\mathbb{P}}^2$ such that the intersection $\ell \cap C$ is constituted of exactly $k$ points over ${\mathbb{F}}_{q^N}$, and we interpret this as the number of lines such that the univariate polynomial $F_{|{\ell}}$ has exactly $k$ linear factors over ${\mathbb{F}}_{q^N}$. Notice that every univariate polynomial $H$ of degree $d$ over ${\mathbb{F}}_q^{N}$ determines a partition $\pi_H$ of $d$, namely a tuple $\pi_H = (\alpha_1, \dotsc, \alpha_s)$ such that $\alpha_1 + \dotsb + \alpha_s = d$ and $\alpha_1 \leq \dotsb \leq \alpha_s$. Such partition is obtained by factoring $H$ over ${\mathbb{F}}_{q^N}$ into irreducible factors $H_1, \dotsc, H_s$ and then setting $\alpha_i = \deg(H_i)$. Then, the number of lines we are interested in can be computed as the sum, over the set of partitions $\pi$ of $d$ with exactly $k$ ones, of the number of lines $\ell$ such that the partition associated to $F_{|\ell}$ is $\pi$. Chebotarev theorem provides a formula for the probability for a line to determine a given partition.
We set $G$ to be the Galois group of the field extension $K \subseteq L$. By definition, this coincides with the Galois group of the curve $C$ as in Definition \[def:galois\_curve\]. Notice that, in our situation, the intersection $L \cap {\mathbb{F}}$, where ${\mathbb{F}}$ is an algebraic closure of ${\mathbb{F}}_q$, coincides with ${\mathbb{F}}_q$. This implies that the subgroup $$G_0 := \bigl\{ g \in G \, \colon \, g_{|_{{\mathbb{F}}_q}}(x) = x \text{ for all } x \in
{\mathbb{F}}_q \bigr\}$$ coincides with $G$. Similarly, if for every $\nu \geq 1$ we set $$G_\nu := \bigl\{ g \in G \, \colon \, g_{|_{{\mathbb{F}}_q}}(x) = x^{q^\nu} \text{ for
all } x \in {\mathbb{F}}_q \bigr\} \, ,$$ then $G_{\nu}$, which in general is a coset of $G_0$ in $G$, coincides with $G$.
As one can see from the definition of $X_d$, its points are intimately related to the probabilities we are interested in. From an algebraic point of view (see [@Mumford1999 Section II.6]) these points correspond to ${\mathbb{F}}_q$-homomorphisms from $S$ to ${\mathbb{F}}$. Moreover, a homomorphism $\Phi \in
{\operatorname{Hom}}_{{\mathbb{F}}_q}(S, {\mathbb{F}})$ such that $\Phi(R) = {\mathbb{F}}_{q^\nu}$ corresponds to a point $(x_1, \dotsc, x_d, [\ell])$ in $X_d$ such that the line $\ell$ is defined over ${\mathbb{F}}_{q^\nu}$. Given such a homomorphism $\Phi$ there always exists an element in $G$, called the *Frobenius element* and denoted $\left[ \frac{S / R}{\Phi} \right]$ such that the following diagram is commutative: $$\label{eq:frobenius}
\xymatrix@C=1.5cm{
S \ar[r]^{\left[ \frac{S / R}{\Phi} \right]} \ar[d]_{\Phi} & S
\ar[d]^{\Phi} \\
{\mathbb{F}}\ar[r]^{\alpha \mapsto \alpha^{q^{\nu}}} & {\mathbb{F}}}$$ In other words, we have the relation $$\Phi \left( \left[ \frac{S / R}{\Phi} \right] \, x \right) =
\Phi(x)^{q^{\nu}} \, .$$ One then can show that $\left[ \frac{S / R}{\Phi} \right] \in G_{\nu}$.
If we fix a line in ${\mathbb{P}}^2$, namely, if we fix an ${\mathbb{F}}_q$-homomorphism $\varphi
\in {\operatorname{Hom}}_{{\mathbb{F}}_q}(R, {\mathbb{F}})$, we can consider all points in $X_d$ “lying over” this line. In other terms, we can consider all homomorphisms $\Phi \in
{\operatorname{Hom}}_{{\mathbb{F}}_q}(S, {\mathbb{F}})$ prolonging $\varphi$. Their corresponding Frobenius elements form one key object in the statement of the Chebotarev theorem. For $\varphi \in {\operatorname{Hom}}_{{\mathbb{F}}_q}(R, {\mathbb{F}})$, we set $$\left( \frac{S / R}{\varphi} \right) := \left\{ \left[ \frac{S /
R}{\Phi} \right] \, \colon \, \Phi \text{ prolongs } \varphi \right\} \, .$$ In our setting, since $G_0 = G$ one can show that $\left( \frac{S / R}{\varphi}
\right)$ is a conjugacy class in $G$. Now we are ready to state Chebotarev theorem (see [@Andrade2015 Theorem A.4]): let $Z \subseteq G$ be a conjugacy class and let $\nu \geq 1$; define $$P_{\nu, Z} := \frac{\left| \bigl\{ \varphi \in {\operatorname{Hom}}_{{\mathbb{F}}_q}(R, {\mathbb{F}}) \text{
such that } \varphi(R) = {\mathbb{F}}_{q^{\nu}} \text{ and } \left( \frac{S / R}{\varphi}
\right) = Z \bigr\} \right|}{\left| \bigl\{ \varphi \in {\operatorname{Hom}}_{{\mathbb{F}}_q}(R, {\mathbb{F}})
\text{ such that } \varphi(R) = {\mathbb{F}}_{q^{\nu}} \bigr\} \right|} \, .$$ Then there exists a constant $\delta$ independent of $q$ such that $$P_{\nu, Z} = \frac{|Z|}{|G|} + \frac{\delta}{\sqrt{q}} \, .$$ In order to use Chebotarev theorem for our purposes, we have to understand what does the condition $ \left( \frac{S / R}{\varphi} \right) = Z$ correspond to in our setting. Suppose that $C$ has simple tangency. then we know by Proposition \[prop:symmetric\_galois\] that $G$ is the symmetric group $S_d$. Notice that to every conjugacy class $Z$ of $S_d$ we can associate a partition $\pi_Z$ of $d$, obtained from the cycle structure of permutations belonging to $Z$. On the other hand, given a line $\ell = \{ y = ax + b\}$, we can consider the restriction of the equation $F$ of $C$ to $\ell$, namely the univariate polynomial $F_\ell = F(x, ax + b)$. This polynomial defines a partition $\pi_\ell$ of $d$ by considering its factorization over ${\mathbb{F}}_{q^\nu}$: the partition $\pi_\ell$ has as many $1$ as the linear factors of $F_\ell$, as many $2$ as the quadratic factors of $F_\ell$, and so on.
If $Z \subseteq S_d$ is a conjugacy class of permutations, then the set $$I_{\nu, Z} := \bigl\{ \varphi \in {\operatorname{Hom}}_{{\mathbb{F}}_q}(R, {\mathbb{F}}) \text{
such that } \varphi(R) = {\mathbb{F}}_{q^{\nu}} \text{ and } \left( \frac{S /
R}{\varphi}
\right) = Z \bigr\}$$ corresponds to the set of lines in ${\mathbb{P}}^2$ defined over ${\mathbb{F}}_{q^{\nu}}$ such that $\pi_\ell = \pi_Z$.
Let $\varphi \in I_{\nu, Z}$ and let $\Phi \in {\operatorname{Hom}}(S, {\mathbb{F}})$ be a homomorphism prolonging $\varphi$. Let $\ell = \{ y = \bar{a} x + \bar{b} \}$ be the line in ${\mathbb{P}}^2_{{\mathbb{F}}_{q^{\nu}}}$ corresponding to $\varphi$. Then from the explicit description of $K$ and $L$ we provided at the beginning of the section, it follows that $M := \Phi(S)$ is a splitting field of the polynomial $F_\ell =
F(x, \bar{a} x + \bar{b})$. By definition of the Frobenius element, we have the commutative diagram $$\xymatrix@C=1.5cm{
L = \frac{K[t_1, \dotsc, t_d]}{\bigl(
F(t_i, a t_i + b) \text{ for } i \in \{1, \dotsc, d \}
\bigr)} \ar[d] \ar[r]^{\left[ \frac{S / R}{\Phi} \right]} &
L = \frac{K[t_1, \dotsc, t_d]}{\bigl(
F(t_i, a t_i + b) \text{ for } i \in \{1, \dotsc, d \}
\bigr)} \ar[d] \\
M = \frac{K[u_1, \dotsc, u_d]}{\bigl(
F(u_i, \bar{a} u_i + \bar{b}) \text{ for } i \in \{1, \dotsc, d \}
\bigr)} \ar[r]^{\alpha \mapsto \alpha^{q^\nu}} &
M = \frac{K[u_1, \dotsc, u_d]}{\bigl(
F(u_i, \bar{a} u_i + \bar{b}) \text{ for } i \in \{1, \dotsc, d \}
\bigr)}
}$$ which is just the extension to $L$ of the diagram in Equation . From the commutativity of this diagram, we see that the permutation action of $\left[ \frac{S / R}{\Phi} \right]$ on the classes $[t_1], \dotsc, [t_d]$ is the same as the action of the map $\alpha
\mapsto \alpha^{q^\nu}$ on the classes $[u_1], \dotsc, [u_d]$. Since the $\{
[u_i] \}$ are the roots of $F(x, \bar{a} x + \bar{b})$, and the latter is a polynomial with coefficients in ${\mathbb{F}}_{q^\nu}$, which are hence preserved by the map $\alpha \mapsto \alpha^{q^\nu}$, it follows that the structure of factors of $F_{\ell}$ over ${\mathbb{F}}_{q^{nu}}$ is the same as the cycle structure of $\left[
\frac{S / R}{\Phi} \right]$. This concludes the proof.
As a corollary, we obtain that the set of lines in ${\mathbb{P}}^2$ over ${\mathbb{F}}_{q^\nu}$ intersecting $C$ in exactly $k$ points corresponds to the set $$\bigcup_{Z \text{ has exactly } k \text{ fixed points}} I_{\nu, Z} \,.$$ The cardinality of this set is given by the so-called *rencontres numbers*, see [@Riordan2002]. We have hence: $$\left| \bigcup_{Z \text{ has exactly } k \text{ fixed points}} I_{\nu, Z}
\right| = d! \sum_s=k^d \frac{(-1)^{k+s}}{s!} \binom{s}{k} \,.$$ Using Chebotarev theorem we then conclude the proof of Theorem \[thm:formula\].
As the reader can see, there is nothing particularly special in considering the setting of plane curves. In fact, the concept of simple tangency (see \[def:simple\_tangency\]) is applicable to curves in arbitrary projective space: an absolutely irreducible curve $C$ in ${\mathbb{P}}^n(\overline{{\mathbb{F}}}_q)$ for $n
\in {\mathbb{N}}$ has simple tangency if there exists a hyperplane $H \subseteq
{\mathbb{P}}^n(\overline{{\mathbb{F}}}_q)$ intersecting $C$ in $d-1$ smooth points of $C$ such that $H$ intersects $C$ transversely at $d-2$ points and has intersection multiplicity $2$ at the remaining point. Also the concepts of Galois group of a curve and probabilities of intersections generalize similarly by considering hyperplanes instead of lines.
The generalized statement for the situation of irreducible curves is the following.
\[prop:space\_formulas\] Let $C$ be an absolutely irreducible algebraic curve of degree $d$ in ${\mathbb{P}}^n$, $n\in{\mathbb{N}}$, over ${\mathbb{F}}_q$, where $q$ is a prime power. Suppose that $C$ has simple tangency. Then for every $k \in \{0, \dotsc, d\}$ we have $$p_k(C) = \sum_{s=k}^d \frac{(-1)^{k+s}}{s!} \binom{s}{k}.$$ In particular, $p_{d-1}(C) = 0$ and $p_d(C) = 1 / d!$.
Using Proposition \[prop:space\_formulas\], we can compute the probabilities of intersection of a given plane curve $C$ with a random plane curve $E$ of degree $e$. In fact, via the *Veronese map* we can reduce this situation to the one of Proposition \[prop:space\_formulas\]. Let us start by defining the probabilities of intersection of a given curve $C$ with a random curve $E$ in the plane in exactly $k$ points, for $k \in \{ 0, \dotsc, de\}$: $$\begin{gathered}
p_k^N(C, e) :=
\frac{\Bigl|
\bigl\{
\text{curves } E \subseteq {\mathbb{P}}^2 ({\mathbb{F}}_{q^N}) \text{ of degree } e \, : \,
|E({\mathbb{F}}_{q^N}) \cap C({\mathbb{F}}_{q^N}) | = k
\bigr\}
\Bigr|}{q^{\binom{e+2}{2}N} + \dotsb + q^{2N}+q^N+1} \,, \\
p_k(C,e) := \lim_{N \to \infty} p_k^N(C, e) \quad \text{when the limit exists} \,.\end{gathered}$$ Recall now that for every $r \in {\mathbb{N}}$, the Veronese map of degree $e$ is an algebraic morphism embedding ${\mathbb{P}}^r$ into a larger projective space, so that hypersurfaces of degree $e$ get mapped to hyperplane sections of the image of the map. In this sense, the Veronese map operates a sort of “linearization” of the problem. In the case of ${\mathbb{P}}^2$, which is the one that interests us, it is given by $$\begin{array}{lccc}
v_e \colon &
{\mathbb{P}}^2 & \longrightarrow & {\mathbb{P}}^{\binom{e+2}{2}-1} \\
& (x:y:z) & \mapsto & \bigl( \{ x^a y^b z^c \}_{a+b+c = e} \bigr)
\end{array} \;.$$ The following lemma ensures that if we start from a plane curve that has simple tangency and we apply the Veronese map, we obtain a curve that has simple tangency.
Let $C$ be a plane curve of degree $d$ with simple tangency and let $e\in {\mathbb{N}}$. Then the image ${\widetilde}{C}=v_{e}(C)$ of $C$ under the Veronese map of degree $e$ has also simple tangency.
Let $\ell_1$ be a line witnessing simple tangency for $C$. Select lines $\ell_2, \dots ,\ell_e$ in ${\mathbb{P}}^2$ such that each of them intersects $C$ in $d$ distinct points and $\ell_i \cap \ell_j \cap C$ is empty for all $i \not= j$. Define $E$ as the zero set of the product $\ell_1 \cdots \ell_e$. The Veronese map sends $E$ to a hyperplane section of the Veronese surface; let ${\widetilde}{H}$ be the corresponding hyperplane. Then, by construction, ${\widetilde}{H}$ witnesses simple tangency for ${\widetilde}{C}$.
Since the Veronese map of degree $e$ defines a bijection between plane curves of degree $e$ and hyperplanes in ${\mathbb{P}}^{\binom{e+2}{2}-1}$, determining the probabilities $p_k(C,e)$ of intersection of a given plane absolutely irreducible curve $C$ with a random curve of degree $e$ in $k$ points is equivalent to compute the corresponding probabilities of intersection of the image ${\widetilde}{C}$ of $C$ under the Veronese map with hyperplanes. We sum up what we obtained in the following:
\[prop:higher\_formulas\] Let $C$ be an absolutely irreducible algebraic curve of degree $d$ in ${\mathbb{P}}^2$ over ${\mathbb{F}}_q$, where $q$ is a prime power. Suppose that $C$ has simple tangency. Let $e \in {\mathbb{N}}$ be a natural number. Then for every $k \in \{0, \dotsc, de\}$ we have $$p_k(C,e) = \sum_{s=k}^{de} \frac{(-1)^{k+s}}{s!} \binom{s}{k}.$$
Acknowledgements
================
Mehdi Makhul, Josef Schicho and Matteo Gallet are supported by the Austrian Science Fund (FWF): W1214-N15, Project DK9. Matteo Gallet has been supported by the Austrian Science Fund (FWF): P26607 and P25652 and P31061.
[EBW[[$^{+}$]{}]{}43]{}
Julio C. Andrade, Lior Bary-Soroker, and Zeev Rudnick, *Shifted convolution and the [T]{}itchmarsh divisor problem over [$\mathbb{F}_q[t]$]{}*, Philos. Trans. Roy. Soc. A **373** (2015), no. 2040.
Lior Bary-Soroker and Moshe Jarden, *On the [B]{}ateman-[H]{}orn conjecture about polynomial rings*, Münster J. Math. **5** (2012), 41–57.
Andreas O. Bender and Olivier Wittenberg, *A potential analogue of [S]{}chinzel’s hypothesis for polynomials with coefficients in [$\mathbb{F}_q[t]$]{}*, Int. Math. Res. Not. (2005), no. 36, 2237–2248.
Antoine [Chambert-Loir]{}, *[A field guide to algebra.]{}*, Springer, New York, 2005.
Claus Diem, *On the discrete logarithm problem for plane curves*, J. Théor. Nombres Bordeaux **24** (2012), no. 3, 639–667.
Zeev Dvir, Swastik Kopparty, Shubhangi Saraf, and Madhu Sudan, *Extensions to the method of multiplicities, with applications to [K]{}akeya sets and mergers*, SIAM J. Comput. **42** (2013), no. 6, 2305–2328.
Zeev Dvir, *On the size of [K]{}akeya sets in finite fields*, J. Amer. Math. Soc. **22** (2009), no. 4, 1093–1097.
Alexei Entin, *Monodromy of Hyperplane Sections of Curves and Decomposition Statistics over Finite Fields*, available at <https://arxiv.org/abs/1805.05454>.
Paul Erd[ö]{}s, Richard Bellman, Hubert S. Wall, James Singer, and Victor Th[é]{}bault, *Problem 4065*, Amer. Math. Monthly **50** (1943), 65–66.
Tibor Gallai, *Solution to problem 4065*, Amer. Math. Monthly **51** (1944), 169–171.
Larry Guth and Nets H. Katz, *On the [E]{}rd[ö]{}s distinct distances problem in the plane*, Ann. of Math. (2) **181** (2015), no. 1, 155–190.
Ben Green and Terence Tao, *On sets defining few ordinary lines*, Discrete Comput. Geom. **50** (2013), no. 2, 409–468.
Ulrich Görtz and Torsten Wedhorn, *Algebraic geometry [I]{}*, Advanced Lectures in Mathematics, Vieweg + Teubner, 2010.
Robin Hartshorne, *Algebraic geometry*, Graduate Texts in Mathematics, no. 52, Springer-Verlag, New York-Heidelberg, 1977.
Vijaya Kumar Murty and John Scherk, *Effective versions of the [C]{}hebotarev density theorem for function fields*, C. R. Acad. Sci. Paris Sér. I Math. **319** (1994), no. 6, 523–528.
Serge Lang and André Weil, *Number of points of varieties in finite fields*, Amer. J. Math. **76** (1954), 819–827.
Eberhard Melchior, *[Ü]{}ber [V]{}ielseite der projektiven [E]{}bene*, Deutsche Math. **5** (1941), 461–475.
David Mumford, *The red book of varieties and schemes*, Lecture Notes in Mathematics, vol. 1358, Springer-Verlag, Berlin, 1999.
Jürgen Rathmann, *The uniform position principle for curves in characteristic [$p$]{}*, Math. Ann. **276** (1987), no. 4, 565–579.
John Riordan, *An introduction to combinatorial analysis*, Dover Publications, Inc.,, 2002, Reprint of the 1958 original \[Wiley, New York; MR0096594 (20 \#3077)\].
Louis Rowen, *Graduate [A]{}lgebra: [C]{}ommutative [V]{}iew*, American Mathematical Society, Providence, RI, 2006.
Igor R. Shafarevich, *Basic algebraic geometry 1. [V]{}arieties in projective space*, third ed., Springer, Heidelberg, 2013.
Shubhangi Saraf and Madhu Sudan, *An improved lower bound on the size of [K]{}akeya sets over finite fields*, Anal. PDE **1** (2008), no. 3, 375–379.
József Solymosi and Miloš Stojaković, *Many collinear [$k$]{}-tuples with no [$k+1$]{} collinear points*, Discrete Comput. Geom. **50** (2013), no. 3, 811–820.
The [Stacks Project Authors]{}, *[Stacks Project]{}*, <http://stacks.math.columbia.edu>, 2017.
James J. Sylvester, *[Mathematical Question 11851]{}*, Educational Times **59** (1893), 385–394.
Terence Tao, *Algebraic combinatorial geometry: the polynomial method in arithmetic combinatorics, incidence combinatorics, and number theory*, EMS Surv. Math. Sci. **1** (2014), no. 1, 1–46.
Ravi Vakil, *Schubert induction*, Ann. of Math. (2) **164** (2006), no. 2, 489–512.
[^1]: We thank two anonymous referees for pointing us to the relevant literature.
[^2]: [ https://terrytao.wordpress.com/2012/08/31/the-lang-weil-bound/](
https://terrytao.wordpress.com/2012/08/31/the-lang-weil-bound/)
[^3]: See also the answer by Sandor Kovács at [
https://mathoverflow.net/questions/86221/fibre-cardinality-of-an-unramified-morp
hism](
https://mathoverflow.net/questions/86221/fibre-cardinality-of-an-unramified-morp
hism).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Here we prove that the homological dimension of the category of sheaves on a topological space satisfying some suitable conditions is finite. In particular, we find conditions to bound the homological dimension of o-minimal and subanalytic sheaves.'
title: '**[ON THE HOMOLOGICAL DIMENSION OF O-MINIMAL AND SUBANALYTIC SHEAVES]{}**'
---
Introduction
============
In [@lucap] we studied the category ${\mathrm{Mod}}(k_{X_{sa}})$ of sheaves on the subanalytic site $X_{sa}$ associated to a real analytic manifold $X$. We defined the subcategory of quasi-injective sheaves (i.e. $F$ is quasi-injective if the restriction $\Gamma(U;F)\to \Gamma(V;F)$ is surjective for each $U \supseteq V$ open subanalytic relatively compact) and we saw that quasi-injective are injective with respect to the functors of direct image, proper direct image and ${\mathcal{H}\mathit{om}}(G,\cdot)$ when $G$ is a ${\mathbb{R}}$-constructible sheaf on $X$. Moreover we proved that the quasi-injective dimension of ${\mathrm{Mod}}(k_{X_{sa}})$ is finite, which implies that the cohomological dimension of the above functors is finite. However we had no answer concerning the homological dimension of ${\mathrm{Mod}}(k_{X_{sa}})$ (see Remark 2.3.5 of [@lucap]). The aim of this paper is to show that under some conditions (concerning the cardinality $2^{\aleph_0}$ of the reals) it is possible to bound this dimension. The key point is the fact that locally the category of subanalytic sheaves is equivalent to the category of sheaves on a topological space $\widetilde{X}$. We are reduced to work with $k$-sheaves on a topological space, and in this case the homological dimension is equal to the flabby dimension. Hence we are reduced to bound the cohomological dimension of $\Gamma(U;\cdot)$, for any open subset $U$ of $\widetilde{X}$. In order to do that we need to assume that the cardinality of subanalytic subsets (which is equal to $2^{\aleph_0}$) is smaller than $\aleph_k$, $k<\infty$.\
In more details the contents of this paper are as follows.
In $\S$\[0\] we recall some notions as the definitions of injective and flabby sheaf and the homological and flabby dimension of the category of sheaves. We refer to [@ks1] for more details.
In $\S$\[1\] we study the general case of sheaves on a topological space $X$ with a basis ${\mathcal{T}}$ whose elements are Lindelöf and stable under finite unions and intersections. We define the subcategory of ${\mathcal{T}}$-flabby sheaves on $X$ by saying that $F$ is ${\mathcal{T}}$-flabby if the restriction $\Gamma(U;F)\to \Gamma(V;F)$ is surjective for each $U,V\in{\mathcal{T}}$ with $U \supseteq V$. Since ${\mathcal{T}}$ forms a basis of the topology of $X$, we have that each open $U$ of $X$ has a covering $\{U_i\}_{i\in I}$ with $U_i\in{\mathcal{T}}$. Hence $$R\Gamma(U;F) \simeq R{\underset{i\in I}{\underleftarrow{\lim}}}R\Gamma(U_i;F),$$ and to bound the cohomological dimension of the functor $\Gamma(U;\cdot)$ is sufficient to bound the cohomological dimension of the functors $\Gamma(V;\cdot)$, $V \in {\mathcal{T}}$ and the cohomological dimension of the projective limit. The cohomological dimension of the projective limit is bounded if the cardinality of the index set $I$ is smaller than $\aleph_k$ for some $k<\infty$, (see [@jen] or [@fp] in the more general setting of quasi-abelian categories). Hence it is bounded if the cardinality of ${\mathcal{T}}$ is smaller than $\aleph_k$, $k<\infty$.
In $\S$\[2\] we consider an o-minimal structure ${\cal M}=(M,<,\ldots)$. Let $\widetilde{X}$ be the o-minimal spectrum of a definable space $X$. In this case ${\mathcal{T}}$ is the family of open constructible subsets of $\widetilde{X}$. In the case of an o-minimal expansion of an ordered group, by a result of [@ejp], the ${\mathcal{T}}$-flabby dimension of $\widetilde{X}$ is finite. Moreover the cardinality of ${\mathcal{T}}$ is bounded by the product of the cardinality of $M$ and the cardinality of the language of the structure ${\cal M}$.
The case of subanalytic sheaves on a real analytic manifold is studied in $\S$\[3\]. We first reduce to ${\mathrm{Mod}}(k_{U_{X_{sa}}})$ where $U$ is a relatively compact subanalytic subset of $X$ isomorphic to ${\mathbb{R}}^N$ endowed with the Grothendieck topology induced by $X$. This category is equivalent to ${\mathrm{Mod}}(k_{\widetilde{U}})$, where $\widetilde{U}$ is the o-minimal spectrum of $U$. In this case ${\mathcal{T}}$ is the family of open globally subanalytic subsets of $U$, and its cardinality is $2^{\aleph_0}$. Hence if we assume that $2^{\aleph_0}$ is smaller or equal than $\aleph_k$, $k<\infty$ (in the case $k=1$ this is the continuum hypothesis), then the homological dimension of ${\mathrm{Mod}}(k_{\widetilde{U}})$ (and hence of ${\mathrm{Mod}}(k_{U_{X_{sa}}})$) is bounded.
We end this work with an example showing that in $\widetilde{X}$ there are open subsets which do not admit countable covers.
Some preliminaries {#0}
==================
We introduce some fundamental results about sheaves we will use in the rest of the paper. We refer to [@ks1] for more details. Let $X$ be a topological space and let $k$ be a field. As usual, we denote by ${\mathrm{Mod}}(k_X)$ the category of sheaves of $k$-vector spaces.
Let $F \in {\mathrm{Mod}}(k_X)$.
- $F$ is injective if the functor ${\mathrm{Hom}}(\cdot,F)$ is exact on ${\mathrm{Mod}}(k_X)$.
- $F$ is flabby if for any open subset $U$ of $X$ the restriction morphism $\Gamma(X;F) \to \Gamma(U;F)$ is surjective.
In general injective $\Rightarrow$ flabby. When $k$ is a field we have injective $\Leftrightarrow$ flabby ([@ks1], Exercise II.10).
The homological (resp. flabby) dimension of the category ${\mathrm{Mod}}(k_X)$ is the smallest $N \in {\mathbb{N}}\cup \{\infty\}$ such that for any $F \in
{\mathrm{Mod}}(k_X)$ there exists an exact sequence $$0 \to F \to {\mathrm{I}}^0 \to \cdots \to I^N \to 0$$ with $I^j$ injective (resp. flabby) for $0 \leq j \leq N$.
We shall need the following results.
- The homological dimension of ${\mathrm{Mod}}(k_X)$ is finite if and only if there exists $N \in {\mathbb{N}}$ such that $R^j{\mathrm{Hom}}(G,F)=0$ for any $F,G \in {\mathrm{Mod}}(k_X)$ and any $j > N$.
- The flabby dimension of ${\mathrm{Mod}}(k_X)$ is finite if and only if there exists $N \in {\mathbb{N}}$ such that for any open subsets $U$ of $X$, any $F \in {\mathrm{Mod}}(k_X)$ and any $j > N$ we have $R^j\Gamma(U;F)=0$ (i.e. if the functor $\Gamma(U;\cdot)$ has finite cohomological dimension).
In particular, when $k$ is a field, since injective $\Leftrightarrow$ flabby, the homological dimension is finite if and only if the functor $\Gamma(U;\cdot)$ has finite cohomological dimension for any open subsets $U$ of $X$.
${\mathcal{T}}$-flabby sheaves and homological dimension {#1}
========================================================
A Lindelöf space is a topological space in which every open cover has a countable subcover.
The Lindelöf degree $l(X)$ of a topological space $X$ is the smallest cardinal $\kappa$ such that every open cover of the space X has a subcover of size at most $\kappa$. In this notation, X is Lindelöf iff $l(X) = \aleph_0$.
Let $k$ be a field. Let $X$ be a topological space and suppose that $X$ admits a family ${\mathcal{T}}$ of open subsets such that
- each element of ${\mathcal{T}}$ is Lindelöf,
- ${\mathcal{T}}$ is stable under finite unions and intersections,
- ${\mathcal{T}}$ forms a basis for the topology of $X$.
A sheaf $F \in {\mathrm{Mod}}(k_X)$ is said to be ${\mathcal{T}}$-flabby if the restriction morphism $\Gamma(U;F) \to \Gamma(V;F)$ is surjective for each $U,V \in {\mathcal{T}}$ with $V \subseteq U$.
Remark that flabby $\Rightarrow$ ${\mathcal{T}}$-flabby.
The ${\mathcal{T}}$-flabby dimension of ${\mathrm{Mod}}(k_X)$ is the smallest $N \in {\mathbb{N}}\cup \infty$ such that for any $F \in
{\mathrm{Mod}}(k_X)$ there exists an exact sequence $$0 \to F \to {\mathrm{I}}^0 \to \cdots \to I^N \to 0$$ with $I^j$ ${\mathcal{T}}$-flabby for $0 \leq j \leq N$.
\[prop:soft\_section\_exact\] Let ${0 \to {F'} \to {F} \to {F''} \to 0}$ be an exact sequence in ${\mathrm{Mod}}(k_X)$ with $F'$ ${\mathcal{T}}$-flabby. Then for any open subset $U$ which is Lindelöf the morphism $$\Gamma(U;F) \to \Gamma(U;F'')$$ is surjective.
[**Proof.**]{} We first consider a section $s''\in \Gamma(U;F'')$. Since $F \to F''$ is surjective, we may find a covering $U = \bigcup_{i\in{\mathrm{I}}} U_i$, $U_i \in {\mathcal{T}}$ and $s_i \in
\Gamma(U_i;F)$ whose image is $s''|_{U_i}$. Since $U$ is Lindelöf, we may find a countable subcover $U = \bigcup_{n\in{\mathbb{N}}} U_n$, $U_n \in {\mathcal{T}}$ and $s_n \in
\Gamma(U_n;F)$ whose image is $s''|_{U_n}$.
Set $V_n = \bigcup_{i=1}^nU_i$. We prove by induction on $n$ that there exists a section $t_{n+1} \in \Gamma(V_{n+1};F)$ whose image is $s''|_{V_{n+1}}$ and $t_{n+1} |_{V_n}= t_n$.
This is clear for $n=0,1$ and we assume it is proved for $n$. By the induction hypothesis there exists a section $t_n \in \Gamma(V_n;F)$ whose image is $s''|_{V_n}$ and $t_n |_{V_m}
= t_m$ if $m<n$. We set for short $V_n=W_1$ and $U_{n+1}=W_2$. We have seen that there exist $t_j \in
\Gamma(W_j;F)$ whose image is $s''|_{W_j}$ for $j=1,2$. On $W_1
\cap W_2$ $t_1-t_2$ defines a section of $\Gamma(W_1 \cap W_2;F')$ which extends to $t' \in \Gamma(V_{n+1};F')$ because $F'$ is ${\mathcal{T}}$-flabby. Replace $t_2$ with $t_2+t'$. We may suppose that $t_1=t_2$ on $W_1 \cap W_2$. Then there exists $t_{n+1} \in \Gamma(V_{n+1};F)$ such that $t_{n+1}|_{W_j}=t_j$, $j=1,2$. The $t_n$’s glue together into a section $$s \in {\underset{n\in{\mathbb{N}}}{\underleftarrow{\lim}}}\Gamma(V_n;F) \simeq \Gamma(U;F)$$ which is sent to $s''$, which proves the surjectivity of the morphism. [ $\Box$\
]{}
\[www\] Let $F',F,F'' \in {\mathrm{Mod}}(k_X)$, and consider the exact sequence $${0 \to {F'} \to {F} \to {F''} \to 0}.$$ Suppose that $F'$ is ${\mathcal{T}}$-flabby. Then $F$ is ${\mathcal{T}}$-flabby if and only if $F''$ is ${\mathcal{T}}$-flabby.
[**Proof.**]{} Let $U,V\in {\mathcal{T}}$ with $V \subseteq U$ and let us consider the diagram below $$\xymatrix{0 \ar[r] & \Gamma(U;F') \ar[d]^\alpha \ar[r] & \Gamma(U;F) \ar[d]^\beta \ar[r] & \Gamma(U;F'') \ar[d]^\gamma \ar[r] & 0\\
0 \ar[r] & \Gamma(V;F') \ar[r] & \Gamma(V;F) \ar[r] & \Gamma(V;F'') \ar[r] & 0}$$ where the row are exact by Proposition \[prop:soft\_section\_exact\] and the morphism $\alpha$ is surjective since $F'$ is ${\mathcal{T}}$-flabby. It follows from the five lemma that $\beta$ is surjective if and only if $\gamma$ is surjective. [ $\Box$\
]{}
\[TUinj\] Let $F$ be ${\mathcal{T}}$-flabby. Then $F$ is $\Gamma(U;\cdot)$-injective for each open $U\subseteq X$ which is Lindelöf.
[**Proof.**]{} The family of ${\mathcal{T}}$-flabby sheaves contains injective sheaves, hence it is cogenerating. Then the result follows from Propositions \[prop:soft\_section\_exact\] and \[www\]. [ $\Box$\
]{}
Now let us consider the case of open subsets of Lindelöf degree $\aleph_k$, $k<\infty$. In order to do that we need the following result.
\[lprodim\] Let $X$ be a topological space admitting a family ${\mathcal{T}}$ satisfying T1-T3. Suppose that $X$ has finite ${\mathcal{T}}$-flabby dimension. Let $U$ be an open subset of $X$ of Lindelöf degree $\aleph_k$, $k<\infty$. Then the cohomological dimension of $\Gamma(U;\cdot)$ is finite.
[**Proof.**]{} We have $$R\Gamma(U ; F) \simeq R{\underset{i \in I}{\underleftarrow{\lim}}}R\Gamma(U_i; F)$$ where $U=\bigcup_{i\in I}U_i$ with $U_i \in {\mathcal{T}}$ and $\sharp I$ smaller than $\aleph_k$ . Since $X$ has finite ${\mathcal{T}}$-flabby dimension and using Proposition \[TUinj\] for each $i$ we may replace $R\Gamma(U_i;F)$ with $\Gamma(U_i;I^\bullet)$, where $I^\bullet$ is a ${\mathcal{T}}$-flabby resolution of $F$ of length $N<\infty$. Since cohomological dimension of ${\underset{i \in I}{\underleftarrow{\lim}}}$ is finite if $\sharp I$ is smaller than $\aleph_k$, $k<\infty$ (see [@jen; @fp]), then the $j$-th cohomology of $R{\underset{i \in I}{\underleftarrow{\lim}}}R\Gamma(U_i; F)$ vanishes for $j>N+M$. Since $M,N<\infty$ are independent of $F$ and $i$, the result follows. [ $\Box$\
]{}
\[homdim\] Let $X$ be a topological space admitting a family ${\mathcal{T}}$ satisfying T1-T3. Suppose that $X$ has finite ${\mathcal{T}}$-flabby dimension. Suppose that there exists $k<\infty$ such that every open subset of $X$ has Lindelöf degree $\aleph_k$. Then ${\mathrm{Mod}}(k_X)$ has finite homological dimension.
[**Proof.**]{} By Proposition \[lprodim\] $X$ has finite flabby dimension. In the case of sheaves of $k$-vector spaces we have flabby $\Leftrightarrow$ injective and the result follows. [ $\Box$\
]{}
Homological dimension of o-minimal sheaves {#2}
==========================================
Let ${\cal M}=(M,<,\ldots)$ be an o-minimal structure. In [@ejp] the authors studied sheaf cohomology of sheaves on $\widetilde{X}$, the o-minimal spectrum of a definable space $X$. The category of sheaves on $\widetilde{X}$ is equivalent to the category of sheaves on the o-minimal site, consisting of open definable subsets of $X$ and coverings admitting a finite refinement.
Let ${\mathcal{T}}$ be the family of open constructible subsets of $\widetilde{X}$. Then ${\mathcal{T}}$ satisfies T1-T3, indeed
- ${\mathcal{T}}$ forms a basis for the topology of $\widetilde{X}$,
- every element of ${\mathcal{T}}$ is quasi-compact,
- ${\mathcal{T}}$ is stable under finite unions and intersections.
For each open subset $U$ of $\widetilde{X}$ and each $F \in {\mathrm{Mod}}(k_{\widetilde{X}})$ we have $$R\Gamma(U;F) \simeq R{\underset{i\in I}{\underleftarrow{\lim}}}R\Gamma(U_i;F)$$ where $U=\bigcup_{i \in I}U_i$ and $U_i \in {\mathcal{T}}$ for each $i \in I$. In order to apply Theorem \[homdim\] we need that:
- the ${\mathcal{T}}$-flabby dimension of $\widetilde{X}$ is finite,
- $\sharp I \leq \aleph_k$, $k<\infty$.
If every open constructible subset $U$ of $X$ is normal (this is true in the case of an o-minimal expansion of an ordered group), then by Proposition 4.2 of [@ejp] we have $R^j\Gamma(U;F)=0$ for all $j>{\rm dim}U={\rm dim}X$.
Concerning $\sharp I$, we have that $\sharp I \leq \sharp {\mathcal{T}}$, and $\sharp {\mathcal{T}}\leq \sharp {\cal L} \cup M=\sharp {\cal L} \cdot \sharp M$, where ${\cal L}$ denotes the language of the structure ${\cal M}$.
Homological dimension of subanalytic sheaves {#3}
============================================
Let $X$ be a real analytic manifold, let $X_{sa}$ the associated subanalytic site and denote by ${\mathrm{Mod}}(k_{X_{sa}})$ the category of sheaves of $k$-vector spaces on $X_{sa}$. We refer to [@ks2; @lucap] for the theory of subanalytic sheaves. We will see that under suitable hypothesis the homological dimension of ${\mathrm{Mod}}(k_{X_{sa}})$ is finite, i.e. there exists $N \in {\mathbb{N}}$ such that $R^j{\mathrm{Hom}}(G,F)=0$ for any $F,G \in {\mathrm{Mod}}(k_{X_{sa}})$ and any $j > N$.
Let $X$ be a real analytic manifold. Then $X$ has an atlas $(U_n,\phi_n)_{n\in{\mathbb{N}}}$, $\phi_n:U_n {\stackrel{\sim}{\to}}{\mathbb{R}}^M$, $U_n$ relatively compact open subanalytic subset. Let us consider the subanalytic site $X_{sa}$ associated to $X$. In order to bound ${\mathrm{RHom}}(F,G)$ for any $F,G \in {\mathrm{Mod}}(k_{X_{sa}})$ it suffices to bound ${\mathit{R}\mathcal{H}\mathit{om}}(F,G)$. Indeed we have ${\mathrm{RHom}}(F,G) \simeq R\Gamma(X;{\mathit{R}\mathcal{H}\mathit{om}}(F,G))$ and $\Gamma(X;\cdot)$ has finite cohomological dimension (see [@lucap]).
We have $R^j{\mathcal{H}\mathit{om}}(F,G)=0$ if $R^j{\mathcal{H}\mathit{om}}(F,G)|_{U_n}=0$ for each $n\in{\mathbb{N}}$. Then we may reduce to the case $U_n$ with the Grothendieck topology induced by the subanalytic site, i.e. the category of globally subanalytic open subsets of ${\mathbb{R}}^M$ with coverings admitting a finite subcover. Globally subanalytic geometry is defined by the o-minimal structure given by the ordered field of real numbers expanded by globally analytic functions ${\mathbb R}_{{\rm an}}=({\mathbb R}, <, 0,1, +, \cdot, (f)_{f\in {\rm an}})$, where $f$ is a restriction to $[-1,1]^m$ of a convergent power series on some neighborhood of $[-1,1]^m$.
Set $X=U_n$ and let $\widetilde{X}$ be the o-minimal spectrum of $X$. The family ${\mathcal{T}}$ of open sets $\widetilde{U}$ where $U$ is open globally subanalytic satisfies T1-T3. Moreover the ${\mathcal{T}}$-flabby dimension of ${\mathrm{Mod}}(k_{\widetilde{X}} )$ is finite. Indeed by Proposition 4.2 of [@ejp] the cohomological dimension of $\Gamma(\widetilde{U};\cdot)$ is finite for each $\widetilde{U} \in {\mathcal{T}}$. Let us see that $\sharp{\mathcal{T}}\leq 2^{\aleph_0}$. In order to see that we have to show that $\sharp {\cal L} = 2^{\aleph_0}$, where ${\cal L}$ denotes the language of ${\mathbb{R}}_{{\rm an}}$. This is true if $\sharp\{(f)_{f\in{\rm an}}\}=2^{\aleph_0}$. This is seen by identifying $f=\sum_{I \in {\mathbb{N}}^m}a_I x^I$ with $(a_I)_{I \in {\mathbb{N}}^m} \in {\mathbb{R}}^{{\mathbb{N}}^m}$. Then $\sharp \{(f)_{f\in{\rm an}}\} \leq \sharp {\mathbb{R}}^{{\mathbb{N}}^m} = (2^{\aleph_0})^{\aleph_0} = 2^{\aleph_0 \cdot \aleph_0} = 2^{\aleph_0}$ (see [@jech]).
Let $U$ be an open subset of $\widetilde{X}$. Then $U =\bigcup_{i\in I} U_i$ with $U_i \in {\mathcal{T}}$ and $\sharp I$ is smaller than $\sharp {\mathcal{T}}= 2^{\aleph_0}$. Hence every open subset of $X$ has Lindelöf degree at most $2^{\aleph_0}$. If we assume that $2^{\aleph_0}$ is smaller than $\aleph_k$ for some $k<\infty$ (if we suppose $\leq \aleph_1$ it is the continuum hypothesis), then Theorem \[homdim\] implies that the homological dimension of ${\mathrm{Mod}}(k_{\widetilde{X}} )$ is finite.
\[lindim\] In $\widetilde{X}$ there are open subsets which are not Lindelöf. For example let us consider the open set (in $\widetilde{X}=\widetilde{{\mathbb{R}}^2}$ with respect to ${\mathbb{R}}_{\mathrm{an}}$) $$U=\bigcup_{r \in (0,1) \setminus {\mathbb{Q}}}\widetilde{V}_r,$$ where $$V_r=\{(x,y) \in {\mathbb{R}}^2\;;\;0<x<1,\;r<y<r+x\}.$$ We prove that $\{V_r\}_{r \in (0,1)\setminus{\mathbb{Q}}}$ has no countable subcover.\
We argue by contradiction. Suppose that there exists a countable subcover $\{\widetilde{V}_{r_n}\}_{n\in{\mathbb{N}}}$ of $U$ in $\widetilde{X}$. Let $0^+$ be the ultrafilter defined by $0^+=\{S \supset (0,\varepsilon)\}$, where $\varepsilon>0$ and $S$ is globally subanalytic. Let $\pi:{\mathbb{R}}^2 \to {\mathbb{R}}$ be the projection onto the first coordinate. We have $$U \cap {\widetilde{\pi}^{-1}}(0^+) = \bigsqcup_{r\in (0,1)\setminus {\mathbb{Q}}} \left( \widetilde{V}_r \cap {\widetilde{\pi}^{-1}}(0^+) \right).$$ Indeed, let $x \in \widetilde{V}_r \cap {\widetilde{\pi}^{-1}}(0^+)$, and let $s \neq r$. Let $\varepsilon < |r-s|$, then $V_s \cap {\pi^{-1}}((0,\varepsilon)) \cap V_r = \emptyset$ and $\widetilde{V_s} \cap {\widetilde{\pi}^{-1}}(0^+) \subset \widetilde{V_s} \cap {\widetilde{\pi}^{-1}}(\widetilde{(0,\varepsilon)})$ since $0^+ \subset \widetilde{(0,\varepsilon)}$ for any $\varepsilon>0$. Hence $\widetilde{V}_r \cap {\widetilde{\pi}^{-1}}(0^+) \cap \widetilde{V}_s = \emptyset$ if $r \neq s$. Moreover $$\bigsqcup_{r\in (0,1)\setminus {\mathbb{Q}}} \left( \widetilde{V}_r \cap {\widetilde{\pi}^{-1}}(0^+) \right) \subset \bigcup_{n\in{\mathbb{N}}}\widetilde{V}_{r_n}=U.$$ Since $\sharp((0,1)\setminus{\mathbb{Q}})=2^{\aleph_0}>\aleph_0$, there exists $n\in{\mathbb{N}}$ such that $$\sharp\left\{t \in (0,1)\setminus{\mathbb{Q}}\;;\; \widetilde{V_t} \cap {\widetilde{\pi}^{-1}}(0^+) \subseteq \widetilde{V}_{r_n} \right\} \geq 2.$$ But $\widetilde{V_t} \cap {\widetilde{\pi}^{-1}}(0^+) \subseteq \widetilde{V}_{r_n}$ implies that $V_t=V_{r_n}$, so $t=r_n$ which is a contradiction.
More generally, as in Remark \[lindim\], given an o-minimal expansion of an ordered group $({\mathbb{R}},<,+,\ldots)$, for any definable manifold $X$ of dimension $\geq 2$, the space $\widetilde{X}$ has open subsets which have Lindelöf degree $2^{\aleph_0}$.
It seems that the right setting is that of locally definable spaces in an o-minimal structure. In that setting the o-minimal (definable) case and the subanalytic case would be treated uniformly. However, to work on that setting one has to take into account the (special case of the) theory locally semi-algebraic spaces developed in [@D3; @dk5].
[XX]{}
M. Coste [*An introduction to o-minimal geometry*]{} Dip. Mat. Univ. Pisa, Dottorato di Ricerca in Matematica, Istituti Editoriali e Poligrafici Internazionali, Pisa (2000).
H. Delfs [*Homology of locally semialgebraic spaces*]{} Lecture Notes in Math. $\bf{1484}$ Springer-Verlag (1991).
H. Delfs and M. Knebusch [*Locally semi-algebraic spaces*]{} Lecture Notes in Math. $\bf{1173}$ Springer-Verlag (1985).
M. Edmundo, G. Jones and N. Peatfield [*Sheaf cohomology in o-minimal structures*]{} J. Math. Logic **6** pp. 163-179 (2006).
T. Jech [*Introduction to set theory*]{} Dekker (1984).
M. Kashiwara and P. Shapira [*Sheaves on manifolds*]{} Grundlehren der Math. $\bf{292}$ Springer Verlag (1990).
M. Kashiwara and P. Shapira [*Ind-sheaves*]{} Astérisque **271** (2001).
C. U. Jensen [*Les foncteurs dérivés de ${\underleftarrow{\lim}}$ et leurs applications en théorie des modules*]{} Lecture Notes in Math. $\bf{254}$ Springer-Verlag (1972).
L. Prelli [*Sheaves on subanalytic sites*]{} Rend. Sem. Mat. Univ. Padova **120** pp. 167-216 (2008).
F. Prosmans [*Derived limits in quasi-abelian categories*]{} Bull. Soc. Roy. Sci. Liège **68** pp. 335-401 (1999).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Effective demagnetizing factors that connect the sample magnetic moment with the applied magnetic field are calculated numerically for perfectly diamagnetic samples of various non-ellipsoidal shapes. The procedure is based on calculating total magnetic moment by integrating the magnetic induction obtained from a full three dimensional solution of the Maxwell equations using adaptive mesh. The results are relevant for superconductors (and conductors in AC fields) when the London penetration depth (or the skin depth) is much smaller than the sample size. Simple but reasonably accurate approximate formulas are given for practical shapes including rectangular cuboids, finite cylinders in axial and transverse field as well as infinite rectangular and elliptical cross-section strips.'
author:
- 'R. Prozorov'
- 'V. G. Kogan'
date: 'Submitted: 16 December 2017; accepted in Phys. Rev. Applied: 26 June 2018'
title: |
Effective Demagnetizing Factors of\
Diamagnetic Samples of Various Shapes
---
Introduction
============
Correcting results of magnetic measurements for the distortion of the magnetic field inside and around a finite sample of arbitrary shape is not trivial, but necessary part of experimental studies in magnetism and superconductivity. The internal magnetic field is uniform only in ellipsoids (see Fig. \[fig2\]) for which demagnetizing factors can be calculated analytically [@Landau1984; @Osborn1945]. In general, however, magnetic field is highly non-uniform inside and outside of finite samples of arbitrary (non-ellipsoidal) shapes and various approaches were used to handle the problem [@Joseph1965; @Chen1991; @Aharoni1998; @Sato89; @Pardo2004; @Smith2010; @Brandt2001]. As discussed below, the major obstacle has been that so far the total magnetic moment of arbitrary shaped samples could not be calculated and approximations and assumptions had to be made. As a result, various approximate demagnetizing factors were introduced. For example, so-called “magnetometric" demagnetizing factor, $N_{m}$, is based on equating magnetostatic self-energy to the energy of a fully magnetized ferromagnetic prism or, more generally, considering volume-average magnetization in magnetized [@Chen1991; @Aharoni1998; @Sato89] or perfectly diamagnetic [@Pardo2004] media. Similarly, so-called “fluxmetric" or “ballistic" demagnetizing factor, $N_{f}$, is based on the average magnetization in the sample mid-plane [@Chen1991; @Pardo2004]. In these formulations, micromagnetic calculations are used to find the distribution of surface magnetic dipoles density that satisfies the boundary conditions and the assumptions made. Then the average magnetization is calculated and used to compute the $N$ factors using formulas similar to those used in this work. One common, but generally incorrect, assumption is that the sum of demagnetizing factors along three principal axes equals to one. This is true only for ellipsoids. Notably, E. H. Brandt has used a different approach by numerically calculating the slope, $dm/dH_0$, of the magnetic moment $m$ vs. applied magnetic field $H_0$ in a perfect superconductor in the Meissner state to compute the approximate $N-$factors for a 2D situation of infinitely long strips of rectangular cross-section in perpendicular field and he extended these results to finite 3D cylinders (also of rectangular cross-section) in the axial magnetic field. We find an excellent agreement between our calculations and Brandt’s results for these geometries. Also, in our earlier work, the 2D numerical solutions of the Maxwell equations obtained using finite element method were generalized to 3D cylinders and brought similar results [@Prozorov2000]).
Yet, despite multiple attempts, results published so far do not describe three dimensional finite samples of arbitrary non-ellipsoidal shapes to answer an important practical question: *What is the **total magnetic moment** of a three dimensional sample of a particular shape in a fixed **applied** magnetic field, $\bm H_{0}$?* We answer this question by finding a way to calculate total magnetic moment from the first principles with no assumptions and introducing the *effective* demagnetizing factors without referring to the details of the spatial distribution of the magnetic induction. We will first consider how these effective demagnetizing factors depend on finite magnetic permeability, $\mu_r$, which highlights the difference between ellipsoidal and non-ellipsoidal shapes. Complete treatment of finite $\mu_r $ requires separate papers in which we will focus on (a) the London-Meissner state in superconductors of arbitrary shape with finite London penetration depth and (b) demagnetizing corrections in local and linear magnetic media with arbitrary $\mu_r$.
In this work we focus on perfectly diamagnetic samples, the magnetic induction $B=0$ inside, which allows studying pure effects of sample shape. The results can be used for the interpretation of magnetic measurement of superconductors when London penetration depth is much smaller than the sample dimensions (a good approximation almost up to $0.95T_c$) or in conducting samples subject to AC magnetic field when the skin depth is small. Our goal is to find simple to use, but accurate enough, approximate formulas suitable for the calculations of the demagnetizing correction for many shapes that can approximate realistic samples, such as finite cylinders and cuboids (rectangular prisms).
Definitions
===========
In local and magnetically linear media without demagnetizing effects (infinite slab or cylinder in parallel magnetic field), $$\begin{aligned}
B&=&\mu H=\mu_{0}\left( M+H\right)\,,
\label{muL}\\
M&=&\frac{B}{\mu_{0}}-H=\chi H \label{mlinear}$$ where $\mu_0 = 4 \pi \times 10^{-7}$ \[N/A$^2$ or H/m\] is magnetic permeability of free space; $\mu $ and $\chi$ are linear magnetic permeability and susceptibility (in general these quantities are second rank tensors, but here we consider the isotropic case.)
It follows then that, $$\chi=\frac{\mu }{\mu_{0}}-1=\mu_r-1
\label{mu}$$ where $\mu_r=\mu /\mu_0$ is relative magnetic permeability; $\mu_r=1$ for non-magnetic media and $\mu_r=0$ and $\chi=-1$ for a perfect diamagnet.
For finite samples of ellipsoidal shape, *constant* demagnetizing factors $N$ connect the applied magnetic field $H_{0}$ along certain principle direction with the internal field, $H$,
$$H=H_{0}-N\,M \label{Hdemag}$$
and in terms of an applied field the magnetization is: $$M=\frac{\chi}{1+\chi N}H_{0} \label{mdemagchiL}$$
In arbitrary shaped samples this simple description breaks down and we have to introduce similarly structured effective equations, albeit *applicable only for integral quantities*. Namely, upon application of an external field $H_{0}$, a finite sample of a given shape develops a measurable total magnetic moment $m$. We now *define* “effective" (or “integral", or “apparent") magnetic susceptibility, $\chi_0$, and corresponding demagnetizing factor, $N$, by writing relations structurally similar to Eq. (\[mdemagchiL\]) $$m=\chi_0 H_{0} V = \frac{\chi H_0 V}{1+\chi N} \label{mom}$$ which reduces to conventional equations in the case of a linear magnetic material of ellipsoidal shape. Importantly, Eq. (\[mom\]) contains only one property to be determined - intrinsic susceptibility, $\chi$ provided the demagnetizing factor $N$ can be calculated for given geometry. This can be done for model materials with known (assumed) $\chi$ and numerically evaluated $\chi_0$ by inverting Eq. (\[mom\]) to obtain:
$$N=\frac{1}{\chi_0}-\frac{1}{\chi} \,. \label{N}$$
where $-1 \leq \chi \leq\infty$ and $-\infty\leq\chi_0\leq\chi$. To eliminate the influence of the material, for calculations of $N$ we will consider a perfect diamagnet with $\chi=-1$, so that when $\chi_0=\chi=-1$, $N=0$ as it should be in case of no demagnetizing effects (infinite slab or a cylinder in parallel field) and $N\rightarrow1$ for infinite plate in perpendicular field where $\chi_0\rightarrow -\infty$, while $\chi$ is still equals -1.
The main issue in using Eq. (\[N\]) to calculate demagnetizing factors is to calculate the total magnetic moment of a sample of a given (arbitrary) shape. There are two ways of approaching this. In non-magnetic (super)conductors, one first solves Maxwell equations with the help of one of existing numerical software packages (such as COMSOL, [@COMSOL]), to find the transport current density $\bm{j}(\bm{r})$. Then, the total magnetic moment is given by [@Landau1984], $$\bm{m}=\frac{1}{2}\int\left[ \bm{r}\times\bm{j(\bm{r})}\right] dV \label{MdefJ}$$ The integral here can be evaluated over the entire space, but the integrand is non-zero only inside the sample where the currents flow.
The second way to calculate the total magnetic moment, $\bm{m}$, is given by, $$\bm{m}=\alpha\int \left[ \frac{\bm{B}(\bm{r})}{\mu_{0}}-\bm{H}_{0}\right] d^3\bm{r}, \label{MdefB}$$ where $\bm{B}(\bm{r})$ is the actual field and $\bm{H}_0$ is the uniform applied field. Here the integral must be evaluated in a region that includes the sample (can be the entire space). We show in the next section that this integral is not unique, but depends on the way chosen for the integration. This is accounted for by a constant $\alpha$ in Eq. (\[MdefB\]), $\alpha=3/2$ for integration over the large spherical domain that includes the whole sample, whereas $\alpha=1$ for integration domain as a large cylinder with the axis parallel to $\bm H_0$. It turns out that for numerical reasons, the cylindrical domain is preferable and we used it for our numerical work. Equation (\[MdefB\]) is central to the present work, because it allows calculations without using the current distribution. This equation (with $\alpha=3/2$) can be found in Jackson’s textbook, Ref. \[\], Eq. (5.62). A related discussion about the multipole representation of the field outside the region where the field sources are localized is given in Ref. \[\].
For evaluation of $\bm B(\bm r)$, one can use approximation of a fully diamagnetic sample imposing “magnetic shielding" boundary conditions available in the COMSOL software. Employing Eq. (\[MdefB\]) with $\bm{B}(\bm{r})$ simplifies numerical procedure and improves accuracy considerably. However, proving Eq. (\[MdefB\]) is not at all trivial and we derive it analytically in the next section. We also verified the results by calculating total magnetic moment $\bm{m}$ utilizing both approaches evaluating current distribution in superconducting samples using London equations and employing Eq. (\[MdefJ\]), and using COMSOL generated field distribution $\bm{B}(\bm{r})$ and Eq. (\[MdefB\]).
Total magnetic moment $\bm m$
=============================
According to Jackson’s book [@Jackson2007], the magnetic moment $\bm m$ of current distribution induced by an applied uniform field $\bm H_0$ in a finite sample, is related to the distribution of the magnetic induction $\bm{B}(\bm{r})$ by $$\begin{aligned}
{ \bm I }=\int\limits_{\cal R}\left[\frac{{\bm B}(\bm r)}{\mu_0}-{\bm H}_0\right]d^3\bm r =\frac{2}{3}\,\bm{m}\,.
\label{eq0}\end{aligned}$$ where $\cal R$ is a radius of a large sphere containing the whole sample. In particular, ${\cal R}$ can be infinite, e.g. the integral can be extended to the whole space. This relation is central for our calculations, so that we provide a more general derivation than that given in [@Jackson2007]. We show that depending on the way chosen to evaluate the integral $\bm I$ over the whole space, Eq. (\[MdefB\]) can have different forms, parameterized by a factor $\alpha$.
The field $\bm B$ consists of the applied field $\bm{H}_0$ and the field $\bm h$ due to currents $\bm j$ in the sample of a finite volume $V$: $$\begin{aligned}
\frac{{\bm B}}{\mu_0}={\bm H}_0 +\bm h\,,
\label{eq1}\end{aligned}$$ i.e., $\bm I=\int {\bm h}\, d^3\bm r$, where according to Biot-Savart law, $$\begin{aligned}
{\bm h}(\bm r)=\frac{1}{4\pi}\int\limits_V d^3{\bm \rho}\,\frac{{\bm j}(\bm \rho)\times{\bm R}}{R^3}\,,\qquad \bm R=\bm r -\bm \rho\,.
\label{eq2}\end{aligned}$$ Hence, we have $$\begin{aligned}
4\pi\bm I&=& \int\limits_{\cal R} d^3 {\bm r} \int\limits_V d^3{\bm \rho}\frac{{\bm j}(\bm \rho)\times{\bm R}}{R^3}\nonumber\\
&=&
\int\limits_V d^3{\bm \rho} \,{\bm j}(\bm \rho)\times \int\limits_{\cal R} d^3\bm r \frac{{\bm R}}{R^3}\nonumber \\
&=&
\int\limits_V d^3{\bm \rho} \,{\bm j}(\bm \rho)\times{\cal E}(\bm \rho)\,.
\label{eq3a}\end{aligned}$$ where we introduce “pseudo-electric field", ${\cal E}(\bm \rho)= \int d^3\bm r {\bm R}/R^3$, which is analogous to the electrostatic field of a uniform charge distribution with a constant density of $-1$ in the whole space. For $\bm \rho=0$, we must have ${\cal E}=\int d^3\bm r ({\bm r}/r^3)=0$ by symmetry. For such a distribution, the field ${\cal E}$ is not defined uniquely, it *depends on the way one divides the space in charged elements*.
If one uses elements as spherical shells, and applies the Gauss theorem to a sphere of a radius $\bm\rho$ one obtains: $$\begin{aligned}
{\cal E}=-\frac{4\pi}{3}\,\bm\rho \,.
\label{eq8}\end{aligned}$$ Hence, we have $$\begin{aligned}
\bm I= -\frac{ 1}{3 }\int\limits_V d^3{\bm\rho} \,{\bm j}(\bm \rho)\times {\bm \rho}=\frac{2}{3}\, {\bm m} \,,
\label{eq9}\end{aligned}$$ where $\bm m$ is the total magnetic moment. It is worth noting that this formula holds for any current distribution within the finite sample of arbitrary shape.
If one uses integration elements as cylindrical shells parallel to $\bm H_0$, i.e. choose the volume element as $2\pi \rho_1\,d\rho_1dz$ ($\bm \rho_1$ is the cylindrical radius vector), and applies the Gauss theorem to a cylinder of a radius $\rho$ one obtains:
$$\begin{aligned}
{\cal E} =- 2\pi {\bm \rho}_1 \,.
\label{eq8a}\end{aligned}$$
Substituting this in Eq. (\[eq3a\]), one expresses the $z$ component of the integral $\bm I$: $$\begin{aligned}
I_z= {m}_z \,.
\label{eq9a}\end{aligned}$$
It is easy to show that the region where the integral [**I**]{} is evaluated can be taken as a sphere (or a cylinder) of a radius ${\cal R}_1$ that contains the entire sample of interest within this region. Then, if one takes a larger radius ${\cal R}_2$, the layer between spheres (cylinders) ${\cal R}_1$ and ${\cal R}_2$ does not contribute to the effective field ${\cal E}$ because the “electric field" of such a uniformly charged spherical (cylindrical) shell $ {\cal E} (r)=0$ for $r<{\cal R}_1$.
In Appendix A, for demonstration purposes, we evaluate the integral $\bm{I}$ both analytically and numerically for spherical and cylindrical integration volumes for the case of a current ring for which the distribution of $\bm{B}(\bm{r})$ is known.
It is worth mentioning that a similar argument can be applied for evaluation of the dipole moment of a metallic sample of arbitrary shape placed in a uniform electric field $\bm E_0$, $\bm d =\int d^3\bm \rho \,n(\bm \rho) \bm \rho$, ($n(\bm \rho)$ is the charge density, for point charges $\bm d =\sum_\nu e_\nu \bm \rho_\nu$). It is straightforward to see that $$\begin{aligned}
\bm d\propto \int d^3\bm \rho \,\left[\bm E (\bm \rho )- \bm E_0\right] \,.
\label{eq19a}\end{aligned}$$
Here, $\bm E (\bm \rho )$ is the electric field distribution which can be found numerically with the help of a software similar to COMSOL [@COMSOL]. As in the magnetic case, the integration here can be done over a spherical (or cylindrical) region which contains the whole sample. The coefficient of proportionality here is $2 \alpha/ \epsilon_0$, where $\alpha$ is given above and $\epsilon_0$ is the vacuum dielectric constant. This result might be useful in problems like those considered in [@polarization].
In Appendix B we provide a derivation of Eq.(\[MdefB\]) in Gaussian units for readers who prefer CGS in general electromagnetic problems.
Numerical calculations
======================
Numerical calculations of three dimensional distribution of vector $\bm{B}(\bm{r})$, were performed with COMSOL software [@COMSOL] using adaptive finite element solution of the Maxwell equations in the form applicable to many different situations, including external currents $\bm{j}_{ext}$ and electrical conductivity $\sigma$ in case of conducting materials.
$$\begin{aligned}
&\nabla \times \bm{H}=\bm{j},\\
&\bm{B}=\nabla \times \bm{A},\\
&\bm{j}=\sigma \bm{E} + \bm{j}_{ext},\\
&\bm{B}=\mu_0 \mu_r \bm{H},
\end{aligned}
\label{ampere}$$
We used finite $\mu_r$ below to illustrate $\mu_r$-dependent effective demagnetizing factor. Otherwise, throughout the manuscript, $\mu_r=0$, $\sigma=0$ and $\bm{j}_{ext}=0$. For details the reader is referred to extensive documentation available on COMSOL web site [@COMSOL]. “AC/DC magnetic field" COMSOL module in a stationary DC study was used to model perfect diamagnetic material. Frequency-dependent AC study was used to formulate London equations with complex frequency-dependent conductivity. In the limit of perfect diamagnetic material both approaches gave identical results.
The main numerical difficulty is to construct the proper adaptive mesh, which should be fine enough to resolve surface currents, but still give solutions in reasonable time. Various strategies were emploied to optimize the process, utilizing symmetries, periodic boundary conditions, perfect magnetic shielding, and various adaptive sweeps and batch modes. Each geometry was solved for by using several different approaches and different meshes to make sure final results are model-independent. Geometries for which analytical solutions are known (ellipsoids and cylinders) were used to verify numerical schemes and gave nearly perfect agreement. All calculations were done in SI, so that factor $\mu_{0}$ was properly taken into account where required.
![(top panel) sphere and cube in a full three dimensional meshed model. (bottom panel) corresponding 3D solutions showing surfaces of constant amplitude magnetic induction of 0.45 G and 1.47 G. Applied external field is 1 Oe.[]{data-label="fig1"}](fig1.png){width="1\linewidth"}
To illustrate the method, top panel of Fig.\[fig1\] shows three-dimensional meshed sphere (right) and cube (left). The meshes used in actual calculations were much finer and contained various adaptive refinements and layers. (they would be irresolvably dark if shown here). Bottom panel of Fig.\[fig1\] shows two surfaces of constant magnetic induction around these samples. With the applied field of $1$ Oe $\approx 79.58$ A/m, one surface with $0.45$ G $= 0.045$ mT corresponds to diamagnetic shielding outside the sample, while $1.47$ G $= 0.147$ mT corresponds to enhancement due to demagnetization. Clearly, cube provides more shielding, $\chi_0=\chi_{cube}=-1.64$, compared to the sphere, $\chi_0=\chi_{sphere}=-1.5$, and this is reflected in a larger demagnetizing factor, $N_{cube}=0.39$ compared to $N_{sphere}=1/3$. Already here, it is obvious that, due to symmetry, the sum of demagnetizing factors in three principal directions for a cube is $\sum N_{i}=3\times0.39=1.17>1$.
Finite magnetic permeability
============================
Unfortunately, complications arise in non-ellipsoidal samples with finite magnetic permeability. While demagnetizing factors are constants independent of $\mu_r$ in ellipsoidal samples, they become $\mu_r$-dependent otherwise. Hence, effective demagnetizing factors are no longer purely geometric parameters. It is still possible to provide some practical approximation of this behavior, but it will require a separate paper.
![$B_{z}$ component of the magnetic induction across the sample in the $x-$ direction (left panels) and in the $z-$ direction (right panels) for two values of relative magnetic permeability, $\mu_r=0.6$ (diamagnetic, dashed lines) and $\mu_r=1.4$ (paramagnetic, solid lines). Top panels are for an oblate spheroid and bottom panels are for a cylinder of the same aspect ratio (see insets). Note constant field inside a spheroid and strongly non-uniform magnetic induction inside a cylinder.[]{data-label="fig2"}](fig2.png){width="1\linewidth"}
Here we outline all the steps of calculating effective demagnetizing factors. First we use COMSOL to calculate three dimensional distribution of the vector $\bm{B}(\bm{r})$ inside and outside the sample. Here we do it for two values of magnetic permeability corresponding to diamagnetic and paramagnetic materials. Figure \[fig2\] shows $B_{z}$ component of the magnetic induction across the sample in the $x$ direction, $B_{z}\left( x,z=0\right)$, (left panels) and in the $z$ direction $B_{z}\left( x=0,z\right) $ (right panels) for two values of relative magnetic permeability, $\mu_r=0.6$ (diamagnetic, dashed lines) and $\mu_r=1.4$ (paramagnetic, solid lines). Top panels are for an oblate spheroid and bottom panels are for a cylinder of the same aspect ratio (see insets). Note constant magnetic induction inside a spheroid and a very non-uniform induction inside a cylinder.
The next step is to use Eq.(\[MdefB\]) with spherical or cylindrical integration volumes to compute total magnetic moment in a fixed applied field of 1 Oe. Finally, we use Eq.(\[N\]) to evaluate the effective demagnetizing factor.
![Effective demagnetizing factor $N$ for a finite cylinder in an axial magnetic field for three different aspect ratios as a function of relative magnetic permeability $\mu_r$. Inset shows the difference $N\left( \mu_r\right) - N\left( 0\right)$.[]{data-label="fig3"}](fig3.png){width="1\linewidth"}
Figure \[fig3\] shows $\mu_r$ dependence of the effective demagnetizing factor $N (\mu_r)$ of a finite cylinder in longitudinal magnetic field for three different values of the thickness to radius aspect ratio. The inset shows the variation of the difference $N\left( \mu_r\right) -N\left(
0\right)$. Expectedly, (see Eq.(\[N\])), for a strongly paramagnetic material with $\mu_r>5-10$ the variation in $N$ is not too substantial. However, for materials of practical interest, $0\leq\mu_r \leq10$ the dependence of $N$ on $\mu_r$ is quite strong. We will attempt to provide a simplified description of $N\left( \mu_r \right) $ for various non-ellipsoidal shapes elsewhere.
Perfect Diamagnets
==================
For now we will focus on a perfect diamagnetic material with constant $B=0$ inside.
General ellipsoid
-----------------
Throughout this paper we adapt uniform designation of sample dimensions $2a\times2b\times2c$ along Cartesian $x,y$ and $z$ axes with external magnetic field applied along the $z-$ axis, parallel to the side $c$ of the sample. Also, we will always use the dimensionless ratios, $b/a$ and $c/a$.
For completeness, it is useful to show here the analytical solution for the ellipsoid with semi-axes, $a$, $b$ and $c$ given in Ref. , Eqs.(4.5) and (4.25). Osborn also gives analytical solutions of this case expressed via the differences of incomplete elliptic integrals and written for a restricting case of $a\geq b\geq c$ [@Osborn1945] (Eqs.(2.1-2.6)). It turns out formulas given in Landau’s book are much easier to compute numerically and they work for any ratio of the dimensions [@Landau1984]. The demagnetizing factor along the $c$ axis is:
$$N_{ellipsoid}=\frac{1}{2}\frac{b}{a}\frac{c}{a}\displaystyle\int\limits_{0}^{\infty}
\frac{ds}{\left( s+\frac{c^2}{a^2}\right) R\left( s\right) }
\label{Nellipsoid}$$
where, $$R\left( s\right) =\sqrt{\left( s+1\right) \left( s+\frac{b^2}{a^2}\right) \left( s+\frac{c^2}{a^2}\right)}
\label{Rs}$$ Demagnetizing factors along other two directions have similar form with $\left( s+(c/a)^2\right) $ in the denominator of Eq.(\[Nellipsoid\]) replaced by $\left( s+1\right)$ or $\left( s+(b/a)^2\right)$, along $a$ axis or $b$ axis, respectively. We verified our numeric approach by calculating $N$ for ellipsoids and found a perfect agreement with Eq.(\[Nellipsoid\]).
Rectangular cuboid
------------------
Brick-shaped sample most commonly encountered in research laboratories, because many single crystals tend to grow in this shape, cutting and polishing procedure also favors this type of samples.
![Effective demagnetizing factor of a rectangular cuboid as function of its two aspect ratios. Inset shows schematic geometric arrangement.[]{data-label="fig4"}](fig4.png){width="1\linewidth"}
Three-dimensional surface of the effective demagnetizing factor, $N$, of a rectangular cuboid as function of two aspect ratios, $c/a$ and $b/a$, is shown in Fig.\[fig4\]. Analysis of the numerical data led us to suggest the simple formula, $$N_{cuboid}\approx\frac{4ab}{4ab+3c\left( a+b\right) }
\label{Ncuboid}$$ *This is an important result of this work*, because it describes most frequently used sample shape. Note that Eq.(\[Ncuboid\]) is an interpolation between limiting cases of infinitely thin sample ( $c\rightarrow0,$ $N\rightarrow1$) and $N\rightarrow0$ in the opposite case.
![Top panel: the difference between numerically calculated $N$ and approximation given by Eq.(\[Ncuboid\]). Bottom panel: the difference between numerically computed $N$ for cuboid and that of an ellipsoid, Eq.(\[Nellipsoid\]).[]{data-label="fig5"}](fig5.png){width="1\linewidth"}
Top panel of Fig. \[fig5\] shows the difference between numerically calculated $N$ for a rectangular cuboid (Fig. \[fig4\]) and approximation given by Eq.(\[Ncuboid\]). For comparison, bottom panel of Fig.\[fig5\] shows the difference between numerically computed $N$ for cuboid and analytical solution for an ellipsoid, Eq.(\[Nellipsoid\]). While the latter shows deviation upward of $0.08$ (considering that $N$ can only vary between $0$ and $1$), the former remains a much closer function approximating numerical results.
The relative error in determining magnetic moment (therefore apparent susceptibility) of a finite sample due to non-exact demagnetizing factor is readily derived from Eqs. \[mdemagchiL\]) and (\[N\]), $$\frac{m(N)-m(N_{approx})}{m(N)}=\frac{N_{approx}-N}{N_{approx}+\chi^{-1}}
\label{chiErr}$$ where $N$ is the exact and $N_{approx}$ is the approximate demagnetizing factors, respectively and $\chi$ is intrinsic magnetic susceptibility. Estimates using Eq.(\[chiErr\]) show that most of the diagram (except for very thin samples) in Fig. \[fig5\](a) results in errors within 10%. The error becomes larger for $b/a \gtrsim 1$ and $c/a \lesssim 0.1$ (lower right) corner of Fig. \[fig5\](a). Indeed, if better precision is needed, full calculations are required.
![Comparison of the effective demagnetizing factors for a square base cuboid: this work (solid red line), “magnetometric", $N_m$, from Aharoni [@Aharoni1998] (dashed black line), “magnetometric", $N_m$ (solid squares) and “fluxmetric", $N_f$, (dash double dotted line) from Pardo [*et al.*]{} [@Pardo2004], and magnetometric $N$ - factors, - approximate (dotted grey), exact (circles) from Sato and Ishii [@Sato89]. Inset zooms at thinner samples. Of all calculations, magnetometric $N$ factors calculated by Pardo [*et al.*]{} [@Pardo2004] are in a very good agreement with out results for $c \leq a$.[]{data-label="fig6"}](fig6.png){width="1\linewidth"}
By matching magnetostatic self-energy to total magnetic energy of a saturated ferromagnetic prism, Aharoni has provided formulas for so-called “magnetometric" demagnetizing factor, $N_m$ of the rectangular prism [@Aharoni1998]. Similarly, Pardo [*et al.*]{} calculated both “magnetometric" factor using volume average magnetic field and “fluxmetric" factors using mid-plane average magnetic fields [@Pardo2004]. Both postulated that the sum of three demagnetizing factor must be $1$ which is not justified. However, when Pardo [*et al.*]{} relaxed this constraint calculating their “magnetometric" $N$-factors, their results agreed perfectly with our numerics for $c \leq a$ [@Pardo2004]. Sato and Ishii provided very simple approximate formulas for square cuboids and circular cylinders of finite thickness in axial magnetic field [@Sato89] that they obtained by the analysis of the solutions by Aharoni and co-workers. They, therefore, agree well with Aharoni [@Aharoni1998], but disagree with our unconstrained numerical results. This is shown graphically in Fig. \[fig6\] where various effective demagnetizing $N$ - factors are shown for a square base cuboid as function of thickness to side ratio, $c/a$.
Finite cylinder in axial and transverse magnetic fields
-------------------------------------------------------
### Finite circular cylinder in an axial magnetic field
This is another typical shape of practical importance and interest. Often it is a piece of a round wire, part of superconducting magnet winding or various cables and transmission lines. They may be subject to either axial or transverse field (or a combination of the two). For the axial case (magnetic field along the cylinder axes) and circular cross-section, the inverse demagnetizing factor is shown in Fig. \[fig7\] and compared to square base cuboid and a spheroid of similar aspect ratio. For comparison, Fig. \[fig7\] also shows rectangular cuboid and a spheroid as a function of $c/a$ ratio. Clearly, for a cylinder, Eq.(\[NaxialCylinder\]) works quite well. Also shown is a comparison with formulas given by Sato and Ishii [@Sato89]. They approach our results in the thin limit, but the general trend is quite different.
![Inverse of the effective demagnetizing factor, $N^{-1}$, as a function of thickness to diameter aspect ratio for a cylinder. Open symbols are our numerical results and solid line is our approximations, Eq. (\[NaxialCylinder\]). For comparison, E. H. Brandt’s theory is shown by the dotted line and simplified approximations of M. Sato and Y. Ischii by the dashed line. For completeness, comparative results for a cuboid are also shown by solid line (this work) and dashed line Ref. [@Sato89]. Inset zooms on to smaller aspect ratio region showing excellent agreement of our numerical results and Brandt’s formula [@Brandt2001]. []{data-label="fig7"}](fig7.png){width="1\linewidth"}
As shown in Fig. \[fig7\], analysis of the numerical results shows that the simplest curves are obtained for the inverse of the effective demagnetizing factor $N^{-1}$ as a function of the aspect ratio. Indeed, this was also noted in many previous works, for example M. Sato and Y. Ishii [@Sato89] and E. H. Brandt [@Brandt2001]. In case of a finite cylinder in axial magnetic field, we obtain a simple approximate formula for the effective inverse demagnetizing factor: $$N^{-1}_{axial} \approx 1+1.6\frac{c}{a}
\label{NaxialCylinder}$$
We note that our very early work suggested similar approximation with a crude estimate of a numerical factor of 2 in place of 1.6 in Eq. \[NaxialCylinder\] [@Prozorov2000]. Our results can be compared with the numerical simulations of finite superconducting samples by E. H. Brandt [@Brandt2001]. He extended his calculations of infinite rectangular strips in perpendicular magnetic field to finite disks of rectangular cross-section. According to Brandt, for a disk of height $2c$ and diameter $2a$ [@Brandt2001],
$$N^{-1}_{disk}=1+\frac{1}{q}\frac{c}{a}
\label{Brandt01d}$$
where $$q=\frac{4}{3\pi}+\frac{2}{3\pi} \tanh{\left[1.27 \frac{c}{a} \ln{\left(1+\frac{a}{c}\right)}\right]}
\label{Brandt01diskq}$$
Figure \[fig7\] shows excellent agreement between our numerical results and Brandt’s equations lending further support to our calculations. On the other hand, simplified approximations of M. Sato and Y. Ishii [@Sato89] for these geometries do not agree at all with our and Brandt’s results.
![Finite cylinder in transverse magnetic field as a function of ratio of diameter to length. Inset shows small $a/b$ values.[]{data-label="fig8"}](fig8.png){width="1\linewidth"}
### Finite circular cylinder in a transverse magnetic field
We now consider a finite cylinder of circular cross-section ($a=c$) in a magnetic field applied perpendicular to its axis. (Please, note that we change the designations of the dimensions compared to the previous subsection to follow the uniform naming scheme of this paper). Figure \[fig8\] shows inverse demagnetizing factor $N^{-1}$ as function of the $a/b$ ratio. Notice that this ratio is reciprocal to that of Eq.(\[NaxialCylinder\]). For small enough ratio of a diameter to length, $a/b$ a good approximation for the demagnetizing factor in this case is: $$N^{-1}_{transv} \approx 2+\frac{1}{\sqrt{2}}\frac{a}{b} \label{NtransCyl}$$
Notice that Eq.(\[NtransCyl\]) gives correct value of $N=1/2$ for an infinite cylinder in transverse field, $a/b\rightarrow 0$, and correct $N=0$ when $a/b\rightarrow \inf$.
### Infinite rectangular and elliptical cross-section strips in a transverse field
Another important case, which is a partial case of the general cuboid is an infinite strip of a rectangular cross-section in a perpendicular field. This geometry is quite relevant for the superconducting tapes as parts of cables or magnet winding. Demagnetizing correction here is an important ingredient of design optimization. It has been considered before by using similar finite element numerical approach as used here, but in two dimensions [@Prozorov2000] and also in a different way using highly nonlinear $E(j)$ characteristics applicable for superconductors by H. Brandt [@Brandt2001].
![Inverse demagnetizing factor, $N^{-1}$, of an infinite strip of a rectangular cross-section as function of thickness to width ratio, $c/a$. Two simple approximate formulas are shown. Also show more elaborate formula by H. Brandt [@Brandt2001], which agrees with the numerical results quite well up to $c/a=10$. Inset shows smaller range of $c/a$.[]{data-label="fig9"}](fig9.png){width="1\linewidth"}
Figure \[fig9\] shows an inverse demagnetizing factor, $N^{-1}$, of an infinite strip of a rectangular cross-section as function of thickness to width ratio, $c/a$. Inset shows smaller range of $c/a$.Two simple approximate formulas are, $$N^{-1}_{inf-rect-strip}=\left\{
\begin{array}
[c]{c}1+\frac{2}{3}\frac{c}{a},\;\text{large }\frac{c}{a} \gtrsim 5\\
\\
1+\frac{3}{4}\frac{c}{a},\;\text{small }\frac{c}{a} \lesssim 5
\end{array}
\right.
\label{NinfStrip}$$
Using numerical simulations of finite superconducting samples, E. H. Brandt gives [@Brandt2001],
$$N^{-1}_{inf-rect-strip}=1+\frac{1}{q}\frac{c}{a}
\label{Brandt01}$$
where $$q=\frac{\pi}{4}+0.64 \tanh{\left[0.64 \frac{c}{a} \ln{\left(1.7+1.2 \frac{a}{c}\right)}\right]}
\label{Brandt01q}$$
Notably, for a square cross-section ($a=c$) infinite along $b-$direction strip, Brand obtained $N_a=N_c=0.538$, also noting that the sum of demagnetizing factors is greater than 1 for non ellipsoidal shapes (the third, $N_b=0$ for infinite strip).
![Inverse demagnetizing factor of an infinite strip of elliptical cross-section compared to a strip of rectangular cross-section. Both are infinite in $b-$direction.[]{data-label="fig10"}](fig10.png){width="1\linewidth"}
Next we consider an infinite strip of elliptical cross-section, compared to a rectangular strip in Fig.\[fig10\]. It turns out, the elliptical cross-section has the simplest approximate equation for the effective demagnetizing factor of all considered cases. Here, $$N^{-1}_{inf-ell-strip} \approx 1+\frac{c}{a} \label{NinfEllStrip}$$
Exotic geometries
-----------------
Described numerical method allows for the calculation of the effective demagnetizing correction for samples of any shape. For an illustration, let us consider a pyramid in shape of the Great Pyramid of Giza, a cone inscribed in this pyramid and a slab enclosing it, all three shown in Fig.\[fig11\]. In all these cases, the ratio of its height to the side is $c/a=2/\pi\approx0.64$. Most likely a pure coincidence, but demagnetizing factor of the Great Pyramid (and of the inscribed cone), $N=0.64$, is the same as the ratio of height to side. For a cuboid of the same $c$ and $a$, $N=0.49$ is smaller owing this to a larger volume compared to the cross-sectional area responsible for the magnetic field distortion around the sample. This adds yet another puzzle for Egyptologists.
![“Exotic” sample shapes: The Great Pyramid of Giza, inscribed cone of the same height and a cuboid enclosing the pyramid. Effective demagnetizing factor of the pyramid (and the same for the inscribed cone), $N=0.64$, is the same as the ratio of height to side, $c/a=2/\pi=0.64.$. For a cuboid with the same height and side, $N=0.49$.[]{data-label="fig11"}](fig11.png){width="1\linewidth"}
Conclusions
===========
We introduced a direct, assumptions free, efficient way to estimate the effective (or “integral") demagnetizing factors of arbitrarily shaped diamagnetic samples. The key equation, Eq.(\[MdefB\]), was generalized and re-derived in form of Eq.(\[MdefJ\]) where vector ${{\cal E}}(\bm \rho)= \int d^3\bm r \, {\bm R}/R^3$ can be interpreted as a pseudo-electric field ${\cal E}$ produced by a charge density -1 uniformly distributed in the entire space, and particular form of ${\cal E}(\bm r)$ depends on the shape of the integration domain. This answers the question posed in the introduction. Namely, it allows calculating total magnetic moment of arbitrarily shaped sample if the volume distribution of the magnetic induction is known. In this work, the latter is obtained from the adaptive finite element full 3D numerical calculations using COMSOL 5.3 software. We provide simple approximate, yet accurate (e.g., for cuboid within $\Delta N < 0.05$, see Fig. [fig5]{}(a)) analytical expressions to estimate effective demagnetizing factors of samples of commonly used non-ellipsoidal shapes, which are summarized in Table I.
This work was supported by the U.S. Department of Energy (DOE), Office of Science, Basic Energy Sciences, Materials Science and Engineering Division. Ames Laboratory is operated for the U.S. DOE by Iowa State University under contract DE-AC02-07CH11358.
L. D. Landau and E. M. Lifshitz, Electrodynamics of continuous media, 2$^{nd}$ ed., Ed. E. M. Lifshitz, L. P. Pitaevskii, (translated by J. B. Sykes, J. S. Bell and M. J Kearsley), vol. 8 (Pergamon Press, 1984).
J. A. Osborn, Demagnetizing Factors of the General Ellipsoid, Phys. Rev. **67**, 351 (1945).
R. I. Joseph and E. Schlmann, Demagnetizing Field in Nonellipsoidal Bodies, J. App. Phys. **36**, 1579 (1965).
D. X. Chen, J. A. Brug, and R. B. Goldfarb, Demagnetizing factors for cylinders, IEEE Trans. Mag. **27**, 3601 (1991).
A. Aharoni, Demagnetizing factors for rectangular ferromagnetic prisms, J. App. Phys. **83**, 3432 (1998).
E. Pardo, D.-X. Chen, and A. Sanchez, Demagnetizing factors for completely shielded rectangular prisms, J. App. Phys. **96**, 5365 (2004).
M. Sato and Y. Ishii, Simple and approximate expressions of demagnetizing factors of uniformly magnetized rectangular rod and cylinder, J. App. Phys. **66**, 983 (1989). A. Smith, K. K. Nielsen, D. V. Christensen, C. R. H. Bahl, R. Bjrk, and J. Hattel, The demagnetizing field of a nonuniform rectangular prism, J. App. Phys. **107**, 103910 (2010).
E. H. Brandt, Geometric edge barrier in the Shubnikov phase of type-II superconductors, Low Temp. Phys. **27**, 723 (2001);
R. Prozorov, R. W. Giannetta, A. Carrington, F. M. Araujo-Moreira, Meissner-London state in superconductors of rectangular cross section in a perpendicular magnetic field, Phys. Rev. B **62**, 115 (2000).
J. D. Jackson, Classical electrodynamics, 3$^{rd}$ Ed., John Wiley & Sons, New York, 1999.
P. M. Morse and H. Feshbach, Methods of theoretical Physics, McGrow Hill, New York, 1953: part II, Section 10.3.
COMSOL version 5.3, Multiphysics reference manual, and AC/DC module user’s guide, www.comsol.com, COMSOL Inc. (2017).
W. T. Doyle and I. S. Jacobs, The influence of particle shape on dielectric enhancement in metal-insulator composites, J. App. Phys. **71**, 3926 (1992).
Current ring
============
To demonstrate how calculations of the total magnetic moment work, it is instructive to consider an example of a ring of a radius $a=1\,$m in plane $z=0$ with current $J/4\pi=1\,$A. The total magnetic moment is $m_z=J \pi a^2 =4\pi^2\,$Am$^2$. The field $B_z$ around the ring according to Landau is [@Landau1984] $$\begin{aligned}
B_z=\frac{2}{ \sqrt{(1+r)^2+z^2}}\left[ \bm{K} +\frac{ 1-r ^2-z^2}{(1-r)^2+z^2}\bm{E} \right],\qquad
\label{r1}\end{aligned}$$ where $r,z$ are cylindrical coordinates, the complete elliptic integrals $\bm{K}, \bm{E}$ are functions of $k^2=4r/[(1+r)^2+z^2]$. The Jackson theorem for this case with no applied field reads: $$\begin{aligned}
\frac{2}{3}\,{m}=\int B_zdV,
\label{jack}\end{aligned}$$ where the integration is over the whole space. Hence, one has to check the equality $$\begin{aligned}
\int\limits B_zd^3\bm{r} = \frac{8\pi^2}{3} \,.
\label{jack1}\end{aligned}$$ One can perform the volume integration in spherical coordinates $R,\theta$: $r=R\sin\theta$, $z=R\cos\theta$, $dV=2\pi R^2\sin\theta dR\,d\theta$: $$\begin{aligned}
B_z&=&\frac{2}{ \sqrt{1+2R\sin\theta +R^2}} \Big[ \bm{K} +\frac{ 1-R^2 }{1-2R\sin\theta +R^2}\bm{E} \Big], \nonumber\\
k^2&=&\frac{4R\sin\theta}{1+2R\sin\theta +R^2}\,.\qquad
\label{Bz1}\end{aligned}$$ The integral is readily calculated numerically: $ \int B_z\, d^3\bm{r} \approx 26.32$ coincides with the analytic value $8\pi^2/3$ and gives $m=3I_z/2=4\pi^2\approx 39.48$.\
Another way of evaluating the integral $\bm I$ in the whole space is to choose cylindrical integration elements with axis along the applied field ${\bf H_0}= H_0 {\hat {\bm z}}$. To this end we use Eq. (\[r1\]) where $B_z$ is given in cylindrical coordinates. The numerical integration yields $I_z=39.48$ which coincides with the result for spherical integration multiplied by 3/2. Eq.(\[eq9a\]) gives $m_z=I_z=4\pi^2\approx 39.48$. Thus, although the results of numerical evaluation of the integral over the whole space differs from the spherical method, the magnetic moment value for two methods comes out the same.\
To illustrate how our numerical calculations reproduce non-trivial analytical results presented here, Fig.\[fig12\] shows evaluation of the $z-$component of the integral $I$, Eq. (\[eq0\]), using spherical and cylindrical shells. The inset in Fig.\[fig12\] shows three dimensional pie-cut picture of the absolute value of the magnetic induction around the ring. The ring radius, $=1/\sqrt{\pi}$ m, is chosen to have ring area to give total magnetic moment of 1 Am$^2$ with 1 A current in the ring. With such choice, the value of the integral $\bm I$ will be equal to $1/\alpha$ of the Eq. (\[MdefB\]). Note that numerical calculations here and everywhere in this paper are carried out in SI units. We obtain that the spherical shell integration tends to the value of $\alpha^{-1}=2/3$ and cylindrical shell integration to $\alpha^{-1}=1$, exactly as shown analytically. Moreover, both integrals stop changing as soon as the sample current is fully enclosed in the integration volume, again as expected from the analytical calculations. Therefore, integration of space outside the sample does not contribute to the integral $\bm{I}$. It does not mean, however, that the outer space can by truncated for numerical calculations. It should still be much larger compared to the sample size in order to solve for (very long - range) magnetic fields distribution correctly.
![Integral $I_z$, of Eq. (\[eq0\]), calculated numerically using spherical shells and cylindrical shells, respectively. The calculations are done in SI units and ring radius ($=1/\sqrt(\pi)$ m) is chosen to give total magnetic moment of 1 Am$^2$ for current of 1 A. This way the integral value is just $1/\alpha$ of Eq. (\[MdefB\]). Inset shows three dimensional pie-cut picture of the absolute value of the magnetic induction around the ring.[]{data-label="fig12"}](fig12.png){width="1\linewidth"}
Identity Eq. (\[MdefB\]) in Gaussian units
==========================================
The derivation of this identity is given in the main text in the “recommended" SI system, which was also used in COMSOL to solve the Maxwell equations for the field distribution in SI. This, however, makes the derivation unnecessary cumbersome. Many researchers working in the fields of superconductivity and magnetism prefer CGS and it is also used in magnetometers such as *Quantum Design* MPMS. We therefore provide here the same derivation in Gaussian system, which, in our opinion, makes the derivation more transparent.\
The total magnetic moment $\bm m$ of a finite-size sample is proportional to the volume integral of the total field distribution $\bm B(\bm r)$ over the sphere of large enough radius ${\cal R}_1$ such that the whole sample is situated inside the sphere. This statement is proven in Jackson’s book [@Jackson2007] which gives: $$\begin{aligned}
{ \bm I }=\int_{r<{\cal R} }\left[{\bm B}(\bm r)-{\bm H}_0\right]d^3\bm r =\frac{8\pi}{3}\,\bm m\,.
\label{eq0}\end{aligned}$$ Here, ${\bm H}_0$ is the applied field. In particular, one can take ${\cal R} \to\infty$, i.e. the integral can cover the whole space. This identity is proven in Ref. by expanding the field outside a sphere $\cal R$ which contains the sample (the currents) in spherical harmonics. From the point of view of a finite-element numerical method it is is more convenient to deal with cylindrical integration domains. We thus provide a proof not related to a particular shape of the integration region.
The field $\bm B$ consists of the applied field and the field $\bm b$ due to currents $\bm J$ in the sample of a finite volume $V$: $$\begin{aligned}
{\bm B}={\bm H}_0 +\bm b\,,
\label{eq1}\end{aligned}$$ i.e., $\bm I=\int {\bm b}\, d^3\bm r$, where according to Biot-Savart $$\begin{aligned}
{\bm b}(\bm r)=\frac{1}{c}\int_V d^3{\bm\rho}\,\frac{{\bm J}(\bm \rho)\times{\bm R}}{R^3}\,,\qquad \bm R=\bm r -\bm \rho\,.\qquad
\label{eq2}\end{aligned}$$ Hence, we have $$\begin{aligned}
c\bm I&=& \int d^3\bm r \int_V d^3{\bm\rho}\frac{{\bm J}(\bm \rho)\times{\bm R}}{R^3}\nonumber\\
&=&
\int_V d^3{\bm\rho} \,{\bm J}(\bm \rho)\times \int d^3\bm r \frac{{\bm R}}{R^3} \,,
\label{eq3}\end{aligned}$$ where the integration over $\bm r$ is extended to the whole space.\
The vector ${\bf{ \cal E}}(\bm \rho)= \int d^3\bm r {\bm R}/R^3$ is analogous to the electrostatic field of a charge distribution with density of $-1$ in the whole space. For such a distribution, the field ${ \cal E}$ is not defined uniquely, it depends on the way one divides the space in charged elements.\
For $\bm\rho=0$, we must have ${\bf { \cal E}}=\int d^3\bm r ({\bm r}/r^3)=0$ by symmetry. If one uses elements as spherical shells, and applies the Gauss theorem to a sphere of a radius $\rho$ one obtains: $$\begin{aligned}
{ \cal E}=-\frac{4\pi}{3}\,\bm\rho \,.
\label{eq8}\end{aligned}$$ Hence, we have $$\begin{aligned}
\bm I= -\frac{ 4\pi }{3c}\int_V d^3{\bm\rho} \,{\bm J}(\bm \rho)\times {\bm \rho}=\frac{8\pi}{3} {\bm m} \,,
\label{eq9}\end{aligned}$$ where $\bm m$ is the total magnetic moment. It is worth noting that this formula holds for any current distribution within the finite sample of arbitrary shape.\
If one uses elements as cylindrical shells parallel to $\bm H_0$, i.e. choose the volume element as $2\pi \rho_1\,d\rho_1dz$ ($\bm \rho_1$ is the cylindrical radius vector), and applies the Gauss theorem to a cylinder of a radius $\rho$ one obtains: $$\begin{aligned}
{ \cal E} =- 2\pi {\bm \rho}_1 \,.
\label{eq8a}\end{aligned}$$ where $\bm\rho_1$ is now the cylindrical radius vector. Substituting this in Eq.(\[eq3\]), one expresses the $z$ component of the integral $\bm I$: $$\begin{aligned}
I_z= 4\pi \, m_z \,.
\label{eq9a}\end{aligned}$$
[m[4cm]{}m[4cm]{}m[10cm]{}]{} & &\
Ellipsoid (exact) &{width="2.2cm"} & $\large{N}=\frac{1}{2}\frac{b}{a}\frac{c}{a}\displaystyle\int\limits_{0}^{\infty}\frac{ds}{(s+\frac{c^2}{a^2})\sqrt{\left( s+1\right) \left( s+\frac{b^2}{a^2}\right) \left( s+\frac{c^2}{a^2}\right)}}$\
Rectangular cuboid &{width="2.2cm"} &\
Strip, rectangular &{width="2.2cm"} & ,\
Strip, elliptical &{width="2.2cm"} &\
Cylinder, axial &{width="2.2cm"} &\
Cylinder, transverse &{width="2.2cm"} &\
Great pyramid\
(cone, cuboid) &{width="2.7cm"} & $N=0.64$ $(0.64, 0.49)$\
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper, we investigate the robust beamforming schemes for a multi-user multiple-input-single-output (MU-MISO) system with per-antenna power constraints and quantized channel direction information (CDI) feedback. Our design objective is to maximize the expectation of the weighted sum-rate performance by means of controlling the interference leakage and properly allocating the power among user equipments (UEs). First, we prove the optimality of the non-robust zero-forcing (ZF) beamforming scheme in the sense of generating the minimum amount of average inter-UE interference under quantized CDI. Then we derive closed-form expressions of the cumulative density function (CDF) of the interference leakage power for the non-robust ZF beamforming scheme, based on which we adjust the leakage thresholds and propose two robust beamforming schemes under per-antenna power constraints with an iterative process to update the per-UE power allocations using the geometric programming (GP). Simulation results show the superiority of the proposed robust beamforming schemes compared with the existing schemes in terms of the average weighted sum-rate performance.'
author:
- '[Ming Ding$^{*}$, ]{}*Member, IEEE*[, Meng Zhang, Hanwen Luo, and Wen Chen, ]{}*Senior Member, IEEE*[^1][^2][^3][^4]'
title: 'Leakage-Based Robust Beamforming for Multi-Antenna Broadcast System with Per-Antenna Power Constraints and Quantized CDI'
---
multi-user, robust beamforming, per-antenna power constraints, quantized CDI, zero-forcing, leakage.
Introduction
============
Multi-antenna broadcast systems have gained considerable attention as they can offer both spatial multiplexing and multi-user (MU) diversity gains [@IntroBC]. In the multi-antenna broadcast channel (BC) model, the multiplexing gain can be achieved by simultaneously serving multiple user equipments (UEs) by space division multiple access schemes with the dirty paper coding (DPC) [@DPC] or low-complexity linear transmit beamforming, e.g., zero-forcing (ZF) beamforming [@ZF]. Moreover, when the UE number is large, the capacity of the BC system also grows with the UE number according to a double logarithm scaling law due to the MU diversity gain [@multi-UE; @diversity; @gain]. However, all these promising results are based on the assumption of perfect channel direction information (CDI) available at the base station (BS), which is too ideal for practical systems, especially for the frequency division duplex (FDD) system such as the fourth generation (4G) cellular network, e.g., the long term evolution advanced (LTE-A) FDD system [@LTE-A]. The imperfectness of CDI is mainly resulted from the limited-bit CDI quantization process performed by the UE [@Limited-bit; @CSI].
The existence of CDI quantization error motivates the design of robust beamforming, which takes the uncertain channel distortions into account [\[]{}7-19[\]]{}. In [@robust_minPower1], the authors proposed robust beamforming schemes for an MU multi-antenna BC system to minimize the BS transmission power while maintaining certain quality-of-service (QoS) requirements for the worst case model, i.e., treating channel errors as norm bounded matrices, and the stochastic model, i.e., assuming certain statistical properties of channel errors. The authors of [@basic_robust_SLNR] addressed a more general problem by considering additional constraints, such as keeping the interference under a preset tolerable level and individually shaping the beamforming vectors. In [@robust_MMSE0], the authors investigated robust beamforming schemes to minimize the sum of UEs’ mean squared errors (MSEs). Considering inter-UE interference leakage [@origSINR], the authors of [@robust_AveSLNR] designed a robust beamforming scheme to maximize a lower bound of each UE’s average signal-to-leakage-plus-noise ratio (SLNR). Recently, in [@robust_ProbLeak1] and [@robust_ProbLeak_ICC12], the authors proposed another leakage-based robust transmit beamforming scheme, which optimizes the average signal-to-interference-plus-noise ratio (SINR) performance implicitly by maximizing the average signal power subject to probabilistic leakage and noise power constraints. Besides, in [@robust_interfchannel], by taking finite-rate CDI feedback into account the authors investigated the transceiver design for a two-UE multiple-input multiple-output (MIMO) interference channel, where each precoder or equalizer is divided into outer and inner parts to eliminate the cross-link interference. In [@robust_MSE_new], a joint design of the channel estimator and the quantization function was proposed based on the criterion of MSE minimization. Furthermore, robust beamforming schemes have been extended to more sophisticated models such as the MIMO relay networks [@robust_extRelay1], [@Prof_chen_relatWork1], [@Prof_chen_relatWork2] and the multi-cell coordinated beamforming operations [@robust_maxSINR1], etc.
In this paper, we further investigate the robust beamforming design based on the approach of leakage control. In particular, we consider the optimization of the weighted sum-rate performance and more realistic power constraints, which limit the BS transmission power on a per-antenna basis. Compared with the existing robust beamforming schemes, our assumption on the transmission power is more practical since each antenna of a multi-antenna BS [@Per_ante_PC] is normally equipped with an individual power amplifier at its analog front-end. Our contributions are two folds:
1. We prove the optimality of the non-robust ZF beamforming scheme in the sense of generating the minimum amount of average inter-UE interference under limited-bit CDI, and derive closed-form expressions of the cumulative density function (CDF) of the interference leakage power for the non-robust ZF beamforming scheme.
2. We adjust the leakage thresholds based on the derived CDF of the leakage power and propose a minimum average leakage control (MALC) and a relaxed average leakage control (RALC) beamforming schemes under per-antenna power constraints, together with an iterative process to update the power allocation among UEs using the geometric programming (GP) to maximize the weighted sum-rate performance.
$\quad$Although in this paper we mainly treat the MU multiple-input-single-output (MISO) model, our results can be extended to other BC models, such as MIMO relay networks [@Future_ext_MIMOrelay2] and multi-cell joint transmissions [@Future_ext_JT].
*Notations*: $\left(\cdot\right)^{\textrm{T}}$, $\left(\cdot\right)^{\textrm{H}}$, $\left(\cdot\right)^{-1}$, $\left(\cdot\right)^{\dagger}$, $\textrm{tr}\left\{ \mathbf{\cdot}\right\} $ and $\textrm{rank}\left\{ \mathbf{\cdot}\right\} $ stand for the transpose, conjugate transpose, inverse, pseudo-inverse, trace and rank of a matrix, respectively. $\mathbf{I}_{N}$ stands for an $N\times N$ identity matrix. $\mathbf{A}_{i,:}$, $\mathbf{A}_{:,j}$ and $\mathbf{A}_{i,j}$ respectively denote the $i$-th row, $j$-th column and $\left(i,j\right)$-th entry of matrix $\mathbf{A}$. Besides, $\mathbf{A}\succeq0$ and $\mathbf{A}\in\mathbb{\mathbb{H}}_{N}^{\mathit{+}}$ mean that matrix $\mathbf{A}$ is positive semi-definite and $\mathbf{A}$ is an $N$ by $N$ positive semi-definite Hermitian matrix, respectively. $\left\Vert \mathbf{a}\right\Vert $ and $\mathbf{a}_{i}$ denotes the Euclidean norm and $i$-th element of vector $\mathbf{a}$. $\mathbb{E}_{\left[\mathbf{x}\right]}\left\{ \cdot\right\} $ and $\textrm{Re}\left\{ \cdot\right\} $ denote the expectation operation over a random vector $\mathbf{x}$ and the real part of a complex value. $\textrm{C}_{j}^{i}$ counts the combinations of choosing $i$ elements from a set of $j$ elements. $\mathcal{N}\left(\mathbf{0},\mathbf{X}\right)$ represents a circularly symmetric complex Gaussian distribution with mean of zero vector and covariance matrix $\mathbf{X}$. Finally, we denote $\Pr\left(x\right)$ as the probability of event $x$.
System Model
============
We consider a downlink MU-MISO system with limited-bit CDI feedback as shown in Fig. \[fig:Sys\_model\], where a BS is equipped with $N$ transmit antennas, and $K$ single-antenna UEs are served simultaneously. To support $K$ independent data streams, it requires $N\geq K$. However, our results can be easily extended to the case of $N<K$ with UE selection performed at the BS [@Prof_Chen_UE_selection].
![\[fig:Sys\_model\]Illustration of a downlink MU-MISO system with limited-bit CDI.](Fig1.eps "fig:"){width="5cm"}\
In Fig. \[fig:Sys\_model\], the large-scale channel attenuation coefficient and the small-scale base-band channel vector between the BS and the $k$-th $\left(k\in\left\{ 1,2,\ldots,K\right\} \right)$ UE are denoted as $\xi_{k}$ and $\mathbf{h}_{k}\in\mathbb{C}^{1\times N}$, respectively. Let $\mathbf{w}_{j}\in\mathbb{C}^{\mathit{N}\times1}$ be the beamforming vector for UE $j$. Then the signal received at UE $k$ can be described by
$$y_{k}=\xi_{k}\mathbf{h}_{k}\sum_{j=1}^{K}\mathbf{w}_{j}x_{j}+n_{k}=\xi_{k}\mathbf{h}_{k}\mathbf{\mathbf{W}}\mathbf{\mathbf{x}}+n_{k},\label{eq:yk}$$
where $\mathbf{x}=\left[x_{1},\ldots,x_{k},\ldots,x_{K}\right]^{\textrm{T}}$ and $x_{k}$ is the data symbol intended for UE $k$. Without loss of generality, we assume that $\mathbf{x}$ satisfies $\mathbb{E}_{\left[\mathbf{x}\right]}\left\{ \mathbf{x\mathbf{x^{\textrm{H}}}}\right\} =\mathbf{I}_{K}$. The channel coefficients in $\mathbf{h}_{k}$ are assumed to experience independently identical distribution (i.i.d.) Rayleigh flat fading and remain unchanged during the downlink MU-MISO transmission. $n_{k}$ is a zero-mean circularly symmetric complex Gaussian (ZMCSCG) noise variable with $\mathbb{E}_{\left[n_{k}\right]}\left\{ n_{k}n_{k}^{\textrm{H}}\right\} =N_{0}$. In addition, $\mathbf{\mathbf{W}}=\left[\mathbf{w}_{1},\mathbf{w}_{2},\ldots,\mathbf{w}_{K}\right]$ and it is subject to an average per-antenna transmit power constraint expressed as
$$\mathbb{E}_{\left[\mathbf{x}\right]}\left\{ \left[\mathbf{\mathbf{W}}\mathbf{\mathbf{x}}\left(\mathbf{\mathbf{W}}\mathbf{\mathbf{x}}\right)^{\textrm{H}}\right]_{n,n}\right\} =\left[\mathbf{\mathbf{W}}\mathbf{\mathbf{W}}^{\textrm{H}}\right]_{n,n}\leq P_{n},\label{eq:PAPC}$$
where $P_{n}$ is the maximum transmission power of the $n$-th $\left(n\in\left\{ 1,2,\ldots,N\right\} \right)$ antenna. In addition, we denote the BS’s maximum transmission power as $P=\sum_{n=1}^{N}P_{n}$. By stacking the received signals of all UEs, we have
$$\mathbf{y}=\mathbf{H}\mathbf{\mathbf{W}}\mathbf{\mathbf{x}}+\mathbf{n},\label{eq:vec_y}$$
where $\mathbf{y}=\left[y_{1},y_{2},\ldots,y_{K}\right]^{\textrm{T}}$, $\mathbf{H}=\left[\xi_{1}\mathbf{h}_{1}^{\textrm{T}},\xi_{2}\mathbf{h}_{2}^{\textrm{T}},\ldots,\xi_{K}\mathbf{h}_{K}^{\textrm{T}}\right]^{\textrm{T}}$ and $\mathbf{n}=\left[n_{1},n_{2},\ldots,n_{K}\right]^{\textrm{T}}$.
The information of $\mathbf{H}$, i.e., the channel state information (CSI), is composed of two parts, which are the channel direction information (CDI) and channel magnitude information (CMI). The CDI is the normalized base-band channel vector of UE $k$ denoted as $\mathbf{\tilde{h}}_{k}=\frac{\mathbf{h}_{k}}{\left\Vert \mathbf{h}_{k}\right\Vert }$ and the CMI is expressed by $\xi_{k}\left\Vert \mathbf{h}_{k}\right\Vert $ or $\xi_{k}^{2}\left\Vert \mathbf{h}_{k}\right\Vert ^{2}$. In practice, perfect CSI is usually not available at the BS side. Hence, we first assume imperfect CDI for the interested MU-MISO system shown in Fig. \[fig:Sys\_model\], where each UE quantizes its CDI and feeds it back to the BS with $B$ bits. The quantized CDI is defined as the index of a vector $\mathbf{\hat{h}}_{k}$ chosen from a random vector quantization (RVQ) codebook $\mathbf{C}_{k}=\left\{ \mathbf{c}_{k,1},\mathrm{\mathbf{c}}_{k,2},\ldots,\mathrm{\mathbf{c}}_{k,2^{B}}\right\} $ [@RVQ] to match $\mathbf{\tilde{h}}_{k}$. To be more specific, $\mathbf{C}_{k}$ consists of $2^{B}$ unit vectors $\mathbf{c}_{k,n}$ $\left(n\in\left\{ 1,2,\ldots,2^{B}\right\} \right)$ isotropically distributed in $\mathbb{C}^{1\times M}$ and $\mathbf{\hat{h}}_{k}$ is selected as
$$\mathbf{\hat{h}}_{k}=\underset{\mathbf{c}_{k,n}\in\mathbf{C}_{k}}{\arg\max}\left|\mathbf{c}_{k,n}\mathbf{\tilde{h}}_{k}^{\textrm{H}}\right|.\label{eq:hk_hat}$$
Then $\mathbf{\tilde{h}}_{k}$ can be decomposed as
$$\mathbf{\tilde{h}}_{k}=\cos\left(\angle\left(\mathbf{\tilde{h}}_{k},\mathbf{\hat{h}}_{k}\right)\right)\mathbf{\hat{h}}_{k}+\sin\left(\angle\left(\mathbf{\tilde{h}}_{k},\mathbf{\hat{h}}_{k}\right)\right)\mathbf{e}_{k},\label{eq:decomp_hk_tilde}$$
where $\mathbf{e}_{k}$ is a quantization error vector orthogonal to $\mathbf{\hat{h}}_{k}$.
Regarding the CMI, $\xi_{k}$ or $\xi_{k}^{2}$ can be inferred from UE $k$’s pathloss information, i.e., $\frac{1}{\xi_{k}^{2}}$, which is implicitly reported to the BS for mobility management in existing cellular networks [@On_LTE]. In addition, $\left\Vert \mathbf{h}_{k}\right\Vert $ or $\left\Vert \mathbf{h}_{k}\right\Vert ^{2}$can be easily quantized using $M$ bits as long as its CDF is known. Since we consider the Rayleigh fading channels in this paper, $\left\Vert \mathbf{h}_{k}\right\Vert ^{2}$ follows a chi-squared distribution with its CDF expressed as $P_{R}\left(r\right)=1-\textrm{e}^{-r}\sum_{l=0}^{N-1}\frac{r^{l}}{l!}$ [@Proakis; @book]. Thus we can divide $P_{R}\left(r\right)$ into $2^{M}$ segments and select the midpoint of each segment to construct the codebook for the quantization of $\left\Vert \mathbf{h}_{k}\right\Vert ^{2}$ as $\left\{ \left.T_{i}=P_{R}^{-1}\left(\frac{2i+1}{2^{M+1}}\right)\right|i\in\left\{ 0,1,\ldots,2^{M}-1\right\} \right\} $. Then the quantized $\left\Vert \mathbf{h}_{k}\right\Vert ^{2}$ is given by
$$\hat{A}_{k}=\underset{T_{i}}{\arg\min}\left(\left|\left\Vert \mathbf{h}_{k}\right\Vert ^{2}-T_{i}\right|\right).\label{eq:hk_CF_CMI}$$ In the sequel, $\xi_{k}^{2}$ will be referred to as the pathloss CMI (PL-CMI) and $\left\Vert \mathbf{h}_{k}\right\Vert ^{2}$ as the channel fading CMI (CF-CMI). The PL-CMI is assumed to be perfect because it changes very slowly, and hence its quantization accuracy can be consistently improved over a long period of time. The CF-CMI, on the other hand, should be subject to quantization errors since it varies as fast as the CDI. Fortunately, in our simulations, we will show that even with the average CF-CMI only, i.e., $\mathbb{E}_{\left[\mathbf{h}_{k}\right]}\left\{ \left\Vert \mathbf{h}_{k}\right\Vert ^{2}\right\} $, the performance of the interested beamforming schemes is comparable with that achieved by the perfect CF-CMI, i.e., $\left\Vert \mathbf{h}_{k}\right\Vert ^{2}$. For notational brevity, we denote $A_{k}^{\textrm{ave}}=\mathbb{E}_{\left[\mathbf{h}_{k}\right]}\left\{ \left\Vert \mathbf{h}_{k}\right\Vert ^{2}\right\} $ hereafter. Note that $A_{k}^{\textrm{ave}}$ can be obtained through analytical calculation or numerical simulation based on the assumption of the channel model. Thus, $A_{k}^{\textrm{ave}}$ is not required to be fed back to the BS, i.e., $M=0$ for the average CF-CMI case. In the following, we will concentrate on the quantized CDI $\mathbf{\hat{h}}_{k}$ with $B>0$ and assume average CF-CMI at the BS. We will explicitly state if quantized CF-CMI $\hat{A}_{k}$ is considered.
Non-Robust Beamforming Schemes
==============================
For non-robust beamforming schemes, the channel uncertainties due to the CDI quantization errors are ignored so that at the BS side each UE’s channel vector is directly replaced by $\mathbf{\check{h}}_{k}=\sqrt{A_{k}^{\textrm{ave}}}\mathbf{\hat{h}}_{k}$. First, we discuss the non-robust ZF beamforming scheme with per-antenna power constraints, which results will serve as the benchmark for performance comparison in our simulations.
The Non-Robust ZF Beamforming Scheme with Per-antenna Power Constraints
-----------------------------------------------------------------------
A commonly used beamforming scheme for the downlink MU-MISO system is based on the ZF approach, which aims to completely mitigate the inter-UE interference. Under the sum power constraint, the non-robust ZF beamforming vector for UE $k$ can be derived as [@ZF]
$$\mathbf{w}_{k}^{\textrm{ZF}}=\sqrt{\tilde{P}_{k}}\mathbf{\tilde{w}}_{k}^{\textrm{ZF}}=\sqrt{\tilde{P}_{k}}\frac{\mathbf{\check{H}}_{:,k}^{\dagger}}{\left\Vert \mathbf{\check{H}}_{:,k}^{\dagger}\right\Vert },\label{eq:NR_ZF_wk}$$
where $\tilde{P}_{k}$ is the transmission power for UE $k$ and it satisfies the sum power constraint $\sum_{k=1}^{K}\tilde{P}_{k}\leq P$. Besides, $\mathbf{\tilde{w}}_{k}^{\textrm{ZF}}$ is the normalized precoding vector and $\mathbf{\check{H}}^{\dagger}$ is the pseudo-inverse of $\mathbf{\check{H}}=\left[\xi_{1}\mathbf{\check{h}}_{1}^{\textrm{T}},\xi_{2}\mathbf{\check{h}}_{2}^{\textrm{T}},\ldots,\xi_{K}\mathbf{\check{h}}_{K}^{\textrm{T}}\right]^{\textrm{T}}$. In the following, the non-robust ZF beamforming scheme based on will be referred to as the non-robust ZF scheme or the ZF scheme for short.
It is obvious that $\mathbf{w}_{k}^{\textrm{ZF}}$ may violate the per-antenna power constraints shown in . In [@ZF_per_ante_PC], the authors investigated the ZF beamforming with per-antenna power constraints and demonstrated that the optimal solution depends on the specific objective function. Considering the maximization of the weighted sum-rate for a MISO system, the authors of [@ZF_per_ante_PC] addressed that the non-robust beamforming solution can be found by solving a standard semi-definite program (SDP) problem shown as
$$\begin{aligned}
\underset{\mathbf{Q}_{k}\in\mathbb{\mathbb{H}}_{N}^{\mathit{+}}}{\max} & & f\left(\left\{ \mathbf{Q}_{k}\right\} \right)=\sum_{k=1}^{K}\alpha_{k}\log_{2}\left(1+\frac{\xi_{k}^{2}\mathbf{\check{h}}_{k}\mathbf{Q}_{k}\mathbf{\check{h}}_{k}^{\textrm{H}}}{N_{0}}\right)\nonumber \\
\textrm{s.t.} & & \textrm{tr}\left\{ \mathbf{Q}_{k}\mathbf{\mathbf{\check{h}}}_{j}^{\textrm{H}}\mathbf{\mathbf{\check{h}}}_{j}\right\} =0,\quad\forall j\neq k;\nonumber \\
& & \sum_{k=1}^{K}\left[\mathbf{Q}_{k}\right]_{n,n}\leq P_{n},\quad\forall n,\label{eq:Problem_ZF_PA}\end{aligned}$$
where $\mathbf{Q}_{k}\in\mathbb{\mathbb{H}}_{N}^{\mathit{+}}$ denotes that $\mathbf{Q}_{k}$ is an $N$ by $N$ positive semi-definite Hermitian matrix. The first and second sets of constraints in problem represent the requirements of zero interference among UEs and per-antenna power limitations, respectively. Problem is a convex optimization problem and its numerical solution can be obtained by the use of standard mathematical softwares [@cvx_software]. Note that it has been proven in [@ZF_per_ante_PC] that problem always admits a solution with rank-one matrices. Thus, the rank-one constraints on $\left\{ \mathbf{Q}_{k}\right\} $ for beamforming operations have been omitted. Suppose that $\left\{ \mathbf{Q}_{k}^{\textrm{ZF-PA}}\right\} $ is the solution to problem and $\mathbf{Q}_{k}^{\textrm{ZF-PA}}=\mathbf{q}_{k}^{\textrm{ZF-PA}}\left(\mathbf{q}_{k}^{\textrm{ZF-PA}}\right)^{\textrm{H}}$, then the non-robust ZF beamforming vector for UE $k$ under per-antenna power constraints becomes
$$\mathbf{w}_{k}^{\textrm{ZF-PA}}=\mathbf{q}_{k}^{\textrm{ZF-PA}}.\label{eq:wk_ZF-PA}$$
The non-robust beamforming scheme under per-antenna power constraints with the solution of will be called as the ZF-PA scheme for abbreviation. It should be noted that both the ZF and ZF-PA schemes will suffer from large performance degradation due to unknown inter-UE interference resulted from inevitable CDI quantization errors [@robust_AveSLNR], [@robust_ProbLeak_ICC12].
The Non-Robust SLNR Beamforming Scheme with Per-UE Power Constraints
--------------------------------------------------------------------
The ZF approach merely focuses on the minimization of the inter-UE interference power, whereas the signal power usually suffers from considerable power loss due to forced vector steering away from the subspace spanned by other UEs’ channel vectors. Therefore, an alternative SLNR approach has been proposed in [@origSINR], which suggests maximizing the ratio of the signal power over the interference leakage power plus the noise power. The non-robust SLNR beamforming scheme under per-UE power constraints, referred to as the SLNR scheme for short, generates the beamforming vector for UE $k$ as [@origSINR]
$$\mathbf{w}_{k}^{\textrm{SLNR}}=\sqrt{\tilde{P}_{k}}\boldsymbol{\textrm{v}}_{\textrm{max}}\left\{ \left(\frac{N_{0}}{\tilde{P}_{k}}\mathbf{I}_{N}+\mathbf{\mathbf{\bar{\check{H}}}_{\mathit{k}}^{\textrm{H}}\mathbf{\bar{\check{H}}}_{\mathit{k}}}\right)^{-1}\left(\xi_{k}^{2}\mathbf{\check{h}}_{k}^{\textrm{H}}\mathbf{\check{h}}_{k}\right)\right\} ,\label{eq:wk_SLNR}$$
where $\boldsymbol{\textrm{v}}_{\textrm{max}}\left\{ \mathbf{Q}\right\} $ denotes the eigen-vector associated with the largest eigen-value of matrix $\mathbf{Q}$ and $\mathbf{\bar{\check{H}}}_{\mathit{k}}$ is an extended channel matrix excluding $\xi_{k}\mathbf{\mathbf{\check{h}}}_{k}$ from $\check{\mathbf{H}}$ ,i.e., $\mathbf{\bar{\check{H}}}_{\mathit{k}}=\left[\xi_{1}\mathbf{\check{h}}_{1}^{\textrm{T}},\ldots,\xi_{k-1}\mathbf{\mathbf{\check{h}}}_{k-1}^{\textrm{T}},\xi_{k+1}\mathbf{\mathbf{\check{h}}}_{k+1}^{\textrm{T}},\ldots,\xi_{K}\mathbf{\check{h}}_{K}^{\textrm{T}}\right]^{\textrm{T}}$.
The Proposed Leakage-based Robust Beamforming with Per-antenna Power Constraints
================================================================================
As for robust beamforming schemes, we should take the imperfect CDI into account when designing UEs’ beamforming vectors, i.e., $\tilde{\mathbf{h}}_{k}$ will be considered as a random vector $\tilde{\mathbf{h}}_{k}^{\diamond}$ isotropically distributed around $\hat{\mathbf{h}}_{k}$. Similar to , $\tilde{\mathbf{h}}_{k}^{\diamond}$ can be written as [@Limited-bit; @CSI]
$$\tilde{\mathbf{h}}_{k}^{\diamond}=\sqrt{1-Z}\mathbf{\hat{h}}_{k}+\sqrt{Z}\mathbf{e}_{k}^{\diamond},\label{eq:decomp_hk_tilde_BS}$$
where $\mathbf{e}_{k}^{\diamond}$ is isotropically distributed in the $\left(N-1\right)$-dimensional nullspace of $\mathbf{\hat{h}}_{k}$ and the random variable $Z$ is defined as $Z=\sin^{2}\left(\angle\left(\tilde{\mathbf{h}}_{k}^{\diamond},\mathbf{\hat{h}}_{k}\right)\right)$.
Based on the idea of controlling the interference leakage to implicitly maximize UEs’ SINRs [@origSINR], recently in [@basic_robust_SLNR], [@robust_AveSLNR], [@robust_ProbLeak1] and [@robust_ProbLeak_ICC12], the authors proposed new robust beamforming schemes for MU-MISO systems under quantized CDI and per-UE power constraints. In [@robust_AveSLNR], the SLNR maximization problem with regard to $\mathbf{w}_{k}$ was transformed to a Rayleigh quotient problem with a lower bound solution by utilizing Jensen’s inequality to maximize each UE’s average SLNR shown as
$$\begin{aligned}
\underset{\mathbf{w}_{k}}{\max} & & \frac{\mathbb{E}_{\left[\tilde{\mathbf{h}}_{j}^{\diamond}\left|\hat{\mathbf{h}}_{j}\right.j=1,2,\ldots,K\right]}\left\{ S_{k}^{\diamond}\right\} }{\mathbb{E}_{\left[\tilde{\mathbf{h}}_{j}^{\diamond}\left|\hat{\mathbf{h}}_{j}\right.j=1,2,\ldots,K\right]}\left\{ L_{k}^{\diamond}\right\} +N_{0}}\nonumber \\
\textrm{s.t.} & & \textrm{tr}\left\{ \mathbf{w}_{k}\mathbf{w}_{k}^{\textrm{H}}\right\} \leq\tilde{P}_{k},\label{eq:Problem_R_SLNRk}\end{aligned}$$
where the signal power $S_{k}^{\diamond}$ and leakage power $L_{k}^{\diamond}$ are defined as $S_{k}^{\diamond}=A_{k}^{\textrm{ave}}\left|\xi_{k}\tilde{\mathbf{h}}_{k}^{\diamond}\mathbf{w}_{k}\right|^{2}$ and $L_{k}^{\diamond}=\sum_{j=1,j\neq k}^{K}A_{j}^{\textrm{ave}}\left|\xi_{j}\tilde{\mathbf{h}}_{j}^{\diamond}\mathbf{w}_{k}\right|^{2}$, respectively. For brevity, we omit the subscription $\left[\tilde{\mathbf{h}}_{j}^{\diamond}\left|\hat{\mathbf{h}}_{j}\right.j=1,2,\ldots,K\right]$ of $\mathbb{E}$ hereafter. According to [@robust_AveSLNR], $\mathbb{E}\left\{ S_{k}^{\diamond}\right\} $ can be computed as
$$\begin{aligned}
\mathbb{E}\left\{ S_{k}^{\diamond}\right\} & = & \xi_{k}^{2}A_{k}^{\textrm{ave}}\mathbf{w}_{k}^{\textrm{H}}\mathbf{U}_{k}\mathbf{w}_{k},\label{eq:E{Sk}}\end{aligned}$$
where $\mathbf{U}_{k}=\left(1-\frac{N\eta}{N-1}\right)\mathbf{\hat{h}}_{k}^{\textrm{H}}\mathbf{\hat{h}}_{k}+\frac{\eta}{N-1}\mathbf{I}_{N}$ and $\eta$ is computed as $\eta=2^{B}\beta\left(2^{B},\frac{N}{N-1}\right)$ [@Limited-bit; @CSI]. Here, $\beta\left(x,y\right)$ is the beta function defined as $\beta\left(x,y\right)=\frac{\Gamma\left(x\right)\Gamma\left(y\right)}{\Gamma\left(x+y\right)}$ [@integral_table; @book], where $\Gamma\left(\cdot\right)$ denotes the gamma function [@integral_table; @book]. It should be noted that $\mathbf{U}_{k}$ is positive definite because it is easy to verify
$$1-\frac{N\eta}{N-1}>0,\quad\textrm{for }N>1,B\geq0.\label{eq:ineq_posi_def}$$
Similar to , $\mathbb{E}\left\{ L_{k}^{\diamond}\right\} $ can be derived as
$$\mathbb{E}\left\{ L_{k}^{\diamond}\right\} =\mathbf{w}_{k}^{\textrm{H}}\mathbf{\bar{U}}_{k}\mathbf{w}_{k},\label{eq:E{Lk}}$$
where $\mathbf{\bar{U}}_{k}=\sum_{j=1,j\neq k}^{K}\xi_{j}^{2}A_{j}^{\textrm{ave}}\mathbf{U}_{j}$. Then the solution to problem can be written in a similar expression as in with $\mathbf{\mathbf{\bar{\check{H}}}_{\mathit{k}}^{\textrm{H}}\mathbf{\bar{\check{H}}}_{\mathit{k}}}$ and $\mathbf{\check{h}}_{k}^{\textrm{H}}\mathbf{\check{h}}_{k}$ respectively replaced by $\mathbf{\bar{U}}_{k}$ and $A_{k}^{\textrm{ave}}\mathbf{U}_{k}$ [@robust_AveSLNR]. The scheme based on problem is named the average SLNR (ASLNR) scheme by the authors of [@robust_AveSLNR].
In [@basic_robust_SLNR], the optimization of SLNR was interpreted as keeping the expectation of leakage power below a threshold while maximizing the expectation of signal power. The optimization problem with regard to $\mathbf{w}_{k}$ can be formulated as
$$\begin{aligned}
\underset{\mathbf{w}_{k}}{\max} & & \mathbb{E}\left\{ S_{k}^{\diamond}\right\} \nonumber \\
\textrm{s.t.} & & \mathbb{E}\left\{ L_{k}^{\diamond}\right\} \leq\gamma_{k},\textrm{ and }\textrm{tr}\left\{ \mathbf{w}_{k}\mathbf{w}_{k}^{\textrm{H}}\right\} \leq\tilde{P}_{k},\label{eq:Problem_LC}\end{aligned}$$
where $\gamma_{k}$ is a design parameter of the leakage power threshold. Problem is a non-convex quadratically constrained quadratic program (QCQP) problem [@vec_randm]. However, it can be transformed to an equivalent SDP problem as
$$\begin{aligned}
\underset{\mathbf{Q}_{k}\in\mathbb{\mathbb{H}}_{N}^{\mathit{+}}}{\max} & & \textrm{tr}\left\{ \mathbf{Q}_{k}\mathbf{U}_{k}\right\} \nonumber \\
\textrm{s.t.} & & \textrm{tr}\left\{ \mathbf{Q}_{k}\mathbf{\bar{U}}_{k}\right\} \leq\gamma_{k},\textrm{ and }\textrm{tr}\left\{ \mathbf{Q}_{k}\right\} \leq\tilde{P}_{k};\nonumber \\
& & \textrm{rank}\left\{ \mathbf{Q}_{k}\right\} =1.\label{eq:Problem_LC_SDP}\end{aligned}$$
Problem is still a non-convex problem due to the non-convex rank-one constraint. Applying the SDP relaxation technique [@vec_randm] by omitting the rank-one constraint, we get
$$\begin{aligned}
\underset{\mathbf{Q}_{k}\in\mathbb{\mathbb{H}}_{N}^{\mathit{+}}}{\max} & & \textrm{tr}\left\{ \mathbf{Q}_{k}\mathbf{U}_{k}\right\} \nonumber \\
\textrm{s.t.} & & \textrm{tr}\left\{ \mathbf{Q}_{k}\mathbf{\bar{U}}_{k}\right\} \leq\gamma_{k},\textrm{ and }\textrm{tr}\left\{ \mathbf{Q}_{k}\right\} \leq\tilde{P}_{k}.\label{eq:Problem_LC_SDPR}\end{aligned}$$
Now problem is a standard convex SDP problem and can be solved efficiently using the mathematical software package [@cvx_software]. According to [@basic_robust_SLNR], the solution of an SDP problem like problem is always rank-one if it has at most three constraints. Since problem has only two constraints, we can conclude that problem and are equivalent. Thus the solution for the original problem can be extracted from the solution to problem .
Obviously, in problem $\gamma_{k}$ should be carefully chosen to make sure that the leakage power constraints are neither too tight nor too loose to achieve a good balance between the maximization of $\mathbb{E}\left\{ S_{k}^{\diamond}\right\} $ and the minimization of $\mathbb{E}\left\{ L_{k}^{\diamond}\right\} $. It is worthwhile to note that recently in [@robust_ProbLeak1] and [@robust_ProbLeak_ICC12], the authors proposed a probabilistic approach in the control of interference leakage, i.e., to keep the event of large leakage below a certain probability. The scheme will be referred to as the probabilistic leakage control (PLC) scheme hereafter. In the PLC scheme, the leakage constraint in problem is regulated as
$$\Pr\left(\mathbb{E}\left\{ L_{k}^{\diamond}\right\} \geq\gamma_{k}\right)\leq p_{k},\label{eq:Prob_leakage_constraint}$$
where $p_{k}$ is a given probability of the event that $\mathbb{E}\left\{ L_{k}^{\diamond}\right\} $ exceeds $\gamma_{k}$. According to [@robust_ProbLeak1] and [@robust_ProbLeak_ICC12], by invoking Markov’s inequality, can be nicely tightened by a new constraint $\mathbb{E}\left\{ L_{k}^{\diamond}\right\} \leq p_{k}\gamma_{k}$, which generates a new problem shown as
$$\begin{aligned}
\underset{\mathbf{w}_{k}}{\max} & & \mathbb{E}\left\{ S_{k}^{\diamond}\right\} \nonumber \\
\textrm{s.t.} & & \mathbb{E}\left\{ L_{k}^{\diamond}\right\} \leq p_{k}\gamma_{k},\textrm{ and }\textrm{tr}\left\{ \mathbf{w}_{k}\mathbf{w}_{k}^{\textrm{H}}\right\} \leq\tilde{P}_{k}.\label{eq:Problem_PLC}\end{aligned}$$
Problem is essentially equivalent to problem if we define $p_{k}\gamma_{k}$ as a new threshold $\tilde{\gamma}_{k}$. Therefore, the issue still remains regarding the appropriate proposal of $\gamma_{k}$.
Minimum Average Leakage Power
-----------------------------
In order to get some insights on the design of $\gamma_{k}$, we will start with the comparison between the average leakage power in problem and that of the non-robust ZF beamforming given by . Our result is summarized in Theorem \[thm:Theorem\_1\].
\[thm:Theorem\_1\]Let $\mathbf{w}_{k}\left(k\neq j\right)$ be a general beamforming vector expressed as $\mathbf{w}_{k}=\sqrt{\tilde{P}_{k}}\mathbf{\tilde{w}}_{k}$, where $\mathbf{\tilde{w}}_{k}$ is the normalized vector of $\mathbf{w}_{k}$. Then for arbitrary CF-CMI $A_{j}$, the expected interference leakage from UE $k$ to $j$ is lower bounded by
$$\begin{aligned}
\mathbb{E}\left\{ \left|\xi_{j}\sqrt{A_{j}}\tilde{\mathbf{h}}_{j}^{\diamond}\mathbf{w}_{k}\right|^{2}\right\} & \geq & \mathbb{E}\left\{ \left|\xi_{j}\sqrt{A_{j}}\tilde{\mathbf{h}}_{j}^{\diamond}\mathbf{w}_{k}^{\textrm{ZF}}\right|^{2}\right\} \nonumber \\
& = & \frac{\tilde{P}_{k}\xi_{j}^{2}A_{j}\eta}{\left(N-1\right)}.\label{eq:Theorem_1}\end{aligned}$$
See Appendix I.
From Theorem \[thm:Theorem\_1\], we can state that the non-robust ZF beamforming scheme is optimal in the sense of generating the minimum amount of average inter-UE interference under limited-bit CDI. In other words, the average leakage power of the non-robust ZF beamforming scheme can serve as a lower bound for the interference leakage in problem .
The Proposed Minimum Average Leakage Control Beamforming with Per-UE Power Constraints
--------------------------------------------------------------------------------------
According to Theorem \[thm:Theorem\_1\], under the assumptions of average CF-CMI and the Rayleigh channel fading, we can set the leakage power threshold $\gamma_{k}$ to the derived lower bound as
$$\gamma_{k}^{\textrm{MALC}}=\frac{\tilde{P}_{k}N\eta}{\left(N-1\right)}\sum_{j=1,j\neq k}^{K}\xi_{j}^{2},\label{eq:gamk_ZF_1_aveCMI}$$
where $A_{j}$ in Theorem \[thm:Theorem\_1\] has been replaced by $A_{j}^{\textrm{ave}}=N$, considering that for Rayleigh fading channels $\left\Vert \mathbf{h}_{j}\right\Vert ^{2}$ follows a chi-squared distribution with $2N$ degrees of freedom and its mean is $N$ [@Proakis; @book]. On the other hand, when the BS has quantized CF-CMI, we can substitute $A_{j}$ with $\hat{A}_{j}$ and choose $\gamma_{k}^{\textrm{MALC}}$ as
$$\gamma_{k}^{\textrm{MALC}}=\frac{\tilde{P}_{k}\eta}{\left(N-1\right)}\sum_{j=1,j\neq k}^{K}\xi_{j}^{2}\hat{A}_{j}.\label{eq:gamk_ZF_2_quantCMI}$$
Then we can re-write problem as
$$\begin{aligned}
\underset{\mathbf{w}_{k}}{\max} & & \mathbb{E}\left\{ S_{k}^{\diamond}\right\} \nonumber \\
\textrm{s.t.} & & \mathbb{E}\left\{ L_{k}^{\diamond}\right\} \leq\gamma_{k}^{\textrm{MALC}},\textrm{ and }\textrm{tr}\left\{ \mathbf{w}_{k}\mathbf{w}_{k}^{\textrm{H}}\right\} \leq\tilde{P}_{k}.\label{eq:Problem_R_ZF}\end{aligned}$$
The beamforming scheme based on problem will be called the minimum average leakage control (MALC) scheme since it controls the expected leakage power according to the minimum average leakage given by Theorem \[thm:Theorem\_1\].
The Proposed Relaxed Average Leakage Control Beamforming with Per-UE Power Constraints
--------------------------------------------------------------------------------------
From Theorem \[thm:Theorem\_1\], we have found the minimum threshold value for the average interference leakage shown in /. Next, we want to raise the leakage threshold so that the weighted sum-rate can be increased. The intuition is that interference minimization, e.g., ZF precoding, is generally not the optimal strategy for throughput maximization or MSE minimization. Allowing some more interference leakage can increase the weighted sum-rate due to signal power boosting. Of course if the tolerated leakage is set to be too large, the rate will inevitably decrease. Our intention is to find an appropriate leakage level, which can generate a good weighted sum-rate performance.
Under the assumption of average CF-CMI, we propose that the upper half segment of the CDF of $\left|\xi_{j}\sqrt{A_{j}^{\diamond}}\tilde{\mathbf{h}}_{j}^{\diamond}\mathbf{w}_{k}^{\textrm{ZF}}\right|^{2}$ should serve as the analytical reference for determining how large $\gamma_{k}$ should be, where $A_{j}^{\diamond}$ is the randomly reconstructed CF-CMI at the BS based on $A_{j}^{\textrm{ave}}$ and the channel-fading model. Our proposal is based on the fact that the CDF of $\left|\xi_{j}\sqrt{A_{j}^{\diamond}}\tilde{\mathbf{h}}_{j}^{\diamond}\mathbf{w}_{k}^{\textrm{ZF}}\right|^{2}$ contains tractable information on the average leakage level of a conservative beamformer $\mathbf{w}_{k}^{\textrm{ZF}}$ with minimum generation of inter-UE interference. As a result, $\gamma_{k}^{\textrm{MALC}}$ will be loosened to another threshold $\gamma_{k}^{\textrm{RALC}}$ and we can design a relaxed average leakage control (RALC) beamforming scheme with per-UE power constraints. Since $\left|\xi_{j}\sqrt{A_{j}^{\diamond}}\tilde{\mathbf{h}}_{j}^{\diamond}\mathbf{w}_{k}^{\textrm{ZF}}\right|^{2}=\tilde{P}_{k}\xi_{j}^{2}\left|\sqrt{A_{j}^{\diamond}}\tilde{\mathbf{h}}_{j}^{\diamond}\mathbf{\tilde{w}}_{k}^{\textrm{ZF}}\right|^{2}$, we can derive the CDF of $\left|\sqrt{A_{j}^{\diamond}}\tilde{\mathbf{h}}_{j}^{\diamond}\mathbf{\tilde{w}}_{k}^{\textrm{ZF}}\right|^{2}$ and scale it by a factor of $\tilde{P}_{k}\xi_{j}^{2}$ to obtain the CDF of $\left|\xi_{j}\sqrt{A_{j}^{\diamond}}\tilde{\mathbf{h}}_{j}^{\diamond}\mathbf{w}_{k}^{\textrm{ZF}}\right|^{2}$. In the following we present our result on the CDF of $\left|\sqrt{A_{j}^{\diamond}}\tilde{\mathbf{h}}_{j}^{\diamond}\mathbf{\tilde{w}}_{k}^{\textrm{ZF}}\right|^{2}$ for Rayleigh fading channels in Theorem \[thm:Theorem\_2\].
\[thm:Theorem\_2\]Let $D=\left|\sqrt{A_{j}^{\diamond}}\tilde{\mathbf{h}}_{j}^{\diamond}\mathbf{\tilde{w}}_{k}^{\textrm{ZF}}\right|^{2}$. Then for Rayleigh fading channels the CDF of $D$ is shown in , where $\textrm{E}_{1}\left(x\right)=\int_{x}^{\infty}\frac{\textrm{e}^{-t}}{t}dt$ is the exponential integral function .
$$\begin{aligned}
P_{D}\left(d\right) & = & 1+\frac{\textrm{e}^{-d}}{\left(N-2\right)!}\left\{ \sum_{n=0}^{N-2}\sum_{m=1}^{2^{B}}\textrm{C}_{N-2}^{n}\textrm{C}_{2^{B}}^{m}\frac{\left(-1\right)^{n+m}m}{mN-\left(m+n\right)}\sum_{l=0}^{N-1-n}l!\textrm{C}_{N-1-n}^{l}d^{N-1-l}\right\} \nonumber \\
& & -\frac{\textrm{e}^{-d}}{\left(N-2\right)!}\left\{ \sum_{n=0}^{N-2}\textrm{C}_{N-2}^{n}\textrm{C}_{2^{B}}^{1}\frac{\left(-1\right)^{n+1}}{N-1-n}d^{N-1}\right\} \nonumber \\
& & -\frac{\textrm{e}^{-d}}{\left(N-2\right)!}\left\{ \sum_{n=0}^{N-2}\sum_{m=2}^{2^{B}}\textrm{C}_{N-2}^{n}\textrm{C}_{2^{B}}^{m}\frac{\left(-1\right)^{n+m}m}{mN-\left(m+n\right)}\left[\sum_{l=1}^{\left(m-1\right)\left(N-1\right)-1}\frac{\left(-1\right)^{l-1}d^{N-1+l}}{l!\textrm{C}_{\left(m-1\right)\left(N-1\right)-1}^{l}}\right]\right\} \nonumber \\
& & -\frac{\textrm{E}_{1}\left(d\right)}{\left(N-2\right)!}\left\{ \sum_{n=0}^{N-2}\sum_{m=2}^{2^{B}}\textrm{C}_{N-2}^{n}\textrm{C}_{2^{B}}^{m}\frac{\left(-1\right)^{n+m}m}{mN-\left(m+n\right)}\frac{\left(-1\right)^{\left(m-1\right)\left(N-1\right)-1}d^{m\left(N-1\right)}}{\left(\left(m-1\right)\left(N-1\right)-1\right)!}\right\} .\label{eq:Theorem_2}\end{aligned}$$
------------------------------------------------------------------------
See Appendix II.
If the BS has quantized CF-CMI $\hat{A}_{j}$, then $A_{j}^{\diamond}$ should take the value of $\hat{A}_{j}$. Consequently, we can get the CDF of $\left|\xi_{j}\sqrt{\hat{A}_{j}}\tilde{\mathbf{h}}_{j}^{\diamond}\mathbf{w}_{k}^{\textrm{ZF}}\right|^{2}$ by scaling that of $\left|\tilde{\mathbf{h}}_{j}^{\diamond}\mathbf{\tilde{w}}_{k}^{\textrm{ZF}}\right|^{2}$ by a factor of $\tilde{P}_{k}\xi_{j}^{2}\hat{A}_{j}$. In Theorem \[thm:Theorem\_3\], we show our result on the CDF of $\left|\tilde{\mathbf{h}}_{j}^{\diamond}\mathbf{\tilde{w}}_{k}^{\textrm{ZF}}\right|^{2}$.
\[thm:Theorem\_3\]Let $V=\left|\tilde{\mathbf{h}}_{j}^{\diamond}\mathbf{\tilde{w}}_{k}^{\textrm{ZF}}\right|^{2}$. Then the CDF of $V$ is
$P_{V}\left(v\right)=1+\left(N-1\right)\times$
$$\sum_{n=0}^{N-2}\sum_{m=1}^{2^{B}}\textrm{C}_{N-2}^{n}\textrm{C}_{2^{B}}^{m}\left(-1\right)^{n+m}m\frac{v^{n}-v^{m\left(N-1\right)}}{mN-\left(m+n\right)}.\label{eq:Theorem_3}$$
See Appendix III.
Based on Theorem \[thm:Theorem\_2\] or \[thm:Theorem\_3\], we can relax the control target of the interference leakage to the $\delta_{k}$ percent point of the leakage power of the non-robust ZF beamforming scheme. To be more specific, considering average CF-CMI, we can set $\gamma_{k}^{\textrm{RALC}}$ according to Theorem \[thm:Theorem\_2\] as
$$\gamma_{k}^{\textrm{RALC}}=\tilde{P}_{k}\sum_{j=1,j\neq k}^{K}\xi_{j}^{2}P_{D}^{-1}\left(\delta_{k}\right),\label{eq:gamk_RZF_1_aveCMI}$$
where $P_{D}^{-1}\left(\delta_{k}\right)=\underset{d}{\arg}\left\{ P_{D}\left(d\right)=\delta_{k}\right\} $. When quantized CF-CMI is available at the BS, we can recall Theorem \[thm:Theorem\_3\] and choose $\gamma_{k}^{\textrm{RALC}}$ as
$$\gamma_{k}^{\textrm{RALC}}=\tilde{P}_{k}\sum_{j=1,j\neq k}^{K}\xi_{j}^{2}\hat{A}_{j}P_{V}^{-1}\left(\delta_{k}\right),\label{eq:gamk_RZF_2_quantCMI}$$
where $P_{V}^{-1}\left(\delta_{k}\right)=\underset{v}{\arg}\left\{ P_{V}\left(v\right)=\delta_{k}\right\} $. Although there is no closed-form expression for the computation of $P_{D}^{-1}\left(\delta_{k}\right)$ or $P_{V}^{-1}\left(\delta_{k}\right)$, it can be conveniently found by the bisection method for a given $\delta_{k}$ since $P_{D}\left(d\right)$ or $P_{V}\left(v\right)$ is a bounded and monotonically increasing function. In the bisection method, considering $P_{D}\left(d\right)$ as an example, first we choose a small value $d_{1}$ and a large one $d_{2}$ to construct an interval $\left[P_{D}\left(d_{1}\right),P_{D}\left(d_{2}\right)\right]$ that contains $\delta_{k}$, then we continually compare $P_{D}\left(\frac{d_{1}+d_{2}}{2}\right)$ with $\delta_{k}$ and update $d_{1}$ or $d_{2}$ with $\frac{d_{1}+d_{2}}{2}$ on condition that $\delta_{k}$ stays inside the interval $\left[P_{D}\left(d_{1}\right),P_{D}\left(d_{2}\right)\right]$. The bisection searching stops when $\left[P_{D}\left(d_{1}\right),P_{D}\left(d_{2}\right)\right]$ is sufficiently narrow, and then we can obtain $\delta_{k}=\frac{d_{1}+d_{2}}{2}$. As for the choice of $\delta_{k}$ in the proposed RALC scheme, since we want to get a relaxed leakage threshold compared with $\gamma_{k}^{\textrm{MALC}}$, it is intuitive to look beyond $\gamma_{k}^{\textrm{MALC}}$ and hinge the leakage power threshold to some point on the upper half segment of $P_{D}\left(d\right)$ or $P_{V}\left(v\right)$, i.e.,
$$\delta_{k}\in\begin{cases}
\left(P_{D}\left(\frac{\gamma_{k}^{\textrm{MALC}}}{\tilde{P}_{k}\sum_{j\neq k}\xi_{j}^{2}}\right),1\right), & \textrm{for }\gamma_{k}^{\textrm{MALC}}\textrm{ from }\eqref{eq:gamk_ZF_1_aveCMI};\\
\left(P_{V}\left(\frac{\gamma_{k}^{\textrm{MALC}}}{\tilde{P}_{k}\sum_{j\neq k}\xi_{j}^{2}\hat{A}_{j}}\right),1\right), & \textrm{for }\gamma_{k}^{\textrm{MALC}}\textrm{ from }\eqref{eq:gamk_ZF_2_quantCMI}.
\end{cases}$$
Based on or , the original problem becomes
$$\begin{aligned}
\underset{\mathbf{w}_{k}}{\max} & & \mathbb{E}\left\{ S_{k}^{\diamond}\right\} \nonumber \\
\textrm{s.t.} & & \mathbb{E}\left\{ L_{k}^{\diamond}\right\} \leq\gamma_{k}^{\textrm{RALC}},\textrm{ and }\textrm{tr}\left\{ \mathbf{w}_{k}\mathbf{w}_{k}^{\textrm{H}}\right\} \leq\tilde{P}_{k}.\label{eq:Problem_R_RZF}\end{aligned}$$
The beamforming scheme based on problem will be referred to as the relaxed average leakage control (RALC) scheme because it relaxes the leakage threshold compared with the MALC scheme. It should be noted that though $\delta_{k}$ is a design parameter chosen to allow more interference leakage, the leakage threshold can be immediately found in an efficient way once the system parameters $\left\{ N,K,B,\delta_{k}\right\} $ are provided. On the other hand, the PLC scheme [@robust_ProbLeak1], [@robust_ProbLeak_ICC12] employs an empirical method to determine the leakage threshold, which needs to exhaustively search the appropriate leakage threshold by a great amount of simulations for each parameter set $\left\{ N,K,B\right\} $.
The Proposed Robust Beamforming Schemes with Per-antenna Power Constraints\[sub:The-Proposed-Scheme\]
-----------------------------------------------------------------------------------------------------
Problems and with the proposed leakage power thresholds are subject to power constraints on a per-UE basis. However, the appropriate choices of $\tilde{P}_{k}$s for those problems that can optimize the system performance measure, e.g., the weighted sum-rate, under per-antenna power constraints are still unclear. In this paper, we propose a two-stage algorithm to alternately update the per-UE power allocations and beamforming vectors in order to maximize the expected weighted sum-rate performance under per-antenna power constraints.
Suppose that the precoding matrix in the $l$-th step is $\mathbf{W}^{\left(l\right)}=\left[\sqrt{\tilde{P}_{1}^{\left(l\right)}}\tilde{\mathbf{w}}_{1}^{\left(l\right)},\sqrt{\tilde{P}_{2}^{\left(l\right)}}\tilde{\mathbf{w}}_{2}^{\left(l\right)},\ldots,\sqrt{\tilde{P}_{K}^{\left(l\right)}}\mathbf{\tilde{w}}_{K}^{\left(l\right)}\right]$, where $\tilde{P}_{k}^{\left(l\right)}$ and $\tilde{\mathbf{w}}_{k}^{\left(l\right)}$ are respectively the beamforming power and normalized beamforming vector for UE $k$. Then the straightforward way to optimize the expected weighted sum-rate is to find new $\tilde{P}_{k}^{\left(l+1\right)}$ and $\tilde{\mathbf{w}}_{k}^{\left(l+1\right)}$ that can maximize $f\left(\mathbf{W}^{\left(l+1\right)}\right)=\mathbb{E}\left\{ \sum_{k=1}^{K}\alpha_{k}\log_{2}\left(1+\frac{S_{k}^{\diamond\left(l+1\right)}}{I_{k}^{\diamond\left(l+1\right)}+N_{0}}\right)\right\} $, where $S_{k}^{\diamond\left(l+1\right)}$ and $I_{k}^{\diamond\left(l+1\right)}$ are computed as $S_{k}^{\diamond\left(l+1\right)}=A_{k}^{\textrm{ave}}\left|\xi_{k}\tilde{\mathbf{h}}_{k}^{\diamond}\sqrt{\tilde{P}_{k}^{\left(l+1\right)}}\tilde{\mathbf{w}}_{k}^{\left(l+1\right)}\right|^{2}$ and $I_{k}^{\diamond\left(l+1\right)}=\sum_{j=1,j\neq k}^{K}A_{k}^{\textrm{ave}}\left|\xi_{k}\tilde{\mathbf{h}}_{k}^{\diamond}\sqrt{\tilde{P}_{j}^{\left(l+1\right)}}\tilde{\mathbf{w}}_{j}^{\left(l+1\right)}\right|^{2}$, respectively.
*1) Updating Per-UE Powers*
First, we concentrate on the update of $\mathbf{\widetilde{P}}^{\left(l+1\right)}=\left(\tilde{P}_{1}^{\left(l+1\right)},\tilde{P}_{2}^{\left(l+1\right)},\ldots,\tilde{P}_{K}^{\left(l+1\right)}\right)$ with the objective to maximize $f\left(\mathbf{W}^{\left(l+1\right)}\right)$ on condition of $\tilde{\mathbf{w}}_{k}^{\left(l+1\right)}=\tilde{\mathbf{w}}_{k}^{\left(l\right)}$. However, $f\left(\mathbf{W}^{\left(l+1\right)}\right)$ is hard to handle because it has no explicit expression. Instead, we treat $\tilde{f}\left(\mathbf{W}^{\left(l+1\right)}\right)\left|_{\tilde{\mathbf{w}}_{k}^{\left(l+1\right)}=\tilde{\mathbf{w}}_{k}^{\left(l\right)}}\right.$, where $\tilde{f}\left(\mathbf{W}^{\left(l+1\right)}\right)=\sum_{k=1}^{K}\alpha_{k}\log_{2}\left(1+\frac{\mathbb{E}\left\{ S_{k}^{\diamond\left(l+1\right)}\right\} }{\mathbb{E}\left\{ I_{k}^{\diamond\left(l+1\right)}\right\} +N_{0}}\right)$, as an approximated metric of the weighted sum-rate. In a similar way as in , we can obtain
$$\mathbb{E}\left\{ S_{k}^{\diamond\left(l+1\right)}\left|_{\tilde{\mathbf{w}}_{k}^{\left(l+1\right)}=\tilde{\mathbf{w}}_{k}^{\left(l\right)}}\right.\right\} =\tilde{P}_{k}^{\left(l+1\right)}\lambda_{k,k}^{\left(l\right)},\label{eq:E{Sk(l+1)}}$$
where $\lambda_{k,k}^{\left(l\right)}=\xi_{k}^{2}A_{k}^{\textrm{ave}}\tilde{\mathbf{w}}_{k}^{\left(l\right)\textrm{H}}\mathbf{U}_{k}\tilde{\mathbf{w}}_{k}^{\left(l\right)}$. Also we can get
$$\mathbb{E}\left\{ I_{k}^{\diamond\left(l+1\right)}\left|_{\tilde{\mathbf{w}}_{k}^{\left(l+1\right)}=\tilde{\mathbf{w}}_{k}^{\left(l\right)}}\right.\right\} =\sum_{j=1,j\neq k}^{K}\tilde{P}_{j}^{\left(l+1\right)}\lambda_{j,k}^{\left(l\right)},\label{eq:E{Ik(l)}}$$
where $\lambda_{j,k}^{\left(l\right)}=\xi_{k}^{2}A_{k}^{\textrm{ave}}\tilde{\mathbf{w}}_{j}^{\left(l\right)\textrm{H}}\mathbf{U}_{k}\tilde{\mathbf{w}}_{j}^{\left(l\right)}$. Note that $\lambda_{j,k}^{\left(l\right)}>0$ because of . Then with some mathematical manipulation, the optimization problem to maximize $\tilde{f}\left(\mathbf{W}^{\left(l+1\right)}\right)\left|_{\tilde{\mathbf{w}}_{k}^{\left(l+1\right)}=\tilde{\mathbf{w}}_{k}^{\left(l\right)}}\right.$ with leakage control and per-antenna power constraints can be formulated as
$$\begin{aligned}
\underset{\mathbf{\widetilde{P}}^{\left(l+1\right)}}{\min} & & \prod_{k=1}^{K}\left(\frac{\sum_{j=1,j\neq k}^{K}\tilde{P}_{j}^{\left(l+1\right)}\lambda_{j,k}^{\left(l\right)}+N_{0}}{\sum_{j=1}^{K}\tilde{P}_{j}^{\left(l+1\right)}\lambda_{j,k}^{\left(l\right)}+N_{0}}\right)^{\alpha_{k}}\nonumber \\
\textrm{s.t.} & & \mathbb{E}\left\{ L_{k}^{\diamond\left(l+1\right)}\left|_{\tilde{\mathbf{w}}_{k}^{\left(l+1\right)}=\tilde{\mathbf{w}}_{k}^{\left(l\right)}}\right.\right\} \leq\gamma_{k}^{\left(l\right)},\quad\forall k;\nonumber \\
& & \left[\sum_{k=1}^{K}\tilde{P}_{k}^{\left(l+1\right)}\tilde{\mathbf{w}}_{k}^{\left(l\right)}\tilde{\mathbf{w}}_{k}^{\left(l\right)\textrm{H}}\right]_{n,n}\leq P_{n},\quad\forall n,\nonumber \\
& & \tilde{P}_{k}^{\left(l+1\right)}\geq0,\quad\forall k,\label{eq:Problem_power_update_GP_nonstand}\end{aligned}$$
where $\mathbb{E}\left\{ L_{k}^{\diamond\left(l+1\right)}\left|_{\tilde{\mathbf{w}}_{k}^{\left(l+1\right)}=\tilde{\mathbf{w}}_{k}^{\left(l\right)}}\right.\right\} =\sum_{j=1,j\neq k}^{K}\tilde{P}_{k}^{\left(l+1\right)}\lambda_{k,j}^{\left(l\right)}$ and $\gamma_{k}^{\left(l\right)}$ is the proposed leakage power threshold in the $l$-th step and it can be updated based on $\gamma_{k}^{\textrm{MALC}}$ or $\gamma_{k}^{\textrm{RALC}}$ with $\tilde{P}_{k}$ replaced by $\tilde{P}_{k}^{\left(l\right)}$ in / or in /. Besides, the second set of constraints reflects the per-antenna power control requirement. Unfortunately, problem is non-convex, and searching of the global optimal solution is of high-complexity. To our best knowledge, till now geometric programming (GP) is one of the most efficient methods to solve this specific power allocation problem [@GP_problem]. According to [@GP_problem], the denominator of the objective function in problem can be lower-bounded by the geometric inequality shown as
$$\prod_{k=1}^{K}\left(\sum_{j=1}^{K}\tilde{P}_{j}^{\left(l+1\right)}\lambda_{j,k}^{\left(l\right)}+N_{0}\right)^{\alpha_{k}}\geq\prod_{k=1}^{K}\prod_{j=0}^{K}\left(\frac{m_{j,k}}{\mu_{j,k}}\right)^{\mu_{j,k}\alpha_{k}},\label{eq:ineq_GP_LB}$$
where $m_{j,k}=\begin{cases}
N_{0}, & j=0\\
\tilde{P}_{j}^{\left(l+1\right)}\lambda_{j,k}^{\left(l\right)}, & j=1,2,\ldots,K
\end{cases}$ and $\mu_{j,k}=\frac{m_{j,k}}{\sum_{j=1}^{K}m_{j,k}}.$ Substituting the denominator of the objective function in problem with the right-hand side in , we can obtain a standard GP problem formulated as
$$\begin{aligned}
\underset{\mathbf{\widetilde{P}}^{\left(l+1\right)}}{\min} & & \frac{\prod_{k=1}^{K}\left(\sum_{j=0,j\neq k}^{K}m_{j,k}\right)^{\alpha_{k}}}{\prod_{k=1}^{K}\prod_{j=0}^{K}\left(\frac{m_{j,k}}{\mu_{j,k}}\right)^{\mu_{j,k}\alpha_{k}}}\nonumber \\
\textrm{s.t.} & & \frac{\mathbb{E}\left\{ L_{k}^{\diamond\left(l+1\right)}\left|_{\tilde{\mathbf{w}}_{k}^{\left(l+1\right)}=\tilde{\mathbf{w}}_{k}^{\left(l\right)}}\right.\right\} }{\gamma_{k}^{\left(l\right)}}\leq1,\quad\forall k;\nonumber \\
& & \frac{\left[\sum_{k=1}^{K}\tilde{P}_{k}^{\left(l+1\right)}\tilde{\mathbf{w}}_{k}^{\left(l\right)}\tilde{\mathbf{w}}_{k}^{\left(l\right)\textrm{H}}\right]_{n,n}}{P_{n}}\leq1,\quad\forall n,\nonumber \\
& & \tilde{P}_{k}^{\left(l+1\right)}\geq0,\quad\forall k.\label{eq:Problem_power_update_GP_stand}\end{aligned}$$
Problem can be solved iteratively by using the mathematical software [@cvx_software] to obtain $\mathbf{\widetilde{P}}^{\left(l+1\right)}$ followed by updating $\left\{ \mu_{j,k}\right\} $ to form the new optimization problem regarding $\mathbf{\widetilde{P}}^{\left(l+1\right)}$ [@GP_problem]. The iteration is terminated when the power difference metric $PD\textrm{\_}metric=\frac{\left|\mathbf{\widetilde{P}}_{\left(i\right)}^{\left(l+1\right)}-\mathbf{\widetilde{P}}_{\left(i-1\right)}^{\left(l+1\right)}\right|}{\left|\mathbf{\widetilde{P}}_{\left(i-1\right)}^{\left(l+1\right)}\right|}$ falls below a pre-determined threshold $\epsilon$, where $i$ is the iteration index in the GP algorithm. Note that according to [@GP_problem], the GP algorithm can handle problem very efficiently and output a nearly-optimal solution to the original problem .
*2) Updating Per-UE Beamforming Vectors*
Next, we fix $\mathbf{\widetilde{P}}^{\left(l+1\right)}$ and update $\left\{ \mathbf{w}_{k}^{\left(l+1\right)}\right\} $. Here, we propose to add per-antenna power constraints into problem and , and optimize the beamforming vectors in an order according to the order of $\tilde{P}_{k}^{\left(l+1\right)}$. To be more specific, suppose that $\left\{ \tilde{P}_{k}^{\left(l+1\right)}\right\} $ is arranged in a descending order as $\tilde{P}_{\pi\left(1\right)}^{\left(l+1\right)}\geq\tilde{P}_{\pi\left(2\right)}^{\left(l+1\right)}\geq\ldots\geq\tilde{P}_{\pi\left(K\right)}^{\left(l+1\right)}$, then $\mathbf{w}_{\pi\left(k\right)}^{\left(l+1\right)}$ shall be the $k$-th beamforming vector for optimization and the power of the $n$-th antenna should be lower than the per-antenna power headroom left by the previous $k-1$ UEs, which is
$$u_{n,\pi\left(k\right)}^{\left(l+1\right)}=P_{n}-\sum_{i=1}^{k-1}\left[\mathbf{w}_{\pi\left(i\right)}^{\left(l+1\right)}\mathbf{w}_{\pi\left(i\right)}^{\left(l+1\right)\textrm{H}}\right]_{n,n}.\label{eq:power_headroom_update}$$
The philosophy behind our design is that, a larger $\tilde{P}_{k}^{\left(l+1\right)}$ generally indicates greater significance of treating the beamforming vector of UE $k$ in the maximization of the weighed sum-rate, and hence $\mathbf{w}_{k}^{\left(l+1\right)}$ should be optimized sequentially according to the order $\pi\left(k\right)$. Based on the above discussion, we can re-formulate problem and as
$$\begin{aligned}
\underset{\mathbf{w}_{\pi\left(k\right)}^{\left(l+1\right)}}{\max} & & \mathbb{E}\left\{ S_{\pi\left(k\right)}^{\diamond\left(l+1\right)}\right\} \nonumber \\
\textrm{s.t.} & & \mathbb{E}\left\{ L_{\pi\left(k\right)}^{\diamond\left(l+1\right)}\right\} \leq\gamma_{\pi\left(k\right)}^{\left(l+1\right)};\nonumber \\
& & \textrm{tr}\left\{ \mathbf{w}_{\pi\left(k\right)}^{\left(l+1\right)}\mathbf{w}_{\pi\left(k\right)}^{\left(l+1\right)\textrm{H}}\right\} \leq\tilde{P}_{\pi\left(k\right)}^{\left(l+1\right)};\nonumber \\
& & \left[\mathbf{w}_{\pi\left(k\right)}^{\left(l+1\right)}\mathbf{w}_{\pi\left(k\right)}^{\left(l+1\right)\textrm{H}}\right]_{n,n}\leq u_{n,\pi\left(k\right)}^{\left(l+1\right)},\quad\forall n,\label{eq:Problem_LC_PA}\end{aligned}$$
where $\gamma_{\pi\left(k\right)}^{\left(l+1\right)}$ is computed in a similar way as $\gamma_{k}^{\left(l\right)}$ with $\tilde{P}_{k}^{\left(l\right)}$ replaced by $\tilde{P}_{\pi\left(k\right)}^{\left(l+1\right)}$. In order to solve problem , we transform it to an SDP problem similar to problem , and then we apply the SDP relaxation technique [@vec_randm] by dropping the rank-one constraints so as to get the following convex SDP problem.
$$\begin{aligned}
\underset{\mathbf{Q}_{\pi\left(k\right)}^{\left(l+1\right)}\in\mathbb{\mathbb{H}}_{N}^{\mathit{+}}}{\max} & & \textrm{tr}\left\{ \mathbf{Q}_{\pi\left(k\right)}^{\left(l+1\right)}\mathbf{U}_{\pi\left(k\right)}\right\} \nonumber \\
\textrm{s.t.} & & \textrm{tr}\left\{ \mathbf{Q}_{\pi\left(k\right)}^{\left(l+1\right)}\bar{\mathbf{U}}_{\pi\left(k\right)}\right\} \leq\gamma_{\pi\left(k\right)}^{\left(l+1\right)};\nonumber \\
& & \textrm{tr}\left\{ \mathbf{Q}_{\pi\left(k\right)}^{\left(l+1\right)}\right\} \leq\tilde{P}_{\pi\left(k\right)}^{\left(l+1\right)};\nonumber \\
& & \left[\mathbf{Q}_{\pi\left(k\right)}^{\left(l+1\right)}\right]_{n,n}\leq u_{n,\pi\left(k\right)}^{\left(l+1\right)},\quad\forall n.\label{eq:Problem_LC_PA_SDPR}\end{aligned}$$
Problem has $N+2$ constraints, which is larger than three for meaningful cases with $N>1$. Thereby, it is not guaranteed that the solution $\mathbf{Q}_{\pi\left(k\right)}^{\left(l+1\right)}$ is always rank-one [@basic_robust_SLNR]. Here we resort to the randomization technique [@vec_randm] and obtain an approximate solution with rank-one matrix. Denote the solution to problem as $\mathbf{Q}_{\pi\left(k\right)}^{\left(l+1\right)\textrm{*}}$ and suppose that $\textrm{rank}\left\{ \mathbf{Q}_{\pi\left(k\right)}^{\left(l+1\right)\textrm{*}}\right\} >1$, then we generate a random vector $\tilde{\mathbf{q}}_{\pi\left(k\right)}^{\left(l+1\right)}\sim\mathcal{N}\left(\mathbf{0},\mathbf{Q}_{\pi\left(k\right)}^{\left(l+1\right)\textrm{*}}\right)$ and scale it by a factor $\rho$ to make sure that all the constraints in problem are satisfied, i.e., $\mathbf{q}_{\pi\left(k\right)}^{\left(l+1\right)}=\rho\tilde{\mathbf{q}}_{\pi\left(k\right)}^{\left(l+1\right)}.$ The vector randomization process is repeated by $L_{\textrm{rand}}$ times and we select the vector with the largest performance measure for problem as the solution to problem , i.e.,
$$\mathbf{w}_{\pi\left(k\right)}^{\left(l+1\right)}=\underset{\mathbf{q}_{\pi\left(k\right)}^{\left(l+1\right),\left(i\right)},i\in\left\{ 1,2,\ldots,L_{\textrm{rand}}\right\} }{\arg\max}\left(\textrm{tr}\left\{ \mathbf{Q}_{\pi\left(k\right)}^{\left(l+1\right),\left(i\right)}\mathbf{U}_{\pi\left(k\right)}\right\} \right),\label{eq:wk_update}$$
where $\mathbf{Q}_{\pi\left(k\right)}^{\left(l+1\right),\left(i\right)}=\mathbf{q}_{\pi\left(k\right)}^{\left(l+1\right),\left(i\right)}\left(\mathbf{q}_{\pi\left(k\right)}^{\left(l+1\right),\left(i\right)}\right)^{\textrm{H}}$.
*3) The Proposed Iterative Algorithm*
With the updated $\mathbf{w}_{\pi\left(k\right)}^{\left(l+1\right)}$, we can compute $\tilde{f}\left(\mathbf{W}^{\left(l+1\right)}\right)$ as the performance measure for the $\left(l+1\right)$-th step. In order to find a locally optimal solution, the two-stage algorithm can be iteratively operated with $L_{\textrm{algo1}}$ times and the beamforming vectors associated with the largest performance measure will be output as the final solutions. The proposed robust beamforming with per-antenna power constraints based on threshold $\gamma_{k}^{\textrm{MALC}}$ or $\gamma_{k}^{\textrm{RALC}}$ will be respectively referred to as the MALC-PA or RALC-PA scheme in the following, which is summarized in Algorithm \[alg:algo\_MALC-RALC\]. In Algorithm \[alg:algo\_MALC-RALC\], we initialize the beamforming vectors as those of the ZF-PA scheme for simplicity.
**
$\textrm{{\color{white}xx}}$****: $\left\{ \mathbf{w}_{k}^{\textrm{algo1}}\right\} $
****$\textrm{{\color{white}x}}$Initialize $\mathbf{w}_{k}^{\textrm{algo1}}=\mathbf{w}_{k}^{\textrm{ZF-PA}}$ $\left(k\in\left\{ 1,2,\ldots,K\right\} \right)$ using ;\
****$\textrm{{\color{white}x}}$Obtain $\tilde{P}_{k}^{\left(0\right)}$ and $\tilde{\mathbf{w}}_{k}^{\left(0\right)}$ based on $\mathbf{w}_{k}^{\textrm{algo1}}$;\
****$\textrm{{\color{white}x}}$Set $perf\textrm{\_}metric=0$;\
****$\textrm{{\color{white}x}}$**** $l:=0$ to **$L_{\textrm{algo1}}$** ******\
****$\textrm{{\color{white}xxx}}$Fix $\tilde{\mathbf{w}}_{k}^{\left(l\right)}$ and update $\tilde{P}_{k}^{\left(l+1\right)}$ by solving problem ;\
****$\textrm{{\color{white}xxx}}$Sort $\tilde{P}_{k}^{\left(l+1\right)}$ in a descendant order $\pi\left(k\right)$;\
****$\textrm{{\color{white}xxx}}$Fix $\tilde{P}_{k}^{\left(l+1\right)}$ and update $\mathbf{w}_{\pi\left(k\right)}^{\left(l+1\right)}$ by solving problem ;\
****$\textrm{{\color{white}xxx}}$Compute the performance measure $\tilde{f}\left(\mathbf{W}^{\left(l+1\right)}\right)$;\
****$\textrm{{\color{white}xxx}}$**** $\tilde{f}\left(\mathbf{W}^{\left(l+1\right)}\right)>perf\textrm{\_}metric$ ****\
****$\textrm{{\color{white}xxx}}$$\textrm{{\color{white}xxx}}\textrm{Update}$ $perf\textrm{\_}metric=\tilde{f}\left(\mathbf{W}^{\left(l+1\right)}\right)$;\
****$\textrm{{\color{white}xxx}}$$\textrm{{\color{white}xxx}}\textrm{Update}$ $\mathbf{w}_{k}^{\textrm{algo1}}=\mathbf{w}_{k}^{\left(l+1\right)}$, $\left(k\in\left\{ 1,2,\ldots,K\right\} \right)$;\
**** $l=l+1$;
Simulation Results and Discussions
==================================
In this section, we present simulation results to compare the performance of the interested beamforming schemes. In our simulations, the system parameters are configured as $\left(N,K\right)=\left(4,4\right)$, $\left\{ \alpha_{k}\right\} =\left[\frac{3}{2},\frac{3}{2},1,1\right]$, $\xi_{k}=1$ and $P_{k}=\frac{P}{N}$. In the proposed MALC-PA and RALC-PA schemes, $\delta_{k}$ equals to 0.8 for the computation of $\gamma_{k}^{\textrm{MALC}}$ and $\gamma_{k}^{\textrm{RALC}}$. Besides, $L_{\textrm{rand}}=1000$, $\epsilon=0.01$ and $L_{\textrm{algo1}}=1\sim10$. Moreover, we define the BS’s SNR as $SNR=P/N_{0}$. All channels are assumed to experience uncorrelated Rayleigh fading and the entries of $\mathbf{h}_{k}$ are i.i.d. ZMCSCG random variables with unit variance. The results are averaged over 5,000 independent channel realizations.
Verifications of Theorem 2 and 3
--------------------------------
Before discussing the performance of various beamforming schemes, in Fig. \[fig:verification\_P(d)\] and \[fig:verification\_P(v)\] we respectively compare the numerical results of $P_{D}\left(d\right)$ and $P_{V}\left(v\right)$ with our analytical results in Theorem \[thm:Theorem\_2\] and \[thm:Theorem\_3\], when $\left(N,B\right)=\left(2,2\right)$, $\left(2,4\right)$, $\left(4,4\right)$ or $\left(4,6\right)$. As can be seen from Fig. \[fig:verification\_P(d)\] and \[fig:verification\_P(v)\], the simulation results perfectly agree with our analysis, which provides theoretical foundations for the parameter configuration of $\gamma_{k}^{\textrm{MALC}}$ and $\gamma_{k}^{\textrm{RALC}}$ in the proposed MALC-PA and RALC-PA schemes.
![\[fig:verification\_P(d)\]Simulation and analytical results of $P_{D}\left(d\right)$ with different $\left(N,B\right)$.](Fig2.eps){width="7.8cm"}
![\[fig:verification\_P(v)\]Simulation and analytical results of $P_{V}\left(v\right)$ with different $\left(N,B\right)$.](Fig3.eps){width="7.8cm"}
Average Weighted Sum-rate Performance with Per-UE Power Constraints
-------------------------------------------------------------------
In this sub-section, we compare the average weighted sum-rate performance of the ZF, SLNR, ASLNR, PLC schemes, together with the proposed MALC and RALC schemes since they are all based on per-UE power constraints. For simplicity, equal power allocation among UEs i.e., $\tilde{P}_{k}=\frac{P}{K}$, is employed. Besides, we assume average CF-CMI and 6-bit CDI. For the PLC scheme, $\gamma_{k}$ and $p_{k}$ are respectively set to 0.9 and 0.05 as in [@robust_ProbLeak_ICC12]. The performance is exhibited in Fig. \[fig:Perf\_perUEPC\], from which we can see that the ZF scheme gives the performance lower bound and the SLNR scheme performs rather poorly in high SNR regime while the ASLNR scheme manages to recover a large portion of the performance loss for the SLNR scheme. Though the PLC scheme exhibits comparable performance with the proposed MALC scheme when the SNR is high, the proposed MALC and RALC schemes significantly outperform the PLC scheme in low to medium SNR regimes, which shows the great significance of appropriate choices of the leakage threshold. It is interesting to note that the proposed MALC scheme performs better than the proposed RALC scheme in high SNR regime since inter-UE interference dominates the performance when the SNR is large and the strategy of minimizing the average leakage prevails. The opposite observation can be drawn for low SNR regime. It should also be noted that although the proposed MALC and RALC schemes only show comparable performance with the ASLNR scheme respectively in high and low SNR regimes, they are more flexible in handling the per-antenna power constraints, which leads to the proposed MALC-PA and RALC-PA schemes to be evaluated in the following sub-sections.
![\[fig:Perf\_perUEPC\]Average weighted sum-rate performance with per-UE power constraints.](Fig4.eps){width="7.8cm"}
Impact of CF-CMI quantization on the Performance
------------------------------------------------
The average weighted sum-rate performance with different bits of CF-CMI quantization is provided in Fig. \[fig:weighted\_sumrate\_diffCMI\] for the ZF-PA scheme and the proposed MALC-PA and RALC-PA schemes to show the impact of CF-CMI quantization on the system performance. We assume $L_{\textrm{algo1}}=10$, 6-bit CDI and average/2-bit/perfect CF-CMI in our simulations. For the perfect CF-CMI case, $A_{k}^{\textrm{ave}}$ is replaced by $\left\Vert \mathbf{h}_{k}\right\Vert ^{2}$ in corresponding formulations. As can be seen from Fig. \[fig:weighted\_sumrate\_diffCMI\], the performance curves of the average CF-CMI are very close to those of the perfect CF-CMI. Besides, the 2-bit quantized CF-CMI is sufficient to achieve almost the same performance as the perfect CF-CMI. Considering the minor effectiveness of quantizing the CF-CMI on the performance, we only consider the average CF-CMI case in the following simulations.
![\[fig:weighted\_sumrate\_diffCMI\]Performance comparison with different bits of CF-CMI quantization.](Fig5.eps){width="7.8cm"}
Convergence of the Proposed Algorithm
-------------------------------------
In this sub-section, we investigate the convergence behavior of the proposed Algorithm \[alg:algo\_MALC-RALC\]. We assume average CF-CMI, 6-bit CDI and $SNR=$ 10 or 20 dB. First, we check the convergence of the power updating based on the GP algorithm addressed in sub-section \[sub:The-Proposed-Scheme\]. The mean value of $PD\textrm{\_}metric$ when $L_{\textrm{algo1}}=1$ is plotted in dB scale in Fig. \[fig:GP Convergence\], from which we can find that the per-UE power allocation coverges rapidly, e.g., only 5\~15 iterations are needed for $PD\textrm{\_}metric<\epsilon$ depending on the working SNR.
![\[fig:GP Convergence\]Convergence of the power updating based on GP.](Fig6.eps){width="7.8cm"}
Next, we investigate the convergence of the proposed Algorithm \[alg:algo\_MALC-RALC\]. In Fig. \[fig:Algorithm Convergence\], we plot the mean value of $perf\textrm{\_}metric$ defined in Algorithm \[alg:algo\_MALC-RALC\] when $L_{\textrm{algo1}}\in\left\{ 1,2,\ldots,10\right\} $. As can be seen from Fig. \[fig:Algorithm Convergence\], both the MALC-PA and RALC-PA schemes quickly converge to their final solutions after merely 2 or 3 iterations, which makes the proposed beamforming schemes feasible for practical uses.
![\[fig:Algorithm Convergence\]Convergence of the Algorithm \[alg:algo\_MALC-RALC\].](Fig7.eps){width="7.8cm"}
Average Weighted Sum-rate Performance with Per-antenna Power Constraints
------------------------------------------------------------------------
In Fig. \[fig:weighted\_sumrate\_averageCMI\], we show the average weighted sum-rate performance with per-antenna power constraints of the proposed schemes with $L_{\textrm{algo1}}=3$ as well as the ZF-PA scheme for the cases of average CF-CMI and 6/12-bit CDI. To show the upper bound of the performance, we also plot the performance of the ZF-PA scheme with perfect CSI. As can be seen from Fig. \[fig:weighted\_sumrate\_averageCMI\], the proposed MALC-PA and RALC-PA schemes achieve considerably larger weighted sum-rate than the ZF-PA scheme when the SNR is medium to high, and their performance approaches that of the ZF-PA scheme with perfect CSI at a relatively fast pace. It should be noted that the RALC-PA scheme shows its advantage over the MALC-PA scheme in all SNR regimes. This is because that the interference minimization strategy is generally not an optimal one and tolerating an appropriate amount of inter-UE interference with proper UE power allocation is beneficial to optimize the performance objective. Also it should be noted that $L_{\textrm{algo1}}$ is set to 3 because the proposed algorithm converges rather quickly with regard to $L_{\textrm{algo1}}$ as have been shown in Fig. \[fig:Algorithm Convergence\].
![\[fig:weighted\_sumrate\_averageCMI\]Average weighted sum-rate performance with per-antenna power constraints.](Fig8.eps){width="7.8cm"}
As future works, closed-form evaluation of the objective function in the optimization problem, analysis on the relationship between the performance and feedback size, the impact of CDI feedback delay, as well as more practical fading channel model and non-RVQ CDI codebook will be considered for the proposed beamforming schemes. In addition, the extensions to more sophisticated systems such as MIMO relay networks and multi-cell cooperative broadcast channels will be investigated in the future.
Conclusions
===========
In this paper, leakage-based robust beamforming for MU-MISO system with per-antenna power constraints and quantized CDI is studied. Based on our derived CDF of the leakage power for the non-robust ZF beamforming scheme, we propose the MALC-PA and RALC-PA schemes using a two-stage algorithm to alternately update the per-UE power allocations and beamforming vectors in order to maximize the expected weighted sum-rate performance under per-antenna power constraints. Simulation results show that the proposed schemes can achieve better performance than the ZF-PA scheme in terms of average weighted sum-rate performance.
Appendix I: Proof of Theorem 1 {#appendix-i-proof-of-theorem-1 .unnumbered}
==============================
We decompose $\mathbf{w}_{k}\left(k\neq j\right)$ as $\mathbf{w}_{k}=\sqrt{\tilde{P}_{k}}\mathbf{\tilde{w}}_{k}=\sqrt{\tilde{P}_{k}}\left(\beta_{k,j}\mathbf{\hat{h}}_{j}^{\textrm{H}}+\sqrt{1-\left|\beta_{k,j}\right|^{2}}\mathbf{v}_{k,j}\right)$, where $\beta_{k,j}=\mathbf{\hat{h}}_{j}\mathbf{\tilde{w}}_{k}$ and $\mathbf{v}_{k,j}$ is a unit-norm random vector representing the projection of $\mathbf{\tilde{w}}_{k}$ onto the nullspace of $\mathbf{\hat{h}}_{j}^{\textrm{H}}$. Then, considering , we can take the expectation of $\left|\tilde{\mathbf{h}}_{j}^{\diamond}\mathbf{w}_{k}\right|^{2}$ with respect to $\tilde{\mathbf{h}}_{j}^{\diamond}$ and $\mathbf{\tilde{w}}_{k}$, and get
$\mathbb{E}\left\{ \left|\tilde{\mathbf{h}}_{j}^{\diamond}\mathbf{w}_{k}\right|^{2}\right\} $
$=\tilde{P}_{k}\mathbb{E}\left\{ \left(1-Z\right)\left|\beta_{k,j}\right|^{2}+Z\left(1-\left|\beta_{k,j}\right|^{2}\right)\left|\mathbf{e}_{j}^{\diamond}\mathbf{v}_{k,j}\right|^{2}\right\} $
$\;+\tilde{P}_{k}\mathbb{E}\left\{ 2\textrm{Re}\left\{ \sqrt{\left(1-Z\right)Z\left(1-\left|\beta_{k,j}\right|^{2}\right)}\beta_{k,j}\mathbf{e}_{j}^{\diamond}\mathbf{v}_{k,j}\right\} \right\} $
$\overset{\left(\textrm{a}\right)}{=}\tilde{P}_{k}\mathbb{E}\left\{ \left(1-Z\right)\left|\beta_{k,j}\right|^{2}+Z\left(1-\left|\beta_{k,j}\right|^{2}\right)\left|\mathbf{e}_{j}^{\diamond}\mathbf{v}_{k,j}\right|^{2}\right\} $
$\overset{\left(\textrm{b}\right)}{=}\tilde{P}_{k}\mathbb{E}\left\{ \left(1-Z\right)\left|\beta_{k,j}\right|^{2}+Z\left(1-\left|\beta_{k,j}\right|^{2}\right)\frac{1}{N-1}\right\} $
$=\tilde{P}_{k}\left[\frac{\eta}{N-1}+\left(1-\frac{N}{N-1}\eta\right)\mathbb{E}\left\{ \left|\beta_{k,j}\right|^{2}\right\} \right].$
$$\qquad\qquad\qquad\qquad\qquad\qquad\quad\qquad\label{eq:E{|hjwk|_squared}}$$
Equation (a) is obtained because $Z$, $\beta_{k,j}$, $\mathbf{e}_{j}^{\diamond}$ and $\mathbf{v}_{k,j}$ are independently distributed. Equation (b) holds because $\mathbf{e}_{j}^{\diamond}$ and $\mathbf{v}_{k,j}^{\textrm{H}}$ are i.i.d. isotropic vectors located in the $\left(N-1\right)$-dimensional nullspace of $\mathbf{\hat{h}}_{j}$. Hence $\left|\mathbf{e}_{j}^{\diamond}\mathbf{v}_{k,j}\right|^{2}$ follows a $\textrm{beta}\left(1,N-2\right)$ distribution, and the mean value of which equals to $\frac{1}{N-1}$ [@integral_table; @book]. From and , we can see that $\mathbb{E}\left\{ \left|\tilde{\mathbf{h}}_{j}^{\diamond}\mathbf{w}_{k}\right|^{2}\right\} $ is a monotonically increasing affine function with respect to $\mathbb{E}\left\{ \left|\beta_{k,j}\right|^{2}\right\} $. Thus, $\mathbb{E}\left\{ \left|\tilde{\mathbf{h}}_{j}^{\diamond}\mathbf{w}_{k}\right|^{2}\right\} $ achieves its minimum value
$$\begin{aligned}
\min\left(\mathbb{E}\left\{ \left|\tilde{\mathbf{h}}_{j}^{\diamond}\mathbf{w}_{k}\right|^{2}\right\} \right) & = & \frac{\tilde{P}_{k}\eta}{\left(N-1\right)},\label{eq:min(E{|hjwk|_squared})}\end{aligned}$$
when $\mathbb{E}\left\{ \left|\beta_{k,j}\right|^{2}\right\} =0$. It implies that $\beta_{k,j}=0$ since $\left|\beta_{k,j}\right|^{2}\geq0$ is always true. On the other hand, $\mathbb{E}\left\{ \left|\tilde{\mathbf{h}}_{j}^{\diamond}\mathbf{w}_{k}^{\textrm{ZF}}\right|^{2}\right\} $ can be calculated as
$\mathbb{E}\left\{ \left|\tilde{\mathbf{h}}_{j}^{\diamond}\mathbf{w}_{k}^{\textrm{ZF}}\right|^{2}\right\} $
$$\begin{aligned}
& = & \mathbb{E}\left\{ \left|\left(\left(\sqrt{1-Z}\mathbf{\hat{h}}_{j}+\sqrt{Z}\mathbf{e}_{j}^{\diamond}\right)\right)\left\Vert \mathbf{w}_{k}^{\textrm{ZF}}\right\Vert \mathbf{\tilde{w}}_{k}^{\textrm{ZF}}\right|^{2}\right\} \nonumber \\
& = & \tilde{P}_{k}\mathbb{E}\left\{ Z\right\} \mathbb{E}\left\{ \left|\mathbf{e}_{j}^{\diamond}\mathbf{\tilde{w}}_{k}^{\textrm{ZF}}\right|^{2}\right\} \nonumber \\
& \overset{\left(\textrm{a}\right)}{=} & \frac{\tilde{P}_{k}\eta}{\left(N-1\right)}.\label{eq:E{interf_ZF}}\end{aligned}$$
Equation (a) comes from the fact that $\left|\mathbf{e}_{j}^{\diamond}\mathbf{\tilde{w}}_{k}^{\textrm{ZF}}\right|^{2}$ also conforms to a $\textrm{beta}\left(1,N-2\right)$ distribution [@Limited-bit; @CSI], and the mean value of which is $\frac{1}{N-1}$. Our proof is completed by combining and to form an inequality with multiplication of $\xi_{j}^{2}A_{j}$ to both sides since CMI of UE $j$ is independent of \~.
$$\begin{aligned}
P_{D}\left(d\right) & = & \Pr\left(RV\leq d\right)\nonumber \\
& = & \Pr\left(V\leq\frac{d}{R}\right)\nonumber \\
& = & \int_{0}^{\infty}P_{V}\left(\frac{d}{r}\right)p_{R}\left(r\right)dr\nonumber \\
& = & \int_{0}^{d}p_{R}\left(r\right)dr+\int_{d}^{\infty}P_{V}\left(\frac{d}{r}\right)p_{R}\left(r\right)dr\nonumber \\
& = & 1+\frac{1}{\left(N-2\right)!}\left[\sum_{n=0}^{N-2}\sum_{m=1}^{2^{B}}\textrm{C}_{N-2}^{n}\textrm{C}_{2^{B}}^{m}\frac{\left(-1\right)^{n+m}md^{n}}{mN-\left(m+n\right)}\int_{d}^{\infty}r^{\left(N-1-n\right)}\textrm{e}^{-r}dr\right]\nonumber \\
& & -\frac{1}{\left(N-2\right)!}\left[\sum_{n=0}^{N-2}\sum_{m=1}^{2^{B}}\textrm{C}_{N-2}^{n}\textrm{C}_{2^{B}}^{m}\frac{\left(-1\right)^{n+m}md^{m\left(N-1\right)}}{mN-\left(m+n\right)}\int_{d}^{\infty}r^{-\left(m-1\right)\left(N-1\right)}\textrm{e}^{-r}dr\right]\nonumber \\
& \overset{\left(\textrm{a}\right)}{=} & 1+\frac{1}{\left(N-2\right)!}\left\{ \sum_{n=0}^{N-2}\sum_{m=1}^{2^{B}}\textrm{C}_{N-2}^{n}\textrm{C}_{2^{B}}^{m}\frac{\left(-1\right)^{n+m}md^{n}}{mN-\left(m+n\right)}\left.\left[-\textrm{e}^{-r}\sum_{l=0}^{N-1-n}l!\textrm{C}_{N-1-n}^{l}r^{N-1-n-l}\right]\right|_{r=d}^{r=\infty}\right\} \nonumber \\
& & -\frac{1}{\left(N-2\right)!}\left\{ \sum_{n=0}^{N-2}\textrm{C}_{N-2}^{n}\textrm{C}_{2^{B}}^{1}\frac{\left(-1\right)^{n+1}d^{N-1}}{N-1-n}\left.\left[-\textrm{e}^{-r}\right]\right|_{r=d}^{r=\infty}\right\} \nonumber \\
& & -\frac{1}{\left(N-2\right)!}\left\{ \sum_{n=0}^{N-2}\sum_{m=2}^{2^{B}}\textrm{C}_{N-2}^{n}\textrm{C}_{2^{B}}^{m}\frac{\left(-1\right)^{n+m}md^{m\left(N-1\right)}}{mN-\left(m+n\right)}\left.\left[\sum_{l=1}^{\left(m-1\right)\left(N-1\right)-1}\frac{\left(-1\right)^{l}\textrm{e}^{-r}}{l!\textrm{C}_{\left(m-1\right)\left(N-1\right)-1}^{l}r^{\left(m-1\right)\left(N-1\right)-l}}\right]\right|_{r=d}^{r=\infty}\right\} \nonumber \\
& & -\frac{1}{\left(N-2\right)!}\left\{ \sum_{n=0}^{N-2}\sum_{m=2}^{2^{B}}\textrm{C}_{N-2}^{n}\textrm{C}_{2^{B}}^{m}\frac{\left(-1\right)^{n+m}md^{m\left(N-1\right)}}{mN-\left(m+n\right)}\frac{\left(-1\right)^{\left(m-1\right)\left(N-1\right)-1}}{\left(\left(m-1\right)\left(N-1\right)-1\right)!}\int_{d}^{\infty}\frac{\textrm{e}^{-r}}{r}dr\right\} .\label{eq:proof_theo2}\end{aligned}$$
------------------------------------------------------------------------
Appendix II: Proof of Theorem 2 {#appendix-ii-proof-of-theorem-2 .unnumbered}
===============================
According to and , we can get $\left|\tilde{\mathbf{h}}_{j}^{\diamond}\mathbf{\tilde{w}}_{k}^{\textrm{ZF}}\right|^{2}=\left|\left(\sqrt{1-Z}\mathbf{\hat{h}}_{j}+\sqrt{Z}\mathbf{e}_{j}^{\diamond}\right)\mathbf{\tilde{w}}_{k}^{\textrm{ZF}}\right|^{2}=Z\left|\mathbf{e}_{j}^{\diamond}\mathbf{\tilde{w}}_{k}^{\textrm{ZF}}\right|^{2}.$ Let $G=\left|\mathbf{e}_{j}^{\diamond}\mathbf{\tilde{w}}_{k}^{\textrm{ZF}}\right|^{2}$. Then we define a random variable $V$ as $V=\left|\tilde{\mathbf{h}}_{j}^{\diamond}\mathbf{\tilde{w}}_{k}^{\textrm{ZF}}\right|^{2}=ZG$. Furthermore, we define a random variable $D$ as $D=RV$, where the random variable $R=A_{j}^{\diamond}$. According to [@Proakis; @book], $R$ is chi-square distributed with $2N$ degrees of freedom, each with variance $\frac{1}{2}$. Hence its PDF and CDF can be respectively written as $p_{R}\left(r\right)=\frac{1}{\Gamma\left(N\right)}r^{N-1}\textrm{e}^{-r}$ and $P_{R}\left(r\right)=1-\textrm{e}^{-r}\sum_{l=0}^{N-1}\frac{r^{l}}{l!}$. Then, based on the result in Theorem \[thm:Theorem\_3\], the CDF of $D=RV$ can be derived as , in which equation (a) is obtained according to [@integral_table; @book], where $\int r^{i}\textrm{e}^{-r}dr=-\textrm{e}^{-r}\sum_{l=0}^{i}l!\textrm{C}_{i}^{l}r^{i-l},\left(i\geq0\right)$ and $\int r^{-i}\textrm{e}^{-r}dr=\sum_{l=1}^{i-1}\frac{\left(-1\right)^{l}\textrm{e}^{-r}}{l!\textrm{C}_{i-1}^{l}r^{i-l}}+\frac{\left(-1\right)^{i-1}}{\left(i-1\right)!}\int\frac{\textrm{e}^{-r}}{r}dr,\left(i>0\right).$ With some mathematical manipulations on , we can get the final form of $P_{D}\left(d\right)$ shown in of Theorem \[thm:Theorem\_2\].
Appendix III: Proof of Theorem 3 {#appendix-iii-proof-of-theorem-3 .unnumbered}
================================
According to the definitions in Appendix II, $V=\left|\tilde{\mathbf{h}}_{j}^{\diamond}\mathbf{\tilde{w}}_{k}^{\textrm{ZF}}\right|^{2}=ZG$, where $G=\left|\mathbf{e}_{j}^{\diamond}\mathbf{\tilde{w}}_{k}^{\textrm{ZF}}\right|^{2}$. According to [@Limited-bit; @CSI], the probability density function (PDF) and CDF of $Z$ can be expressed as
$p_{Z}\left(z\right)=$
$$\left(N-1\right)\sum_{m=1}^{2^{B}}\textrm{C}_{2^{B}}^{m}\left(-1\right)^{m-1}mz^{m\left(N-1\right)-1},\left(z\in\left[0,1\right]\right),$$
and
$P_{Z}\left(z\right)=1-\left(1-z^{N-1}\right)^{2^{B}}$
$$=-\sum_{m=1}^{2^{B}}\textrm{C}_{2^{B}}^{m}\left(-1\right)^{m}z^{m\left(N-1\right)},\left(z\in\left[0,1\right]\right).$$
As explained earlier, $G$ follows a $\textrm{beta}\left(1,N-2\right)$ distribution since $\mathbf{e}_{j}^{\diamond}$ and $\left(\mathbf{\tilde{w}}_{k}^{\textrm{ZF}}\right)^{\textrm{H}}$ are i.i.d. isotropic vectors in the $\left(N-1\right)$-dimensional nullspace of $\mathbf{\hat{h}}_{j}$ [@Limited-bit; @CSI]. Thus, the CDF of $G$ is
$P_{G}\left(g\right)=1-\left(1-g\right)^{N-2}$
$$=-\sum_{n=1}^{N-2}\textrm{C}_{N-2}^{n}\left(-1\right)^{n}g^{n},\left(g\in\left[0,1\right]\right).\qquad\label{eq:PG(g)}$$
Therefore, the CDF of $V=ZG$ can be deduced as , which concludes our proof.
$$\begin{aligned}
P_{V}\left(v\right) & = & \Pr\left(ZG\leq v\right)\nonumber \\
& = & \Pr\left(G\leq\frac{v}{Z}\right)\nonumber \\
& = & \int_{0}^{1}P_{G}\left(\frac{v}{z}\right)p_{Z}\left(z\right)dz\nonumber \\
& = & \int_{0}^{v}p_{Z}\left(z\right)dz+\int_{v}^{1}P_{G}\left(\frac{v}{z}\right)p_{Z}\left(z\right)dz\nonumber \\
& = & P_{Z}\left(v\right)+\int_{v}^{1}\left[-\sum_{n=1}^{N-2}\textrm{C}_{N-2}^{n}\left(-1\right)^{n}\left(\frac{v}{z}\right)^{n}\right]\left[-\left(N-1\right)\sum_{m=1}^{2^{B}}\textrm{C}_{2^{B}}^{m}\left(-1\right)^{m}mz^{m\left(N-1\right)-1}\right]dz\nonumber \\
& = & P_{Z}\left(v\right)+\left(N-1\right)\sum_{n=1}^{N-2}\sum_{m=1}^{2^{B}}\textrm{C}_{N-2}^{n}\textrm{C}_{2^{B}}^{m}\left(-1\right)^{n+m}mv^{n}\left.\left[\frac{z^{mN-\left(m+n\right)}}{mN-\left(m+n\right)}\right]\right|_{z=v}^{z=1}\nonumber \\
& = & 1+\left(N-1\right)\sum_{n=0}^{N-2}\sum_{m=1}^{2^{B}}\textrm{C}_{N-2}^{n}\textrm{C}_{2^{B}}^{m}\left(-1\right)^{n+m}m\frac{v^{n}-v^{m\left(N-1\right)}}{mN-\left(m+n\right)}.\label{eq:proof_theo3}\end{aligned}$$
------------------------------------------------------------------------
[10]{} D. Gesbert, M. Kountouris, R. Heath, C. Chae, and T. Salzer, From single-user to multiuser communications: shifting the MIMO paradigm, *IEEE Signal Process. Mag.*, vol. 24, no. 5, pp. 36-46, Sep. 2007.
G. Caire and S. Shamai, On the achievable throughput of a multiantenna Gaussian broadcast channel, *IEEE Trans. Inform. Theory*, vol. 49, no. 7, pp. 1691-1706, Jul. 2003.
T. Yoo and A. Goldsmith, On the optimality of multianntenna broadcast scheduling using zero-forcing beamforming, *IEEE J. Select. Areas Commun.*, vol. 24, no. 3, pp. 528-541, Mar. 2006.
M. Sharif and B. Hassibi, “A comparison of time-sharing, DPC, and beamforming for MIMO broadcast channels with many users,” *IEEE Trans. Commun.*, vol. 55, no. 1, pp. 11-15, Jan. 2007.
3GPP TS 36.213 V10.0.1, Evolved universal terrestrial radio access (E-UTRA): physical layer procedures, *Tech. Spec.*, Mar. 2011.
N. Jindal, MIMO broadcast channels with finite-rate feedback, *IEEE Trans. Inform. Theory*, vol. 52, no. 11, pp. 5045-5060, Nov. 2006.
M. Botros and T. Davidson, “Convex conic formulations of robust downlink precoder designs with quality of service constraints,” *IEEE J. Select. Topics Signal Process*, vol. 1, no. 4, pp. 714-724, Dec. 2007.
Y. Huang and D. Palomar, Rank-constrained separable semidefinite programming With applications to optimal beamforming, *IEEE Trans. on Signal Processing*, vol. 58, no. 2, pp. 664-678, Feb. 2010.
C. Zhang, W. Xu, and M. Chen, “Robust MMSE beamforming for multiuser MISO systems with limited feedback,” *IEEE Signal Process. Lett.*, vol. 16, no. 7, pp. 588-591, Jul. 2009.
M. Sadek, A. Tarighat, and A. Sayed, “A leakage-based precoding scheme for downlink multi-user MIMO channels,” *IEEE Trans. Wireless Commun.*, vol. 6, no. 5, pp. 1711-1721, May 2007.
B. Dai, W. Xu, and C. Zhao, “Multiuser beamforming optimization via maximizing modified SLNR with quantized CSI feedback,” in *2011 IEEE Wireless Communications, Networking and Mobile Computing (WiCOM 2011)*, pp. 1-5, Sep. 2011.
H. Du and P. Chung, “A probabilistic approach for robust leakage-based MU-MIMO downlink beamforming with imperfect channel state information,” *IEEE Trans. Wireless Commun.*, vol. 11, no. 3, pp. 1239-1247, Mar. 2012.
P. Chung and H. Du, “Robust SLNR downlink beamforming based on Markov’s inequality,” in *2012 IEEE International Conference on Commun. (ICC 2012)*, pp. 3680-3684, Jun. 2012.
K. Huang and R. Zhang, “Cooperative Precoding with Limited Feedback for MIMO Interference Channels,” *IEEE Trans. Wireless Commun.*, vol. 11, no.3, pp. 1012-1021, Mar. 2012.
M. Joham, P. Castro, W. Utschick and L. Castedo, L, “Robust Precoding With Limited Feedback Design Based on Precoding MSE for MU-MISO Systems,” *IEEE Trans. on Signal Processing*, vol. 60, no. 6, pp. 3101-3111, Jun. 2012.
J. Zou, W. Liu, M. Ding, H. Luo and H. Yu, “Transceiver design for AF MIMO two-way relay systems with imperfect channel estimation,” in *2012 IEEE Global Telecommunications Conference (GLOBECOM 2011)*, pp. 1-5, Dec. 2012.
Z. Wang, W. Chen, and J. Li, “Efficient beamforming for MIMO Relaying broadcast channel with imperfect channel estimation,” *IEEE Transactions on Vehicular Technology*, vol. 61, no. 1, pp. 419-426, Jan. 2012.
Z. Wang, W. Chen, F. Gao, and J. Li, “Capacity performance of efficient relay beamformings for dual-hop MIMO multi-relay networks with imperfect R-D CSI at relays,” *IEEE Transactions on Vehicular Technology*, vol. 60, no. 6, pp. 2608-2619, Jun. 2011.
A. Tajer, N. Prasad, and X. Wang, “Robust linear precoder design for multi-cell downlink transmission,” *IEEE Trans. Signal Process.*, vol. 59, no. 1, pp. 235-251, Jan. 2011.
W. Yu and T. Lan, “Transmitter optimization for the multi-antenna downlink with per-antenna power constraints,” *IEEE Trans. Signal Process.*, vol. 55, no. 6, pp. 2646-2660, Jun. 2007.
H. Wan, W. Chen, and J. Ji, “Efficient linear transmission strategy for MIMO relaying broadcast channels with direct links,” *IEEE Wireless Communications Letters*, vol. 1, no. 1, pp. 14-17, Jan. 2012.
M. Ding, J. Zou, Z. Yang, H. Luo, and W. Chen, “Sequential and incremental precoder design for joint transmission network MIMO systems with imperfect backhaul,” *IEEE Transactions on Vehicular Technology*, vol. 61, no. 6, pp. 2490-2503, Jul. 2012.
Z. Wang, and W. Chen, Regularized zero-forcing for multiantenna broadcast channels with user selection, *IEEE Wireless Communications Letters*, vol. 1, no. 2, pp. 129-132, Feb. 2012.
W. Santipach and M. Honig, Asymptotic capacity of beamforming with limited feedback,” in *2004 IEEE Int. Symp. on Inf. Theory (ISIT 2004)*. pp. 289, Jun. 2004.
S. Sesia, I. Toufik, M. Baker, *LTE: The UMTS Long Term Evolution*. John Wiley and Sons, USA, 2009.
J. Proakis, *Digital Communications* *(Third Ed.)*, New York: McGraw-Hill, 1995.
A. Wiesel, Y. Elda, and S. Shamai, “Zero-forcing precoding and generalized inverses,” *IEEE Trans. on Signal Processing*, vol. 56, no. 9, pp. 4409-4418, Sep. 2008.
M. Grant, S. Boyd, and Y. Ye, CVX users guide for cvx version 1.1 (build 565), Nov. 2007.
I. Gradshteyn and I. Ryzhik, *Table of Integrals, Series, and Products* *(Seventh Ed.)*. Elsevier Inc., 2007.
Z. Luo, W. Ma, A. So, Y. Ye, S. Zhang, “Semidefinite relaxation of quadratic optimization problems,” *IEEE Signal Processing Magazine*, vol. 27, no. 3, pp. 20-34, May 2010.
M. Chiang, “Geometric Programming for Communication Systems,” *Short Monograph in Foundations and Trends in Communication and Information Theory*, vol. 2, Jul. 2005.
[^1]: Copyright (c) 2012 IEEE. Personal use of this material is permitted. However, permission to use this material for any other purposes must be obtained from the IEEE by sending a request to [[email protected]](http://[email protected]).
[^2]: Ming Ding is with Sharp Laboratories of China Co., Ltd. (E-mail: [email protected]).
[^3]: Meng Zhang, Hanwen Luo and Wen Chen are with the Dept. of Electronic Engineering, Shanghai Jiao Tong University, Shanghai, P. R. China. (E-mail: {mengzhang, hwluo, wenchen}@sjtu.edu.cn).
[^4]: This work is sponsored by Sharp Laboratories of China Co., Ltd.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A Rademacher-type convergent series formula which generalizes the Hardy-Ramanujan-Rademacher formula for the number of partitions of $n$ and the Zuckerman formula for the Fourier coefficients of $\vartheta_4(0 \mid \tau)^{-1}$ is presented.'
address: 'Department of Mathematical Sciences, Georgia Southern University, Statesboro, Georgia, 30460-8093 USA'
author:
- 'Andrew V. Sills'
title: A RADEMACHER TYPE FORMULA FOR PARTITIONS AND OVERPARTITIONS
---
Background
==========
Partitions
----------
A *partition* of an integer $n$ is a representation of $n$ as a sum of positive integers, where the order of the summands (called *parts*) is considered irrelevant. It is customary to write the parts in nonincreasing order. For example, there are three partitions of the integer $3$, namely $3$, $2+1$, and $1+1+1$. Let $p(n)$ denote the number of partitions of $n$, with the convention that $p(0)=1$, and let $f(x)$ denote the generating function of $p(n)$, i.e. let $$f(x) := \sum_{n=0}^\infty p(n) x^n .$$
Euler [@le1748] was the first to systematically study partitions. He showed that $$\label{egf}
f(x) = \prod_{m=1}^\infty \frac{1}{1-x^m} .$$ Euler also showed that $$\label{epnt} \frac{1}{f(x)} = \sum_{n=-\infty}^\infty (-1)^n x^{n(3n-1)/2},$$ and since the exponents appearing on the right side of are the pentagonal numbers, Eq. is often called “Euler’s pentagonal number theorem."
Although Euler’s results can all be treated from the point of view of formal power series, the series and infinite products above (and indeed all the series and infinite products mentioned in this paper) converge absolutely when $|x|<1$, which is important for analytic study of these series and products.
Hardy and Ramanujan were the first to study $p(n)$ analytically and produced an incredibly accurate asymptotic formula [@hr1918 p. 85, Eq. (1.74)], namely $$\begin{gathered}
\label{HRPofN}
p(n) = \frac{1}{2\pi\sqrt{2}}
\sum_{k=1}^{\lfloor \alpha\sqrt{n} \rfloor} \sqrt{k}
\underset{(h,k)=1}{\sum_{0\leqq h < k}} \omega(h,k)
e^{-2 \pi i h n/k}
\frac{d}{dn} \left( \frac{ \exp\left( \frac{\pi}{k} \sqrt{\frac 23 \left( n-\frac{1}{24} \right)} \right) }
{ \sqrt{n-\frac{1}{24}}}\right) \\ + O(n^{-1/4}),\end{gathered}$$ where $$\omega(h,k) = \exp \left( \pi i \sum_{r=1}^{k-1} \frac rk \left(
\frac{hr}{k} - \lfloor \frac{hr}{k} \rfloor - \frac 12 \right) \right),$$ $\alpha$ is an arbitrary constant, and here and throughout $(h,k)$ is an abbreviation for $\gcd(h,k)$.
Later Rademacher [@har1937] improved upon by finding the following convergent series representation for $p(n)$: $$\label{RadPofN}
p(n) = \frac{1}{\pi\sqrt{2}} \sum_{k=1}^\infty \sqrt{k}
\underset{(h,k)=1}
{ \sum_{0 \leqq h < k}} \omega(h,k) e^{- 2\pi i nh/k}
\frac{d}{dn} \left( \frac{ \sinh \left( \frac{\pi}{k} \sqrt{\frac23 \left( n-\frac{1}{24}\right)}\right)}{\sqrt{n-\frac{1}{24}}} \right).$$
Rademacher’s method was used extensively by many practitioners, including Grosswald [@eg1958; @eg1960], Haberzetle [@mh1941], Hagis [@ph1962; @ph1963; @ph1964a; @ph1964b; @ph1965a; @ph1965b; @ph1965c; @ph1966; @ph1971], Hua [@lkh1942], Iseki [@si1959; @si1960; @si1961], Lehner [@jl1941], Livingood [@jl1945], Niven [@in1940], and Subramanyasastri [@vvs1972] to study various restricted partitions functions.
Recently, Bringmann and Ono [@bo2009] have given exact formulas for the coeffcients of all harmonic Maass forms of weight $\leqq \frac 12$. The generating functions considered herein are weakly holomorphic modular forms of weight $-\frac 12$, and thus they are harmonic Maass forms of weight $\leqq \frac 12$. Accordingly, the results of this present paper could be derived from the general theorem in [@bo2009]. However, here we opt to derive the results via classical method of Rademacher.
Overpartitions
--------------
Overpartitions were introduced by S. Corteel and J. Lovejoy in [@cl2004] and have been studied extensively by them and others including Bringmann, Chen, Fu, Goh, Hirschhorn, Hitczenko, Lascoux, Mahlburg, Robbins, R[ø]{}dseth, Sellers, Yee, and Zho [@bl2007; @czj2005; @cgh2006; @ch2004; @cl2004; @cly2004; @cm2007; @fl2005; @hs2005; @hs2006; @jl2003; @jl2004a; @jl2004b; @jl2005a; @jl2005b; @jl2007; @km2004; @nr2003; @rs2005].
An *overpartition* of $n$ is a representation of $n$ as a sum of positive integers with summands in nonincreasing order, where the last occurrence of a given summand may or may not be overlined. Thus the eight overpartitions of $3$ are $3$, $\bar{3}$, $2+1$, $\bar{2}+1$, $2+\bar{1}$, $\bar{2}+\bar{1}$, $1+1+1$, $1+1+\bar{1}$.
Let $\bar{p}(n)$ denote the number of overpartitions of $n$ and let $\bar{f}(x)$ denote the generating function $\sum_{n=0}^\infty \bar{p}(n) x^n$ of $\bar{p}(n)$. Elementary techniques are sufficient to show that $$\bar{f}(x) = \prod_{m=1}^\infty \frac{1+x^m}{1-x^m} = \frac{f(x)^2}{f(x^2)}.$$
Note that $$\frac{1}{\bar{f}(x)} = \sum_{n=-\infty}^{\infty} (-1)^n x^{n^2}$$ via an identity of Gauss [@gea1976 p. 23, Eq. (2.2.12)], so that the reciprocal of the generating function for overpartitions is a series wherein a coefficient is nonzero if and only if the exponent of $x$ is a perfect square, just as the reciprocal of the generating function for partitions is a series wherein a coefficient is nonzero if and only if the exponent of $x$ is a pentagonal number.
Hardy and Ramanujan, writing more than 80 years before the coining of the term “overpartition," stated [@hr1918 p. 109–110] that the function which we are calling $\bar{p}(n)$ “has no very simple arithmetical interpretation; but the series is none the less, as the direct reciprocal of a simple $\vartheta$-funciton, of particular interest." They went on to state that $$\label{HRPBarofN}
\bar{p}(n) = \frac{1}{4\pi} \frac{d}{dn}\left( \frac{ e^{\pi\sqrt{n}}}{\sqrt{n}} \right) +
\frac{\sqrt{3}}{2\pi} \cos\left( \frac 23 n \pi - \frac 16 \pi \right) \frac{d}{dn}
\left( e^{\pi\sqrt{n}/3} \right) + \cdots + O(n^{-1/4}).$$
In fact, was improved to the following Rademacher-type convergent series by Zuckerman [@hsz1939 p. 321, Eq. (8.53)]: $$\label{overptnformula}
\bar{p}(n) = \frac{1}{2\pi} \underset{2\nmid k}{\sum_{k\geqq 1}} \sqrt{k}
\underset{(h,k)=1}{\sum_{0\leqq h < k}} \frac{\omega(h,k)^2 }{ \omega(2h,k) }
e^{-2\pi i nh /k } \frac{d}{dn} \left( \frac{\sinh \left( \frac{\pi \sqrt{n}}{k} \right) }{\sqrt{n} } \right).$$ A simplified proof of Eq. was given by L. Goldberg in this Ph.D. thesis [@lg1981].
Partitions where no odd part is repeated
----------------------------------------
Let $pod(n)$ denote the number of partitions of $n$ where no odd part appears more than once. Let $g(x)$ denote the generating function of $pod(n)$, so we have $$g(x) = \sum_{n=0}^\infty pod(n) x^n = \prod_{m=1}^\infty \frac{1+x^{2j-1}}{1-x^{2j}}
= \frac{f(x)f(x^4)}{f(x^2)}.$$
Via another identity of Gauss [@gea1976 p. 23, Eq. (2.2.13)], it turns out that $$\frac{1}{g(x)} = \sum_{n=0}^\infty (-x)^{n(n+1)/2} = \sum_{n=-\infty}^\infty (-1)^n x^{2n^2-n},$$ so in this case the reciprocal of the generating function under consideration has nonzero coefficients at the exponents which are triangular (or equivalently, hexagonal) numbers.
The analogous Rademacher-type formula for $pod(n)$ is as follows. $$\begin{gathered}
\label{norepoddsformula}
pod(n) = \frac{2}{\pi} \sum_{k\geqq 1}
\sqrt{k \left( 1-(-1)^k + \lfloor \frac{(k,4)}{4} \rfloor \right)} \\ \times
\underset{(h,k)=1}{\sum_{0\leqq h < k}} \frac{\omega(h,k) \ \omega\left( \frac{4h}{(k,4)}, \frac{k}{(k,4)} \right) }{ \omega\left( \frac{2h}{(k,2)} ,\frac{k}{(k,2)} \right) }
e^{-2\pi i nh /k } \frac{d}{dn} \left( \frac{\sinh \left( \frac{\pi \sqrt{(k,4) (8n-1)}}{4k} \right) }{\sqrt{8n-1} } \right).
\end{gathered}$$
Eq. is the case $r=2$ of Theorem \[MainFormula\] to be presented in the next section.
A common generalization
=======================
Let us define $$\label{frdef} f_r (x) := \frac{f(x) f(x^{2^r})}{f(x^2)},$$ where $r$ is a nonnegative integer. Thus, $$\begin{aligned}
f_0 (x) &= \bar{f}(x) = \sum_{n=0}^\infty \bar{p}(n) x^n, \\
f_1 (x) &= f(x) = \sum_{n=0}^\infty p(n) x^n, \\
f_2 (x) &= g(x) = \sum_{n=0}^\infty pod(n) x^n.\end{aligned}$$
Let $p_r (n)$ denote the coefficient of $x^n$ in the expansion of $f_r(n)$, i.e. $$f_r(x) = \sum_{n=0}^\infty p_r(n) x^n.$$
Notice that $f_r(x) $ can be represented by several forms of equivalent infinite products, each of which has a natural combinatorial interpretation: $$\begin{aligned}
f_r(x) & = \prod_{m=1}^\infty \frac{1+x^m}{1-x^{2^r m}} \label{interp1}\\
& = \prod_{m=1}^\infty \frac{1}{(1-x^{2m-1})(1-x^{2^r m})}\label{interp2}\\
& = \prod_{m=1}^\infty \frac{1}{1-x^{2^{r-1}m} } \prod_{\lambda=1}^{2^{r-1}-1} (1+x^{2^{r-1}m+\lambda}).
\label{interp3}\end{aligned}$$
Thus, $p_r(n)$ equals each of the following:
- the number of overpartitions of $n$ where nonoverlined parts are multiples of $2^r$ (by );
- the number of partitions of $n$ where all parts are either odd or multiples of $2^r$ (by ), provided $r\geqq 1$;
- the number of partitions of $n$ where where nonmultiples of $2^{r-1}$ are distinct (by ), provided $r\geqq 1$.
\[MainFormula\] For $r = 0, 1, 2, 3, 4$, $$\begin{gathered}
p_r (n) = \frac{2^{(r+1)/2} \sqrt{3}}{\pi}
\underset{ (k, 2^{\max(r,1)})=1}{ \sum_{k\geqq 1} \sqrt{k}}
\underset{(h,k)=1}{\sum_{0\leqq h < k}} e^{-2\pi i n h/k} \frac{\omega(h,k) \omega(2^r h, k)}
{\omega(2h,k)} \\ \qquad\times \frac{d}{dn} \left\{ \frac{\sinh\left(
\frac{\pi \sqrt{(24n-2^r+1)(1+2^{r-1})}}{2^{r/2}\cdot 6k} \right)}{\sqrt{24n-2^r+1}} \right\}\\
+
\frac{ \sqrt{3}}{\pi} \sum_{j=1+\lfloor \frac r2 \rfloor}^r 2^{(2-j+r)/2}
\underset{ (k, 2^r)=2^j}{ \sum_{k\geqq 1} \sqrt{k}}
\underset{(h,k)=1}{\sum_{0\leqq h < k}} e^{-2\pi i n h/k} \frac{\omega(h,k)
\omega(2^{r-j} h, 2^{-j}k)}
{\omega(h, \frac k2)} \\ \qquad \times \frac{d}{dn} \left\{ \frac{\sinh\left(
\frac{\pi \sqrt{(24n-2^r+1)(-1+2^{2j-r})}}{ 6k} \right)}{\sqrt{24n-2^r+1}} \right\}.\end{gathered}$$
A Proof of Theorem \[MainFormula\] {#OverPtnPf}
==================================
The method of proof is based on Rademacher’s proof of in [@har1943] with the necessary modifications. Additional details of Rademacher’s proof of are provided in [@har1955], [@har1973 Ch. 14] and [@tma1990 Ch. 5].
Of fundamental importance is the path of integration to be used. In [@har1943], Rademacher improved upon his original proof of given in [@har1937], by altering his path of integration from a carefully chosen circle to a more complicated path based on Ford circles, which in turn led to considerable simplifications later in the proof.
Farey fractions
---------------
The sequence ${\mathcal{F}}_N$ of *proper Farey fractions of order $N$* is the set of all $h/k$ with $(h,k)=1$ and $0\leqq h/k <1$, arranged in increasing order. Thus, e.g., ${\mathcal{F}}_4 = \left\{ \frac 01, \frac 14, \frac 13, \frac 12, \frac 23, \frac 34\right\}.$
For a given $N$, let $h_p$, $h_s$, $k_p$, and $k_s$ be such that $\frac{h_p}{k_p}$ is the immediate predecessor of $\frac hk $ and $\frac{h_s}{k_s}$ is the immediate successor of $\frac hk$ in ${\mathcal{F}}_N$. It will be convenient to view each ${\mathcal{F}}_N$ cyclically, i.e. to view $\frac 01$ as the immediate successor of $\frac {N-1}{N}$.
Ford circles and the Rademacher path
------------------------------------
Let $h$ and $k$ be integers with $(h,k) = 1$ and $0\leqq h < k$. The *Ford circle* [@lrf1938] $C(h,k)$ is the circle in $\mathbb C$ of radius $\frac{1}{2k^2}$ centered at the point $$\frac{h}{k} + \frac{1}{2k^2} i.$$
The *upper arc $\gamma(h,k)$ of the Ford circle $C(h,k)$* is those points of $C(h,k)$ from the initial point $$\label{AlphaI}
\alpha_I(h,k):= \frac hk - \frac{k_p}{k(k^2+k_p^2)} + \frac{1}{k^2+k_p^2} i$$ to the terminal point $$\label{AlphaT}
\alpha_T(h,k):= \frac hk + \frac{k_s}{k(k^2+k_s^2)} + \frac{1}{k^2+k_s^2} i,$$ traversed in the clockwise direction.
Note that we have $$\alpha_I(0,1) = \alpha_T(N-1,N) .$$
Every Ford circle is in the upper half plane. For $\frac{h_1}{k_1}, \frac{h_2}{k_2} \in {\mathcal{F}}_N$, $C(h_1, k_1)$ and $C(h_2, k_2)$ are either tangent or do not intersect.
The *Rademacher path* $P(N)$ of order $N$ is the path in the upper half of the $\tau$-plane from $i$ to $i+1$ consisting of $$\label{RadPath} \bigcup_{\frac hk \in {\mathcal{F}}_N} \gamma(h,k)$$ traversed left to right and clockwise. In particular, we consider the left half of the Ford circle $C(0,1)$ and the corresponding upper arc $\gamma(0,1)$ to be translated to the right by 1 unit. This is legal given then periodicity of the function which is to be integrated over $P(N)$.
Set up the integral
-------------------
Let $n$ and $r$ be fixed, with $n > (2^r - 1)/24$.
Since $$f_r(x) = \sum_{n=0}^\infty p_r(n) x^n,$$ Cauchy’s residue theorem implies that $$\label{Cauchy}
p_r(n) = \frac{1}{2\pi i} \int_{\mathcal C} \frac{f_r(x)}{x^{n+1}}\ dx,$$ where $\mathcal C$ is any simply closed contour enclosing the origin and inside the unit circle. We introduce the change of variable $$x = e^{2 \pi i \tau}$$ so that the unit disk $|x| \leqq 1$ in the $x$-plane maps to the infinitely tall, unit-wide strip in the $\tau$-plane where $0 \leqq \Re\tau \leqq 1$ and $\Im\tau \geqq 0$. The contour $\mathcal C$ is then taken to be the preimage of $P(N)$ under the map $x\mapsto e^{2\pi i \tau}$.
Better yet, let us replace $x$ with $e^{2 \pi i \tau}$ in to express the integration in the $\tau$-plane: $$\begin{aligned}
\label{TauIntegral}
p_r(n) &= \int_{{P}(N)} {f_r}(e^{2\pi i \tau}) e^{-2\pi i n\tau} d\tau \nonumber\\
&= {\sum_{\frac hk \in {\mathcal{F}}_N}}
\int_{\gamma(h,k)} {f_r}(e^{2\pi i \tau}) e^{-2\pi i n \tau} d\tau \nonumber\\
&={\sum_{k=1}^{N}} \underset{(h,k)=1}{ \sum_{0 \leqq h < k}} \int_{\gamma(h,k)}
{f_r}(e^{2\pi i \tau}) e^{-2\pi i n \tau} d\tau
\nonumber\end{aligned}$$
Another change of variable
--------------------------
Next, we change variables again, taking $$\label{ZToTau} \tau = \frac {iz + h}{k},$$ so that $$\label{TauToZ} z = -ik \left(\tau - \frac hk\right).$$ Thus $C(h,k)$ (in the $\tau$-plane) maps to the clockwise-oriented circle $K^{(-)}_k$ (in the $z$-plane) centered at $1/2k$ with radius $1/2k$.
So we now have $$\begin{gathered}
p_r(n)
=i{ \sum_{k=1}^{N}} k^{-1} \underset{(h,k)=1}
{ \sum_{0 \leqq h < k}} e^{-2\pi i n h/k}
\underset{\mbox{\tiny arc of $K_k^{(-)} $} }
{\int_{z_I(h,k)}^{z_T(h,k)} }e^{2n\pi z/k} f_r( e^{2\pi i h/k - 2\pi z/k} ) \ dz
, \label{IntegralZ}\end{gathered}$$ where $z_I(h,k)$ (resp. $z_T(h,k)$) is the image of $\alpha_I(h,k)$ (see ) (resp. $\alpha_T(h,k)$ \[see \]) under the transformation .
So the transformation maps the upper arc $\gamma(h,k)$ of $C(h,k)$ in the $\tau$-plane to the arc on $K^{(-)}_k$ which initiates at $$\label{ZI}
z_I(h,k) = \frac{k}{k^2 + k_p^2} + \frac{ k_p}{k^2+ k_p^2} i$$ and terminates at $$\label{ZT}
z_T(h,k) = \frac{k}{k^2+k_s^2} - \frac{ k_s}{k^2+k_s^2} i.$$
Exploiting a modular transformation
-----------------------------------
From the theory of modular forms, we have the transformation formula [@hr1918 p. 93, Lemma 4.31] $$\begin{gathered}
\label{FFunctionalEq}
f\left(
\exp \left( \frac {2\pi i h}{k} - \frac{2\pi z}{k} \right) \right)\\ =
\omega(h,k) \exp\left(\frac{\pi(z^{-1}-z)}{12k} \right)\sqrt{z}
f\left( \exp\left(2\pi i \frac{ iz^{-1}+H}{k}\right) \right),\end{gathered}$$ where $\sqrt{z}$ is the principal branch, $(h,k)=1$, and $H$ is a solution to the congruence $$hH \equiv -1\pmod{k} .$$ From , we deduce the analogous transformation for $f_r (x) $.
The transformation formula is a piecewise defined function with $r+1$ cases corresponding to $j=0,1,2,\dots, r$, where $(k, 2^r) = 2^j$. $$\begin{gathered}
\label{frFunctionalEq}
f_r\left(
\exp \left( \frac {2\pi i h}{k} - \frac{2\pi z}{k} \right) \right)\\ =
\frac{ \omega(h,k) \omega(2^{r-j}h, 2^{-j} k ) }{ \omega\left( \frac{2h}{2-\delta_{j0}},
\frac{k}{2-\delta_{j0}} \right) }
\exp\left( \frac{\pi(2+2^{2j-r+1} - (2-\delta_{j0})^2)}{24kz} + \frac{\pi(1-2^r)z}{12k} \right) \\
\times \sqrt{z \ 2^{r-j-1} (2-\delta_{j0})}
\frac{ f\left( \exp\left(\frac{-2\pi}{kz} + \frac{2H_j\pi i}{k} \right)\right)
f\left( \exp\left(\frac{-2^{2j-r+1}\pi}{kz} + \frac{2^{2j-r+1}H_j\pi i}{k} \right)\right) }
{ f\left( \exp\left(\frac{-\pi(2-\delta_{j0})^2}{kz} + \frac{ H_j\pi (2-\delta_{j0})^2 i}{k} \right)\right) }
,\end{gathered}$$ where $H_j$ is divisible by $2^{r-j}$ and is a solution to the congruence $hH_j \equiv -1\pmod{k}$, and $$\delta_{j0} = \left\{ \begin{array}{ll} 1 & \mbox{if $j=0$} \\ 0 & \mbox{if $j\neq 0$}
\end{array} \right.$$ is the Kronecker $\delta$-function.
Notice that in particular, for $\lfloor \frac r2 \rfloor \leqq j \leqq r$, simplifies to $$\begin{gathered}
\label{frFunctionalEqSimp}
f_r\left(
\exp \left( \frac {2\pi i h}{k} - \frac{2\pi z}{k} \right) \right)\\ =
\frac{ \omega(h,k) \omega(2^{r-j}h, 2^{-j} k ) }{ \omega\left( h,
\frac{k}{2} \right) }
\exp\left( \frac{\pi}{12k} \left[ (2^{2j-r}-1)z^{-1} + (1-2^r)z\right] \right) \\
\times \sqrt{z \ 2^{r-j} }
f_{2j-r}\left( \exp\left(\frac{-2\pi}{kz} + \frac{2H_j\pi i}{k} \right)\right) .\end{gathered}$$
Since the $r=0$ case was established by Zuckerman, and the $r=1$ case by Rademacher, we will proceed with the assumption that $r>1$.
Apply to to obtain $$\begin{gathered}
p_r(n)
=i \sum_{j=0}^{r} \underset{(k,2^r)=j}{ \sum_{k=1}^{N}} k^{-1} \underset{(h,k)=1}
{ \sum_{0 \leqq h < k}} e^{-2\pi i n h/k}
\underset{\mbox{\tiny arc of $K_k^{(-)} $} }
{\int_{z_I(h,k)}^{z_T(h,k)} }
\frac{ \omega(h,k) \omega(2^{r-j}h, 2^{-j} k ) }{ \omega\left( \frac{2h}{2-\delta_{j0}},
\frac{k}{2-\delta_{j0}} \right) } \\ \times
\exp\left( \frac{\pi(2+2^{2j-r+1} - (2-\delta_{j0})^2)}{24kz} + \frac{\pi(24n+1-2^r)z}{12k} \right) \\
\times \sqrt{z \ 2^{r-j-1} (2-\delta_{j0})} \\ \times
\frac{ f\left( \exp\left(\frac{-2\pi}{kz} + \frac{2H_j\pi i}{k} \right)\right)
f\left( \exp\left(\frac{-2^{2j-r+1}\pi}{kz} + \frac{2^{2j-r+1}H_j\pi i}{k} \right)\right) }
{ f\left( \exp\left(\frac{-\pi(2-\delta_{j0})^2}{kz} + \frac{H_j\pi (2-\delta_{j0})^2 i}{k} \right)\right) } \ dz
. $$
Normalization
-------------
Next, introduce a normalization $\zeta = zk$. (This is not strictly necessary, but it will allow us in the sequel to quote various useful results directly from the literature.)
$$\begin{gathered}
p_r(n)
=i \sum_{j=0}^{r} \underset{(k,2^r)=j}{ \sum_{k=1}^{N}} k^{-5/2} \underset{(h,k)=1}
{ \sum_{0 \leqq h < k}} e^{-2\pi i n h/k}
\underset{\mbox{\tiny arc of $K^{(-)} $} }
{\int_{\zeta_I(h,k)}^{\zeta_T(h,k)} }
\frac{ \omega(h,k) \omega(2^{r-j}h, 2^{-j} k ) }{ \omega\left( \frac{2h}{2-\delta_{j0}},
\frac{k}{2-\delta_{j0}} \right) } \\ \times
\exp\left( \frac{\pi(2+2^{2j-r+1} - (2-\delta_{j0})^2)}{24\zeta} + \frac{\pi(24n+1-2^r)\zeta}{12k^2} \right) \\
\times \sqrt{\zeta \ 2^{r-j-1} (2-\delta_{j0})} \\ \times
\frac{ f\left( \exp\left(\frac{-2\pi}{\zeta} + \frac{2H_j\pi i}{k} \right)\right)
f\left( \exp\left(\frac{-2^{2j-r+1}\pi}{\zeta} + \frac{2^{2j-r+1}H_j\pi i}{k} \right)\right) }
{ f\left( \exp\left(\frac{-\pi(2-\delta_{j0})^2}{\zeta} + \frac{H_j\pi (2-\delta_{j0})^2 i}{k} \right)\right) }
\ d\zeta
, \label{IntegralZeta}\end{gathered}$$
where $$\label{ZetaI}
\zeta_I(h,k) = \frac{k^2}{k^2 + k_p^2} + \frac{k k_p}{k^2+ k_p^2} i$$ and $$\label{ZetaT}
\zeta_T(h,k) = \frac{k^2}{k^2+k_s^2} - \frac{k k_s}{k^2+k_s^2} i.$$
Let us now rewrite as $$\begin{gathered}
\label{RewrittenIntegralZeta}
p_r(n)
=i \sum_{j=0}^{r} \underset{(k,2^{\max(r,1)})=2^j}{ \sum_{k=1}^{N}} k^{-5/2} \underset{(h,k)=1}
{ \sum_{0 \leqq h < k}} e^{-2\pi i n h/k}
\frac{ \omega(h,k) \omega(2^{r-j}h, 2^{-j} k ) }{ \omega\left( \frac{2h}{2-\delta_{j0}},
\frac{k}{2-\delta_{j0}} \right) } \\ \times
\left( {\mathcal{I}}_{j,1} + {\mathcal{I}}_{j,2} \right)
$$ where $$\begin{gathered}
{\mathcal{I}}_{j,1}:=
\underset{\mathrm{arc}}{\int_{\zeta_I(h,k)}^{\zeta_T(h,k)} }
\exp\left( \frac{\pi(2+2^{2j-r+1} - (2-\delta_{j0})^2)}{24\zeta} + \frac{\pi(24n+1-2^r)\zeta}{12k^2} \right) \\ \times \sqrt{\zeta \ 2^{r-j-1} (2-\delta_{j0})}
\left\{ -1 +
\frac{ f\left( \exp\left(\frac{-2\pi}{\zeta} + \frac{2H_j\pi i}{k} \right)\right)
f\left( \exp\left(\frac{-2^{2j-r+1}\pi}{\zeta} + \frac{2^{2j-r+1}H_j\pi i}{k} \right)\right) }
{ f\left( \exp\left(\frac{-\pi(2-\delta_{j0})^2}{\zeta} + \frac{H_j\pi (2-\delta_{j0})^2 i}{k} \right)\right) }
\right\} \ d\zeta,\end{gathered}$$ and $$\begin{gathered}
{\mathcal{I}}_{j,2}:=
\underset{\mathrm{arc}}{\int_{\zeta_I(h,k)}^{\zeta_T(h,k)} }
\exp\left( \frac{\pi(2+2^{2j-r+1} - (2-\delta_{j0})^2)}{24\zeta} + \frac{\pi(24n+1-2^r)\zeta}{12k^2}
\right) \\ \times \sqrt{\zeta \ 2^{r-j-1} (2-\delta_{j0})} \ d\zeta.\end{gathered}$$
Estimation
----------
It will turn out that as $N\to\infty$, only ${\mathcal{I}}_{j,2}$ for $j=0$ and $\lfloor r/2 \rfloor < j \leqq r$ ultimately make a contribution (provided $r<5$). Note that all the integrations in the $\zeta$-plane occur on arcs and chords of the circle $K$ of radius $\frac 12$ centered at the point $\frac 12$. So, inside and on $K$, $0<\Re\zeta\leqq 1$ and $\Re \frac 1\zeta \geqq 1$.
### Estimation of ${\mathcal{I}}_{j,2}$ for $1\leqq j \leqq \lfloor r/2 \rfloor$
The regularity of the integrand allows us to alter the path of integration from the arc connecting $\zeta_I(h,k)$ and $\zeta_T(h,k)$ to the directed segment. By [@tma1990 p. 104, Thm. 5.9], the length of the path of integration does not exceed $2\sqrt{2} k/N$, and on the segment connecting $\zeta_I(h,k)$ to $\zeta_T(h,k)$, $|\zeta|< \sqrt{2} k /N$. Thus, the absolute value of the integrand, $$\begin{aligned}
& \phantom{=} \left| \exp\left( \frac{\pi(2^{2j-r} - 1)}{12\zeta} + \frac{\pi(24n+1-2^r)\zeta}{12k^2}
\right) \sqrt{\zeta \ 2^{r-j}} \right| \\
&= |\zeta|^{1/2} 2^{(r-j)/2} \exp\left(\frac{(24n+1-2^r)\pi \Re\zeta}{12k^2} \right)
\exp\left( \frac{\pi(2^{2j-r}-1 )}{12} \Re\frac{1}{\zeta} \right)\\
& \leqq |\zeta|^{1/2} 2^{r/2} \exp(2\pi n) .
\end{aligned}$$ Thus, for $1\leqq j \leqq \lfloor r/2 \rfloor$, $$| {\mathcal{I}}_{j,2} | \leqq \frac{2\sqrt{2} k}{N} \left( \frac{\sqrt{2} k }{N} \right)^{1/2} 2^{r/2} e^ {2\pi n} \leqq
C'_j k^{3/2} N^{-3/2},$$ for a constant $C'_j$ (recalling that $n$ and $r$ are fixed).
### Estimation of ${\mathcal{I}}_{j,1}$ for $1\leqq j \leqq \lfloor r/2 \rfloor$
We have the absolute value of the integrand: $$\begin{aligned}
& \left| \sqrt{\zeta \ 2^{r-j}} \exp\left( \frac{\pi(2^{2j-r} - 1)}{12\zeta} + \frac{\pi(24n+1-2^r)\zeta}{12k^2} \right) \right| \\ & \qquad \times \left|
-1 +
\frac{ f\left( \exp\left(\frac{-2\pi}{\zeta} + \frac{2H_j\pi i}{k} \right)\right)
f\left( \exp\left(\frac{-2^{2j-r+1}\pi}{\zeta} + \frac{2^{2j-r+1}H_j\pi i}{k} \right)\right) }
{ f\left( \exp\left(\frac{-4\pi}{\zeta} + \frac{4 H_j\pi i}{k} \right)\right) }
\right| \\
=& \left| \sqrt{\zeta \ 2^{r-j}} \exp\left( \frac{\pi(2^{2j-r} - 1)}{12\zeta} + \frac{\pi(24n+1-2^r)\zeta}{12k^2} \right) \right| \\ & \qquad \times \left|
-1 + f_{2j-r} \left( \exp\left( \frac{-2\pi}{\zeta} + \frac{2H_j\pi i}{k} \right) \right)
\right| \\
= &|\zeta|^{1/2} 2^{(r-j)/2} \exp\left( \frac{ (24n+1-2^r) \pi \Re\zeta}{12k^2} \right)
\exp\left( \frac{(2^{2j-r}-1)\pi}{12} \Re \frac 1\zeta \right)\\ & \qquad \times
\left| \sum_{m=1}^\infty p_{2j-r} (m)
\exp\left( \frac{-2\pi m}{\zeta} + \frac{2H_j\pi i m}{k} \right)\right| \\
\leqq &|\zeta|^{1/2} 2^{(r-j)/2} \exp\left( \frac{ (24n+1-2^r) \pi }{12} \right)
\sum_{m=1}^\infty p_{2j-r} (m)
\exp\left( -\frac{\pi}{12} (24m - 2^{2j-r}+ 1) \right) \\
\leqq &|\zeta|^{1/2} 2^{r/2} e^{2\pi n}
\sum_{m=1}^\infty p_{0} (m) e^{-2\pi m} \\
= & c_j |\zeta|^{1/2},\end{aligned}$$ for a constant $c_j$. So, for $1\leqq j \leqq \lfloor r/2 \rfloor$, $$|{\mathcal{I}}_{j,1}| \leqq \frac{2\sqrt{2} k}{N} \left( \frac{\sqrt{2} k }{N} \right)^{1/2} c_j
< C_j k^{3/2} N^{-3/2}$$ for a constant $C_j$.
### Estimation of $I_{0,1}$
Let $p^*_r(x)$ be defined by $$\sum_{n=0}^\infty p^*_r(n) x^n = \frac{ f(x^{2^r}) f(x)}{ f(x^{2^{r-1}}) }.$$
Again, the regularity of the integrand allows us to alter the path of integration from the arc connecting $\zeta_I(h,k)$ and $\zeta_T(h,k)$ to the directed segment.
With this in mind, we estimate the absolute value of the integrand: $$\begin{aligned}
&\phantom{=} \left | \exp\left( \frac{\pi(1+2^{1-r}) }{24\zeta} + \frac{\pi(24n+1-2^r)\zeta}{12k^2} \right) \sqrt{\zeta \ 2^{r-1} } \right| \\ &\qquad\times
\left| -1 +
\frac{ f\left( \exp\left(\frac{-2\pi}{\zeta} + \frac{2H_0\pi i}{k} \right)\right)
f\left( \exp\left(\frac{-2^{1-r}\pi}{\zeta} + \frac{2^{1-r} H_0\pi i}{k} \right)\right) }
{ f\left( \exp\left(\frac{-\pi}{\zeta} + \frac{H_0\pi i}{k} \right)\right) }
\right| \\
&= \left | \exp\left( \frac{\pi(1+2^{1-r}) }{24\zeta} + \frac{\pi(24n+1-2^r)\zeta}{12k^2} \right) \sqrt{\zeta \ 2^{r-1} } \right| \\ &\qquad\times
\left| \sum_{m=1}^\infty p^*(m) \exp \left( \frac{-2^{1-r} \pi m}{\zeta} +
\frac{2^{1-r}H_0 \pi i m}{k} \right)
\right| \\
&= \exp\left( \frac{\pi(1+2^{1-r}) }{24} \Re\frac{1}{\zeta} \right)
\exp\left( \frac{\pi(24n+1-2^r)\Re\zeta}{12k^2} \right)
|\zeta|^{1/2} 2^{(r-1)/2} \\ &\qquad\times
\left| \sum_{m=1}^\infty p^*(m) \exp \left( \frac{-2^{1-r} \pi m}{\zeta} \right)
\exp\left( \frac{2^{1-r}H_0 \pi i m}{k} \right)
\right| \\
&\leqq e^{2\pi n}
|\zeta|^{1/2} 2^{(r-1)/2} \\ &\qquad\times
\sum_{m=1}^\infty | p^*(m) |
\exp\left( \frac{\pi(1+2^{1-r}) }{24} \Re\frac{1}{\zeta}-2^{1-r} \pi m\Re\frac{1}{\zeta} \right)\\
&= e^{2\pi n}
|\zeta|^{1/2} 2^{(r-1)/2} \sum_{m=1}^\infty | p^*(m) |
\exp\left( -\frac{\pi }{24\cdot 2^{r-1}} \Re\frac{1}{\zeta} \left( 24m - 1 - 2^{r-1} \right) \right) \\
&\leqq e^{2\pi n}
|\zeta|^{1/2} 2^{(r-1)/2}
\sum_{m=1}^\infty | p^*(m) |
\exp\left( -\frac{\pi }{24} \left( 24m - 1 - 2^{r-1} \right) \right) \\
&< e^{2\pi n}
|\zeta|^{1/2} 2^{(r-1)/2}
\sum_{m=1}^\infty | p^*(24m-1-2^{r-1}) |
y^{24m - 1 - 2^{r-1}} \quad\mbox{(where $y=e^{-\pi/24}$)} \\
& =c_0 |\zeta|^{1/2},
\end{aligned}$$ for a constant $c_0$. So, $$|{\mathcal{I}}_{0,1}| \leqq \frac{2\sqrt{2} k}{N} \left( \frac{\sqrt{2} k }{N} \right)^{1/2} c_0
< C_0 k^{3/2} N^{-3/2}$$ for a constant $C_0$.
### Estimation of ${\mathcal{I}}_{j,1}$ for $1+\lfloor r/2 \rfloor \leqq j \leqq r \leqq 4$ {#EstIj1-2}
Again, the regularity of the integrand allows us to alter the path of integration from the arc connecting $\zeta_I(h,k)$ and $\zeta_T(h,k)$ to the directed segment.
With this in mind, $$\begin{aligned}
&\phantom{=}\left| \exp\left( \frac{\pi(2^{2j-r} -1) }{12\zeta} + \frac{\pi(24n+1-2^r)\zeta}{12k^2} \right) \sqrt{\zeta \ 2^{r-j} } \right| \notag \\
& \qquad\times
\left| -1 +
f_{2j-r}\left( \exp\left(\frac{-2\pi}{\zeta} + \frac{2H_j\pi i}{k} \right)\right)
\right| \notag\\
&= \left| \exp\left( \frac{\pi(2^{2j-r} -1) }{12\zeta} \right)
\exp\left( \frac{\pi(24n+1-2^r)\zeta}{12k^2} \right) \sqrt{\zeta \ 2^{r-j} } \right| \notag\\
& \qquad\times
\left| \sum_{m=1}^\infty p_{2j-r}(m) \exp\left( \frac{-2\pi m}{\zeta} \right)
\exp\left(\frac{2H_j\pi i m}{k} \right) \right| \notag \\
&= \exp\left( \frac{\pi(2^{2j-r} -1) }{12}\Re\frac{1}{\zeta} \right)
\exp\left( \frac{\pi(24n+1-2^r)\Re\zeta}{12k^2} \right) |\zeta|^{1/2} 2^{(r-j)/2} \notag\\
& \qquad\times
\left| \sum_{m=1}^\infty p_{2j-r}(m) \exp\left( {-2\pi m} \Re\frac{1}{\zeta} \right)
\exp\left(\frac{2H_j\pi i m}{k} \right) \right| \notag \\
&\leqq
e^{2\pi n} |\zeta|^{1/2} 2^{r/2}
\sum_{m=1}^\infty p_{2j-r}(m)
\exp\left(
-\frac{\pi}{12}\Re\frac{1}{\zeta} (24m- 2^{2j-r} +1 ) \right) \label{later}\\
&\leqq
e^{2\pi n} |\zeta|^{1/2} 2^{r/2}
\sum_{m=1}^\infty p_{0}(m)
\exp\left(
-\frac{\pi}{12}\Re\frac{1}{\zeta} (24m- 2^{r} +1 ) \right)\notag\\
&= c_j |\zeta|^{1/2} \notag\end{aligned}$$ for a constant $c_j$. So, $$|{\mathcal{I}}_{j,1}| \leqq \frac{2\sqrt{2} k}{N} \left( \frac{\sqrt{2} k }{N} \right)^{1/2} c_j
< C_j k^{3/2} N^{-3/2}$$ for a constant $C_j$, when $1+\lfloor r/2 \rfloor\leqq j \leqq r$.
### Combining the estimates
$$\begin{aligned}
&\left|
i \underset{(k,2^r)=1}{\sum_{k=1}^N} k^{-5/2} \underset{(h,k)=1}{ \sum_{0\leqq h<k}}
e^{-2\pi i n h/k} \frac{\omega(h,k) \omega(2^r h,k)}{\omega(2h,k)}
{\mathcal{I}}_{0,1} \right. \\
&\left. + i \sum_{j=1}^{\lfloor r/2 \rfloor}
\underset{(k,2^r)=2^j}{\sum_{k=1}^N} k^{-5/2} \underset{(h,k)=1}{ \sum_{0\leqq h<k}}
e^{-2\pi i n h/k} \frac{\omega(h,k) \omega(2^{r-j}h, 2^{-j}k)}
{\omega(h,k/2)} \left( {\mathcal{I}}_{j,1} + {\mathcal{I}}_{j,2} \right) \right. \\
&\left. + i \sum_{j=1+\lfloor r/2 \rfloor}^{r}
\underset{(k,2^r)=2^j}{\sum_{k=1}^N} k^{-5/2} \underset{(h,k)=1}{ \sum_{0\leqq h<k}}
e^{-2\pi i n h/k} \frac{\omega(h,k) \omega(2^{r-j}h, 2^{-j}k)}
{\omega(h,k/2)} {\mathcal{I}}_{j,1} \right| \\
&< \sum_{j=0}^r \sum_{k=1}^N \sum_{h=0}^{k-1} C_j k^{-1} N^{-3/2}
+ \sum_{j=1}^{\lfloor r/2 \rfloor}\sum_{k=1}^N \sum_{h=0}^{k-1} C'_j k^{-1} N^{-3/2}\\
& \leqq C'' N^{-3/2} \sum_{k=1}^N 1,\qquad \mbox{(where $C'' = \sum_{j=0}^r C_j
+ \sum_{j=1}^{\lfloor r/2 \rfloor} C'_j$)} \\
&= O(N^{-1/2}).\end{aligned}$$
Thus, we may revise to $$\begin{gathered}
p_r(n) =
i \sum_{k=1}^N k^{-5/2} \underset{(h,k)=1}{\sum_{0\leqq h<k} } e^{-2\pi i n h/k}
\frac{\omega(h,k) \omega(2^{r-j}h, 2^{-j} k)}{\omega(2h,k)} {\mathcal{I}}_{0,2} \\
+ i \sum_{j=1+\lfloor \frac r2 \rfloor}^r
\sum_{k=1}^N k^{-5/2} \underset{(h,k)=1}{\sum_{0\leqq h<k} } e^{-2\pi i n h/k}
\frac{\omega(h,k) \omega(2^{r-j}h, 2^{-j} k)}{\omega(h,k/2)} {\mathcal{I}}_{j,2}
+ O(N^{-1/2}).
\label{IntegralZetaO}\end{gathered}$$
Evaluation of ${\mathcal{I}}_{j,2}$ for $1+\lfloor \frac r2 \rfloor \leqq j \leqq r $
-------------------------------------------------------------------------------------
Write ${\mathcal{I}}_{j,2}$ as $$\begin{gathered}
{\mathcal{I}}_{j,2} = \int_{K^{(-)}}
\exp\left( \frac{\pi(2^{2j-r} -1)}{12\zeta} + \frac{\pi(24n+1-2^r)\zeta}{12k^2}
\right) \sqrt{\zeta \ 2^{r-j} } \ d\zeta \\
- {\mathcal{I}}_{j,3}- {\mathcal{I}}_{j,4}, \label{FullCircle}\end{gathered}$$ where $${\mathcal{I}}_{j,3} := \int_{0}^{\zeta_I(h,k)}, $$ \_[j,4]{} := \_[\_T(h,k)]{}\^[0]{}, and ${\mathcal{I}}_{j,3}$ and ${\mathcal{I}}_{j,4}$ have the same integrand as .
### Estimation of ${\mathcal{I}}_{j,3}$ and ${\mathcal{I}}_{j,4}$
We note that the length of the arc of integration in ${\mathcal{I}}_{j,3}$ is less than $\frac{\pi k}{\sqrt{2}N}$, and on this arc $|\zeta| < \sqrt{2} k / N$. [@har1973 p. 272]. Also, $\Re \frac 1\zeta = 1$ on $K$ [@har1973 p. 271, Eq. (120.2)]. Further, $0<\Re \zeta < 2k^2/N^2$ [@har1973 p. 271, Eq. (119.6)]. The absolute value of the integrand is thus $$\begin{aligned}
&|2^{r-j} \zeta |^{1/2} \exp\left( \frac{ (24n+1-2^r) \pi \Re\zeta}{12k^2}
+ \frac{(2^{2j-r}-1)\pi}{12}\Re \frac 1\zeta\right)\\
<& 2^{(r-j)/2} 2^{1/4} k^{1/2} N^{-1/2}
\exp\left( \frac{(24n+1-2^r)\pi}{6 N^2} + \frac{(2^{2j-r}-1)\pi}{12} \right)$$ so that $$\begin{aligned}
|{\mathcal{I}}_{j,2}|
&< \pi k 2^{-1/2} N^{-1} 2^{(r-j)/2} 2^{1/4} k^{1/2} N^{-1/2}
\exp\left( \frac{(24n+1-2^r)\pi}{6 N^2} + \frac{(2^{2j-r}-1)\pi}{12} \right) \\
&= \pi k^{3/2} N^{-3/2}
2^{(2r-2j-1)/4}
\exp\left( \frac{(24n+1-2^r)\pi}{6 N^2} + \frac{(2^{2j-r}-1)\pi}{12} \right) \\
&= O\left(k^{3/2} N^{-3/2} \exp\left( \frac{(24n+1-2^r)\pi}{6 N^2}
\right) \right) . \end{aligned}$$ By the same reasoning, $|{\mathcal{I}}_{j,3}| =
O\left(k^{3/2} N^{-3/2} \exp\left( \frac{(24n+1-2^r)\pi}{6 N^2}
\right) \right) . $
We may therefore revise to $$\begin{gathered}
p_r(n) =
i \sum_{k=1}^N k^{-5/2} \underset{(h,k)=1}{\sum_{0\leqq h<k} } e^{-2\pi i n h/k}
\frac{\omega(h,k) \omega(2^{r-j}h, 2^{-j} k)}{\omega(2h,k)} \\ \times
\int_{K^{(-)}}
\sqrt{\zeta 2^{r-j-1}}
\exp\left\{ \frac{\pi(2^{1-r}+1)}{24\zeta} + \frac{\pi(24n+1-2^r)\zeta}{12k^2}
\right\} \ d\zeta\\
+ i \sum_{j=1+\lfloor \frac r2 \rfloor}^r
\sum_{k=1}^N k^{-5/2} \underset{(h,k)=1}{\sum_{0\leqq h<k} } e^{-2\pi i n h/k}
\frac{\omega(h,k) \omega(2^{r-j}h, 2^{-j} k)}{\omega(h,k/2)} \\ \times
\int_{K^{(-)}}
\sqrt{\zeta 2^{r-j}}
\exp\left\{ \frac{\pi(2^{2j-r}-1)}{12\zeta} + \frac{\pi(24n+1-2^r)\zeta}{12k^2}
\right\} \ d\zeta
+ O(N^{-1/2})
\label{IntegralZetaO2}\end{gathered}$$ and upon letting $N$ tend to infinity, obtain $$\begin{gathered}
p_r(n) =
i \sum_{k=1}^\infty k^{-5/2} \underset{(h,k)=1}{\sum_{0\leqq h<k} } e^{-2\pi i n h/k}
\frac{\omega(h,k) \omega(2^{r-j}h, 2^{-j} k)}{\omega(2h,k)} \\ \times
\int_{K^{(-)}}
\sqrt{\zeta 2^{r-j-1}}
\exp\left\{ \frac{\pi(2^{1-r}+1)}{24\zeta} + \frac{\pi(24n+1-2^r)\zeta}{12k^2}
\right\} \ d\zeta\\
+ i \sum_{j=1+\lfloor \frac r2 \rfloor}^r
\sum_{k=1}^\infty k^{-5/2} \underset{(h,k)=1}{\sum_{0\leqq h<k} } e^{-2\pi i n h/k}
\frac{\omega(h,k) \omega(2^{r-j}h, 2^{-j} k)}{\omega(h,k/2)} \\ \times
\int_{K^{(-)}}
\sqrt{\zeta 2^{r-j}}
\exp\left\{ \frac{\pi(2^{2j-r}-1)}{12\zeta} + \frac{\pi(24n+1-2^r)\zeta}{12k^2}
\right\} \ d\zeta
\label{IntegralZetaNoO}\end{gathered}$$
The final form
--------------
We may now introduce the change of variable $$\zeta = \frac{\pi\left( (2+2^{2j-r+1}-(2-\delta_{j0})^2) \right)}{24t},$$ (where the first summation in is the $j=0$ term separated out for clarity), which allows the integral to be evaluated in terms of $I_{3/2}$, the Bessel function of the first kind of order $3/2$ with purely imaginary argument [@ww1927 p. 372, §17.7] when we bear in mind that a “bent" path of integration is allowable according to the remark preceding Eq. (8) on p. 177 of [@gnw1944]. See also [@tma1990 p. 109]. The final form of the formula is then obtained by using the fact that Bessel functions of half-odd integer order can be expressed in terms of elementary functions.
We therefore have $$\begin{gathered}
p_r(n) = \frac{\pi}{(24n-2^r+1)^{3/4}} \left\{
\frac{(1+2^{r-1})^{3/4}}{2^{(r-2)/4}} \underset{(k,2^{\max(r,1)})=1)}{\sum_{k\geqq 1}}
k^{-1} \right. \\ \times \underset{(h,k)=1}{\sum_{0\leqq h<k}} e^{-2\pi i n h/k}
\frac{\omega(h,k) \omega(2^r h,k)}{\omega(2h,k)} I_{3/2}
\left( \frac{\pi \sqrt{(24n-2^r+1)(1+2^{r-1})}}{2^{r/2}\cdot6k} \right) \\
+ \sum_{j=1+\lfloor \frac r2 \rfloor}^r (2^{2j-r}-1 )^{3/4} 2^{(2-j+r)/2}
\underset{(k,2^r)=2^j}{\sum_{k\geqq 1}}
k^{-1}\\ \left.
\times \underset{(h,k)=1}{\sum_{0\leqq h<k}} e^{-2\pi i n h/k}
\frac{\omega(h,k) \omega(2^{r-j} h, 2^{-j}k)}{\omega(h,k/2)} I_{3/2}
\left( \frac{\pi \sqrt{(24n-2^r+1)(1+2^{2j-r})}}{6k} \right) \right\}
\end{gathered}$$ which, after application of the formula [@tma1990 p. 110] $$I_{3/2} (z) = \sqrt{ \frac{2z}{\pi}} \frac{d}{dz} \left( \frac{\sinh z}{z} \right),$$ is equivalent to Theorem \[MainFormula\]. $\Box$
Acknowledgment {#acknowledgment .unnumbered}
==============
The author thanks the anonymous referee for bringing the work of Zuckerman [@hsz1939] and Goldberg [@lg1981] to his attention. This, in turn, led the author to seek the more general result presented here in this final version of the paper.
[00]{} G.E. Andrews, [The Theory of Partitions]{}, Encyclopedia of Mathematics and its Applications, vol. 2, Addison-Wesley, 1976. Reissued, Cambridge, 1998.
T.M. Apostol, [Modular Functions and Dirichlet Series in Number Theory]{}, Graduate Texts in Mathematics, vol. 41, 2nd ed., Springer-Verlag, 1990.
K. Bringmann, J. Lovejoy, Dyson’s rank, overpartitions, and weak Maass forms, Int. Math. Res. Not. IMRN 2007, no. 19, 34 pp.
K. Bringmann and K. Ono, Coefficients of harmonic Maass forms, Proceedings of the 2008 University of Florida Conference on Partitions, $q$-series, and modular forms, to appear.
O-Y. Chan, Some asymptotics for cranks, Acta Arith. 120 (2005) 107–143.
W.Y.C. Chen, J.J.Y. Zhao, The Gaussian coefficients and overpartitions, Discrete Math. 305 (2005) 350–353.
S. Corteel, W.M.Y. Goh, P. Hitczenko, A local limit theorem in the theory of overpartitions, Algorithmica 46 (2006) 329–343.
S. Corteel, P. Hitczenko, Multiplicity and number of parts in overpartitions, Ann. Comb. 8 (2004) 287–301.
S. Corteel, J. Lovejoy, Overpartitions, Trans. Amer. Math. Soc. 356 (2004) 1623–1635.
S. Corteel, J. Lovejoy, A. J. Yee, Overpartitions and generating functions for generalized Frobenius partitions. Mathematics and computer science. III, 15–24, Trends Math., Birkhäuser, Basel, 2004.
S. Corteel, Sylvie, O. Mallet, Overpartitions, lattice paths, and Rogers-Ramanujan identities, J. Combin. Theory Ser. A 114 (2007) 1407–1437.
L. Euler, [Introductio in Analysin Infinatorum]{}, Marcum-Michaelem Bousquet, Lausanne, 1748.
L.R. Ford, Fractions, American Math. Monthly 45 (1938) 586–601.
A.M. Fu, A. Lascoux, $q$-identities related to overpartitions and divisor functions, Electron. J. Combin. 12 (2005) \#R38, 7 pp.
E. Grosswald, Some theorems concerning partitions, Trans. Amer. Math. Soc. 89 (1958) 113–128.
E. Grosswald, Partitions into prime powers, Mich. Math. J. 7 (1960) 97–122.
L. Goldberg, Transformation of theta-functions and analogues of Dedekind sums, Ph.D. thesis, University of Illinois (1981).
M. Haberzetle, On some partition functions, Amer. J. Math. 63 (1941) 589–599.
P. Hagis, A problem on partitions with a prime modulus $p\geq 3$, Trans. Amer. Math. Soc. 102 (1962) 30–62.
P. Hagis, Partitions into odd summands, Amer. J. Math. 85 (1963) 213–222.
P. Hagis, On a class of partitions with distinct summands, Trans. Amer. Math. Soc. 112 (1964) 401–415.
P. Hagis, Partitions into odd and unequal parts, Amer. J. Math. 86 (1964) 317–324.
P. Hagis, Partitions with odd summands–some comments and corrections, Amer. J. Math. 87 (1965) 218–220.
P. Hagis, A correction of some theorems on partitions, Trans. Amer. Math. Soc. 118 (1965) 550.
P. Hagis, On partitions of an integer into distinct odd summands, Amer. J. Math. 87 (1965) 867–873.
P. Hagis, Some theorems concerning partitions into odd summands, Amer. J. Math. 88 (1966) 664–681.
P. Hagis, Partitions with a restriction on the multiplicity of summands, Trans. Amer. Math. Soc. 155 (1971) 375–384.
G.H. Hardy, S. Ramanujan, Asymptotic formulae in combinatory analysis, Proc. London Math. Soc. (2) 17 (1918) 75–115.
M.D. Hirschhorn, J.A. Sellers, Arithmetic relations for overpartitions, J. Combin. Math. Combin. Comput. 53 (2005) 65–73.
M.D. Hirschhorn, J.A. Sellers, Arithmetic properties of overpartitions into odd parts, Ann. Comb. 10 (2006) 353–367.
L.K. Hua, On the number of partitions into unequal parts, Trans. Amer. Math. Soc. 51 (1942) 194–201.
S. Iseki, A partition functin with some congruence condition, Amer. J. Math. 81 (1959) 939–961.
S. Iseki, On some partition functions, J. Math. Soc. Japan 12 (1960) 81–88.
S. Iseki, Partitions in certain arithmetic progressions, Amer. J. Math. 83 (1961) 243–264.
D.M. Kane, Resolution of a conjecture of Andrews and Lewis involving cranks of partitions, Proc. Amer. Math. Soc. 132 (2004) 2247–2256.
J. Lehner, A partition function connected with the modulus five, Duke Math. J. 8 (1941) 631–655.
J. Livingood, A partition function with prime modulus $p>3$, Amer. J. Math. 67 (1945) 194–208.
J. Lovejoy, Gordon’s theorem for overpartitions, J. Combin. Theory Ser. A 103 (2003) 393–401.
J. Lovejoy, Overpartitions and real quadratic fields, J. Number Theory 106 (2004) 178–186.
J. Lovejoy, Overpartition theorems of the Rogers-Ramanujan type, J. London Math. Soc. (2) 69 (2004) 562–574.
J. Lovejoy, A theorem on seven-colored overpartitions and its applications, Int. J. Number Theory 1 (2005) 215–224.
J. Lovejoy, Rank and conjugation for the Frobenius representation of an overpartition, Ann. Comb. 9 (2005) 321–334.
J. Lovejoy, Partitions and overpartitions with attached parts, Arch. Math. (Basel) 88 (2007) 316–322.
K. Mahlburg, The overpartition function modulo small powers of 2, Discrete Math. 286 (2004), no. 3, 263–267.
I. Niven, On a certain partition function, Amer. J. Math. 62 (1940) 353–364.
H. Rademacher, On the partition function $p(n)$, Proc. London Math. Soc. (2) 43 (1937) 241–254.
H. Rademacher, On the expanson of the partition function in a series, Ann. Math. (2) 44 (1943) 416–422.
H. Rademacher, Lectures on Analytic Number Theory, Tata Institute, Bombay, 1954-1955.
H. Rademacher, [Topics in Analytic Number Theory]{}, Die Grundelhren der mathematischen Wissenschaften, Bd. 169, Springer-Verlag, 1973.
N. Robbins, Some properties of overpartitions. JP J. Algebra Number Theory Appl. 3 (2003) 395–404.
. R[ø]{}dseth, J. A. Sellers, On $m$-ary overpartitions, Ann. Comb. 9 (2005) 345–353.
V.V. Subramanyasastri, Partitions with congruence conditions, J. Indian Math. Soc. 11 (1972) 55–80.
G.N. Watson, [A Treatise on the Theory of Bessel Functions]{}, 2nd ed., Cambridge, 1944.
E.T. Whittaker and G.N. Watson, [A Course of Modern Analysis]{}, 4th ed., Cambridge, 1927.
H. S. Zuckerman, [On the coefficients of certain modular forms belonging to subgroups of the modular group]{}, Trans. Amer. Math. Soc. 45 (1939) 298–321.
| {
"pile_set_name": "ArXiv"
} |
harvmac
John Preskill
California Institute of Technology
Pasadena, CA 91125
and
Alexander Vilenkin[^1]
California Institute of Technology
Pasadena, CA 91125
and
Lyman Laboratory of Physics
Harvard University
Cambridge, MA 02138
.3in We systematically analyze the decay of metastable topological defects that arise from the spontaneous breakdown of gauge or global symmetries. Quantum–mechanical tunneling rates are estimated for a variety of decay processes. The decay rate for a global string, vortex, domain wall, or kink is typically suppressed compared to the decay rate for its gauged counterpart. We also discuss the decay of global texture, and of semilocal and electroweak strings.
Topological defects arise as stable solutions of classical field equations in a variety of models with spontaneously broken symmetries. The higher symmetry usually characterizes the high–temperature phase of the model, and the symmetry breaking corresponds to a phase transition. The type of defects formed at a phase transition depends on the topology of the vacuum manifold, $M=G/H$, where $G$ and $H$ are, respectively, the symmetry groups before and after the symmetry breaking .
Linear defects, or strings, are formed if the first homotopy group is nontrivial, $\pi_1(M) \neq I$; point defects, or monopoles, are formed if $\pi_2 (M) \neq I$, and sheet–like defects, or domain walls, are formed if $\pi_0(M) \neq I$. These defects are stable in the sense that “unwinding” the topological knots associated with the defects would require going over an infinitely high potential barrier. The physical properties of defects crucially depend on whether the broken symmetry is gauge or global. For example, the mass of a global monopole and the mass per unit length of a global string are infrared–divergent, while the corresponding quantities for gauge defects are finite.
The purpose of this paper is to give a systematic account of the decay of [*metastable*]{} defects in relativistic field theories. A metastable defect is a stable solution to the classical field equations, stable in the sense that small vibrations about the solution have non-negative frequency squared. But a metastable defect can be unwound by going over a [*finite*]{} potential barrier; hence, it can decay quantum mechanically. In the limit of small $\hbar$, the decay rate approaches zero like $e^{-B/\hbar}$. We will describe how the “tunneling action” $B$ can be calculated for various types of metastable defects.
One type of metastable defect can arise in models with a sequence of phase transitions, Defects will be formed if the manifolds $M_1 = G/H_1$ and $M_2 =
H_1/H_2$ have nontrivial homotopy groups. However, these defects will not be topologically stable if $M= G/H_2$ has trivial topology. Consider, for example, the sequence with $\pi_0(G) = \pi_1 (G) = I$. Since $\pi_1(G/Z_2) = Z_2$ and $\pi_0 (Z_2) = Z_2$, the first phase transition gives rise to strings and the second to domain walls. However, it can be shown that strings formed at the first phase transition become boundaries of the walls formed at the second phase transition. Closed and infinite walls without boundaries can also be formed, but they are not topologically stable: an infinite planar wall decays by spontaneous nucleation of circular holes bounded by strings. Quite similarly, the sequence with $\pi_2(G) = \pi_1 (G) = I$ leads to formation of monopoles which get connected by strings. The strings formed at the second phase transition are metastable and decay by nucleation of monopole–antimonopole pairs.
These observations are not new. Indeed, the decay rate of a metastable string was estimated in Ref. \[\], and the decay rate of a metastable wall was estimated in Ref. , assuming that the thickness of the defects can be neglected. Our intent is to discuss such tunneling phenomena in a fairly comprehensive way. In various cases, we describe the instanton (or “bounce”) corresponding to the decay, and estimate the tunneling action. While some of our calculations merely rederive familiar results, we also present a number of new results. We consider defects in $D=1,\; 2,$ and $3$ spatial dimensions, and discuss the connection between tunneling phenomena in different dimensions. (Some of these lower–dimensional defects may have applications to condensed matter physics.) We emphasize in particular the differences between defects arising from gauge and global symmetries. Because global defects have long-range interactions mediated by massless Nambu–Goldstone bosons, the decay of a global defect is typically suppressed compared to the decay of its gauged counterpart.
There are also other types of metastable defects that are not associated with a hierarchy of symmetry breakdown of the form eq. . One interesting example is global texture , and we discuss the decay of metastable texture in various dimensions. Another interesting case is the electroweak string \[\] (or vortex), which might occur in realistic extensions of the standard model; we analyze its decay as well.
We outline a general classification of metastable defects in Section 2, and then proceed in the remainder of the paper to discuss various special cases in more detail. Section 3 concerns defects that arise from a hierarchy of gauge symmetry breaking—monopoles, strings, domain walls, and their lower dimensional analogs. Section 4 analyzes the consequences of a hierarchy of global symmetry breaking. Metastable defects arising from the intrinsic breaking of a spontaneously broken global symmetry by a small perturbation (like axion domain walls) are discussed in Section 5. We consider in Section 6 heavy metastable defects that decay to light stable defects. The decay of global texture is treated in Section 7. Electroweak and “semilocal” defects are discussed in Section 8. Section 9 contains our conclusions, including some remarks about the cosmological implications of metastable defects.
In this section, we formulate the general theory of metastable topological defects. This theory will be illuminated later by various examples.
The metastable defects that we will discuss fall into three broad categories. Those in the first category are associated with a hierarchy of (gauge or global) symmetry breakdown. Those in the second category are associated with the intrinsic breaking of a global symmetry by a small perturbation. Those in the third category do not fit into either of the first two categories—they are classically stable but are not prevented from decaying by any topological conservation law. (Examples include global texture and electroweak vortices.)
[**a) Hierarchy of Symmetry Breakdown**]{}
Let us first consider models with a sequence of phase transitions Here G is a finite-dimensional compact Lie group that we may take to be connected. It may be either a global symmetry group or a gauge group. (The distinction between global and gauge symmetry will be discussed later.) The $G$ symmetry breaks to the subgroup $H_1$ at the mass scale $\eta_1$, and then breaks further to $H_2\subset H_1$ at the much lower mass scale $\eta_2$. We wish to address whether topological defects associated with the second stage of symmetry breakdown remain topologically stable when $H_1$ is embedded in $G$, and also the closely related but somewhat different question whether topological defects produced in the first stage survive when the second symmetry breakdown occurs.
[*i) Codimension 1*]{}
By a codimension 1 defect we mean one of dimension $D-1$ in $D$ spatial dimensions—it is a domain wall, or, in $D=1$, a particle or “kink.” Topologically stable codimension 1 defects exist if the vacuum manifold is disconnected. Thus, if the symmetry group $H_1$ breaks to $H_2$ (and assuming no accidental degeneracy), these defects are classified by the homotopy group $\pi_0(H_1/H_2)$. But if $H_1$ is actually embedded in a larger symmetry group $G$ that breaks at a much larger mass scale, then this defect may not be absolutely stable (although it is very long-lived). The domain wall separates two regions in which the order parameter takes values in two distinct connected components of $H_1/H_2$. If these components are connected in the larger manifold $G/H_2$, then the domain wall is metastable. Mathematically, since $H_1/H_2$ is included in $G/H_2$, there is a natural homomorphism Metastable defects of codimension 1 are classified by the nontrivial elements of the kernel of this homomorphism.
Associated with each nontrivial element of this kernel, there is a path in $G$ that begins at the identity and ends at an element of $H_1$ that is not connected to the identity in $H_1$. This path defines a representative of a nontrivial element homotopy class in $\pi_1(G/H_1)$. Associated with this class there is a string or vortex that arises in the symmetry breakdown $G\to
H_1$. The physical interpretation is that the metastable codimension 1 defect can end on a codimension 2 defect.
[*ii) Codimension 2*]{}
A codimension 2 defect is a “string”, or, in D=2, a particle or “vortex.” By reasoning analogous to that above, metastable defects of codimension 2 are classified by the nontrivial elements of the kernel of the homomorphism Associated with each nontrivial element of the kernel, there is a noncontractible closed loop in $H_1$ that can be deformed to a point in $G$. This deformation of the loop defines a nontrivial element of Associated with this element is a magnetic monopole that arises when $G$ breaks to $H_1$. The physical interpretation is that the metastable codimension 2 defect can end on a codimension 3 defect.
[*iii) Codimension 3*]{}
A codimension 3 defect is a “monopole,” a particle in $D=3$. In principle, metastable monopoles are classified by the nontrivial elements of the kernel of the homomorphism However, this kernel is always trivial—metastable monopoles do not exist. Mathematically, this is because $\pi_2(H_1)=I$ for any finite-dimensional compact Lie group $H_1$. (Note that metastable domain walls are associated with nontrivial elements of $\pi_0(H_1)$, and metastable strings are associated with nontrivial elements of $\pi_1(H_1)$.)
We can express this result in more physical terms in the case where $G$ is a gauge symmetry. Then the magnetic monopole that arises when $H_1$ breaks to $H_2$ carries a conserved magnetic charge, which can be detected by measuring the long-range $H_2$ magnetic field of the monopole. Embedding $H_1$ in $G$ does not extinguish that long-range field, or destroy the conservation law—the monopole remains absolutely stable. (Although quantum effects, specifically color confinement, may cause the magnetic field to be screened, these effects do not prevent the charge from being detected at long range, and do not destroy the conservation law .)
[**b) Comments**]{}
[*i) Bianchi Identity*]{}
We could also consider a more intricate symmetry breaking hierarchy of the form One might then wonder if it is possible for a monopole to arise when $G$ breaks to $H_1$ such that the monopole becomes attached to a string when $H_1$ breaks to $H_2$, and the string in turn becomes attached to a wall when $H_2$ breaks to $H_3$. It is easy to see that this is not possible. This conclusion is probably best understood as a consequence of the Bianchi identity—“the boundary of a boundary is zero.” If a string is the boundary of a domain wall, then the string cannot end (on a monopole). In terms of the above homotopy classification, we saw that there are two types of strings that can arise when $H_1$ breaks to $H_2$. A string that ends on a monopole is associated with a noncontractible closed path in $H_1$, while a string that bounds a domain wall is associated with an [*open*]{} path in $H_1$ that begins at the identity and ends at an element of $H_2$ that is not connected to the identity in $H_2$.
Another observation is closely related to the above: It is impossible for a string that ends on a monopole to have nontrivial Aharonov-Bohm interactions that can detect the “quantum hair” \[\] of charged particles. Strings that detect quantum hair carry a magnetic flux that does not lie in the connected component of the unbroken group $H_2$; they are the kind of strings that can bound domain walls, not the kind that can end on monopoles . This is not to say that the Aharonov–Bohm interactions of strings that end on monopoles must be completely trivial; rather, the group element that characterizes the flux trapped in the core of the string must be connected to the identity in $H_2$. An example of a string that ends on a monopole, yet has nontrivial Aharonov–Bohm interactions, is the electroweak string that we will discuss in Section 8 .
[*ii) Bundles*]{}
The analysis in (a) above can be reexpressed in the language of fiber bundles. When the symmetry breaking pattern occurs, we may view the vacuum manifold $G/H_2$ as the total space of a bundle with basespace $G/H_1$, fiber $H_1/H_2$, and structure group $H_1$. Then the topological defects arising in the first stage of the symmetry breakdown are determined by the topology of the basespace, and the defects arising in the second stage are determined by the topology of the fiber. Our criterion for a codimension $n+1$ defect to be metastable, then, is that a mapping that represents a nontrivial element of $\pi_n$ of the fiber is topologically trivial in the total space of the bundle.
[*iii) Survival*]{}
We may also ask a slightly different question than that formulated in (a) above. If a defect arises when $G$ breaks to $H_1$, will that defect “survive” if the symmetry breaks further, to $H_2$? Before, we found the criterion for a monopole to become attached to a single string, or for a string to become attached to a single wall. Now we are asking a more general question, because it is also possible for a monopole to become attached to more than one string, or for a string to become attached to more than one wall.
The criterion for a defect to survive is most simply stated in the fiber bundle language. A codimension $n+1$ topological defect that arises when $G$ breaks to $H_1$ is associated with a nontrivial element of $\pi_n$ of the basespace of the bundle. The defect survives if this element can be “lifted” to $\pi_n$ of the total space of the bundle. That is, the bundle comes equipped with a projection map $\phi:~G/H_2\to G/H_1$, and the defect is characterized by a topologically nontrivial map $f:~S^n\to G/H_1$. The defect survives if there is a continuous map $\tilde f:~S^n\to G/H_2$ such that $f=\phi\circ\tilde f$.
Domain walls always survive, but strings and monopoles need not. Note that it is possible for a monopole to be attached to two (or more) strings where one string is heavy and the other is light. In $D=2$, then, the monopole mediates the decay of a heavy vortex to a light vortex (or several light vortices). If two degenerate strings end on a monopole \[\], then in $D=2$ the monopole is an instanton that allows two degenerate vortices to mix quantum mechanically. Similarly, a string could be attached to a heavy wall and a light wall. Then, in $D=1$, the vortex mediates the decay of a heavy kink to a light kink. If two degenerate walls end on a string, then in $D=1$ the vortex allows degenerate kinks to mix.
[**c) Intrinsic Symmetry Breaking**]{}
Another type of metastable defect can arise when an [*approximate*]{} global symmetry is spontaneously broken. Consider the pattern That is, $G_{\rm approx}$ is spontaneously broken to $H_{\rm approx}$, and is also intrinsically broken by a small perturbation to $G_{\rm exact}$. (Thus $G_{\rm approx}$ must be a global symmetry group; gauge symmetries are always exact.) The unbroken exact symmetry is $H_{\rm exact}$, the intersection of $G_{\rm exact}$ and $H_{\rm approx}$.
If we ignore the intrinsic symmetry breaking, then topological defects of codimension $n+1$ are classified by $\pi_n(G_{\rm approx}/H_{\rm approx})$. We may ask if such a defect can “survive” when the intrinsic symmetry breakdown is taken into account. The criterion for survival can be expressed in the following way. The defect is associated with a topologically nontrivial map from $S^n$ to the approximate vacuum manifold $G_{\rm approx}/H_{\rm approx}$ The defect survives if this mapping can be continuously deformed to one that takes values in the exact vacuum manifold $G_{\rm exact}/H_{\rm exact}$.
If a $G_{\rm approx}/H_{\rm approx}$ domain wall does not survive, then the energy densities on the two sides of the wall are unequal, and a resulting pressure pushes the wall away. If a $G_{\rm approx}/H_{\rm approx}$ string does not survive, then it becomes attached to one or more walls; if the number of walls is exactly one, then there is a metastable wall that can end on a loop of string. If a $G_{\rm approx}/H_{\rm approx}$ (global) monopole does not survive, then it becomes attached to one or more strings; if the number of strings is exactly one, there is a metastable string that can break by nucleating a monopole pair.
The most familiar example of this phenomenon is the axion string, which becomes attached to $N$ axion domain walls . Thus, if $N=1$, an axion domain wall is metastable and can decay by nucleating a loop of axion string. In $D=1$, the axion vortex mediates the decay of an axion kink.
[**d) Other Cases**]{}
There are a few interesting classes of metastable defects that do not arise due to a hierarchy of symmetry breakdown, or because of intrinsic symmetry breaking. These defects are classically stable, but are not forbidden to decay quantum mechanically.
[*i) Global Texture*]{}
If a global symmetry $G$ is spontaneously broken to $H$, then a global texture (or “skyrmion”) is a field configuration that takes values in (or near) the vacuum manifold $G/H$ everywhere. (There is no “restoration” of the spontaneously broken symmetry in its core.) Such configurations, if they have finite energy, are classified in D spatial dimensions by $\pi_D(G/H)$ . A texture has only gradient energy, and for $D\ge 3$, the gradient energy makes it want to collapse. But it can be stabilized if suitable higher derivative terms \[\] are introduced into the action (or if a [*subgroup*]{} of $G$ is gauged \[\]).
For $D=1$, its gradient energy makes a texture want to spread out. It can be stabilized if space is compactified to a circle of finite circumference.
Unlike a kink in one dimension, a (gauge) vortex in two dimensions, or a (gauge) monopole in three dimensions, a global texture in D dimensions is not separated from the vacuum by an infinite energy barrier. Thus, even if it is classically stable, there is no topological conservation law that prevents it from decaying quantum mechanically.
Indeed, in any model that contains a $D$-dimensional global texture, there is also a “global instanton” that mediates the decay of the texture. This instanton is also classified by $\pi_D(G/H)$; it is a pointlike defect in $D+1$-dimensional Euclidean spacetime, with the world line of a texture ending on the instanton. We will discuss texture decay in more detail in Section 7.
[*ii) Semilocal and Electroweak Strings*]{}
“Semilocal” strings \[\] (or vortices) can arise in models that have both gauge and global symmetries that are spontaneously broken, but only if the symmetries “mix;” that is, there must be unbroken global symmetry generators that are nontrivial linear combinations of spontaneously broken gauge symmetry generators and global symmetry generators.
Consider the pattern Here $G_1$ is the gauge group and $G_2$ is a global symmetry group. $H_1$ is the unbroken gauge group, the intersection of $G_1$ and $H$. In this scheme, for $D=2$, there is a topologically conserved magnetic flux that is classified by $\pi_1(G_1/H_1)$, and a natural homomorphism If this homomorphism has a nontrivial kernel (which is possible only if $G_1$ and $G_2$ mix ), then there are field configurations that carry nontrivial $G_1/H_1$ magnetic flux, where the order parameter takes values in the vacuum manifold $[G_1\times G_2]/H$ everywhere.
When such configurations exist, it becomes a dynamical question whether the energy in a given magnetic flux sector is minimized by a localized vortex or by a configuration in which the magnetic flux is spread out over an arbitrarily large area. The answer depends on the details of the Higgs potential.
In some models, there may be vortices that are classically stable, but are kinematically allowed to decay to configurations in which the magnetic flux is spread out. Then the decay is mediated by a global monopole , as we will describe in more detail in Section 8. Similarly, for $D=3$, there may be string that can decay via the nucleation of a global monopole pair.
The simplest example of the semilocal phenomenon is the case $G_1=U(1)$, $G_2=SU(2)$, and $H=U(1)$, which may be regarded as the minimal standard electroweak model in the limit $\sin^2\theta_W=1$.
It is also interesting to consider a semilocal model in which the vortex [*is*]{} stable, and ask what would happen if $G_2$ were gauged. Then the vortex no longer carries any conserved topological charge, and will surely decay. But if the $G_2$ gauge coupling is sufficiently weak, one expects the vortex to remain [*classically*]{} stable . Thus, the vortex is metastable, and its decay is mediated by a magnetic monopole, as we will explain in Section 8. Similarly, for $D=3$, there is a metastable string that decays by nucleating a monopole pair. Because such a metastable string arises in the minimal standard model, for an (unrealistic \[\]) range of values of $\sin^2\theta_W$ and the Higgs mass, we refer to it as an “electroweak string.”
We will now discuss in more detail the decay of metastable defects in a model with the pattern of [*gauge*]{} symmetry breaking .
[**a) Codimension 2**]{}
[*i) $D=3$*]{}
We first consider the decay of metastable gauge strings in three dimensions. If the string can break due to the nucleation of a monopole-antimonopole pair, then there will be a nonzero probability of decay per unit length and time. A long straight string with string tension $\mu$ will tunnel to a configuration of the same energy, in which there is a gap in the string of length $L$, with a monopole at rest at each end. The energy cost of producing the pair is $2m$, where $m$ is the monopole mass; this cost must be balanced by the energy $\mu L$ saved due to the reduction in the string length. Thus, the initial separation of the monopole and antimonopole is $L=2 m/\mu$. The monopoles subsequently pull apart, consuming the string. Since the width of the barrier is $L$ and its height is $2m$, we can crudely estimate that the decay probability will be of the form $P\sim
e^{-B}$, where the WKB tunneling factor is of order $m^2/\mu$ (in units with $\hbar=1$).
To calculate the semiclassical tunneling factor more precisely, it is most convenient to use the Euclidean path integral method \[\]. The decay is described by an instanton (or “bounce”), which is a solution of Euclidean (imaginary–time, $t=-i\tau$) field equations approaching the unperturbed string solution at $\tau \rightarrow \pm
\infty$. The origin of $\tau$ can be chosen so that the instanton is symmetric with respect to $\tau \rightarrow - \tau$. Then the instanton solution at $\tau = 0$ gives the field configuration at the moment of nucleation of the monopole–antimonopole pair. The semiclassical decay probability is where is the difference between the Euclidean actions of the instanton ($S$) and of the unperturbed string ($S_0$).
In the four–dimensional language, the string axis is represented by a two–dimensional world sheet, and the monopole center by a one–dimensional world line. A static straight string oriented along the $z$–axis corresponds to a planar world sheet, $x=y=0$; in the instanton solution this world sheet has a “hole" which is bounded by the closed world line of the monopole (see Fig. 1). In relativistic field theories, the string is invariant with respect to Lorentz boosts in the $z$–direction, which turn into $z-\tau$ rotations after the change $t \rightarrow -i\tau$. The bounce solution preserves this symmetry; the monopole world line is a circle in the $z-\tau$ plane. In the rest of this paper we shall assume relativistic invariance, although our results can be easily generalized to the nonrelativistic case.
The planar string world sheet with a circular hole bounded by the monopole world line can be thought of as a domain wall bounded by a string in four spatial dimensions. Indeed, the conditions for the existence of planar and linear defects in four dimensions are, respectively, $\pi_1 (M) \neq I$ and $\pi_2 (M) \neq I$. From this point of view, our instanton is a time–independent solution of ($4 + 1$)–dimensional field equations describing a planar “wall" with a circular hole bounded by a “string" in a state of unstable equilibrium. If the “string" radius is decreased, the hole begins to shrink, and if it is increased, it starts growing. The bounce solution must have such a negative mode, according to general arguments \[\].
If the radius of the monopole world line is much greater than the string thickness, $R \gg \delta_s$, we can use the thin–string and thin–monopole approximation, in which the actions for the monopole and for the string are proportional to the world line length and world sheet area, respectively, For our instantons, the bounce action is This expression is stationary with respect to $R$ for Hence, the initial separation of the monopole–antimonopole pair is $2m/\mu$, as we anticipated, and the nucleation probability is $P \propto
\exp(-\pi m^2/\mu)$, in agreement with .
The thin–defect approximation is justified if This condition is typically satisfied if the symmetry breaking scale of monopoles is much greater than that of the strings, $\eta_1 \gg \eta_2$. Exceptions to this rule can occur if Higgs or gauge couplings of the model are very small. For $R \; \roughly{<} \;
\delta_s$, the bounce action depends on the details of the model, and no simple estimate can be given in the general case.
[*ii) $D=2$*]{}
In two spatial dimensions, models like give rise to metastable vortices. The vortex can tunnel to a configuration of the same energy, and about the same core size. This configuration has a nonzero magnetic field, but its total magnetic flux is trivial. Thus, there is nothing to prevent the configuration from subsequently relaxing to the vacuum.
The instanton in this case is a monopole–antimonopole pair in unstable equilibrium in three Euclidean dimensions. The Coulombic attraction between the monopole and antimonopole is balanced by the tension of the strings pulling them in opposite directions (see Fig. 2). The theory must have a solution of this form, since we know that models of the type give monopoles connected by strings in three dimensions. The mid-section of the instanton (surface $\Sigma$ in Fig. 2) gives the field configuration of the decaying vortex right after the tunneling.
The bounce action can be roughly estimated as where $m$ is the monopole mass, the second term is the Coulombic energy of the pair and the last term is the energy of the missing piece of string. This expression is stationary for $R \sim e^{-1} \mu^{- 1/2} \sim \delta_s$, where $\delta_s$ is the string thickness. Eq. is only a rough estimate because $R$ is comparable to the thickness of the string; therefore, the Coulomb interaction between the monopoles is significantly distorted by the string. The last two terms in are both of the order $e^{-1}\mu^{1/2}$ and are negligible compared to the first term if the symmetry breaking scale of the monopoles is much greater than that of strings. Hence,
Recall that the “mass” of the monopole actually has the dimensions of action in $D=2$ (in units with $c=1$). In order of magnitude it is $m\sim 4\pi\eta_1/e$, where $\eta_1$ is the expectation value of a scalar field; $\eta_1$ has the dimensions of $({\rm energy})^{1/2}$, and $e^{-1}$ has the dimensions of $({\rm energy})^{1/2}({\rm length})$, in two spatial dimensions.
For $\eta_1>>\eta_2$, the effects on the vortex of the physics at energy scale $\eta_1^2$ can be conveniently incorporated into an effective Lagrangian . Were we to ignore the heavy magnetic monopoles, the low energy effective theory would have an exact topological conservation law that would ensure that the vortex is absolutely stable. But when the monopole instantons are integrated out, operators are induced in the effective Lagrangian that violate this conservation law. Specifically, an operator appears that annihilates (or creates) a vortex, with a coefficient that is proportional to $e^{-m}$. Calculating with this effective Lagrangian, we again find a vortex decay rate of order $e^{-2m}$.
[**b) Codimension 1**]{}
[*i) $D=3$*]{}
The decay of metastable domain walls can be analyzed in a similar way. If the wall can herniate by nucleating a loop of string, then there will be a nonzero decay probability per unit time and area . A planar wall will tunnel to a configuration of the same energy, with a circular hole in the wall bounded by the string loop. The radius of the hole is chosen so that the energy cost of the string loop matches the energy saved due to the missing wall. The string loop appears at rest, and then expands, consuming the wall.
Again, we compute the tunneling action using the Euclidean path integral method. In the thin–defect approximation, the action is where $\mu$ and $\sigma$ are string and wall tensions, respectively. In the instanton solution, the wall world membrane is a three–dimensional hyperplane with a spherical hole bounded by the string world sheet. The bounce action is then which is stationary with respect to $R$ for
[*ii) $D=2$*]{}
In two spatial dimensions, models like give rise to metastable linear defects which decay by nucleation of vortex–antivortex pairs. The decay is described by Eqs. —, where now $m$ stands for the vortex mass and $\mu$ for the tension of the linear defects.
[*iii) $D=1$*]{}
In one spatial dimension, the domain wall becomes a particle, or “kink.” Were it not for the existence of vortices, the kink would carry a conserved topological quantum number, and would be stable. But the vortices enable the kink to decay.
The instanton describing the kink decay is a vortex–antivortex pair, in two Euclidean dimensions, in which the attraction between the vortex and antivortex is balanced by the tension of linear defects (“walls") attached to the vortices. An argument similar to the one that led to Eq. gives for the case when the first symmetry breaking scale is much greater than the second. Here $\mu$ is the vortex action and is given by the same expression as the string tension in the corresponding $(3+1)$–dimensional theory.
The kink decays to what might be called a gauged texture, a configuration that matches the asymptotic behavior of the kink, but in which the Higgs field is a pure gauge that has a nontrivial winding around $G/H_1$. This configuration has a different (1+1-dimensional) “Chern-Simons number” than the kink. The change in the Chern-Simons number is provided by the vortex instanton.
Again, the violation of the topological conservation law can be incorporated into an effective Lagrangian. Integrating out vortices induces an operator that destroys (or creates) a kink, with a coefficient proportional to $e^{-\mu}$.
In the case of global symmetry breaking, the instantons still represent unstable equilibrium configurations of higher–dimensional defects, but there is an important difference. Global defects have long-range interactions mediated by massless Nambu–Goldstone bosons. This leads to an increase in the height of the potential barrier and to a strong suppression of the decay. Moreover, since the dominant part of the energy of global defects is located outside the core, the thin–defect approximation can no longer be used, and the field configuration of the instanton has to be studied in some more detail.
[**a) Codimension 1**]{}
A simple example of a model with metastable domain walls or kinks is where $\varphi_1$ and $\varphi_2$ are complex scalar fields and $$V(\varphi_1,\;\varphi_2) =\lambda_1(\vert \varphi_1\vert^2 -
\eta^2_1)^2 + \lambda_2(\vert \varphi_2\vert^2 - \eta^2_2)^2$$ This model, for $\eta_1>>\eta_2$, realizes eq. as a hierarchy of [*global*]{} symmetry breaking, with $G=U(1)$.
Without the last term in the potential, the model would have a $U(1) \times
U(1)$ symmetry and $V(\varphi)$ would be minimized by with arbitrary $\theta_1$ and $\theta_2$. But the last term breaks the symmetry to $U(1)$, and fixes the value of $\theta_1+2\theta_2$. To simplify the analysis, we shall assume that $\lambda_{12}$ is sufficiently small that it does not affect the magnitudes of the expectation values . (Specifically, one needs $\lambda_1 \eta^2_1 \gg \lambda_{12}\eta^2_2, \;
\lambda_2\eta^2_2 \gg \lambda_{12} \eta^2_1$). Then the effective Lagrangian for the angular variables $\theta_1$ and $\theta_2$ is
We can diagonalize this Lagrangian by introducing the new variables $\tilde\theta_1$ and $\tilde\theta_2$, Then, to leading order in $\eta_1 / \eta_2$ we have $\theta_1 + 2\theta_2 =
2\tilde\theta_2$ and where $m^2 = 4\lambda_{12}\eta^2_1$. The field $\tilde\theta_1$ is a massless Nambu–Goldstone field resulting from the breaking of the global symmetry $\theta_1 \rightarrow \theta_1 + 2
\alpha,\; \theta_2 \rightarrow \theta_2 - \alpha$, while the field $\tilde\theta_2$ is described by a sine–Gordon Lagrangian. The potential for $\tilde\theta_2$ is minimized when $\tilde\theta_2 = n\pi$ ($n={\rm
integer}$), and the kink solution that interpolates between, say, $\tilde\theta_2 = 0$ and $\tilde\theta_2 = \pi$ is
The energy of the kink (or tension in the wall) is
[*i) $D=1$*]{}
In estimating the decay rates, we will start with the case of the kink in one dimension, and then work our way up to $D=3$.
In the kink solution, $\theta_2$ rotates with $\theta_1$ held fixed. The kink can tunnel to a configuration with the same asymptotic behavior and the same energy, in which $\theta_2$ and $\theta_1$ rotate together. This configuration has gradient energy, but no potential energy, so nothing prevents it from spreading, and relaxing to a configuration in which $\theta_1$ and $\theta_2$ are nearly constant.
If, after tunneling, the region in which $\theta_1$ twists is of length $L$, then the gradient energy is of order $\eta_1^2/L$. This configuration will be degenerate with the kink for $L\sim \eta_1^2/\sigma$.
To analyze the tunneling more precisely, and compute the tunneling action, we use the Euclidean path integral method. The instanton describing the kink decay is an unstable defect configuration in two Euclidean dimensions. The two–dimensional theory with the Lagrangian , has vortex solutions in which the phase $\theta_1$ changes by $2\pi$ around the vortex. The potential in is minimized by setting $\theta_2 = - \theta_1/2$, but then the change of $\theta_2$ around the vortex is only $\pi$, and thus the vortex should be attached to a linear defect (a “wall”). The cross–section of the “wall” is identical to the kink and its tension is given by . The instanton describing the kink decay consists of a vortex–antivortex pair held apart by two “walls.” Outside the “walls,” $\theta_1$ is well approximated by $\theta_1 = \phi_1 + \phi_2$, where $\phi_1$ and $\phi_2$ are azimuthal angles defined in Fig. 3. The corresponding pattern for $\theta_2$ near the “walls” and on the mid-section $\Sigma$ is sketched in Fig. 4. The action for this instanton is here the first term is twice the energy $\mu$ of a vortex, $\delta_s \sim \lambda^{-1/2}_1 \eta^{-1}_1$ is the size of the vortex core and $R$ is the vortex–antivortex separation. Since the $R$–dependence of $\mu$ is only logarithmic, the result is not sensitive to whether we choose $R$ or, say, $R/2$ as a cut–off in . The second term in is the energy of the missing “wall". Eq. is stationary for and Note that with $\eta_1 \gg \eta_2$, $R$ is much greater than the “wall" thickness, $\delta_w \sim m^{-1}$. This justifies the thin–wall approximation in .
[*ii) $D=2$*]{}
In two spatial dimensions, the model , gives rise to metastable linear defects. The instanton describing their decay consists of a planar “wall" with a circular hole bounded by a global string in three Euclidean dimensions. Assuming that $R \gg \delta_w$, the bounce action can be written as which is stationary for and We note that can be obtained from the thin–defect result \[see Eq. \] using the “renormalized" global string tension .
[*iii) $D=3$*]{}
Finally, we consider metastable global defects in three dimensions. The model , has metastable domain walls which decay by nucleation of circular loops of string. By the same argument as before, the corresponding tunneling action is given by the thin–defect equations , with $\sigma$ from and $\mu$ from ,
[**b) Codimension 2**]{}
As a simple example of a model that contains metastable global strings or vortices, consider a model with a spontaneously broken $SU(2)$ global symmetry, which has a scalar triplet $\vec \varphi_1$ interacting with a doublet $\varphi_2$, via the potential $$V(\vec\varphi_1,\; \varphi_2) = \lambda_1(\vec\varphi_1^2- \eta^2_1)^2 +
\lambda_2
(\varphi^\dagger_2 \varphi_2 - \eta^2_2)^2$$ For $\eta_1>>\eta_2$, the symmetry breaking is that of eq. with $G= SU(2)$.
[*i) $D=2$*]{}
In two spatial dimensions, the model contains a metastable vortex. In the vortex solution, $\vec\varphi_1$ is essentially a constant, which we may take to be and $\varphi_2$ has the asymptotic behavior (for $\lambda_{12}>0$).
When we consider the decay of this object, there is a subtlety, namely that the energy of an isolated vortex is divergent in an infinite volume. We should therefore imagine that the vortex is actually a member of a distantly separated vortex–antivortex pair (or that a suitable infrared cutoff has been imposed in some other way). The “mass” of the vortex is of order $2\pi
\eta_2^2\ln(R_{\rm
cutoff}/\delta_s)$, where $R_{\rm cutoff}$ is the infrared cutoff and $\delta_s$ is the size of the vortex core (of order $m_2^{-1}$, where $m_2$ is the mass of $\varphi_2$).
The vortex can tunnel to a configuration that has the same asymptotic behavior, but has negligible potential energy. In this configuration, $\vec\varphi_1$ rotates inside a region of radius $R$, and the scalar fields lie close to the vacuum manifold everywhere. The gradient energy is then of order $\eta_1^2+\eta_2^2\ln(R_{\rm
cutoff}/R)$ (assuming that $R_{\rm cutoff}>>R$), so this configuration is degenerate with the vortex for $\ln(R/\delta_s)\sim \eta_1^2/\eta_2^2$.
If a vortex and antivortex are held at fixed positions, the tunneling of one of the two is kinematically forbidden unless the distance between them is truly enormous. If, say, the antivortex tunnels, the fields will eventually relax to a configuration that has a vortex core, but is trivial at spatial infinity, a configuration with a gradient energy of order $\eta_1^2$.
We turn now to the computation of the tunneling action. In a three–dimensional space the model has solutions describing global monopoles attached to global strings, and the instanton describing the vortex decay consists of a monopole–antimonopole pair held apart by the string tension. The energy of this configuration diverges not only due to the infinite length of strings, but also because the energy per unit length of a global string is logarithmically divergent. However, the bounce action can still be expected to be finite, since the monopoles do not significantly affect the field of the string at distances much greater than the monopole separation, $R$. Assuming that $R$ is much larger than the thickness of the string core $\delta_s$, we can write the bounce action as The first term in is the monopole energy, and the second is the energy of the missing part of the string.
Eq. is stationary for and For $\eta_1 \gg \eta_2$ this action is exponentially large, and thus the decay of global vortices is very strongly suppressed. The value of the prefactor in front of the exponential in eq. should not be taken seriously; this prefactor is difficult to estimate, because it is sensitive to terms, subleading in $R$, that have been omitted from eq. . To determine the prefactor more accurately, one would have to solve the field equations for the instanton.
We see that if a vortex and antivortex are held at fixed positions, with separation $R_{\rm pair}$, the tunneling rate remains finite as $R_{\rm pair}$ approaches infinity (while the interaction energy of the pair diverges). There is of course a competing process, in which the vortex and antivortex [*annihilate*]{} due to tunneling. This is the dominant tunneling process for $R_{\rm pair}<< R$ (with $R$ as in eq. ), but is strongly suppressed for $R_{\rm pair}>>R$.
[*ii) $D=3$*]{}
Metastable global strings of the model decay by nucleation of monopole–antimonopole pairs. Again, the energy per unit length of a global string is infrared divergent, so we should consider a very long (but finite) string loop, or impose an infrared cutoff in some other suitable way.
A global monopole and antimonopole attract one another with a force $4\pi\eta_1^2$ that is independent of their separation. Thus, it is not so easy to pay back the energy cost of nucleating the pair by pulling the monopole and antimonopole apart. In fact, it is just the logarithmic infrared divergence in the tension of the global string that makes the tunneling possible—the energy saved by reducing the string length by $L$ is enhanced by a factor of $\ln L$ relative to the energy cost of producing a monopole pair with separation $L$.
The instanton in this case is a planar defect in four dimensions, with a circular hole bounded by a linear defect (representing the monopole world line). The tunneling action is which is stationary with respect to $R$ for With $\eta_1 \gg \eta_2$, $B$ is typically very large, and the strings are essentially stable. (Again, the estimate of the prefactor in front of the exponential in eq. should not be considered reliable.)
If a finite string loop of radius $R_{\rm loop}$ is held in a fixed position, then, as in our discussion of vortex decay, there is a competing process in which the loop annihilates instead of breaking. But the action for this process is proportional to $R_{\rm loop}^2$; it is subdominant for $R_{loop}>>R$ (with $R$ given by eq. ).
As noted in Section 2c, a defect associated with a spontaneously broken global symmetry may fail to “survive” when the symmetry is intrinsically broken by a small perturbation. When a defect of codimension $n+1$ does not survive, it becomes attached to one or more defects of codimension $n$. If the number of codimension $n$ defects attached to the codimension $n+1$ defect is exactly one, then the codimension $n$ defect is metastable, and can decay by nucleating a codimension $n+1$ defect. In this section, we illustrate this phenomenon with a few examples.
[**a) Codimension 1**]{}
Consider the case, in the notation of eq. , $G_{\rm
approx}=U(1)\rightarrow H_{\rm approx}=I$, $G_{\rm exact}=Z_N$. Suppose that, when we ignore the intrinsic symmetry breaking, the $U(1)$ symmetry is spontaneously broken by the condensation of a complex scalar field (where the phase $\theta$ is arbitrary). This model contains a global string, such that $\theta$ advances by $2\pi$ on a large circle that encloses the string core.
When we introduce the symmetry breaking perturbation, however, there are $N$ degenerate vacuum states, with Hence, the string does not “survive.” On a circle surrounding the string core, $\theta$ will choose to stay close to the vacuum manifold $\theta=2\pi
k/N$, except at isolated points where $\theta$ abruptly jumps from one vacuum value to the next. Thus, the string becomes attached to $N$ domain walls . (Conceivably, these walls will attract each other, so that in the configuration of minimal energy $\theta$ jumps by $2\pi$ all in one step. Then the string is attached to a single metastable wall.)
The walls have a thickness of order $m_a^{-1}$, where $m_a$ is the mass of the $U(1)$ pseudo–Goldstone boson (the “axion”). The string tension $\mu$ and the wall tension $\sigma$ are, in order of magnitude where $\delta_s$ is the thickness of the string core.
Note that this precise pattern of symmetry breaking occurs in models that solve the strong $CP$ problem by the Peccei–Quinn mechanism , where $U(1)$ is the Peccei–Quinn symmetry, and the intrinsic symmetry breaking is due to QCD.
In the discussion below, we assume that $N=1$, so that a single “axion domain wall” ends on the “axion string.”
[*i) $D=1$*]{}
In one spatial dimension, this model has a metastable axion “kink.” We can anticipate that this kink, with width of order $m_a^{-1}$, will decay by tunneling to an “unwound” configuration of about the same size.
As in Section 4a.i, the bounce solution is a vortex–antivortex pair in unstable equilibrium, with the attraction between the pair balanced by the pull of the kink world lines that are attached to the vortices. Unlike the discussion in Section 4a.i, though, the separation of the pair is comparable to the wall thickness. Hence, it is difficult to calculate the tunneling action accurately.
If we repeat our previous analysis (even though it is not well justified here), we obtain the crude estimates for the vortex separation, and for the bounce action.
[*ii) $D=2,3$*]{}
In two spatial dimensions, an axion wall decays by nucleating a pair of axion vortices, and in three spatial dimensions an axion wall decays by nucleating a loop of axion string. These decay processes may be analyzed just as in Section 3b, with $\mu$ and $\sigma$ given by eq. . However, to justify the thin-defect approximation used there, we must have
[**b) Codimension 2**]{}
Now consider the case $G_{\rm approx}=S0(3)\rightarrow H_{\rm approx}=U(1)$, with $G_{\rm
exact}=I$. If we ignore the intrinsic symmetry breaking, then this model contains a texture in two spatial dimensions, or a global monopole \[\] in three spatial dimensions. But when the symmetry breaking perturbation is introduced, the order parameter has a unique vacuum value. Hence, the texture collapses to a point singularity, and the long-range field of the monopole collapses to a singular line.
If we introduce a short distance cutoff, like a lattice spacing, then the texture wants to twist in a region with size of order the cutoff. But on that scale, there is really no notion of topology that stabilizes the texture, and there is no reason to expect a metastable defect.
To prevent the texture from shrinking indefinitely, let us introduce into the action of the model a higher-derivative “Skyrme term” . Then there will be a classically stable defect whose decay we can analyze semiclassically.
Roughly, the energy of a texture with radius $R$ is here, the first term is the Skyrme term (with the coupling constant $e_{\rm
sk}^2$ [*defined*]{} by eq. ), the second term is the conventional gradient term, and the third term is the potential energy due to the intrinsic symmetry breaking (where $m$ is mass of the pseudo–Goldstone boson, and $\alpha^2$ is a constant of order one). By minimizing with respect to $R$, we find the size of the texture Note that $R_{\rm texture}\to 0$ if we turn off the Skyrme term ($e_{\rm
sk}^2\to\infty$), and that $R_{\rm texture}\to\infty$ if we turn off the intrinsic symmetry breaking ($m\to 0$). The mass of the texture is the second term can be neglected as the Skyrme term or the intrinsic symmetry breaking turns off.
In eq. we have made the assumption that the order parameter is close to the approximate vacuum manifold $G_{\rm approx}/H_{\rm approx}$. This assumption is reasonable only if the energy density inside the texture is small compared to the energy density of the “false vacuum” in which the $G_{\rm
approx}$ symmetry is restored. The energy density of the false vacuum can be expressed as $\lambda \eta^4$, where $\lambda$ is a scalar self coupling; thus, eq. and for $\mu_{\rm texture}/R_{\rm
texture}^2 <<\lambda\eta^4$, or If this condition is not satisfied, there is no good reason to expect a metastable texture to exist.
If we ignore the perturbations, then the global monopole has a core size of order where $\lambda$ is a scalar self coupling; this is the linear size of the region in which the Higgs field departs significantly from its vacuum value. But the Skyrme term may distort the core significantly. If $\lambda/e_{\rm sk}^2>>1$, then the Skyrme term dominates the gradient energy inside the core, and we find instead The mass of the monopole core, in order of magnitude, is
The ratio of the texture size to the monopole core size is Comparing with the condition eq. , we find that $R_{\rm texture}>>R_{\rm
core}$. It is reasonable to expect a classically stable texture to exist, and to treat the intrinsic symmetry breaking as a small perturbation, only if the texture is large compared to the monopole core. A monopole attached to a “texture string” in three dimensions is illustrated in Fig. 5.
[*i) $D=2$*]{}
In two spatial dimensions, the texture decays by tunneling to an unwound configuration of about the same size. The bounce is a monopole–antimonopole pair in unstable equilibrium, with the attraction of the pair balanced by the pull of the strings that are attached to the monopoles. The typical separation of the pair is comparable to the size of the texture, which makes it difficult to calculate the tunneling action reliably.
Because $R_{\rm texture}>>R_{\rm core}$, the action of the bounce will be dominated by the interaction between the monopoles, rather than by the core action. In order of magnitude, we expect that
[*ii) $D=3$*]{}
In three spatial dimensions, the texture becomes a string, which can decay by nucleating a monopole pair. The bounce solution is a planar string world sheet, punctured by a hole of radius $R$ that is bounded by the world line of a monopole. However, $R$ is comparable to the thickness of the string, so we can not justify the approximation used in Section 3a.i, where we neglected the string thickness.
Very roughly, we can estimate the order of magnitude of the bounce action as with $\mu_{\rm string}$ and $R_{\rm string}$ given by eq. and .
We have seen that, in models with a hierarchy of symmetry breakdown, it is possible for a monopole (or string) that arises at a short distance scale to become the boundary of a string (or wall) that arises at a longer distance scale. In this section, we will comment on another logical possibility. A monopole might connect together two distinct types of string, or a string might connect together two distinct types of wall. Thus, by nucleating a monopole pair, a heavy string might decay to a light string. And by nucleating a string, a heavy wall might decay to a light wall. We will illustrate these possibilities by discussing some particular examples.
[**a) A String Bounding Two Walls**]{}
Consider the sequence of phase transitions where $U(1)$ is a gauge symmetry and $Z_2$ is a global symmetry. This pattern of symmetry breaking occurs in a model with three complex scalar fields $\phi$, $\psi$, and $\chi$, which carry $U(1)$ charges and transform under $Z_2$ as In this model, $\phi$ condenses at the scale $\eta_1$, breaking $U(1)$ to $Z_2'$ such that Then $\psi$ condenses at $\eta_2<<\eta_1$, breaking $Z_2'$. Finally $\chi$ condenses at $\eta_3<<\eta_2$, breaking $Z_2$.
The symmetry breaking at scale $\eta_1$ gives rise to a string that eventually becomes the boundary of both a heavy $\psi$ domain wall at scale $\eta_2$, and a light $\chi$ domain wall at scale $\eta_3$ (see Fig. 6). Since the global $Z_2$ symmetry is spontaneously broken, there is a stable domain wall in this model, the $\chi$ wall. There is also a stable string, which carries twice the $U(1)$ flux of the minimal string that bounds two walls.
[*i) $D=1$*]{}
In one spatial dimension, this model contains a heavy ($\psi$) kink and a light ($\chi$) kink. From the point of view of an effective field theory that describes physics well below the scale $\eta_1$, both kinks appear to carry conserved topological charges, and so should be stable. But in the underlying theory, there is just a single topological conservation law that does not forbid the decay of a $\psi$ kink to a $\chi$ kink. The decay of the heavy kink is mediated by the $U(1)$ vortex. As in the discussion in Section 3b.iii, the action of the bounce is where $\mu$ is the vortex action. (Integrating out the vortex generates an operator that destroys a $\psi$ kink and creates a $\chi$ kink, with a coefficient of order $e^{-\mu}$.)
[*ii) $D=2,3$*]{}
In two spatial dimensions, a heavy $\psi$ wall decays to a light $\chi$ wall by nucleating a pair of vortices, and in three dimensions a $\psi$ wall decays to a $\chi$ wall by nucleating a loop of string. These decay processes can be analyzed as in Section 3b, except that the wall tension $\sigma$ is replaced by $\sigma_\psi-\sigma_\chi$, the [*difference*]{} between the heavy and light tensions.
[**b) A Monopole Bounding Two Strings**]{}
Consider the sequence of phase transitions where $U(1)\times U(1)'$ is generated by the diagonal $SU(3)$ generators and $Z_2$ is generated by $e^{2\pi i Q}=e^{2\pi i Q'}$. This pattern can occur in a model with an $SU(3)$ octet that condenses at scale $\eta_1$, a triplet that condenses at $\eta_2<<\eta_1$, and another octet that condenses at $\eta_3<<\eta_2$.
When $SU(3)$ breaks to $U(1)\times U(1)'$, there are two conserved magnetic charges, and so there will be two distinct types of stable magnetic monopole. The magnetic charges of the stable monopoles are expected to be the minimal charges $(g_D/2,g_D'/2)$ and $(g_D/2,-g_D'/2)$, where $g_D$ and $g_D'$ are the Dirac magnetic charges associated with $U(1)$ and $U(1)'$ respectively. (These charges satisfy the Dirac quantization condition because $U(1)$ and $U(1)'$ have a nontrivial element in common.) The two monopoles need not be degenerate, unless there is a charge conjugation symmetry to enforce the degeneracy. Monopole solutions with charges $(g_D,0)$ and $(0,g_D')$ may also exist, but they are likely to be unstable, since they can decay to minimally charged monopoles, which have lower Coulomb energy.
When the symmetry breaking proceeds further, both monopoles eventually become attached to two strings—a heavy $U(1)'$ string at scale $\eta_2$ and a light $U(1)$ string at scale $\eta_3$. The light string is a stable $Z_2$ string. The model contains no stable magnetic monopole.
[i) $D=2$]{}
In two spatial dimensions, this model contains a heavy vortex that carries $U(1)'$ magnetic flux, and a light vortex that carries $U(1)$ magnetic flux. If we integrate out the monopoles, and ignore their exponentially small effects at low energy, then there are two independent conserved vortex numbers, each taking integer values.
The monopoles break these conservation laws and mediate the decay of a heavy vortex to a light vortex. The decay can be analyzed as in Section 3b.ii, and the bounce action is where $m$ is the monopole action. A heavy vortex can decay to either a light vortex or a light antivortex; which decay is favored depends on which of the two monopole species is lighter.
Since the only exactly conserved vortex quantum number is a $Z_2$ charge, a pair of light vortices (as opposed to a vortex–antivortex pair) must be able to annihilate. The annihilation process involves monopoles of both types, and has a cross section of order $e^{-2(m_1+ m_2)}$, where $m_{1,2}$ denotes the monopole action.
[ii) $D=3$]{}
In three spatial dimensions, this model contains a heavy string that can decay to a light string by nucleating a monopole–antimonopole pair. The decay can be analyzed as in Section 3b.i, but with $\mu$ replaced by $\mu_{\rm
heavy}-\mu_{\rm light}$, the [*difference*]{} between the heavy and light string tensions.
As we noted in Section 2d.i, global texture, even if classically stable, can always decay quantum mechanically. We discuss a few examples in this section.
[**a) $D=1$**]{}
To be concrete, we consider a model of a single complex scalar field $\varphi$ with Lagrangian This model has a spontaneously broken $U(1)$ global symmetry.
In one dimension, a global texture in this model wants to spread out, but we can stabilize it by imposing an infrared cutoff. Suppose we take space to be a circle with circumference $L$. Then a texture with topological charge $n$ has the form and has energy $(2\pi n)^2/L$. We may take the limit $L\to\infty$ with the number of twists per unit length $n/L$ held fixed. In this limit, the texture has a decay rate per unit time and length that can be computed semiclassically.
Obviously, the texture with $n$ twists will decay to a texture with $n-1$ twists, and we can anticipate that the tunneling will take place in a region with a size of order $l\equiv L/n$, the length of a single twist. The instanton that mediates the decay is the global vortex in two Euclidean dimensions; we construct the bounce solution, and compute the tunneling action, by finding a configuration with a pair of vortices in unstable equilibrium, with boundary conditions that fix the topological charge per unit length at $\tau=\pm\infty$.
This problem is actually identical to a problem in two–dimensional electrodynamics, since the (long–range) interaction between vortices is the same as the Coulomb interaction. The boundary conditions place the vortex pair in a constant background electric field. In the unstable solution, the electric force exerted on the vortex by the antivortex is precisely canceled by the background field.
If the separation between the vortices is $R$, then the action is The first term is the vortex–antivortex interaction term, and the second term is due to the interaction of the vortices with the background texture; here, $\delta_s\sim \lambda^{-1/2}\eta^{-1}$ is the size of the vortex core, and $l$, again, is the length of a single twist of the texture. Eq. is stationary for and Eq. is valid for $l>>\delta_s$.
(Note that we can do a related calculation for the case of “gauge texture.” On a circle of finite length, the number of twists of the Higgs field is a topological invariant with a gauge invariant meaning, and coincides with the Chern-Simons number of the gauge field. But since these configurations are pure gauge, the states with different winding numbers are degenerate classical vacuum states, analogous to the “n-vacua” of 3+1–dimensional Yang–Mills theory. The gauge vortex is the two–dimensional instanton that causes these classical ground states to mix \[\]. The mixing splits the degeneracy, giving rise to “$\theta$-vacuum” states.)
[**b) $D=2$**]{}
In two spatial dimensions, a global texture has an arbitrary size, and is marginally stable at the classical level. Hence, we will not discuss its semiclassical decay. The case where the texture is stabilized by suitable perturbations has already been discussed in Section 5b.
[**c) $D=3$**]{}
In a model in three spatial dimensions in which a global $G$ symmetry is broken to $H$, there will be global texture if $\pi_3(G/H)$ is nontrivial . A global texture wants to shrink, but it can be stabilized if a higher–derivative term (“Skyrme term”) is introduced into the action of the theory . As was recently pointed out by Hindmarsh , the Skyrme term can be generated by gauge boson exchange if an appropriate [*subgroup*]{} of $G$ is gauged.
Crudely speaking, the energy of a texture with radius $R$ is of order Here, the first term is the Skyrme term, and $e_{\rm sk}^2$ is the dimensionless Skyrme coupling constant. (In the sort of model considered by Hindmarsh , it is related to a gauge coupling.) The second term is the conventional (two-derivative) kinetic term, and $\eta$ is the expectation value of the order parameter. Minimizing with respect to $R$, we find the size and mass of the texture: In a model that admits global texture, there is always a “global instanton” that mediates the decay of the texture. In four (or more) Euclidean dimensions, this instanton is attached to a global line defect, which we may interpret as the world line of the texture. Inside the core of the instanton, the spontaneously broken global $G$ symmetry is “restored.”
If we suppose that the size $R_{\rm core}$ of the instanton core is small compared to $R_{\rm texture}$, then the action contained inside a sphere of radius $R_{\rm texture}$ centered on the core is, in order of magnitude, The first term is the Skyrme term, the second is the conventional kinetic term, and the third term is the potential energy of the core; $\lambda$ is a scalar self coupling that is [*defined*]{} by eq. . Inside a larger radius $R>>R_{\rm texture}$, the action is dominated by the linear defect, so we have $S_{\rm inst}\sim M_{\rm texture}R$.
Assuming that $\lambda/e_{\rm sk}^2>>1$, we find that $S_{\rm inst}$ is minimized for so that the assumption $R_{\rm core}<<R_{\rm texture}$ is justified, and the instanton action (cut off at $R\sim R_{\rm texture}$) is
When it decays, the texture tunnels to an “unwound” configuration with core size of order $R_{\rm texture}$, which is then free to dissipate. The bounce solution that describes this decay is a pair of global instantons in unstable equilibrium, with the pull of the texture balanced by the instanton–anti-instanton attraction. The separation between the instantons in equilibrium is of order $R_{\rm texture}$, and, for $\lambda/e_{\rm sk}^2>>1$, the action of the bounce is The decay is strongly suppressed in the limit $e_{\rm sk}^2\to 0$, where the texture becomes large. It is also suppressed, much more weakly, as the “barrier height” $\lambda \eta^4$ gets large.
Like the strings that we discussed in Sections 3 and 4, semilocal and electroweak strings can end on magnetic monopoles, not because of a symmetry-breaking hierarchy, but for other reasons. Here we will estimate the tunneling action for the decay of semilocal and electroweak vortices, in two dimensions, and for the breaking of semilocal and electroweak strings, in three dimensions. The analysis is similar in spirit to that described in Sections 3 and 4, but differs in detail. Actually, our estimates will be very crude; to do a better job, one would need to study the interactions of the monopoles in more detail.
[**a) Semilocal Defects**]{}
Recall from the discussion in Section 2d.ii that a semilocal model has a “topologically conserved” magnetic flux (in two dimensions), yet there are configurations of finite energy in which the flux is spread out over an arbitrarily large area. In these configurations, the Higgs field takes values in the vacuum manifold everywhere, and there is no Higgs field potential energy. If the size $R$ of the configuration is very large, the Coulomb energy of the magnetic flux can also be neglected; then the only contribution to the energy is due to Higgs fields gradients. In two dimensions, gradient energy is scale invariant, and so remains finite and nonzero as $R\to\infty$. In the sector with a single quantum of magnetic flux, let us denote the minimum energy in the limit $R\to\infty$ by where $\eta$ is the magnitude of the Higgs field expectation value. Here $\alpha_\infty$ is a numerical factor of order one; it depends on the geometry of the vacuum manifold, but not on any coupling constants or parameters of the theory.
There are also “vortex” configurations, in which the magnetic flux is confined to a core of finite size. The characteristic feature of the vortex is that the stability group of the Higgs field is different at its center than in the vacuum; thus, the vortex carries Higgs field potential energy. Let us suppose that there is a vortex solution to the classical field equations with energy The structure of the core depends on the detailed dynamics of the theory, so $\alpha_s$ has a nontrivial dependence on coupling constants.
Now, if $\alpha_s>\alpha_\infty$, then the vortex is not stable. But it may or may not be metastable. In fact, in the one model that has been studied in detail (the minimal electroweak model in the limit $\sin^2\theta_W=1$), it turns out that the vortex is either absolutely stable or classically unstable . Nevertheless, we will ask what would happen if there is a classically stable vortex with $\alpha_s>\alpha_\infty$.
To understand the decay of semilocal strings and vortices, it is important to recognize that a semilocal string can end on a (global) monopole . In the notation of Section 2d.ii, the defining property of a semilocal model is that a closed path that is noncontractible in the gauge orbit $G_1/H_1$ [*can*]{} be contracted in the full vacuum manifold $[G_1\times G_2]/H$. On a large sphere that surrounds the monopole, a quantum of $G_1/H_1$ magnetic flux enters through the core of a vortex at, say, the south pole. The Higgs field configuration on the sphere excluding the south pole is just a deformation of the nontrivial loop in $G_1/H_1$ to a point; on each line of constant latitude, the Higgs field executes a closed path in the vacuum manifold, which becomes a trivial path at the north pole. Thus, a quantum of confined magnetic flux is converted in the core of the monopole to unconfined flux that spreads and returns to spatial infinity. The energy of the monopole is infrared divergent, for the Higgs field gradient energy inside a sphere of radius $R$ is of order $\alpha_\infty\eta^2 R$ (excluding the energy of the vortex).
To be concrete, let us suppose that the gauge group is $G_1=U(1)$, and is completely broken. We denote the gauge coupling by $g'$, in deference to the analogy with the hypercharge coupling in the standard model.
[*i) $D=2$*]{}
A metastable semilocal vortex will tunnel to a configuration with the same energy that does not have significant Higgs field potential energy stored in its core. If this configuration has radius $R$, we can estimate its energy as $\alpha_\infty\eta^2+4\pi/g'^2 R^2$, where the first term is due to Higgs field gradients and the second is due to the magnetic flux. Equating with the vortex mass $\alpha_s\eta^2$, we find $R\sim
\sqrt{4\pi}(\alpha_s-\alpha_\infty)^{-1/2}(g'\eta)^{-1}$.
To compute the tunneling action, we construct the bounce. It consists of a monopole–antimonopole pair in unstable equilibrium in three–dimensional Euclidean space (see Fig. 7). If we assume that $\alpha_s-\alpha_\infty<<1$, then the pair will be widely separated, and the interaction “energy” can be well approximated by a linear plus Coulomb potential. Thus, if $R$ is the separation, the action is the first term is the core action of the monopoles, the second is the action of the missing string, the third is the linear interaction of the monopoles, and the fourth is the Coulomb term. (We have normalized the gauge coupling so that $4\pi/g'$ is the magnetic flux quantum.) This expression is stationary for as we anticipated. For $\alpha_s-\alpha_\infty$ small, the bounce action is dominated by the core action of the monopoles, We expect $m_{\rm core}\sim 4\pi\eta/g'$, for this is the magnetic self–energy of a monopole with core size of order the vortex width $\delta_s\sim
(g'\eta)^{-1}$.
[*ii) $D=3$*]{}
A metastable semilocal string will decay by nucleating a monopole–antimonopole pair. If $\alpha_s-\alpha_\infty$ is small, then the pair will be sufficiently distantly separated right after the tunneling that the Coulomb interaction between the monopoles can be neglected. If the pair nucleates with separation $L$, then the energy $\alpha_s\eta^2 L$ saved by removing the string must balance the energy $2m_{\rm core}+\alpha_\infty\eta^2 L$ of the pair (where $m_{\rm core}$ is the mass of the monopole core). We conclude that $L\approx 2m_{\rm
core}/(\alpha_s-\alpha_\infty)\eta^2$.
As in our previous calculations of string decay due to monopole pair nucleation, the bounce solution is a planar string world sheet with a circular hole of radius $R$, the hole bounded by the world line of the monopole. If the Coulomb interaction is neglected, we can calculate $R$ and the tunneling action using a minor modification of the method in Section 2a.i. The modification is that the string tension $\mu$ is replaced by $(\alpha_s-\alpha_\infty)\eta^2$, the [*difference*]{} between the string tension and the coefficient in the monopole linear potential, and $m$ is replaced by the core mass $m_{\rm core}$. Then, from eq. and , we find and
[**b) Electroweak Defects**]{}
Now consider the case of a semilocal model with $\alpha_s<\alpha_\infty$, so that a vortex is [*stable*]{}. Let us ask what would happen if we were to gauge the global $G_2$ symmetry. To be concrete, consider an interesting example—the standard electroweak model with gauge group $SU(2)_L\times U(1)_Y$ and a Higgs doublet. If we turn off the $SU(2)_L$ gauge coupling $g$, this becomes a semilocal model; the gauge group $U(1)_Y$ is broken, but a noncontractible loop in the gauge orbit is contractible in the full vacuum manifold. The vortex turns out to be stable, in this limit, if the Higgs mass is less than the $Z^0$ mass .
For $g\ne 0$, there is no longer a topological conservation law, and the vortex is no longer absolutely stable. Only a finite energy barrier separates a vortex with heavy $Z^0$ magnetic flux from a configuration with massless $A$ magnetic flux that is free to spread out. Correspondingly (as Nambu \[\] observed long ago), a $Z^0$ string can end on an [*electromagnetic*]{} monopole (see Fig. 8). The quantity of $Z^0$ flux trapped in the string is and the quantity of $A$ flux emanating from the monopole is (where $\tan\theta\equiv g'/g$).
In the limit $g<<g'$ (or $\sin\theta\approx 1$), the monopole has a core size that is large compared to the thickness of the string $\delta_s\sim
m_Z^{-1}$ (where $m_Z$ is the $Z^0$ mass). Deep inside the core, it resembles the semilocal monopole, with spreading $Z^0$ flux. At a radius $R_{\rm core}$, the $Z^0$ flux is converted to $A$ flux. The core radius is determined by the competition between the linearly divergent energy of the “global” monopole and the magnetic self energy. Roughly, the core energy is where the second term is due to the $Z^0$ flux and the third term is due to the $A$ flux. By minimizing with respect to $R_{\rm core}$, we find (the second equality following from our assumption $g<< g'$), and .
[*i) $D=2$*]{}
In two dimensions, an electroweak vortex decays by tunneling to a configuration in which its $Z^0$ flux has been converted to $A$ flux. In the limit $g<<g'$ the flux $\Phi_A$ given by eq. is much larger than $\Phi_Z$. Therefore, $A$ flux is energetically very costly; to be degenerate with the vortex, the configuration after tunneling must be very large.
The bounce solution is a monopole–antimonopole pair in unstable equilibrium in three Euclidean dimensions, with the Coulomb attraction of the pair compensated by the tension in the electroweak strings. For typical values of the parameters, the separation $R$ of the pair will be comparable to the size $R_{\rm core}$ of the monopole, which makes it difficult to estimate the tunneling action reliably. But if we assume that $R>>R_{\rm core}$, then the action of the configuration is which is stationary for (assuming $\sin^2\theta \approx 1$.) Thus, our assumption $R>>R_{\rm core}$ is justified under the (not very physical) condition The tunneling action is it is dominated by the core action for $\alpha_s<<\alpha_\infty$.
[*ii) $D=3$*]{}
An electroweak string decays by nucleating a monopole–antimonopole pair. Typically, the string will tunnel to a configuration in which the separation of the pair is comparable to the monopole core size, but if we assume that $\alpha_s<<\alpha_\infty$, then the core is sufficiently small that the analysis of Section 3a.i applies, with $m=E_{\rm core}$ and $\mu=\alpha_s\eta^2$. Thus, the monopole world line has radius (for $\sin^2\theta\sim 1$), and the bounce action is So, of course, the decay is heavily suppressed as $g^2\to 0$.
[*i) Summary*]{}
In this paper, we have studied the decay of metastable defects arising from symmetry breaking in relativistic field theories. The decay occurs through quantum tunneling; for example, strings decay by nucleation of monopole-antimonopole pairs, and domain walls decay by nucleation of circular holes bounded by strings. We also studied the decay of defects in one and two-dimensional systems and the decay of non-topological defects, such as global texture and semilocal and electroweak strings.
The decay probability is determined by an instanton which can be found by solving Euclideanized field equations with appropriate boundary conditions. The problem is greatly simplified when the dimensions of the instanton are much greater than the size of the defect core, so that the thin defect approximation can be used. We assumed the validity of this approximation in most of our calculations.
We have estimated the instanton action for the decay of various defects in $D$=1, 2, or 3 spatial dimensions. We found that the decay rate of defects arising from a global symmetry breaking is strongly suppressed compared to the decay rate in the corresponding gauge theory. In particular, the instanton action for the decay of a global string is exponentially large (see eq. ). But even in gauge theories, the tunneling action is large and the decay rate is correspondingly small in the case where two different symmetry breaking scales are widely separated.
In conclusion, we would like to comment on some possible applications and extensions of our results.
[*ii) Thin Defect Approximation*]{}
The thin defect approximation is similar to the thin wall approximation in the vacuum decay problem . This latter approximation is in bad repute, since it is known to apply only to a very limited class of potentials \[\]. On the other hand, the thin-defect approximation applies if the symmetry breaking scales for the two types of defects involved in the decay are substantially different. This is a typical situation in elementary particle theories, and thus we expect our results to have a reasonably wide range of validity.
[*iii) “Embedded” Defects*]{}
Let us briefly consider the behavior of the decay rate when the two symmetry breaking scales are close together. Though the thin–defect approximation does not apply in this case, we can roughly estimate the tunneling action by using our thin–defect formulas. For example, consider the breaking of a string due to monopole nucleation. The monopole mass can be crudely estimated as where $\eta_1$ is the higher symmetry breaking scale and $e$ is the gauge coupling, and the string tension is roughly where $\eta_2$ is the lower symmetry breaking scale. Then the tunneling action eq. becomes Naturally, the behavior of this expression as $\eta_1$ approaches $\eta_2$ (aside from the numerical factor $8\pi^2$, which should not be taken too seriously anyway), could be determined by dimensional analysis, as $e^{-2}$ has the dimensions of $\hbar$.
There is no guarantee that a [*classically stable*]{} string solution will continue to exist as the two symmetry breaking scales approach each other. However, if the Higgs potential obeys suitable conditions, one can argue that there [*is*]{} a static string solution for $\eta_1=\eta_2$, although it may be unstable. This is a special case of the “embedded” string recently described by Vachaspati and Barriola \[\]. Eq. suggests that, if an embedded string is classically stable, its quantum–mechanical decay rate is likely to be small, in a weakly coupled theory. Of course, it is a fact of life that if our semiclassical (small $\hbar$) approximation is reliable, then the tunneling action is large and the decay rate is small.
We can do a similar estimate of the rate for the decay of a domain wall due to nucleation of a string loop. Taking the string tension to be and the wall tension to be where $\lambda$ is a scalar self coupling, eq. becomes If there is a classically stable “embedded” domain wall for $\eta_1=\eta_2$, it will be long–lived at weak coupling.
[*iv) Cosmological Applications*]{}
The very small decay rates for metastable defects do not necessarily mean that such decay processes are observationally irrelevant. Consider for example metastable strings formed in the sequence of cosmological phase transitions . The first phase transition gives rise to monopoles and the second to strings which connect monopole–antimonopole pairs. Even if we disregard the breaking of the string due to the quantum–mechanical decay process, such hybrid defects typically disappear long before the present epoch . The string energy is dissipated by friction and by radiation of gauge quanta and of gravitational waves. As the strings get shorter, the monopoles are pulled closer together; they eventually capture one another into Coulombic bound states and annihilate. In this scenario, the demise of the monopole–string system is so rapid that the slow monopole nucleation process has very little effect.
But the evolution can be quite different if there is a period of inflation between the two phase transitions in . In this case the monopoles can be diluted beyond the present horizon, and the evolution of strings will initially be similar to that of topologically stable strings. However, at some point the string decay by monopole pair nucleation will become important. To estimate the time $t_*$ when this happens, we write the nucleation probability per unit string length per unit time as where $A$ is a dimensionless coefficient and we have used for the tunneling action. The strings stretching across the horizon at cosmic time $t$ have length of order $t$; so the condition for approximately one pair per horizon volume to nucleate on a given string by cosmic time $t_*$ is $ P{t_*}^2
\sim 1$, or With a grand unification symmetry breaking scale for strings, $\mu^{1/2} \sim
10^{16}$ GeV, and assuming that $A$ is not very different from 1, this time is smaller than the present age of the universe if $m^2 /\mu \leq 85$.
At $t \sim t_*$ the strings are cut into pieces of length $\sim t_*$ with monopoles at the ends. For large values of $t_*$ it may take a long time to dissipate the string energy, and oscillating string pieces may still be flying somewhere in the universe. As they are pulled by the strings, the monopoles are accelerated to energies comparable to the energy of the string, $ E
\sim \mu t_*$. If $t_*$ is close to the present time, then, for grand–unification strings this energy corresponds to a mass of order $10^{17}$ solar masses. A monopole will move ultrarelativistically, with its gravitational field concentrated in the transverse plane and the gravitational force decreasing as $r^{-1}$ with the distance from the monopole. Gravitational effects of such supermassive relativistic objects can be quite significant.
[*v) Nonzero Temperatures*]{}
Our results can be easily extended to the case of defect decay at a nonzero temperature, T. The Euclidean path integral in this case is taken over field configurations periodic in imaginary time $\tau$ with a period $\Delta \tau =
T^{-1}$, and the instanton solutions should have the same periodicity. At sufficiently low temperatures, the finite–$T$ instanton is simply the zero–$T$ instanton periodically repeated along the $\tau$–axis . The effects of periodicity become important when $T$ becomes comparable to the inverse size of the instanton, $R^{-1}$. At still higher temperatures, the decay of the defect tends to be dominated by thermal fluctuations rather than quantum fluctuations. The interesting and somewhat unexpected behavior of instantons at higher temperatures will be discussed in a separate paper \[\].
[*vi) Condensed Matter Applications*]{}
Condensed matter systems exhibit a fascinating variety of defects, many of which are metastable, at least in principle. For example, superfluid ${}^3$He-$B$ contains domain walls that can terminate on strings \[\], and both nematic liquid crystals \[\] and ${}^3$He-$A$ \[\] contain strings that can terminate on monopoles. In suitable materials, the decay of walls due to string nucleation and of strings due to monopole nucleation may occur at observable rates. Also, liquid crystals and ${}^3$He can contain textures that may decay either quantum mechanically or due to thermal fluctuations.
In these cases, the simplifying assumptions of Lorentz invariance and the thin–defect approximation may not apply. But the instantons corresponding to the decay processes have the same general structure as the instantons that we have constructed, and the decay rates can be estimated using the methods described here.
[**Acknowledgements**]{}
This work was supported in part by DOE contract DE-AC03-81-ER40050 and by NSF contract 8605578. A. V. gratefully acknowledges the hospitality of Caltech, where he visited as a Sherman Fairchild Distinguished Scholar.
[^1]: On leave from Tufts University, Medford, MA 02155
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
M. Georganopoulos , D. Kazanas ,\
E. S. Perlman
- 'F. W. Stecker'
title: A Probe of the Matter Content of Quasar Jets
---
Introduction
============
The composition of extragalactic jets continues to remain elusive. A number of attempts (e.g. Reynolds et al. 1996, Wardle et al. 1998) have been made over the years toward measuring, or at the least constraining, the matter content of jets, and in particular the fraction of kinetic energy stored in protons and low energy or cold leptons, whose low radiative efficiencies fail to provide direct evidence of their presence.
A direct estimate of the cold lepton content of blazar jets was proposed by Sikora & Madejski (2000): The observed non-thermal blazar emission is thought to be produced at distances $\sim 10^{17}-10^{18}$ cm from the central engine; the jet leptons providing the blazar emission at these distances need to be transported practically cold by a relativistic flow of bulk Lorentz factor $\Gamma \sim 10$ from the black hole vicinity to the blazar emission site; as these cold jet leptons propagate through the blazar broad line region (BLR) they would Compton – scatter the BLR optical-UV photons to energies $ \sim 1$ keV, to produce a black – body type hump in their X-ray spectra. The fact that such a feature is not observed in the inverse-Compton dominated X-ray spectrum of blazars, led the above authors to conclude that the jet power is carried mainly by protons, although cold leptons dominate the number of particles in the jet.
While this idea is well founded and appealing, concrete answers are hindered by unknowns such as the distance at which the jet is formed, its sub-pc scale opening angle and the actual photon energy density of the BLR, as well as by the presence of a strong X-ray non-thermal continuum that apparently could “hide" the proposed bulk-Comptonized component.
Using the CMB as a seed photon source
=====================================
Arguments based on the BC emission can be applied to any astrophysical site involving relativistic flows. One can then obtain more concrete conclusions provided that the flow geometry and the target photon density are better determined. Such a site is presented by the large scale jets of [*Chandra*]{} – detected superluminal quasars such as PKS 0637–752 (Schwartz et al. 2000; Chartas et al. 2001). The jet of PKS 0637–752, which is resolved and is found to be well collimated, propagates through a well understood photon field: the CMB. The source exhibits radio, optical, and X-ray emission from the quasar core and then from well separated knots along the jet at angular distances $\sim 8'' $. The fact that the bridge, the region between the core and the first knot WK7.8, radiates weakly in radio, optical, and X-ray energies is very important because: ([*i*]{}) it shows that most of the leptons propagating in the bridge are cold and ([*ii*]{}) it provides a region free from unwanted contamination by unrelated broad band non-thermal radiation.
The luminosity $L_{BC}$ of the BC emission depends on the power $L_e$ carried by cold leptons in the bridge, the length $l$ of the bridge, the bulk Lorentz factor $\Gamma$ of the flow, and the angle $\theta$ formed between the jet axis and the line of sight. It can be shown (Georganopoulos et al. 2004) that $$L_{BC}\approx 1.4 \;10^{-4} \, l_{100 Kpc} \Gamma_{10}^3 (1+z)^4 L_e,$$ where $z$ is the redshift of the source and we have assumed the typical for superluminal sources $\theta=1/\Gamma$. The BC component peaks in the IR regime $$\nu_{BC}\approx 4 \; 10^{13} \; \Gamma^2_{10} \; {\rm Hz},$$ regardless of $z$. The BC emission, therefore, requires an estimate of the jet power and kinematics. These can be provided by the spectrum and luminosity of the [*Chandra*]{} – detected knots, once the knot X-ray emission mechanism has been established.
The X-ray emission mechanism
----------------------------
Schwartz et al. (2000) noted that the X-ray emission from knots at a projected distance of $\sim 100$ kpc from the core of PKS 0637 – 752 is part of a spectral component separate from the synchrotron radio-optical emission and it is too bright to be explained through synchrotron self Compton SSC emission from electrons in energy equipartition with the jet magnetic field. Tavecchio et al. (2000) and Celotti et al. (2001) argued that the X-ray emission is due to external Compton (EC) scattering of CMB photons off relativistic electrons in the jet, provided that the jet flow is sufficiently relativistic ($\Gamma \sim 10$) to boost the CMB energy density in the flow frame (by $\Gamma^2$) to the level needed to reproduce the observed X-ray flux. This was the first suggestion that powerful jets retain significantly relativistic velocities at large distances from the core, a very important feature because it boosts the level of the anticipated BC emission by $\sim \Gamma^2$. We adopt here this interpretation of the X-ray emission; the reader can find a discussion of the alternatives in Georganopoulos et al. (2004).
Minimum power conditions
------------------------
A set of constraints for the jet power and beaming based on multiwavelength observations of knots has been presented by Dermer and Atoyan (2004, hereafter DA). These authors model the knots as homogeneous sources moving with a Lorentz factor $\Gamma$ at an angle $\theta$ to the line of sight. The knot matter content is described through the ratio of power carried by protons to power carried by a power law lepton distribution. Assuming that the X-rays are due to EC scattering off the CMB and that $\delta=\Gamma$, DA calculate (their Eq. (12)) the Doppler factor $\delta_{min}$ that minimizes the power needed in the knot to produce the X-ray flux. Because the minimized quantity is the total knot power and [*not*]{} the power in relativistic electrons and magnetic field only, [*$\delta_{min}$ depends on the matter content of the jet, and it is higher for $e-p$ jets*]{} (Georganopoulos et al. 2004).
The BC emission of PKS 0637-752
-------------------------------
Using the formalism of DA we derive minimum power Doppler factors $\delta_{min}=17.4$ for a $e^\pm$ composition and $\delta_{min}=27.8$ for an $e-p$ composition of knot WK7.8. These correspond to a jet minimum power $L_{min}=9.7 \times 10^{45} $ erg s$^{-1}$ for a $e^\pm$ jet and $L_{min}=6.3 \times 10^{46} $ erg s$^{-1}$ for an $e-p$ jet. The corresponding lepton power is $L_{lept}=3.7 \times 10^{45} $ erg s$^{-1}$ for a $e^\pm$ jet and $L_{lept}=6.8 \times 10^{44} $ erg s$^{-1}$ for an $e-p$ jet; these numbers are only weakly affected by our choice of the low energy cutoff $\gamma_{min}$ of the electron power law as long as $\gamma_{min} > 10$, a condition imposed by the requirement that EC does not overproduce the observed optical flux. Assuming $\theta=1/\Gamma$, at $z=0.651$ $1''$ corresponds to $6.9$ Kpc and the actual bridge length is $l \approx $ 930 Kpc for a $e^\pm$ jet and $l \approx 1.5$ Mpc for an $e-p$ jet. We calculate the BC flux for an $e-p$ and an $e^{\pm}$ composition for two cases.
[*Case A:*]{} The lepton power $L_{lept}$ required in the knot is provided by the cold leptons in the beam ($L_{e}=L_{lept}$). This requires that only a minority of the leptons get accelerated in the knot, and it is an optimistic estimate of the anticipated BC emission.
[*Case B:*]{} The most conservative case for the anticipated BC emission, according to which the jet provides simply the number of leptons needed in the knot, and the leptons are accelerated in the knot using energy exclusively from other agents such as the magnetic field and/or the jet hadrons.
As can be seen in Figure \[eikona1\], in case A the emission for a $e^\pm$ jet peaks at mid IR energies, while that for an $e-p$ jet peaks at near IR - optical energies. For both compositions the anticipated mid IR flux is above the [*Spitzer*]{} sensitivity limits; the $e-p$ case however violates the [*HST*]{} 3$\sigma$ detection limits for both a $0.5''$ and $0.1''$ thin jet. In case B the BC emission is still above the [*Spitzer*]{} sensitivity limit for the two shorter wavelength bands. However, the existing [*HST*]{} optical limits cannot be used to argue against an $e-p$ jet in this case. At $\lambda=3.6-8.0 \; \mu m$, the most likely band for the BC scattered emission to appear, the [*Spitzer*]{} angular resolution ($\sim 1''-3''$) is considerably smaller than the $\sim 8''$ bridge, and we anticipate that [*Spitzer*]{} will resolve the BC emission along the bridge.
Discussion
==========
Existing [*HST*]{} limits for PKS 0637-752 already disfavor case A $e-p$ models. Additional constraints for pure $e-p$ jets come from the large Lorentz factors required. Although values of $\delta_{min}\sim 30$ are compatible with the superluminal motions observed in some blazars (e.g. Jorstad et al. 2002), the number of such highly relativistic sources should not overproduce the parent (misaligned) population (e.g. Lister 2003). Additionally, the large Doppler factors required for pure $e-p$ jets, suggest that the actual jets are over $1$ Mpc long, a value reached by only the largest radio galaxies.
An assumption made in our calculations is that the Doppler factor of the jet flow in the bridge between the core and the knot is the same as the Doppler factor of the knot. This can happen if the flow does not decelerate substantially at the knot. This seems to be the case in PKS 0637-752, where VLBI observations of superluminal velocities with $v_{app}=17.8\pm 1 \,c $ in the core of the source (Lovell et al. 2000) set limits of $\Gamma>17.8$, $\theta<6^{\circ}.4$, in agreement with the Doppler factor $\delta=17.4$ derived from minimizing the jet power in an $e^{\pm}$ jet.
As was first discussed by Schwartz (2002) the X-ray emission due to EC off the CMB will remain visible at the same flux level independently of redshift. This is also the case for the BC emission. This suggests an exciting possibility for jets that have a very low radiative efficiency past the core: their IR-optical BC emission will be detectable independent of redshift, and it will be the only observable signature of these otherwise invisible jets.
Celotti, A., Ghisellini, G. & Chiaberge M. 2001, , 321, L1
Chartas, G. et al. 2001, ApJ, 542, 655
Dermer, C D., Atoyan, A., ApJ, 611, L9
Georganopoulos, M., Kazanas, D., Perlman, E. S., Stecker, F. W. 2004, ApJ, submitted
Jorstad, S. G., et al. 2002, ApJ, 556, 738
Lister, M. L. 2003, ApJ, 599, L105
Lovell, J. E. J., et al. 2000, in Astrophysical Phenomena Revealed by Space VLBI, ed. H. Hirabayashi, P. G. Edwards, & D. W. Murphy (Sagamihara: ISAS), 215
Reynolds, C. S., Fabian, A. C., Celotti, A. & Rees, M. J. 1996, MNRAS, 283, 873
Schwartz, D. E. et al. 2000, ApJ, 540, L69
Schwartz, D. E. 2002, , 569, L23
Sikora, M. & Madejski, G. 2000, ApJ, 534, 109
Tavecchio, F., Maraschi, L., Sambruna, R. & Urry, C. M., 2000, ApJ, 544, L23
Wardle, J. F. C., Homan, D. C., Ojha, R., Roberts, D. H. 1998, Nature, 395, 457
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study projective surfaces $X \subset \mathbb{P}^r$ (with $r \geq 5$) of maximal sectional regularity and degree $d > r$, hence surfaces for which the Castelnuovo-Mumford regularity ${\operatorname{reg}}(\mathcal{C})$ of a general hyperplane section curve $\mathcal{C} = X \cap \mathbb{P}^{r-1}$ takes the maximally possible value $d-r+3$. We use the classification of varieties of maximal sectional regularity of [@BLPS1] to see that these surfaces are either particular divisors on a smooth rational $3$-fold scroll $S(1,1,1)\subset \mathbb{P}^5$, or else admit a plane $\mathbb{F} = \mathbb{P}^2 \subset \mathbb{P}^r$ such that $X \cap \mathbb{F} \subset \mathbb{F}$ is a pure curve of degree $d-r+3$. We show that our surfaces are either cones over curves of maximal regularity, or almost non-singular projections of smooth rational surface scrolls. We use this to show that the Castelnuovo-Mumford regularity of such a surface $X$ satisfies the equality ${\operatorname{reg}}(X) = d-r+3$ and we compute or estimate various of the cohomological invariants as well as the Betti numbers of such surfaces. We also study the geometry of extremal secant lines of our surfaces $X$, more precisely the closure $\Sigma(X)$ of the set of all proper extremal secant lines to $X$ in the Grassmannian $\mathbb{G}(1, \mathbb{P}^r).$'
address:
- 'Universität Zürich, Institut für Mathematik, Winterthurerstrasse 190, CH – Zürich, Switzerland'
- 'Pukyong National University, Department of applied Mathematics, Daeyeon Campus 45, Yongso-ro, Nam-Gu, Busan, Republic of Korea'
- 'Korea University, Department of Mathematics, Anam-dong, Seongbuk-gu, Seoul 136-701, Republic of Korea'
- 'Martin-Luther-Universität Halle-Wittenberg, Institut für Informatik, Von-Seckendorff-Platz 1, D – 06120 Halle (Saale), Germany'
author:
- 'Markus BRODMANN, Wanseok LEE, Euisung PARK, Peter SCHENZEL'
date: 'Busan, Halle, Seoul and Zürich, '
title: ON SURFACES OF MAXIMAL SECTIONAL REGULARITY
---
Introduction
============
Varieties of maximal sectional regularity {#varieties-of-maximal-sectional-regularity .unnumbered}
-----------------------------------------
In [@BLPS1] we have studied and classified projective varieties $X \subset \mathbb{P}^r$ of dimension $n \geq 2$, of codimension $c \geq 3$ and of degree $d \geq c+3$ which are of maximal sectional regularity, which means that the Castelnuovo-Mumford regularity ${\operatorname{reg}}(\mathcal{C})$ of a general linear curve section $\mathcal{C} = X \cap \mathbb{P}^{c+1} \subset \mathbb{P}^r$ of $X$ takes the maximally possible value $d-c+1$. There are two possible types of such varieties, namely (see also Theorem \[theorem 2.1\] below):\
Either it holds $c=3$ and $X \approx H + (d-3)F$ is a divisor on a rational $(n+1)$-fold scroll $W \subset \mathbb{P}^{n+3}$ with $n-3$-dimensional vertex, where $H \subset W$ is the hyperplane divisor and $F = \mathbb{P}^n \subset W$ is a linear $n$-space;\
or else, there is linear subspace $\mathbb{F} = \mathbb{F}(X) = \mathbb{P}^n \subset \mathbb{P}^r$ such that $X \cap \mathbb{F} \subset \mathbb{F}$ is a hypersurface of degree $d-c+1$.\
If $X$ is of type II, the $n$-space $\mathbb{F}(X)$ is unique and coincides with the so-called extremal variety of $X$, that is the closed union of all lines in $\mathbb{P}^r$ which are $(d-c+1)$-secant to a general curve section of $X$. Moreover, if the (algebraically closed) base field $\Bbbk$ is of characteristic $0$ or if $n = 2$, each variety $X \subset \mathbb{P}^r$ of maximal sectional regularity is sectionally smooth rational (in the sense of Section 2 below) and hence is an almost non-singular linear projection of a rational $n$-fold scroll $\widetilde{X} \subset \mathbb{P}^{d+n-1}$ (see Theorem \[theorem 2.5\]). In addition, this projecting scroll $\widetilde{X}$ is singular if and only if $X$ is a cone.
Surfaces of maximal sectional regularity {#surfaces-of-maximal-sectional-regularity .unnumbered}
----------------------------------------
In this paper we focus on the case in which $n=2$, hence the case in which $X$ is a surface of maximal sectional regularity, and we shall investigate in detail the structure of $X$. In this case, the above two possible types present themselves as follows (see Corollary \[corollary 2.1’\] below):\
*Type I:* It holds $r = 5$ and $X \approx H + (d-3)F$ is a smooth divisor on the smooth rational $3$-fold scroll $W = S(1,1,1) \subset \mathbb{P}^5$, where $H \subset W$ is the hyperplane divisor and $F = \mathbb{P}^2 \subset W$ is a ruling plane.\
*Type II:* There is a plane $\mathbb{F} = \mathbb{P}^2 \subset \mathbb{P}^r$ such that $X \cap \mathbb{F} \subset \mathbb{F}$ is a pure curve of degree $d-r+3$.\
Moreover in this situation either the projecting surface scroll $\widetilde{X} \subset \mathbb{P}^{d+1}$ is smooth or $X$ is a cone over a curve of maximal regularity. It turns out that the surfaces in question have a rich geometric, homological and cohomological structure, which we aim to investigate in this paper.
Preview of results {#preview-of-results .unnumbered}
------------------
*Section 2:* We give a few preliminaries, which mainly rely on results established in [@BLPS1]. We also establish a bound for the Castelnuovo-Mumford regularity of varieties which are almost non-singular projections of varieties satisfying the Green-Lazarsfeld property $N_{2,p}$ (see Theorem \[2.5’ Theorem\]). We shall revisit the extremal secant locus $\Sigma(X)$ of an arbitrary non-degenerate irreducible projective variety $X \subset \mathbb{P}^r$ of degree $d$ and codimension $c \geq 2$, that is the closure of the set of all proper $(d-c+1)$-secant lines to $X$ in the Grassmannian $\mathbb{G}(1,\mathbb{P}^r)$ of all lines in $\mathbb{P}^r$. This locus is particularly interesting if $X$ is a surface of extremal regularity and hence satisfies the inequality $\mathrm{reg}(X) \geq d-r+3$. We also consider the so-called special extremal locus ${}^*\Sigma(X)$ of a variety $X$ of maximal sectional regularity, hence the closure of the set of all lines in $\mathbb{G}(1,\mathbb{P}^r)$ which are $(d-r+3)$-secant to a general curve section of $X$. We show that this latter locus has dimension $2$, if $X$ is a surface of maximal sectional regularity (see Proposition \[prop:dimspecextrloc\]).\
*Section 3:* We study sectionally smooth rational surfaces, hence surfaces whose general curve section is smooth and rational – a property which holds for surfaces of maximal sectional regularity. As sectionally smooth rational surfaces are almost non-singular projections of surface scrolls, they have a number of interesting properties and their cohomology is quite well understood. In particular, they satisfy the conjectural regularity bound of Eisenbud-Gôto [@EG] (see Theorem \[theorem 2.9\]). The results of this section will pave our way for a more detailed investigation of surfaces of maximal sectional regularity.\
*Section 4*: We investigate surfaces of maximal sectional regularity of type I. In particular, we compute their Betti tables (see Theorem \[t1-betti\]) and their cohomological Hilbert functions (see Theorem \[t1-coh\]). Moreover we show that the special extremal secant locus of such surfaces coincides with their extremal locus, and we show that these loci become Veronese Surfaces in a projective $5$-space under the Plücker embedding (see Proposition \[prop:extseclocI\]).\
*Section 5:* We study surfaces of maximal sectional regularity which fall under type II. We notably investigate the cohomological invariants and the cohomology tables of these surfaces (see Theorem \[4.14” Theorem\] and Corollary \[coro:cohomology\]). If $X$ is a variety of maximal sectional regularity of dimension $n \geq 2$, which falls under type II, the union $Y := X \cup \mathbb{F}(X)$ of $X$ with its extremal variety $\mathbb{F}(X) = \mathbb{P}^n \subset \mathbb{P}^r$ plays a crucial role. As an application of Theorem \[4.14” Theorem\], we shall establish a lower bound on the number of defining quadrics of a variety of maximal sectional regularity $X$ of type II with arbitrary dimension $n \geq 2$ – a bound which is sharp if and only $Y$ is arithmetically Cohen-Macaulay (see Corollary \[cor:higherdim\]). Finally, in the surface case, we give a comparison result for the Betti numbers of $X$ and $Y$ (see Proposition \[prop:BettiNumbers\]).\
*Section 6:* We study the index of normality $N(X)$ of a surface $X$ of maximal sectional regularity which falls under type II. In all examples we found, this index is sub-maximal and hence satisfies the inequality $N(X) \leq d-r$. In Theorem \[4.17” Proposition\] we give various conditions which are equivalent to the mentioned sub-maximality of $N(X)$. This sub-maximality notably implies that the homogeneous vanishing ideal of the union $Y = X \cup \mathbb{F}(X)$ is generated in degrees $\leq d-r+2$. This latter property allows to draw conclusions on the geometry of extremal secant lines to $X$ (see Remark \[remark:normality\]). We also revisit surfaces of degree $r+1$ in $\mathbb{P}^r$ and prove, that two of the eleven cases listed in [@B2] and [@BS6] may indeed not occur, as conjectured (see Remark \[3.3’ Remark\]).\
*Section 7:* This section is devoted to examples and open problems. We first provide examples of large families of surfaces of extremal regularity which are not of maximal sectional regularity and whose extremal secant locus is of any dimension in the maximally possible range $\{-1,0,1\}.$ (see Construction and Examples \[7.1 Construction and Examples\]). This is of some interest, as the paper [@GruLPe] let to the expectation that “there are only a few “exceptional“ varieties of extremal regularity without extremal secant lines”. We also provide some examples which make explicit the Betti tables in the case of surfaces of maximal sectional regularity which fall under type I (see Example \[example:t1\]). We also suggest a general construction principle which provides large classes of surfaces of maximal sectional regularity of type II (see Construction and Examples \[7.2 Construction and Examples\]). We use this principle to produce explicit examples for which we compute the Betti tables (see Examples \[7.3 Example\], \[7.4 Example\] and \[7.5 Example\]). Finally, we give some conclusive remarks and suggest a few open problems, which are related to the previously mentioned question on the sub-maximality of the index of normality (see Problems and Remark \[7.6 Problem and Remark\]).
Preliminaries
=============
The classification of varieties of maximal sectional regularity {#the-classification-of-varieties-of-maximal-sectional-regularity .unnumbered}
---------------------------------------------------------------
Let $X \subset \mathbb{P}^r$ be a non-degenerate irreducible projective variety of dimension $n \geq 2$, codimension $c \geq 3$ and degree $d \geq c+3$. We recall the following classification result on varieties of maximal sectional regularity, which was established in [@BLPS1 Theorem 7.1]
\[theorem 2.1\] If either $n = 2$ or $\mathrm{Char}(\Bbbk) = 0$, the variety $X \subset \mathbb{P}^r$ is of maximal sectional regularity if and only if it falls under one of the following two types:
- [***Type I:***]{} $c=3$ and $X$ is a divisor of the $(n+1)$-fold scroll $$W := S(\underbrace{0,\ldots,0}_{(n-2)-\rm{times}},1,1,1) \subset \mathbb{P}^{n+3}$$ with $X \approx H + (d-3)F$, where $H$ is the hyperplane divisor of $W$ and $F \subset W$ is a linear subspace of dimension $n$.
- [***Type II:***]{} There exists an $n$-dimensional linear subspace $\mathbb{F} = \mathbb{P}^n \subset \mathbb{P}^r$ such that $X \cap \mathbb{F} \subset \mathbb{F}$ is a hypersurface of degree $d-c+1$.
\[remark 2.2\] The previous classification result allows to conclude that there exist varieties $X \subset \mathbb{P}^r$ of maximal sectional regularity of dimension $n$, of codimension $c$ and of degree $d$ for any given triplet $(n,c,d)$ with $n \geq 2$, $c \geq
3$ and $d \geq c+3$.\
For the purposes of the present paper, we notice in particular the following result (see also [@BLPS1 Theorem 6.3]):
\[corollary 2.1’\] Let $5 \leq r < d$. Then, a non-degenerate irreducible projective surface $X \subset \mathbb{P}^r$ of degree $d$ is of maximal sectional regularity if and only if it falls under one of the following two types:
- [***Type I:***]{} $r=5$ and $X$ is a divisor of the smooth $3$-fold scroll $W := S(1,1,1) \subset \mathbb{P}^5$ with $X \approx H + (d-3)F$, where $H$ is the hyperplane divisor of $W$ and $F \subset W$ is a ruling plane. In this case, the surface $X$ is smooth.
- [***Type II:***]{} There exists a plane $\mathbb{F} = \mathbb{P}^2 \subset \mathbb{P}^5$ such that $X \cap \mathbb{F} \subset \mathbb{F}$ is a pure curve of degree $d-r+3$. In this case, the surface $X$ is singular.
We now introduce the notion of sectional regularity and characterize surfaces of maximal sectional regularity by means of this invariant.
\[remark and definition 0\] (A) Let $X \subset \mathbb{P}^r = {\operatorname{Proj}}(S := \Bbbk[x_0,x_1,\ldots,x_r])$ be a non-degenerate irreducible projective variety of dimension $n \geq 2$, of codimension $c \geq 2$ and of degree $d$. We introduce the notation $$\mathbb{H}_h := {\operatorname{Proj}}(S/hS) \mbox{ for all } h \in S_1 = \sum_{i=0}^r \Bbbk x_i \mbox{ with } h \neq 0\},$$ and we define the *sectional regularity* of $X$ by $$\mathrm{sreg}(X) := \mathrm{min}\{\mathrm{reg}(X \cap \mathbb{H}_h) \mid h \in S_1 \setminus \{0\}\}.$$ As the regularity is semi-continuous on hyperplane sections, we can say that $$\mathbb{W}(X) := \{h \in S_1 \setminus \{0\} \mid \mathrm{reg}(X \cap \mathbb{H}_h) = \mathrm{sreg}(X)\} \mbox{ is a dense open subset of } S_1 \setminus\{0\}.$$ Now, there is a dense open subset $\mathbb{U} \subseteq \mathbb{W}$ such that $X \cap \mathbb{H}_h$ is an integral scheme, and we denote the largest of these open sets by $\mathbb{U}(X).$\
(B) Now, assume that $4 \leq r < d$ and that $X \subset \mathbb{P}^r = {\operatorname{Proj}}(S := \Bbbk[x_0,x_1,\ldots,x_r])$ is a non-degenerate irreducible projective surface of degree $d$. In this situation $$\mathcal{C}_h := X \cap \mathbb{H}_h \mbox{ is an integral curve of degree $d$ with } \mathrm{reg}(\mathcal{C}_h) = \mathrm{sreg}(X) \mbox{ for all } h \in \mathbb{U}(X).$$ Hence, in particular we see that $$\mathrm{sreg}(X) \leq d-r+3 \mbox{ with equality if and only if } X \mbox{ is of maximal sectional regularity}.$$
Curves of maximal regularity {#curves-of-maximal-regularity .unnumbered}
----------------------------
As the generic linear curve sections of varieties of maximal sectional regularity are curves of maximal regularity, it will be useful for us to keep in mind the following fact.
\[proposition 2.3\] Let $r\geq 4$ and let $\mathcal{C} \subset
\mathbb{P}^r$ be a curve of degree $d \geq r+2$ which is of maximal sectional regularity, so that ${\operatorname{reg}}(\mathcal{C}) = d-r+2$. Then $\mathcal{C}$ admits a unique $(d-r+2)$-secant line $\mathbb{L}$. Moreover, if $d \geq 3r-3,$ then $\mathrm{depth}(\mathcal{C} \cup
\mathbb{L}) = 1.$
The existence of the $(d-r+2)$-secant line $\mathbb{L}$ follows from the classification of curves of maximal regularity given in [@GruLPe]. The uniqueness of the extremal secant line $\mathbb{L}$ follows by [@BS2 (3.1)].\
Assume now that $d \geq 3r-3$. Note that $\mathbb{S} := \mathrm{Join}(\mathbb{L},C) \subset \mathbb{P}^r$ is a rational normal $3$-fold scroll of type $S(0,0,r-1)$ whose vertex $S(0,0) \subset S(0,0,1)$ equals $\mathbb{L}.$ So, in degree $2$, the homogeneous vanishing ideals $I_{\mathcal{C}}$ and $I_{\mathbb{S}}$ of $\mathcal{C}$ respectively of $\mathbb{S}$ in $S := \Bbbk[x_0.x_1,\ldots,x_r]$ satisfy the relation $$\mathrm{dim}_{\Bbbk}(I_{\mathcal{C}})_2 \geq \mathrm{dim}_{\Bbbk}(I_{\mathbb{S}})_2 = \binom{r-2}{2}.$$ Assume now that $\mathrm{depth}(\mathcal{C} \cup \mathbb{L}) \neq 1$, so that $\mathcal{C} \cup \mathbb{L}$ is arithmetically Cohen-Macaulay. Then, by [@BS2 Proposition 3.3] it follows that $$\mathrm{dim}_\Bbbk(I_C)_2 = \binom{r+1}{2}-d-1, \mbox{ whence } \binom{r-2}{2} \leq \binom{r+1}{2}-d-1,$$ and this yields the contradiction that $d \leq 3r-4.$
Sectionally rational varieties {#sectionally-rational-varieties .unnumbered}
------------------------------
It is most important, that each variety $X$ of maximal sectional regularity is *sectionally rational*, which means that its general curve section is rational. If the general curve section of $X$ is even smooth and rational, we say that $X$ is *sectionally smooth rational*. A particularly interesting property of sectionally rational varieties is the fact, that they are birational linear projections of varieties of minimal degree. To make this statement more precise, we first give the following definition.
\[definition and remark 2.4\] (A) We define the *singular locus* of a finite morphism $f: X' \longrightarrow X$ of noetherian schemes by $$\mathrm{Sing}(f) := \{x \in X \mid \mathrm{length}\big(f^{-1}(x)\big) \geq 2\}.$$ Observe, that we also may write $$\mathrm{Sing}(f) = \mathrm{Supp}\big((f_*\mathcal{O}_{X'})/\mathcal{O}_X\big).$$
\(B) We say that the finite morphism $f: X' \longrightarrow X$ is *almost non-singular* if its singular locus $\mathrm{Sing}(f)$ is a finite set.
Now, we have the following result (see [@BLPS1 Theorem 4.1]).
\[theorem 2.5\] Let $X \subset \mathbb{P}^r$ be a non-degenerate irreducible sectionally rational projective variety of dimension $n\geq 2$ and degree $d$. Assume furthermore that $\rm{char}(\Bbbk)=0$ or $n=2$. Then, we may write
- $X = \pi_{\Lambda}(\widetilde{X})$, where $\widetilde{X} \subset \mathbb{P}^{d+n-1}$ is an $n$-dimensional variety of minimal degree,
- $\Lambda = \mathbb{P}^{d+n-r-2} \subset \mathbb{P}^{d+n-1}$ is a linear subspace with $\widetilde{X} \cap \Lambda = \emptyset,$
- $\pi_{\Lambda}:\mathbb{P}^{d+n-1} \setminus \Lambda \longrightarrow \mathbb{P}^r$ is the linear projection map from $\Lambda$ and
- the induced finite morphism $\pi_{\Lambda}:\widetilde{X} \rightarrow X$ is the normalization of $X$.
Moreover, if $d\geq 5$, then $\widetilde{X}$ is a rational $n$-fold scroll. Finally, if $X$ is a sectionally smooth rational surface, the morphism $\pi_{\Lambda}:\widetilde{X} \rightarrow X$ is almost non-singular.
A regularity bound for almost non-singular projections {#a-regularity-bound-for-almost-non-singular-projections .unnumbered}
------------------------------------------------------
We know by Theorem \[theorem 2.5\], that sectionally smooth rational surfaces are almost non-singular linear projections of rational normal scrolls. This will allow us to prove that these surfaces satisfy the conjectural Eisenbud-Goto bound. In this subsection, we shall actually prove a much more general bounding result for the regularity of almost non-singular linear projections.
\[definition 2.6\] (A) Let $p \in \mathbb{N}$. The graded ideal $I \subset S := \Bbbk[x_0,x_1,\ldots,x_r]$ is said to satisfy the *(Green-Lazarsfeld) property* $N_{2,p}$ (see [@GL]) if the Betti numbers of $S/I$ satisfy the condition $$\beta_{i,j} := \beta^S_{i,j}(S/I) = \beta^S_{i-1,j+1}(I) = 0 \mbox{ whenever } i \leq p \mbox{ and } j \neq 1,$$ – hence if the minimal free resolution of $I$ is linear up to the homological degree $p$ – and thus has the form $$\ldots \rightarrow S^{\beta_{p,1}}(-p-1)\rightarrow \ldots \rightarrow S^{\beta_{1,1}}(-2) \rightarrow I \rightarrow 0.$$ (B) The closed subscheme $Z \subset \mathbb{P}^r$ is said to satisfy the *property* $N_{2,p}$ if its homogeneous vanishing ideal $I_Z \subset S$ satisfies the property $N_{2,p}$.
Now, we may prove the announced regularity bound for almost non-singular projections of $N_{2,p}$-varieties.
\[2.5’ Theorem\] Let $r' \geq r$ be integers, let $X' \subset \mathbb{P}^{r'}$ be a non-degenerate projective variety of dimension $ \geq 2$ which satisfies the property $N_{2,p}$ for some $p \geq\max\{2,r'-r+1\}$. Let $\Lambda = \mathbb{P}^{r'-r-1}$ be a subspace such that $X' \cap \Lambda = \emptyset$ and let $\pi_\Lambda:\mathbb{P}^{r'} \setminus \Lambda: \twoheadrightarrow \mathbb{P}^r$ be the linear projection from $\Lambda$. Let $X := \pi_\Lambda(X') \subset \mathbb{P}^r$ and assume that the induced finite morphism $\pi_{\Lambda}: X' \twoheadrightarrow X$ is almost non-singular. Then
- The homogeneous vanishing ideal $I_X \subset S$ of $X$ is generated by homogeneous polynomials of degrees $\leq r'-r+2$.
- ${\operatorname{reg}}(X) \leq \max\{{\operatorname{reg}}(X'), r'-r+2\}$.
Let $I_{X'} \subset S' := \Bbbk[x_0,x_1,\ldots,x_r,x_{r+1},\ldots,x_{r'}] = S[x_{r+1},\ldots,x_{r'}]$ be the homogeneous vanishing ideal of $X' \subset \mathbb{P}^{r'} = {\operatorname{Proj}}(S')$ and let $A':= S'/I_{X'}$ be the homogeneous coordinate ring of $X'$. We assume that $\Lambda = {\operatorname{Proj}}(S'/(x_0,x_1,\ldots,x_r)S')$, consider $A'$ as a finitely generated graded $S$-module and set $t:= r'-r$. As $X'$ satisfies the condition $N_{2,p}$ with $p \geq\max\{2,t+1\}$, it follows by [@AK Theorem 3.6], that the minimal free presentation of $A'$ has the shape $$S^s(-2) \stackrel{v}{\rightarrow} S \oplus S^t(-1) \stackrel{q}{\rightarrow} A' \rightarrow 0$$ for some $s \in \mathbb{N}$. Moreover, the coordinate ring $A = S/I_X$ of $X$ is nothing else than the image $q(S)$ under $q$ of the direct summand $S \subset S\oplus S^t(-1)$. Therefore $$A'/A \cong {\operatorname{Coker}}\big(u:S^s(-2) \rightarrow S^t(-1)\big),$$ where $u$ is the composition of the map $v:S^s(-2) \rightarrow S\oplus S^t(-1)$ with the canonical projection map $w: S\oplus S^t(-1) \twoheadrightarrow S^t(-1)$. Hence, the $S$-module $(A'/A)(1)$ is generated by $t$ homogeneous elements of degree $0$ and related in degree $1$. As ${\operatorname{Sing}}(\pi_{\Lambda})$ is finite, we have $\dim(A'/A) \leq 1$. So, it follows by [@ChFN Corollary 2.4] that ${\operatorname{reg}}\big((A'/A)(1)\big) \leq t-1$, whence ${\operatorname{reg}}(A'/A) \leq t$. Now, the short exact sequence $0\rightarrow A \rightarrow A' \rightarrow A'/A \rightarrow 0$ implies that ${\operatorname{reg}}(A) \leq \max\{{\operatorname{reg}}(A'), t+1\}$. It follows that ${\operatorname{reg}}(X) = {\operatorname{reg}}(A) + 1 \leq \max\{{\operatorname{reg}}(A') + 1, t+2\} = \max\{{\operatorname{reg}}(X'), t+2\} = \max\{{\operatorname{reg}}(X'), r'-r+2\}$. This proves claim (b).\
To prove claim (a), observe that $I_X = {\operatorname{Ker}}(q) \cap S$ occurs in the short exact sequence of graded $S$-modules $0 \rightarrow I_X \rightarrow \mathrm{Im}(v) \stackrel{w\upharpoonright}{\rightarrow} \mathrm{Im}(u) \rightarrow 0$, where ${w\upharpoonright}$ is the restriction of the above projection map $w$. In particular, we may identify ${w\upharpoonright}$ with the canonical map $S^s(-2)/{\operatorname{Ker}}(v) \twoheadrightarrow S^s(-2)/{\operatorname{Ker}}(u)$. It follows, that $I_X \cong {\operatorname{Ker}}(u)/{\operatorname{Ker}}(v)$. In view of the exact sequence $0\rightarrow {\operatorname{Ker}}(u) \rightarrow S^s(-2) \stackrel{u}{\rightarrow}S^t(-1) \rightarrow A'/A \rightarrow 0$, we finally get ${\operatorname{reg}}({\operatorname{Ker}}(u)) \leq t+2 = r'-r+2$. Therefore ${\operatorname{Ker}}(u)$ is generated in degrees $\leq r'-r+2$, and hence so is $I_X$. This proves statement (a).
Extremal secant loci and extremal varieties {#extremal-secant-loci-and-extremal-varieties .unnumbered}
-------------------------------------------
In this subsection, we recall a few facts on the geometry of proper $(d-c+1)$-secant lines to a non-degenerate irreducible projective variety $X \subset \mathbb{P}^r$ of codimension $c$ and degree $d$. We also recall the related notion of extremal secant locus $\Sigma (X)$ of $X$, that is, the closure of the set of all proper $(d-c+1)$-secant lines of $X$ in the Grassmannian $\mathbb{G}(1,\mathbb{P}^r)$.
\[4.1” Notation and Reminder\] (See [@BLPS1 Notation and Reminder 3.1]) (A) Let $X \subset \mathbb{P}^r$ be as above and let $$\Sigma(X) := \overline{\{\mathbb{L} \in \mathbb{G}(1,\mathbb{P}^r) \mid d-c+1 \leq \mathrm{length}(X \cap \mathbb{L}) < \infty\}}$$ denotes the *extremal secant locus* of $X.$ Keep in mind, that setting $n := \dim(X) = r-c$ we can say (see [@BLPS1 Theorem 3.4]) $$\dim\big(\Sigma(X)\big) \leq 2n-2 \mbox{ with equality if and only if } X \mbox{ is of maximal sectional regularity}.$$ (B) Keep the above notations and hypotheses and let $\mathcal{U}(X)$ denote the largest open subset $\mathcal{U} \subset \mathbb{G}(c+1, \mathbb{P}^r )$ such that $$\mathcal{C}_{\Lambda} := X \cap \Lambda \subset \Lambda = \mathbb{P}^{c+1} \mbox{ is an integral curve of maximal regularity for all } \Lambda \in \mathcal{U}.$$ Observe that $$X \mbox{ is of maximal sectional regularity if and only if } \mathcal{U}(X) \neq \emptyset.$$
We introduce a subset of the extremal locus of a variety of maximal sectional regularity, which reflects in a particular way the nature of these varieties. We use this set to define the extremal variety of a variety of maximal sectional regularity.
\[4.2” Notation and Reminder\] (See [@BLPS1 Section 5]) (A) Let the notations and hypotheses be as in Notation and Reminder \[4.1” Notation and Reminder\], and assume that $X$ is of dimension $n \geq 2$ and of maximal sectional regularity, so that $\mathcal{U}(X) \neq \emptyset$. For all $\Lambda \in \mathcal{U}(X)$ let $\mathbb{L}_{\Lambda} \in \Sigma(X)$ denote the unique $(d-c+1)$-secant line to the curve $\mathcal{C}_{\Lambda} \subset \Lambda = \mathbb{P}^{c+1}$ (see Notation and Reminder \[4.1” Notation and Reminder\] (B) and Proposition \[proposition 2.3\]), so that $$\mathbb{L}_{\Lambda} \in \mathbb{G}(1, \mathbb{P}^r) \mbox{ with } \quad \mathrm{length}(\mathcal{C}_{\Lambda} \cap \mathbb{L}_{\Lambda}) = \mathrm{length}(X \cap \mathbb{L}_{\Lambda}) = d-c+1.$$ The $(d-c+1)$-secant lines of the form $\mathbb{L}_{\Lambda}$ with $\Lambda \in \mathcal{U}(X)$ are called *special extremal secant lines*, whereas the set $${}^*\Sigma(X) := \overline{\{\mathbb{L}_{\Lambda} \mid \Lambda \in \mathcal{U}(X)\}} \subseteq \Sigma(X) \subset \mathbb{G}(1,\mathbb{P}^r)$$ is called the *special extremal secant locus* of $X$. If $n=1$, then $4 \leq r < d$ and $X \subset \mathbb{P}^r$ is a curve of maximal sectional regularity, and hence admits a unique extremal secant line $\mathbb{L} \in \mathbb{G}(1,\mathbb{P}^r)$ (see Proposition \[proposition 2.3\]). So, we define ${}^*\Sigma(X) := \{\mathbb{L}\}$ in this case.\
(B) We define the *extremal variety* and the *extended extremal variety* of $X$ respectively by $$\mathbb{F}(X) := \overline{\bigcup_{\Lambda \in \mathcal{U}(X)} \mathbb{L}_{\Lambda}} \quad (\quad \subseteq \quad) \quad \mathbb{F}^{+}(X) := \overline{\bigcup_{\mathbb{L} \in \Sigma(X)} \mathbb{L}}.$$ (C) Keep the previous notations and hypotheses, and assume that $5 \leq r < d$. Let $X \subset \mathbb{P}^r = {\operatorname{Proj}}\big(S := \Bbbk[x_0,x_1,\ldots,x_r]\big)$ be a surface of maximal sectional regularity, so that $c+1 = r-1$ and $\mathrm{sreg}(X) = d-r+3$. Then, in the notations of Remark and definition \[remark and definition 0\] we have $$\mathbb{U}(X) = \{h \in S_1 \setminus\{0\} \mid {\operatorname{Proj}}(S/hS) \in \mathcal{U}(X)\}.$$ Moreover, for each $h \in \mathbb{U}(X)$, the line $\mathbb{L}_h := \mathbb{L}_{\mathbb{H}_h}$ is the unique $(d-r+3)$-secant line to the curve of maximal regularity $\mathcal{C}_h \subset \mathbb{H}_h$, and hence the line defined by the condition $$\mathbb{L}_h \in \mathbb{G}(1, \mathbb{P}^r)\quad \mbox{ with } \quad \mathrm{length}(\mathcal{C}_h \cap \mathbb{L}_h) = \mathrm{length}(X \cap \mathbb{L}_h) = d-r+3.$$ Now, according to [@BLPS1 Theorem 6.3] we can say:
- If $X$ is of type I, then the extremal variety $\mathbb{F}(X)$ and the extended extremal variety $\mathbb{F}^{+}(X)$ of $X$ both coincide with the smooth $3$-fold scroll $W = S(1,1,1) \subset \mathbb{P}^r$ of Corollary \[corollary 2.1’\].
- If $X$ is of type II, then the extremal variety $\mathbb{F}(X)$ of $X$ coincides with the plane $\mathbb{F} = \mathbb{P}^2 \subset \mathbb{P}^r$ of Corollary \[corollary 2.1’\].
The following result says that the extremal secant locus and the special extremal secant locus of a variety of maximal sectional regularity have the same dimension.
\[prop:dimspecextrloc\] Let $c \geq 3$, let $d \geq c+3$, let $n \geq 1$ and $X \subset \mathbb{P}^r$ be a non-degenerate irreducible variety of dimension $n$ and degree $d$ which is of maximal sectional regularity. Then $\dim\big({}^*\Sigma(X)\big) = 2n-2$.
As ${}^*\Sigma(X) \subseteq \Sigma(X)$ it follows by the last observation made in Notation and Reminder \[4.1” Notation and Reminder\] (A) that $\dim\big({}^*\Sigma(X)\big) \leq \dim\big(\Sigma(X)\big) \leq 2n-2.$ It remains to show, that $\dim\big({}^*\Sigma(X)\big) \geq 2n-2.$\
We proceed by induction on $n.$ If $n = 1$, then our claim is clear by the definition of ${}^*\Sigma(X)$ (see Notation and Reminder \[4.2” Notation and Reminder\] (A)). So, let $n > 1$ and let $\mathbb{H} = \mathbb{P}^{r-1} \subset \mathbb{P}^r$ be a general hyperplane. Then $X \cap \mathbb{H} \subset \mathbb{H}$ is a variety of dimension $n - 1$, codimension $c$ and degree $d$, which is of maximal sectional regularity. Moreover, each special extremal secant line $\mathbb{L} \in {}^*\Sigma(X \cap \mathbb{H})$ to $X \cap \mathbb{H}$ is a special extremal secant line to $X$. Therefore, we can say that ${}^*\Sigma(X \cap \mathbb{H}) \subseteq {}^*\Sigma(X) \cap \mathbb{G}(1,\mathbb{H}).$\
By induction, we have $\dim\big({}^*\Sigma(X \cap \mathbb{H})\big) \geq 2(n-1)-2 = 2n-4.$ If we apply [@BLPS1 Lemma 3.2] with $T = {}^*\Sigma(X)$ we obtain $\dim\big({}^*\Sigma(X)\big) \geq \dim\big({}^*\Sigma(X) \cap \mathbb{G}(1,\mathbb{H})) + 2.$ This proves our claim.
Sectionally Smooth Rational Surfaces
====================================
The projecting scroll {#the-projecting-scroll .unnumbered}
---------------------
In this section we investigate sectionally smooth rational surfaces. We do this, because surfaces of maximal sectional regularity are sectionally smooth rational. Let us recall first, that according to Theorem \[theorem 2.5\] we can say.
\[corollary 2.7\] Let $X \subset \mathbb{P}^r$ be a non-degenerate irreducible projective and sectionally smooth rational surface of degree $d$ with $r \geq 4$ and $d \geq r+1$.\
Then, there exists a unique non-negative integer $a \leq \frac{d}{2}$ such that $X = \pi_{\Lambda} (\widetilde{X})$, where $\widetilde{X} = S(a,d-a) \subset \mathbb{P}^{d+1}$ and
- $\Lambda = \mathbb{P}^{d-r} \subset \mathbb{P}^{d+1}$ is a subspace such that $\widetilde{X} \cap \Lambda = \emptyset$,
- $\pi_{\Lambda} : \mathbb{P}^{d+1} \setminus \Lambda \rightarrow \mathbb{P}^r$ is the linear projection map from $\Lambda$,
- the induced finite morphism $\pi_{\Lambda}: \widetilde{X} \rightarrow X$ is almost non-singular and coincides with the normalization of $X$, and
- $\widetilde{X}$ is smooth (or, equivalently, $a > 0$) if and only if $X$ is not a cone.
In the above situation, we call $\widetilde{X}= S(a,d-a)$ the *projecting scroll* of the surface $X$, $\Lambda \subset \mathbb{P}^r$ the *projecting center* for $X$ and $\pi_\Lambda:\widetilde{X} \twoheadrightarrow X$ the *standard normalization* of $X$.\
Algebraic and cohomological properties {#algebraic-and-cohomological-properties .unnumbered}
--------------------------------------
The precise aim of this section is to investigate a few algebraic and geometric properties of sectionally smooth rational surfaces, which are encoded in Corollary \[corollary 2.7\]. We begin with a few preliminaries.
\[2.7’ Notation and Reminder\] Let $X \subset \mathbb{P}^r = {\operatorname{Proj}}(S := \Bbbk[x_0,\ldots,x_r])$ be a non-degenerate irreducible projective surface of degree $d$, homogeneous vanishing ideal $I \subset S$ and homogeneous coordinate ring $A = S/I$.
\(A) (see [@BS6]) For any graded ideal $\mathfrak{b} \subseteq A_+ := S_+A$ and any graded $A$-module $M$ let $D_\mathfrak{b}(M) := \varinjlim {\operatorname{Hom}}_A({\mathfrak b}^n, M)$ denote the $\mathfrak{b}$-transform of $M$, and let $H^i_\mathfrak{b}(M), \quad (i \in \mathbb{N}_0)$ denote the $i$-th local cohomology module of $M$, both furnished with their natural grading. We usulally will write $H^i(M)$ instead of $H^i_{A_+}(M).$\
Let ${\mathfrak a} \subseteq A_+$ be the graded radical ideal which defines the non-Cohen-Macaulay locus $X \setminus \operatorname{CM}(X)$ of $X$. Observe that height ${\mathfrak a} \geq 2,$ so that the $\mathfrak{a}$-transform $$B(A):= D_{\mathfrak a}(A) = \varinjlim {\operatorname{Hom}}_A({\mathfrak a}^n, A) = \bigcup_{n\in \mathbb{N}}(A:_{\rm{Quot}(A)}\mathfrak{a}^n) =
\bigoplus _{n \in {\mathbb Z}} \Gamma (\operatorname{CM}(Z), {\mathcal O}_Z (n))$$ of $A$ is a positively graded finite birational integral extension domain of $A$. In particular $B(A)_0 = \Bbbk$. Moreover $B(A)$ has the second Serre-property $S_2$. As $\mbox{Proj}(B(A))$ is of dimension $2$, it thus is a locally Cohen-Macaulay scheme.
If $E$ is a finite graded integral extension domain of $A$ which satisfies the property $S_2$, we have $A \subset B(A) \subset E$. So $B(A)$ is the least finite graded integral extension domain of $A$ which has the property $S_2$. Therefore, we call $B(A)$ the $S_2$[*-cover of*]{} $A$. We also can describe $B(A)$ as the endomorphism ring $\mbox{End}(K(A), K(A))$ of the canonical module $K(A) = K^3(A) = \mbox{Ext}^{r-2}_S(A, S(-r-1))$ of $A$.
\(B) The inclusion map $A \rightarrow B(A)$ gives rise to a finite morphism $$\pi: \widetilde{X}:= {\operatorname{Proj}}(B(A)) \twoheadrightarrow X, \mbox{ with } {\operatorname{Sing}}(\pi) = X \setminus {\operatorname{CM}}(X).$$ In particular $\pi$ is almost non-singular and hence birational. Moreover, for any finite morphism $\rho: Y \twoheadrightarrow X$ such that $Y$ is locally Cohen-Macaulay, there is a unique morphism $\sigma:Y \rightarrow \widetilde{X}$ such that $\rho = \pi \circ\sigma$. In addition $\sigma$ is an isomorphism if and only if ${\operatorname{Sing}}(\rho) = X \setminus {\operatorname{CM}}(X)$. Therefore, the morphism $\pi:\widetilde{X} \twoheadrightarrow X$ is addressed as the *finite Macaulayfication* of $X$. Keep in mind, that – unlike to what happens with normalization – there may be proper birational morphisms $\tau:Z \twoheadrightarrow X$ with $Z$ locally Cohen-Macaulay, which do not factor through $\pi$ (see [@B0]).
\(C) We also introduce the invariants $${\operatorname{e}}_x(X) := {\operatorname{length}}(H^1_{\mathfrak{m}_{X,x}}(\mathcal{O}_{X,x})),
(x \in X \mbox{ closed }) \; \mbox{ and }
{\operatorname{e}}(X):= \sum_{x\in X, {\rm{closed}}} {\operatorname{e}}_x(X).$$ Note that the latter counts the [*number of non-Cohen-Macaulay points*]{} of $X$ in a weighted way. Keep in mind that $${\operatorname{e}}(X) = h^1 (X, \mathcal{O}_X(j)) = h^2(\mathbb{P}^r,\mathcal{I}_X(j)) \mbox{ for all } j \ll 0.$$
Now, we are ready to prove the following result on sectionally smooth rational surfaces.
\[theorem 2.9\] Let $X \subset \mathbb{P}^r$ be a non-degenerate irreducible projective and sectionally smooth rational surface of degree $d$ with $r \geq 4$ and $d \geq r+2$. Then it holds
- ${\operatorname{reg}}(X) \leq d - r + 3.$
- ${\operatorname{Reg}}(X) = {\operatorname{Nor}}(X) = {\operatorname{CM}}(X)$.
- The homogeneous coordinate ring of the projecting scroll $\widetilde{X}\subset \mathbb{P}^{d+1}$ is the $S_2$-cover $B(A)$ of the homogeneous coordinate ring $A$ of $X \subset \mathbb{P}^r$ and the standard normalization $\pi_{\Lambda}: \widetilde{X} \rightarrow X$ is the finite Macaulayfication of $X$.
- If ${\operatorname{e}}(X) = 0,$ then $h^2(\mathbb{P}^r,\mathcal{I}_X(j)) = 0$ for all $j \in \mathbb{Z}$. If ${\operatorname{e}}(X) \neq 0$, then $$h^2(\mathbb{P}^r,\mathcal{I}_X(j)) \begin{cases} = {\operatorname{e}}(X) &\mbox{ for all } j \leq 0;\\
= {\operatorname{e}}(X) + h^1(\mathbb{P}^r,\mathcal{I}_X(1)) + r - d -1 &\mbox{ for } j=1;\\
\leq \max\{0, h^2(\mathbb{P}^r,\mathcal{I}_X(j-1)) -1 \} &\mbox{ for all } j > 1;\\
\leq 1 &\mbox{ for } j = d-r;\\
= 0 &\mbox{ for all } j \geq d-r+1.
\end{cases}$$
- $h^3(\mathbb{P}^r, \mathcal{I}_X(j)) =
h^2(\widetilde{X},\mathcal{O}_{\widetilde{X}}(j))$ for all $j \in \mathbb{Z}$, thus $$h^3(\mathbb{P}^r, \mathcal{I}_X(j))= \begin{cases} \frac{(j+1)(dj+2)}{2} &\mbox{ for all } j \leq - 2,\\
0 &\mbox{ for all } j \geq -1. \end{cases}$$
(a): The projecting scroll $\widetilde{X} \subset \mathbb{P}^{d+1}$ satisfies the conditions $N_{2,p}$ for all $p \in \mathbb{N}$, so that ${\operatorname{reg}}(\widetilde{X}) = 2$. Therefore, by Theorem \[2.5’ Theorem\] we get ${\operatorname{reg}}(X) \leq (d-1)-r+2 =d-r+3$, and this proves our claim.\
(b): As $X$ is a surface, we have ${\operatorname{Reg}}(X) \subset {\operatorname{Nor}}(X) \subset {\operatorname{CM}}(X)$. Therefore, it suffices to show that ${\operatorname{CM}}(X) \subset {\operatorname{Reg}}(X).$\
So, let $x \in {\operatorname{CM}}(X)$ be a closed point. We always write $\pi := \pi_\Lambda$. Then $\big(\pi_*\mathcal{O}_{\widetilde{X}}\big)_x$ is a finite integral extension domain of the local $2$-dimensional CM ring $(\mathcal{O}_{X,x},\mathfrak{m}_{X,x})$. As the morphism $\pi: \widetilde{X} \rightarrow X$ is almost non-singular, the finitely generated $\mathcal{O}_{X,x}$-module $$\big((\pi_*\mathcal{O}_{\widetilde{X}})/ \mathcal{O}_X\big)_x = \big(\pi_*\mathcal{O}_{\widetilde{X}}\big)_x/ \mathcal{O}_{X,x}$$ is annihilated by some power of $\mathfrak{m}_{X,x}$ and hence contained in the local cohomology module $H^1_{\mathfrak{m}_{X,x}}(\mathcal{O}_{X,x})$. But this latter module vanishes because $\mathcal{O}_{X,x}$ is a local ${\operatorname{CM}}$-ring of dimension $>1$. This shows, that $x \notin \mathrm{Sing}(\pi)$.\
But this means, that $x$ has a unique preimage $\widetilde{x} \in \widetilde{X}$ under the morphism $\pi$ and that $$\mathcal{O}_{\widetilde{X},\widetilde{x}} \cong \mathcal{O}_{X,x}.$$ Assume now, that $x \notin {\operatorname{Reg}}(X).$ Then $\widetilde{x} \notin
{\operatorname{Reg}}(\widetilde{X})$. This means that $\widetilde{X} \subset \mathbb{P}^{d+1}$ is a singular $2$-fold scroll with vertex $\widetilde{x}$. But this implies that the tangent space $\mathrm{T}_{\widetilde{x}}(\widetilde{X})$ of $\widetilde{X}$ at $\widetilde{x}$ has dimension $d+1$. In view of the above isomorphism, we thus get the contradiction that the tangent space $\mathrm{T}_x(X)$ of $X \subset \mathbb{P}^r$ at $x$ has dimension $d+1$. This proves that indeed $x \in {\operatorname{Reg}}(X)$.\
(c): Let $E$ denote the homogeneous coordinate ring of the projecting scroll $\widetilde{X} \subset \mathbb{P}^{d+1}$. As $E$ is a CM ring, we have canonical inclusions of graded rings $$A \subset B(A)=\bigcup_{n\in \mathbb{N}}(A:_{\rm{Quot}(A)}\mathfrak{a}^n) \subset E$$ (see Notation and Reminder \[2.7’ Notation and Reminder\] (A)). Keeping in mind statement (b) and observing that the projection morphism $\pi_\Lambda: \widetilde{X} \longrightarrow X$ provides the normalization of $X$, we thus get $${\operatorname{Proj}}(A/\mathfrak{a}) = X \setminus {\operatorname{CM}}(X) = X \setminus {\operatorname{Nor}}(X) = {\operatorname{Sing}}(\pi_{\Lambda})$$ and hence $E \subset
\bigcup_{n\in \mathbb{N}}(A:_{\rm{Quot}(A)}\mathfrak{a}^n) = B(A)$. Therefore $E = B(A)$ and statement (c) is shown.\
(d): Let $B := B(A)$ as above, let $D:= D_{A_+}(A)$ and consider the short exact sequence of graded $S$-modules $$0 \longrightarrow D \longrightarrow B \longrightarrow C \longrightarrow 0.$$ Observe that $\dim(C) \leq 1$, ${\operatorname{depth}}(C) > 0$ and that $\widetilde{C} \cong \mathcal{F} := \pi_*
\mathcal{O}_{\widetilde{X}}/\mathcal{O}_X$, so that $$\dim_{\Bbbk}(D_{A_+}(C)_j) = {\operatorname{e}}(X) \mbox{ for all } j \in \mathbb{Z}.$$ As $B$ is a Cohen-Macaulay module of dimension $3$, we have $$H^2_{*}(\mathbb{P}^r,\mathcal{I}_X) \cong H^2(A) \cong H^2(D) \cong H^1(C) \cong D_{A_+}(C)/C.$$ Hence, if ${\operatorname{e}}(X) = 0$, we have indeed $h^2(\mathbb{P}^r,\mathcal{I}_X(j)) = 0$ for all $j \in \mathbb{Z}$.\
So, let ${\operatorname{e}}(X) > 0$. As $C_j = 0$ for all $j \leq 0$ we get that $h^2(\mathbb{P}^r,\mathcal{I}_X(j))={\operatorname{e}}(X)$ for all $j \leq 0$. As $\dim_{\Bbbk}(D_1) = r+1 + h^1(\mathbb{P}^r,\mathcal{I}_X(1))$, we have $\dim_{\Bbbk}(C_1) = \dim_{\Bbbk}(B_1) - \dim_{\Bbbk}(D_1) = d+2 - \big(r+1+h^1(\mathbb{P}^r,\mathcal{I}_X(1))\big)$ and hence $$\begin{aligned}
h^2(\mathbb{P}^r,\mathcal{I}_X(1)) &= \dim_\Bbbk(D_{A_+}(C)_1)-\dim_\Bbbk(C_1) = {\operatorname{e}}(X) - \dim_\Bbbk(C_1) = \\ &= {\operatorname{e}}(X) - h^1(\mathbb{P}^r,\mathcal{I}_X(1)) + r - d -1.\end{aligned}$$ Observe that $C$ is a Cohen-Macaulay module of dimension $1$. Moreover, the regularity of $B$ as an $A$-module and as a $B$-module take the same value ${\operatorname{reg}}(B) = {\operatorname{reg}}(\widetilde{X})-1 = 1$, so that the $A$-module $B$ is generated in degree $1$. Therefore, the $A$-module $C$ is generated in degree $1$, and hence $\dim_{\Bbbk}(H^1(C)_j) \leq \max\{0, \dim_{\Bbbk}(H^1(C)_{j-1})-1\}$ for all $j > 1$ (see for example [@BS1]). Therefore we get indeed $h^2(\mathbb{P}^r,\mathcal{I}_X(j)) \leq \max\{0, h^2(\mathbb{P}^r,\mathcal{I}_X(j-1)) -1 \}$ for all $j > 1.$\
By statement (a) we have $h^2(\mathbb{P}^r,\mathcal{I}_X(j)) = 0$ for all $j \geq d-r+1$.\
Finally, let us consider the exact sequence $$0 \longrightarrow \mathcal{I}_X(-1) \longrightarrow \mathcal{I}_X \longrightarrow \mathcal{I}_{\mathcal{C}} \longrightarrow 0,$$ where $\mathcal{C} = X \cap \mathbb{H} \quad (\mathbb{H}\in
\mathbb{G}(r-1,\mathbb{P}^r))$ is a general hyperplane section of $X$. As $\mathcal{C} \subset \mathbb{H} = \mathbb{P}^{r-1}$ is a smooth rational curve of degree $d$, we have $h^1(\mathbb{P}^{r-1},
\mathcal{I}_{\mathcal{C}}(1)) = d-r+1$ and $h^1(\mathbb{P}^{r-1},\mathcal{I}_{\mathcal{C}}(j+1)) \leq \mathrm{max}\{0, h^1(\mathbb{P}^{r-1},\mathcal{I}_{\mathcal{C}}(j))-1\}$ for all $j \geq 1$, so that $h^1(\mathbb{P}^{r-1},\mathcal{I}_{\mathcal{C}}(d-r+1)) \leq 1.$ Applying the above exact sequence and keeping in mind that $h^2(\mathbb{P}^r,\mathcal{I}_X(d-r+1)) = 0$ we get $h^2(\mathbb{P}^r,\mathcal{I}_X(d-r)) \leq 1.$ and this proves our claim.\
(e): Let the notation be as above. As the sheaf $\mathcal{F} :=
\pi_* \mathcal{O}_{\widetilde{X}}/\mathcal{O}_X$ has finite support, the sequence $$0 \longrightarrow \mathcal{O}_X \longrightarrow \pi_* \mathcal{O}_{\widetilde{X}} \longrightarrow \mathcal{F} \longrightarrow 0$$ together with the well known formulas for the cohomology of a rational surface scroll yields that $$h^3(\mathbb{P}^r, \mathcal{I}_X(j)) = h^2(X,\mathcal{O}_X(j))= h^2(\widetilde{X},\mathcal{O}_{\widetilde{X}}(j)) = \begin{cases} \frac{(j+1)(dj+2)}{2}, &\mbox{ if } j \leq - 2,\\
0, &\mbox{ if } j \geq -1. \end{cases}$$ This proves statement (e).
Local properties {#local-properties .unnumbered}
----------------
Finally, we want to give the following result, in which $\mathrm{mult}_z(Z)$ is used to denote the *multiplicity* of the noetherian scheme $Z$ at the point $z \in Z.$
\[prop:loc,properties\] Let $X \subset \mathbb{P}^r$ be a non-degenerate irreducible projective and sectionally smooth rational surface of degree $d$ with $r \geq 4$ and $d \geq r+2$. Let $\pi = \pi_\Lambda: \widetilde{X} \twoheadrightarrow X$ be the standard normalization of $X$. Then it holds:
- If $x \in \mathrm{Sing}(X)$, then $2 \leq \mathrm{length}(\pi^{-1}(x)) \leq \mathrm{max}\{\mathrm{mult}_x(X),\mathrm{e}_x(X)+1\}.$
- If $\mathbb{K} \in \mathbb{G}(k,\mathbb{P}^r)$ with $0 \leq k \leq r-1$ and $\mathrm{dim}(X \cap \mathbb{K}) \leq 0$, then $$\mathrm{length}\big({\operatorname{Reg}}(X) \cap \mathbb{K}\big) + 2\#\big({\operatorname{Sing}}(X) \cap \mathbb{K}\big) \leq d-r+k+2.$$
- If $\mathbb{L} \in \mathbb{G}(1,\mathbb{P}^r)$ is a proper extremal secant line to $X$ with $X \cap \mathbb{L} \subset \mathrm{Reg}(X)$, and $x \in \mathrm{Sing}(X)$, then $\dim(X \cap \langle x,\mathbb{L}\rangle) = 1$.
(a): Indeed, as $x$ is an isolated point of $\mathrm{Sing}(X) = \mathrm{Sing}(\pi_\Lambda)$ and as $\widetilde{X}$ is a Cohen-Macaulay surface, the semilocal ring $\mathcal{O}_{\widetilde{X},x} := \big({\pi}_{*}\mathcal{O}_{\widetilde{X},x}\big)_x$ is a Cohen-Macaulay finite integral extension domain of the local ring $\mathcal{O}_{X,x}$ and $H^1_{\mathfrak{m}_{X,x}}(\mathcal{O}_{X,x}) \cong \mathcal{O}_{\widetilde{X},x}/\mathcal{O}_{X,x}$. As $\pi^{-1}(x) \cong \mathcal{O}_{\widetilde{X},x}/\mathfrak{m}_{X,x}\mathcal{O}_{\widetilde{X},x}$ our claim follows easily.\
(b): Let $\mathbb{K}' := \overline{(\pi'_\Lambda)^{-1}(\mathbb{K})} \in \mathbb{G}(d-r+k+1,\mathbb{P}^{d+1})$ be the closed preimage of $\mathbb{K}$ under the linear projection $\pi'_\Lambda:\mathbb{P}^{d+1} \setminus \Lambda \twoheadrightarrow \mathbb{P}^r$. Then $\widetilde{X} \cap \mathbb{K}' = \pi^{-1}(X \cap \mathbb{K})$, and so the morphism $\pi: \widetilde{X} \twoheadrightarrow X$ induces an isomorphism $$\pi\upharpoonright: \widetilde{X} \cap \mathbb{K}'\setminus \big[\pi^{-1}\big(\mathrm{Sing}(\pi)\cap \mathbb{K}\big)\big] \stackrel{\cong}{\longrightarrow} X \cap \mathbb{K} \setminus \big(\mathrm{Sing}(\pi)\cap \mathbb{K}\big) = \mathrm{Reg}(X) \cap \mathbb{K}.$$ Moreover $\widetilde{X} \cap \mathbb{K}'$ is of dimension $\leq 0.$ As $\widetilde{X} \subseteq \mathbb{P}^{d+1}$ is a surface scroll and $\mathbb{K} \in \mathbb{G}(d-r+k+1,\mathbb{P}^{d+1})$, we get that $\mathrm{length}(\widetilde{X} \cap \mathbb{K}') \leq d-r+k+2.$ Now, our statement follows by the first inequality of statement (a).\
(c): Observe, that $\mathbb{K} := \langle x,\mathbb{L}\rangle$ is a plane. Assume that $\dim(X\cap \mathbb{K}) \leq 0$. Then, by statement (b) we get $$d-r+5 \leq \mathrm{length}(X \cap \mathbb{L}) + 2 \leq \mathrm{length}\big(\mathrm{Reg}(X) \cap \mathbb{K}\big) + 2\#\big(\mathrm{Sing}(X)\cap \mathbb{K}\big)
\leq d-r+4.$$ This contradiction proves our claim.
Surfaces of Type I
==================
The Betti numbers {#the-betti-numbers .unnumbered}
-----------------
In this section we study the surfaces which fall under type I of our classification. We begin by investigating their Betti numbers.
\[convention and remark\] (A) Let $X \subset \mathbb{P}^5$ denote a projective surface contained in a smooth rational three-fold scroll in $\mathbb{P}^5$, hence that $$X \subset W:= S(1,1,1) \subset \mathbb{P}^5.$$ We assume furthermore that the divisor $X \subset W$ satisfies $$X \approx H+(d-3)F \mbox{ for some } d \geq 5,$$ where $H$ is a hyperplane section and $F$ is a plane of $W$. Then it is easy to check that $$\mathrm{deg} (X)=d \quad \mbox{and} \quad \mathrm{reg} (X)=d-2.$$
\(B) With the definition of the Betti numbers and the Betti diagram we follow the notations suggested by D. Eisenbud (see [@E]). So, if $Z \subset \mathbb{P}^r$ is a closed subscheme, with homogeneous vanishing ideal $I_Z \subset S:= \Bbbk[x_0,x_1,\ldots,x_r]$ and homogeneous coordinate ring $A_Z := S/I_Z$, we write $$\beta_{i,j} = \beta_{i,j}(Z) := \mathrm{dim}_\Bbbk
\big(\mathrm{Tor}^S_i(\Bbbk,A_Z)_{i+j}\big) \mbox{ for all } i \in \mathbb{N}_0 \mbox{ and all } j \in\mathbb{Z}.$$ As usually, if $Z$ is non-degenerate, we list this numbers only the range $1\leq i \leq r+1-\mathrm{depth}(Z)$ and $1\leq j < \mathrm{reg}(Z)$
\[t1-betti\] In the previous notation we have the following Betti diagram of $X$ $$\begin{tabular}{|c|c|c|c|c|c|c|}\hline
\multicolumn{2}{|c||}{$i$}& $1$&$2$&$3$&$4$&$5$ \\\cline{1-7}
\multicolumn{2}{|c||}{$\beta_{i,1}$}& $3$&$2$&$0$&$0$&$0$ \\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,2}$}& $0$& $0$&$0$& $0$&$0$ \\\cline{1-1}
\multicolumn{2}{|c||}{$\vdots$}& $\vdots$&$\vdots$&$\vdots$&$\vdots$&$\vdots$\\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,d-4}$}& $0$&$0$&$0$&$0$&$0$\\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,d-3}$}& $\beta_{1,d-3}$&$\beta_{2,d-3}$&$\beta_{3,d-3}$&$\beta_{4,d-3}$&$\beta_{5,d-3}$\\\cline{1-7}
\end{tabular}$$ with the following entries $$\begin{gathered}
\beta_{1,d-3}={{d-1}\choose{2}},\quad \beta_{2,d-3}=2(d-1)(d-3),\quad\beta_{3,d-3}=3(d^2-5d+5) \\
\beta_{4,d-3}=2(d-2)(d-4)\quad \mbox{and} \quad \beta_{5,d-3}={{d-3}\choose{2}}.\end{gathered}$$
As in Convention and Remark \[convention and remark\] (A), we put $W = S(1,1,1)$, and denote the coordinate rings of $X$ and $W$ by $A_X$ and $A_W$ respectively. Then there is a short exact sequence $$0 \to I_X/I_W \to A_W \to A_X \to 0$$ of graded $S$-modules, where $I_W$ denotes the defining ideal of $W \subset \mathbb{P}^5$. In a first step we compute the Hilbert series $$H(I_X/I_W,t) = \sum_{n \in \mathbb{Z}}\dim_{\Bbbk}[I_X/I_W]_n \cdot t^n$$ of $I_X/I_W$. By applying sheaf cohomology to the corresponding short exact sequence $0 \to \mathcal{I}_X/\mathcal{I}_W \to \mathcal{O}_W
\to \mathcal{O}_X \to 0$ we obtain an isomorphism $$[I_X/I_W]_n \cong H^0(W,\mathcal{O}_W(-X +nH))
\cong H^0(W,\mathcal{O}_W((n-1)H - (d-3)F)).$$ Therefore, it follows that $$\dim_{\Bbbk}[I_X/I_W]_n = \begin{cases}
\binom{n+1}{2} (n-d+3) & \text{ if } n \geq d-2 \\
\quad \quad 0 & \text{ if } n < d-2
\end{cases}$$ As an application to the Hilbert series $H(I_X/I_W,t)$ it turns out that $$H(I_X/I_W, t) = \sum_{n \geq d-2}\binom{n+1}{2}(n-d+3) t^n =
\frac{t^{d-2}}{(1-t)^4} \big( \binom{d-1}{2} -(d-1)(d-4)t +
\binom{d-3}{2}t^2 \big).$$ The formula for the expression of the generating function as a rational function might be proven directly or by some Computer Algebra System. The Hilbert series of $A_Y$ is given by $H(A_W,t) = (1+2t)/(1-t)^4$. By the above short exact sequence of graded modules it follows that $$H(A_X,t) = \frac{1}{(1-t)^4} \big( 1+2t - \binom{d-1}{2}t^{d-2}
+(d-1)(d-4)t^{d-1} - \binom{d-3}{2}t^d \big).$$ As a consequence of [@P Remark 4.8 (2)] the Betti diagram of $X$ now must have the shape as indicated in the statement. In particular the first row has the stated form. For the sake of simplicity, we put $\beta_{i,d-3} = \beta_i, i = 1,\ldots,5$. Then by the additivity of the Hilbert series on short exact sequences of graded $S$-modules the Betti diagram implies the following form of the Hilbert series $H(A_X,t)$ $$H(A_X,t) = \frac{1}{(1-t)^6} \big(
1-3t^2-\beta_1t^{d-2}+2t^3+\beta_2t^{d-1} -\beta_3 t^d+
\beta_4t^{d+1}-\beta_5 t^{d+2} \big).$$ By comparing both expressions for our Hilbert series we obtain the desired values for the remaining Betti numbers.
Cohomological properties {#cohomological-properties .unnumbered}
------------------------
We now provide a result which summarizes some cohomological properties of surfaces of type I. It is worth noticing that in this case the values for the sheaf cohomology of $\mathcal{O}_X$ and $\mathcal{I}_X$, and hence also the *index of normality* $$N(X) := \sup\{j\in \mathbb{N} \mid h^1(\mathbb{P}^r,\mathcal{I}_X(j)) \neq 0 \}\quad ( \in \mathbb{N} \cup \{-\infty \})$$ of $X$ are completely determined.
\[t1-coh\] With the previous notation there are the following equalities for the cohomology:
- $h^1(\mathbb{P}^5, \mathcal{I}_X(j)) = \binom{j+1}{2} (d-j-3)$ for $1 \leq j \leq d-4$ and zero else.
- $h^0(X,\mathcal{O}_X(j)) = 1/2(j+1)(dj+2)$ for all $j \geq 0$ and zero else.
- $h^1(X,\mathcal{O}_X(j)) = 0$ for all $j \in \mathbb{Z}$.
- $h^2(X, \mathcal{O}_X(-j)) = 1/2(j-1)(dj-2)$ for all $j \geq 2$ and zero else.
We start with the proof of (a), using the notations introduced in Convention and Remark \[convention and remark\]. Clearly $H^1(\mathbb{P}^5, \mathcal{I}_X(j)) = 0$ for all $j \leq 0$. So let $j \geq 1$. Then we use the short exact sequence $$0 \to \mathcal{I}_W \to \mathcal{I}_X \to \mathcal{O}_W(-X) \to 0.$$ Since $W$ is arithmetically Cohen-Macaulay, this sequence yields that $H^i(\mathbb{P}^5, \mathcal{I}_W(j)) = 0$ for $i = 1,2$ and all $j \in \mathbb{Z}$. Therefore the long exact cohomology sequence induces isomorphisms $H^1(\mathbb{P}^5,\mathcal{I}_X(j)) \cong
H^1(W,\mathcal{O}_W(jH -X))$ for all $j \in \mathbb{Z}$. Because of $X \approx H + (d-3)F$ it follows that $$h^1(\mathbb{P}^5, \mathcal{I}_X(j)) = h^1(W,\mathcal{O}_W((j-1)H- (d-3)F)
= \binom{j+1}{2} h^1(\mathbb{P}^1, \mathcal{O}_{\mathbb{P}^1}(j+2-d)).$$ By duality we get $h^1(\mathbb{P}^1, \mathcal{O}_{\mathbb{P}^1}(j+2-d)) =
h^0(\mathbb{P}^1, \mathcal{O}_{\mathbb{P}^1}(d-j-4))$. This proves statement (a).\
Because of $h^1(\mathbb{P}^5, \mathcal{I}_X(1)) = d - 4$ (as shown in (a)) and because of $\deg(X) = d$, the linearly normal embedding of $X$ implies that $X \subset \mathbb{P}^5$ is isomorphic to the linear projection of a smooth rational normal surface scroll $\tilde{X} \subset \mathbb{P}^{d+1}$. As a consequence we have $H^i(X,\mathcal{O}_X(j)) \cong
H^i(\tilde{X},\mathcal{O}_{\tilde{X}}(j))$ for all $i,j \in
\mathbb{Z}$. Since $\tilde{X}$ is arithmetically Cohen-Macaulay this yields statement in (c).\
The Hilbert function $$j \mapsto h_{A_{\tilde{X}}}(j) := \dim_{\Bbbk}
[A_{\tilde{X}}]_j = h^0(\tilde{X},\mathcal{O}_{\tilde{X}}(j))$$ of the coordinate ring $A_{\tilde{X}}$ is given by $1/2(j+1)(dj+2)$, and this proves statement (b).\
By interchanging $j$ and $-j$ this provides also the proof of the statement in (d).
The extremal secant locus {#the-extremal-secant-locus .unnumbered}
-------------------------
Now, we consider the (special) secant locus of a surface of type I. We first give the following auxiliary result, which shall be of use for us again later.
\[4.12” Lemma\] Let $s > 1$, let $\mathcal{C} \subset \mathbb{P}^s$ be a closed subscheme of dimension $1$ and degree $d$ and let $\mathbb{H} = \mathbb{P}^{s-1} \subset \mathbb{P}^s$ be a hyperplane. Then $$\mathrm{length}(\mathcal{C} \cap \mathbb{H}) \geq d \mbox{ with equality if and only if } {\rm Ass}_{\mathcal{C}}(\mathcal{O}_\mathcal{C}) \cap \mathbb{H} = \emptyset.$$
Let $R = \Bbbk\oplus R_1\oplus R_2\oplus \ldots = \Bbbk[R_1]$ be the homogeneous coordinate ring of $\mathcal{C}$ and let $f \in R_1$ be such that $\mathcal{C}\cap \mathbb{H} = {\operatorname{Proj}}(R/fR)$. Let $H_R(t) = dt + c$ be the Hilbert polynomial of $R$. Then, the two exact sequences $$0\rightarrow fR \rightarrow R \rightarrow R/fR\rightarrow 0 \mbox{ and } 0 \rightarrow(0:_R f)(-1) \rightarrow R(-1) \rightarrow fR \rightarrow 0$$ yield that the Hilbert polynomial of $R/fR$ is given by $$H_{R/fR}(t) = d + H_{(0:_R f)}(t-1).$$ Observe that the polynomial $H_{(0:_R f)}(t-1)$ vanishes if and only if $(0:_R f)_t = 0$ for all $t \gg 0$, hence if and only if $$f\notin \bigcup_{\mathfrak{p} \in {\rm Ass}(R)\setminus \{R_+\}} \mathfrak{p}.$$ But this latter condition is equivalent to the requirement that ${\rm Ass}(\mathcal{C}) \cap \mathbb{H} = \emptyset$.
\[prop:extseclocI\] Let $X \subset \mathbb{P}^5$ be a surface of maximal sectional regularity of degree $d > 5$ which is of type I. Then, in the notations of Convention and Remark \[convention and remark\] we have:
- $\mathbb{F}^+(X) = \mathbb{F}(X) = W = S(1,1,1).$
- ${}^*\Sigma(X) = \Sigma(X).$
- The image $\psi\big(\Sigma(X)\big)$ of $\Sigma(X)$ under the Plücker embedding $\psi: \mathbb{G}(1,\mathbb{P}^5) \rightarrow \mathbb{P}^{14}$ is a Veronese surface in a subspace $\mathbb{P}^5 \subset \mathbb{P}^{14}$.
Statement (a) is a restatement of Notation and Remark \[4.2” Notation and Reminder\] (C)(a).\
(b), (c): We identify $S(1,1,1) = W$ with the image of the Segre embedding $\sigma:\mathbb{P}^1 \times \mathbb{P}^2 \rightarrow \mathbb{P}^5$. Consider the canonical projection $$\varphi:\mathbb{P}^1 \times \mathbb{P}^2 \twoheadrightarrow \mathbb{P}^1 \mbox{ and its restriction } \varphi\upharpoonright:X \twoheadrightarrow \mathbb{P}^1.$$ Let $$\Theta := \{\mathbb{P}^1 \times \{q\} \mid q \in \mathbb{P}^2\} \subset \mathbb{G}(1,\mathbb{P}^5)$$ denote the closed subset of all fibers under the canonical projection $\mathbb{P}^1 \times \mathbb{P}^2 \twoheadrightarrow \mathbb{P}^2$, hence the set of all line sections of $\varphi.$\
Fix a closed point $p \in \mathbb{P}^1$. Then, the fiber $\varphi^{-1}(p) = \{p\}\times \mathbb{P}^2 =: \mathbb{P}_p^2$ is a ruling plane of $W$. As $X$ is smooth (see Corollary \[corollary 2.1’\]) and hence locally Cohen-Macaulay, the fiber $(\varphi\upharpoonright)^{-1}(p) = X \cap \varphi^{-1}(p) = X \cap \mathbb{P}_p^2$ is of pure dimension $1$ and has no closed associated points. Therefore $\mathrm{length}(X \cap \mathbb{L}) = \deg(X \cap \mathbb{P}^2_p)$ for all lines $\mathbb{L} \subset \mathbb{P}^2_p$ not contained in $X$ (see Lemma \[4.12” Lemma\]). Consequently, if $\mathbb{P}^2_p$ would contain a proper extremal secant line to $X$, the curve $X \cap \mathbb{P}_p^2 \subset \mathbb{P}^2_p$ would be pure and of degree $d-2$, so that $X$ would be of type II. This contradiction shows, that no proper extremal secant line to $X$ is contained in a ruling plane $\mathbb{P}^2_p$. Hence each proper secant line to $X$ must be a line section of $W$.\
As $\Theta \subset \mathbb{G}(1,\mathbb{P}^5)$ is closed, it follows that ${}^*\Sigma(X) \subseteq \Sigma(X) \subseteq \Theta,$ so that finally $\psi\big({}^*\Sigma(X)\big) \subseteq \psi\big(\Sigma(X)) \subseteq \psi(\Theta).$ Standard arguments on Plücker embeddings show that $\psi(\Theta)$ is the Veronese surface in some subspace $\mathbb{P}^5 \subset \mathbb{P}^{14}$. As ${}^*\Sigma(X)$ is of dimension $2$ (see Proposition \[prop:dimspecextrloc\]), as $\psi(\Theta)$ is irreducible and as $\psi$ is a closed embedding, statements (c) and (b) follow.
Surfaces of Type II
===================
The cohomological aspect {#the-cohomological-aspect .unnumbered}
------------------------
In this section, we investigate the surfaces of maximal sectional regularity which fall under type II. For the whole section we make the following convention.
\[convention and notation\] Let $5 \leq r < d$ and let $X \subset \mathbb{P}^r$ be a surface of degree $d$ and of maximal sectional regularity of type II which is not a cone. Set $Y := X \cup \mathbb{F}$, where $\mathbb{F} = \mathbb{F}(X) = \mathbb{P}^2$ denotes the extremal plane of $X$. Moreover, let $I$ and $L$ respectively denote the homogeneous vanishing ideal of $X$ and of $\mathbb{F}$ in $S = \Bbbk[x_0,x_1,\ldots,x_r]$.
\[4.14” Theorem\] Let the notations and hypotheses be as in Convention and Notation \[convention and notation\]. Then the following statements hold
- - $\mathrm{reg}(X) = d-r+3$ and $\mathrm{e}(X) \geq \binom{d-r+2}{2}.$
- $h^1(\mathbb{P}^r, \mathcal{I}_X(j)) = 0$ for all $j\leq 1.$ In particular $X$ is linearly normal.
- $h^2(\mathbb{P}^r,\mathcal{I}_X(j)) \begin{cases} = {\operatorname{e}}(X) &\mbox{ for all } j \leq 0;\\
= {\operatorname{e}}(X) + r - d -1 &\mbox{ for } j=1;\\
\leq \max\{0, h^2(\mathbb{P}^r,\mathcal{I}_X(j-1)) -1 \} &\mbox{ for all } j > 1;\\
= 1 &\mbox{ for } j = d-r;\\
= 0 &\mbox{ for all } j \geq d-r+1.
\end{cases}$
- $h^3(\mathbb{P}^r, \mathcal{I}_X(j))= \begin{cases} \frac{(j+1)(dj+2)}{2}, &\mbox{ if } j \leq - 2,\\
0, &\mbox{ if } j \geq -1. \end{cases}$
- - $\mathrm{reg}(Y) \leq d-r+3.$
- $H^1_{*}(\mathbb{P}^r, \mathcal{I}_Y) \cong H^1_{*}(\mathbb{P}^r,\mathcal{I}_X).$
- $h^2(\mathbb{P}^r,\mathcal{I}_Y(j)) = h^2(\mathbb{P}^r,\mathcal{I}_X(j)) - \max\{0,\binom{-j+d-r+2}{2}\}$ for all $j \geq 0$. In particular $h^2(\mathbb{P}^r,\mathcal{I}_Y(1)) = {\operatorname{e}}(X)-\binom{d-r+2}{2}$ and $h^2(\mathbb{P}^r,\mathcal{I}_Y(d-r)) = 0.$
- $h^2(\mathbb{P}^r,\mathcal{I}_Y(j)) \geq h^2(\mathbb{P}^r,\mathcal{I}_Y(j-1))$ for all $j \leq 1$, with equality for $j = 1$.
- $h^2(\mathbb{P}^r,\mathcal{I}_Y(j)) \leq {\rm max}\{0,h^2(\mathbb{P}^r,\mathcal{I}_Y(j-1)) - 1\}$ for all $j > 1$.
- $h^3(\mathbb{P}^r,\mathcal{I}_Y(j)) = 0$ for all $j \geq 0.$
- If $h^2(\mathbb{P}^r, \mathcal{I}_Y) = 0$, then $h^2(\mathbb{P}^r,\mathcal{I}_Y(j)) = 0$ for all $j \in \mathbb{Z}$.
- For the pair $\tau(X) := \big({\operatorname{depth}}(X),{\operatorname{depth}}(Y)\big)$ we have
- $\tau(X) = (2,3)$ if $r+1 \leq d \leq 2r-4$;
- $\tau(X) \in \{(1,1),(2,2),(2,3)\}$ if $2r-3 \leq d \leq 3r-7$;
- $\tau(X) \in \{(1,1),(2,2)\}$ if $3r-6 \leq d$.
- $h^0(\mathbb{P}^r,\mathcal{I}_X(2))\geq \binom{r}{2}-d-1$ with equality if and only if $\tau(X) = (2,3).$
As an immediate application, we get the following information on the *cohomology tables* $$\big(h^i(\mathbb{P}^r,\mathcal{I}_X(j))\big)_{i =1,2,3 \mbox{ and } j \in \mathbb{Z}}, \quad \quad \big(h^i(\mathbb{P}^r,\mathcal{I}_Y(j))\big)_{i =1,2,3 \mbox{ and } j \in \mathbb{Z}}$$ of the sheaves of vanishing ideals $\mathcal{I}_X, \mathcal{I}_Y \subset \mathcal{O}_{\mathbb{P}°r}$ of $X$ and $Y$.
\[coro:cohomology\] Let $X$ and $Y$ be as above. Then the ideal sheaves $\mathcal{I}_X, \mathcal{I}_Y \subset \mathcal{O}_{\mathbb{P}°r}$ of $X$ and $Y$ have the following cohomology tables: $$\begin{tabular}{| c | c | c | c | c | c | c |c | c | c | c | c |c |}
\hline
$j$ & $\cdots$ & $-2$
& $-1$ & $0$ & $1$ & $2$ &$\cdots$ &$\varkappa-1$ &$\varkappa$ &$\varkappa+1$ &$\varkappa+2$ & $\cdots$ \\ \hline
$h^1(\mathbb{P}^r,\mathcal{I}_X(j))$ &$\cdots$ &$0$ &$0$ &$0$ &$0$ &$\ast$ &$\cdots$ &$\ast$ &$\ast$ &$\ast$ &$0$ & $\cdots$\\ \hline
$h^2(\mathbb{P}^r,\mathcal{I}_X(j))$ &$\cdots$ &${\operatorname{e}}$ &${\operatorname{e}}$ &${\operatorname{e}}$ &${\operatorname{e}}-\varkappa-1$ &$\ast$ &$\cdots$ &$\ast$ &$1$ &$0$ &$0$ & $\cdots$\\ \hline
$h^3(\mathbb{P}^r,\mathcal{I}_X(j))$ &$\cdots$ &$\ast$ &$0$ &$0$ &$0$ &$0$ &$\cdots$ &$0$ &$0$ &$0$ &$0$ & $\cdots$\\ \hline
\end{tabular}$$ and $$\begin{tabular}{| c | c | c | c | c | c | c |c | c | c | c | c |c |}
\hline
$j$ & $\cdots$ & $-2$
& $-1$ & $0$ & $1$ & $2$ &$\cdots$ &$\varkappa-1$ &$\varkappa$ &$\varkappa+1$ &$\varkappa+2$ & $\cdots$ \\ \hline
$h^1(\mathbb{P}^r,\mathcal{I}_Y (j))$ &$\cdots$ &$0$ &$0$ &$0$ &$0$ &$\ast$ &$\cdots$ &$\ast$ &$\ast$ &$\ast$ &$0$ & $\cdots$\\ \hline
$h^2(\mathbb{P}^r,\mathcal{I}_Y (j))$ &$\cdots$ &${\operatorname{e}}$ &${\operatorname{e}}$ &${\operatorname{e}}$ &${\operatorname{e}}-\binom{d-r+2}{2}$ &$\ast$ &$\cdots$ &$\ast$ &$0$ &$0$ &$0$ & $\cdots$\\ \hline
$h^3(\mathbb{P}^r,\mathcal{I}_Y (j))$ &$\cdots$ &$\ast$ &$\ast$ &$0$ &$0$ &$0$ &$\cdots$ &$0$ &$0$ &$0$ &$0$ & $\cdots$\\ \hline
\end{tabular}$$ where $\varkappa := d-r, \mathrm{e} :=\mathrm{e}(X)$ and $\ast$ stands for non-specified non-negative integers.
An auxiliary result {#an-auxiliary-result .unnumbered}
-------------------
Before we establish Theorem \[4.14” Theorem\] we prove the following Lemma.
\[4.12” Lemma+\] Let the notations and hypotheses be as in Convention and Notation \[convention and notation\]. In addition let $\mathcal{C} := X \cap \mathbb{F}.$ Then the following statements hold.
- Each line $\mathbb{L} \subset \mathbb{F}$ which is not contained in $X$, satisfies $$\mathrm{length}(\mathcal{C}\cap \mathbb{L}) = \mathrm{length}(X\cap\mathbb{L}) = d-r+3.$$ In particular, $\mathcal{C} \subset \mathbb{F}$ is a curve of degree $d-r+2$ and has no closed associated points.
- $I_{d-r+3} \setminus I \cap L \neq \emptyset$ and for each $f \in I_{d-r+3} \setminus I \cap L$ it holds $I = (I \cap L, f)$
(a): First let $h \in \mathbb{U}(X)$. Then $\mathbb{L}_h \subset \mathbb{F}$ and $$\mathrm{length}(\mathcal{C}\cap \mathbb{L}_h) = \mathrm{length}(X \cap \mathbb{L}_h) = \mathrm{length}(\mathcal{C}_h \cap \mathbb{L}_h) = d-r+3 .$$ This shows, that $\mathcal{C} \subset \mathbb{F}$ is a closed subscheme of dimension $1$ and degree $d-r+3$. Now, let $\mathbb{L} \subset
\mathbb{F}$ be an arbitrary line which is not contained in $X$. As $\mathcal{C} \subset \mathbb{F}$ is of dimension $1$ and of degree $d-r+3$ we have $\mathrm{length}(X \cap \mathbb{L}) =
\mathrm{length}(C \cap \mathbb{L}) \geq d-r+3$. As ${\operatorname{reg}}(X) \leq d-r+3$ (see Theorem \[theorem 2.9\] (a)) we also have $\mathrm{length}(X \cap \mathbb{L}) \leq d-r+3$, so that indeed $\mathrm{length}(X \cap \mathbb{L}) = d-r+3$. Now, it follows by Lemma \[4.12” Lemma\] that $\mathcal{C}$ has no closed associated point.\
(b): According to statement (a), there is a homogeneous polynomial $g \in S_{d-r+3} \setminus L$ such that the homogeneous vanishing ideal $(I + L)^{{\operatorname{sat}}} \subset S$ of $\mathcal{C}$ in $S$ can be written as $(L,g)$. In particular we have $I_{\leq d-r+2} \subset L$. As ${\operatorname{reg}}(X) = d-r+3$, the ideal $I \subset S$ is generated by homogeneous polynomials of degree $\leq d-r+3$. As $g \notin L$ it follows that $I_{d-r+3}$ is not contained in $L$ and hence that $I+L = (L,f)$ for all $f \in I_{d-r+3} \setminus
I \cap L$. Therefore $ I = I\cap(I+L) = I\cap (L,f) = (I\cap L,f)$ for all such $f$.
Proof of Theorem \[4.14” Theorem\]. {#proof-of-theorem-4.14-theorem. .unnumbered}
-----------------------------------
(a)(1): Since $X$ admits $(d-r+3)$-secant lines, we have ${\operatorname{reg}}(X) \geq d-r+3$. On the other hand, ${\operatorname{reg}}(X) \leq d-r+3$ by Theorem \[theorem 2.9\] (a). This proves the stated equality for the regularity. For the moment, we postpone the proof of the stated estimate for the invariant $\mathrm{e}(X)$.\
(a)(2): It is obvious that $h^1(\mathbb{P}^r, \mathcal{I}_X(j)) = 0$ for all $j \leq 0$. Assume $h^1(\mathbb{P}^r, \mathcal{I}_X(1))> 0$. Then $X$ is a regular projection of a surface $X' \subset \mathbb{P}^{r+1}$. Note that $X'$ is again a sectionally smooth rational surface and hence ${\operatorname{reg}}(X') \leq d - r+2$ by Theorem \[theorem 2.9\](a). On the other hand, the preimage $\mathcal{C}'$ of $\mathcal{C} = X \cap \mathbb{F}$ under this regular projection is a plane curve of degree $(d-r+3)$, and hence ${\operatorname{reg}}(X') \geq d - r+3$. This contradiction proves our claim.\
(a)(4): See Theorem \[theorem 2.9\](e)\
(b)(4) and (b)(6): Let $h \in \mathbb{U}(X)$ and consider the induced exact sequence $$0 \rightarrow \mathcal{I}_Y (-1) \rightarrow \mathcal{I}_Y \rightarrow \mathcal{I}_{\mathcal{C}_h \cup \mathbb{L}_h } \rightarrow 0$$ Keep in mind that $H^1 (\mathbb{P}^r , \mathcal{I}_{\mathcal{C}_h \cup \mathbb{L}_h } (j))=0$ for all $j \leq 1$ and $H^2 (\mathbb{P}^r , \mathcal{I}_{\mathcal{C}_h \cup \mathbb{L}_h } (j))=0$ for all $j \geq 1$ by [@BS2 Proposition 2.7(c),(d)]. Both claims now follow easily.\
(b)(5): Assume again that $h \in \mathbb{U}(X)$ and keep in mind that $S/(I\cap L ,h)^{\rm sat}$ is the homogeneous coordinate ring of $\mathcal{C}_h \cup \mathbb{L}_h$ in $S$. By [@BS2 Remark 3.2 B)] the graded $S$-module $$H^1(S/(I\cap L,h)^{\rm sat}) = \bigoplus_{j \in\mathbb{Z}}
H^1(\mathbb{P}^r,\mathcal{I}_{\mathcal{C}_h \cup \mathbb{L}_h}(j))$$ is generated by homogeneous elements of degree $2$. Now, the induced exact sequences of local cohomology modules $$H^1(S/(I \cap L, h)^{\rm sat}) \longrightarrow H^2(S/I\cap L)(-1)
\stackrel{\times h}{\longrightarrow} H^2(S/I\cap L) \longrightarrow
H^2(S/(I \cap L, h)^{\rm sat})$$ proves claim (b)(5), since the multiplication map $\cdot h$ is an epimorphism in all positive degrees and its kernel is generated by homogeneous elements of degree $2$.\
(b)(7): This is an immediate consequence of (b)(4) and (b)(5).\
(b)(2): Keep in mind that $$H^i_{*}(\mathbb{P}^r,\mathcal{I}_X) \cong H^i(S/I) \mbox { and } H^i_{*}(\mathbb{P}^r,\mathcal{I}_Y) \cong H^i(S/I\cap L) \mbox{ for } i = 1,2,3.$$ According to Lemma \[4.12” Lemma+\](b) we have an exact sequence $$0 \rightarrow (S/L)(-d+r-3) \rightarrow S/I \cap L \rightarrow S/I \rightarrow 0.$$ Claim (b)(2) now follows immediately, since $H^i((S/L)(-d+r-3))$ vanishes for $i=1,2$.\
(a)(3): For all $j \neq d-r$ the values of $h^2(\mathbb{P}^r,\mathcal{I}_X(j))$ are as requested by statement (a)(2) and by Theorem \[theorem 2.9\] (d). Observe that $h^2 (\mathbb{P}^r , \mathcal{I}_X (d-r)) \leq 1$ by Theorem \[theorem 2.9\] (d). So, it remains to show that $H^2 (\mathbb{P}^r , \mathcal{I}_X (d-r)) \neq 0$. This follows from the exact sequence $$H^2(S/I)_{d-r} \longrightarrow H^3(S/L)_{-3} \cong \Bbbk \longrightarrow H^3(S/I\cap L)_{d-r} = 0.$$ (b)(3):The first part of this claim follows immediately by (b)(6) and the exact sequence used in the proof of (b)(2). Now, the second part of (b)(3) comes immediately from (a)(3).\
(b)(1): The required vanishing conditions $$h^i (\mathbb{P}^r , \mathcal{I}_Y (d-r+3-i+k))=0 \mbox{ for } i=1,2,3 \mbox{ and all } k \geq 0$$ are obtained respectively by (a)(1), (b)(2), (b)(3) and (b)(6).\
Finally, the inequality for $\mathrm{e}(X)$ stated in (a)(1), follows from (b)(3) applied with $j=1.$\
(c): By (a)(3), we know that ${\operatorname{depth}}(X) \leq 2$. Also (b)(2) implies that, if ${\operatorname{depth}}(X) =1$ then ${\operatorname{depth}}(Y) =1$ and, if ${\operatorname{depth}}(X) =2$ then ${\operatorname{depth}}(Y) =2$ or $3$. Thus we need only to show (c)(1) and (c)(3). If $d \leq 2r-4= 2(r-1) -2$, then $\mathcal{C}_h \cup \mathbb{L}_h$ is arithmetically Cohen-Macaulay by [@BS2 Proposition 3.5]. Thus we have ${\operatorname{depth}}(Y) =3$ and hence ${\operatorname{depth}}(X) =2$. On the other hand, if $d \geq 3r-6$, then ${\operatorname{depth}}(\mathcal{C}_h \cup \mathbb{L}_h)=1$ by Proposition \[proposition 2.3\] and hence ${\operatorname{depth}}(Y) \leq 2$. Therefore either $\tau(X)=(1,1)$ or $\tau (X) = (2,2)$.\
(d): Since $X$ is linearly normal we have $h^0 (\mathbb{P}^r , \mathcal{I}_X (2)) = h^0 (\mathbb{P}^{r-1} , \mathcal{I}_{\mathcal{C}_h} (2))$. Moreover, by [@BS2 Proposition 3.6], we have $$h^0 (\mathbb{P}^{r-1} , \mathcal{I}_{\mathcal{C}_h} (2)) \geq {{r}\choose{2}}-d-1$$ where equality holds if and only if $\mathcal{C}_h \cup \mathbb{L}_h$ is arithmetically Cohen-Macaulay, or equivalently, if and only if ${\operatorname{depth}}(Y) = 3$. Finally, we know by $(c)$ that ${\operatorname{depth}}(Y) = 3$ if and only if $\tau(X)=(2,3)$. This completes the proof.
Simplicity of the socle of the second cohomology {#simplicity-of-the-socle-of-the-second-cohomology .unnumbered}
------------------------------------------------
As a first application and extension of Theorem \[4.14” Theorem\] we show that (in the previous notation) the vanishing condition $h^2(\mathbb{P}^r,\mathcal{I}_Y) = 0$ which occurs in statement (b)(7) of that Theorem is equivalent to the simplicity of the socle of the second total cohomology module $H^2_{*}(\mathbb{P}^r,\mathcal{I}_X) = \bigoplus_{j\in \mathbb{Z}}H^2(\mathbb{P}^r, \mathcal{I}_X(j))$ of $\mathcal{I}_X$. To formulate our result, we recall the following notation.
\[4.16”’ Notation and Reminder\] Let $T = \bigoplus _{n \in \mathbb{Z}} T_n$ be a graded $S$-module. Then, we denote the *socle* of $T$ by ${\rm Soc}(T)$, thus: $${\rm Soc}(T) := (0:_T S_+) \cong {\operatorname{Hom}}_S(\Bbbk,T) = {\operatorname{Hom}}_S(S/S_+,T).$$ Keep in mind that the socle of a graded Artinian $S$-module $T$ is a $\Bbbk$-vector space of finite dimension which vanishes if and only if $T$ does.
\[proposition:soc.eq\] Let the notations and hypotheses be as in Convention and Notation \[convention and notation\]. Then following statements are equivalent:
- $\mathrm{e}(X)$ takes its minimally possible value $\binom{d-r+2}{2}.$
- $h^2(\mathbb{P}^r,\mathcal{I}_Y) = 0.$
- $h^2(\mathbb{P}^r,\mathcal{I}_Y(j)) = 0$ for all $j \in \{0,1,2,\ldots,d-r-1\}.$
- $h^2(\mathbb{P}^r,\mathcal{I}_Y(j)) = 0$ for all $j \in \mathbb{Z}.$
- $h^2(\mathbb{P}^r,\mathcal{I}_X(j)) = \binom{-j+d-r+2}{2}$ for all $j \in \{0,1,2,\ldots,d-r\}.$
- $h^2(\mathbb{P}^r,\mathcal{I}_X(j)) = \mathrm{max}\{0, \binom{-j+d-r+2}{2}\}$ for all $j \geq 0.$
- $\mathrm{dim}_{\Bbbk}\big(\mathrm{Soc}(H^2_{*}(\mathbb{P}^r,\mathcal{I}_X))\big) = 1.$
Indeed, the equivalences (i) $\Leftrightarrow$ (ii) $\Leftrightarrow$ (iii) $\Leftrightarrow$ (iv) $\Leftrightarrow$ (v) $\Leftrightarrow$ (vi) follow by Theorem \[4.14” Theorem\]. It remains to show the equivalence (iv) $\Leftrightarrow$ (vii). Consider the exact sequence of graded $S$-modules $$0 \longrightarrow H^2_{*}(\mathbb{P}^r, \mathcal{I}_Y) \longrightarrow H^2_{*}(\mathbb{P}^r, \mathcal{I}_Y) \longrightarrow H^3_{*}(\mathbb{P}^r,\mathcal{I}_{\mathbb{F}}(-d+r-3)).$$ As $$\begin{aligned}
\mathrm{Soc}\big(H^3_{*}(\mathbb{P}^r,\mathcal{I}_{\mathbb{F}}(-d+r-3))\big) &= \Bbbk(d-r),\\
h^2(\mathbb{P}^r, \mathcal{I}_Y(j)) &= 0 \mbox{ for all } j \geq d-r, \mbox{ and }\\
h^2(\mathbb{P}^r, \mathcal{I}_X(j)) &= 0 \mbox{ for all } j \geq d-r+1\end{aligned}$$ we get an isomorphism of graded $S$-modules $$\mathrm{Soc}\big(H^2_{*}(\mathbb{P}^r, \mathcal{I}_Y)\big) \cong \mathrm{Soc}\big(H^2_{*}(\mathbb{P}^r, \mathcal{I}_X)\big)_{\leq d-r-1}.$$ From this isomorphism, we see that $$H^2_{*}(\mathbb{P}^r, \mathcal{I}_Y) = 0 \mbox{ if and only if } \mathrm{Soc}\big(H^2_{*}(\mathbb{P}^r, \mathcal{I}_X)\big)_{\leq d-r-1} = 0.$$ By Theorem \[4.14” Theorem\] (b)(7) the module $H^2_{*}(\mathbb{P}^r, \mathcal{I}_Y)$ vanishes if and only if the number $h^2(\mathbb{P}^r, \mathcal{I}_Y)$ does. So, condition (iv) holds, if and only if $\mathrm{Soc}\big(H^2_{*}(\mathbb{P}^r, \mathcal{I}_X)\big)$ is concentrated in degrees $\geq d-r$. By Theorem \[4.14” Theorem\] (a)(4) this is the case if and only if condition (vii) holds.
\[remark.eq.cond\] (A) If the above equivalent conditions (i) – (vii) hold, we must have $$\tau(X) \in \{(1,1), (2,3)\}.$$ Moreover $\tau(X) = (2,3)$ implies the above equivalent conditions (i) – (vii).\
(B) Observe, that the above minimality condition (i) describes a *generic situation*. So, it is noteworthy that the simplicity of the socle of $H^2_{*}(\mathbb{P}^r,\mathcal{I}_X)$ occurs in the generic situation, too. Below, we shall see, that in such a generic situation, a number of additional conclusions may be drawn.
\[corollary:coh.gen.\] Let the notations and hypotheses be as in Convention and Notation \[convention and notation\]. Then, the following statements hold:
- $h^1(\mathbb{P}^r, \mathcal{I}_X(2)) \leq h^0(\mathbb{P}^r, \mathcal{I}_X(2))-\binom{r}{2} +d+1.$
- If the equivalent conditions (i)-(vii) of Proposition \[proposition:soc.eq\] hold, the $S$-module $H^1_{*}(\mathbb{P}^r, \mathcal{I}_X)$ is minimally generated by $h^0(\mathbb{P}^r, \mathcal{I}_X(2))-\binom{r}{2} +d+1$ homogeneous elements of degree $2$.
In view of Theorem \[4.14” Theorem\](b)(2) we may replace $X$ by $Y$. We choose $h \in \mathbb{U}(X)$ and apply cohomolgy to the induced exact sequence of sheaves $$0 \rightarrow \mathcal{I}_Y (-1) \stackrel{h}{ \rightarrow} \mathcal{I}_Y \rightarrow \mathcal{I}_{\mathcal{C}_h \cup \mathbb{L}_h } \rightarrow 0$$ in order to end up with an exact secuence of graded $S$-modules $$0 \rightarrow H^1_{*}(\mathbb{P}^r, \mathcal{I}_Y)/hH^1_{*}(\mathbb{P}^r, \mathcal{I}_Y)\rightarrow H^1_{*}(\mathbb{H}_h, \mathcal{I}_{\mathcal{C}_h \cup \mathbb{L}_h }) \rightarrow H^2_{*}(\mathbb{P}^r, \mathcal{I}_Y)$$ (a): By Proposition 3.6 of [@BS2], the $S$-module $H^1_{*}(\mathbb{H}_h, \mathcal{I}_{\mathcal{C}_h \cup \mathbb{L}_h })$ is minimally generated by $$h^0(\mathbb{H}_h, \mathcal{I}_{\mathcal{C}_h}(2))-\binom{r}{2} +d+1$$ homogeneous elements of degree $2$. As $X$ is linearly normal, we have $ h^0(\mathbb{H}_h, \mathcal{I}_{\mathcal{C}_h}(2)) = h^0(\mathbb{P}^r, \mathcal{I}_X(2))$. Now, our claim follows immediately.\
(b): By our hypothesis, the third module in the above sequence vanishes. Now, we get our claim by Nakayama.
\[corollary:Hilb.funct.\] Let the notations and hypotheses be as in Convention and Notation \[convention and notation\]. Then, the following statements hold:
- For all $j \in \mathbb{N}_0$ we have $$h^0(X,\mathcal{O}_X(j)) = d\binom{j+1}{2}+j+1 +h^2(\mathbb{P}^r, \mathcal{I}_X(j))- {\operatorname{e}}(X).$$
- If the equivalent conditions (i)-(vii) of Proposition \[proposition:soc.eq\] hold, then $$h^0(X,\mathcal{O}_X(j)) = \begin{cases}d\binom{j+1}{2}+j+1 + \binom{d-r+2-j}{2}-\binom{d-r+2}{2} &\mbox{ for } 0 \leq j \leq d-r, \\ \\
d\binom{j+1}{2}+j+1 -\binom{d-r+2}{2} &\mbox{ for } d-r < j.\end{cases}$$
(a): Once more, let $B$ be the homogeneous coordinate ring of the projecting scroll $\widetilde{X} \subset \mathbb{P}^{d+1}$ of $X$, let $D := D_{A_+}(A) = \Gamma_{*}(X,\mathcal{O}_X)$ be the $A_+$-transform of $A$ and consider the short exaxt sequence $0 \rightarrow D \rightarrow B \rightarrow C \rightarrow 0$ in which $C$ is a one-dimensional Cohen-Macaulay module of dimension $1$ with $$\dim_\Bbbk(C_j) = {\operatorname{e}}(X) - h^2(\mathbb{P}^r, \mathcal{I}_X(j)) \mbox{ for all } j \in \mathbb{Z}.$$ As $$\dim_\Bbbk(B_j)= \chi(\mathcal{O}_{\widetilde{X}}(j)) = d\binom{j+1}{2}+j+1 \mbox{ for all } j \in \mathbb{N}_0,$$ we get our claim.\
(b): This follows immediately from statement (a) bearing in mind the values of ${\operatorname{e}}(X)$ and of $h^2(\mathbb{P}^r, \mathcal{I}_X(j))$ imposed by the conditions (i) and (vi) of Proposition \[proposition:soc.eq\].
The second deficiency module {#the-second-deficiency-module .unnumbered}
----------------------------
We first remind the notion of deficiency module.
\[reminder.def.mod\] Let $A$ be the homogeneous coordinate ring of the surface $X \subset \mathbb{P}^r$, let $M$ be a finitely generated graded $A$-module and let $i \in \mathbb{N}_0$. Then, the $i$-th *deficiency module* of $M$ is defined by $$K^i(M) := \mathrm{Ext}^{r-i+1}_S(M,S(-r-1)) \cong H^i(M)^{\vee},$$ where $\bullet ^{\vee} := {}^*\mathrm{Hom}_{\Bbbk}(\bullet, \Bbbk)$ denotes the (contravariant exact) graded Matlis duality functor. The module $K^{\dim(M)}(M)$ is called the *canonical module* of $M$.
In this subsection, we are interested in the second deficiency module $$K^2(A) = \mathrm{Ext}^{r-1}_S(A,S(-r-1)) = H^2(A)^{\vee}$$ of the coordinate ring $A$ of $X$ and its induced sheaf $$\mathcal{K}^2_X := \widetilde{K^2(A)}.$$
\[proposition:def.mod\] Let the notations and hypotheses be as in Convention and Notation \[convention and notation\]. Let $\pi = \pi_\Lambda:\widetilde{X} \twoheadrightarrow X$ be the standard normalization of $X$. Let $B$ be the homogeneous coordinate ring of the projecting scroll $\widetilde{X} \subset \mathbb{P}^{d+1}$, let $D := D_{A_+}(A) = \Gamma_{*}(X,\mathcal{O}_X) (\subseteq B)$ and set $\mathcal{F} := \pi_{*}(\mathcal{O}_{\widetilde{X}})/\mathcal{O}_X = \widetilde{B/D}.$ Then, the following statements hold:
- $K^2(A) \cong K^1(B/D)$ is a Cohen-Macaulay module of dimension one, and
- $\mathrm{reg}(K^2(A)) = 0;$
- $\mathcal{K}^2_X \cong \widetilde{\Gamma_{*}(X,\mathcal{F})^{\vee}}.$
- For each closed point $x \in X$, the stalk $\mathcal{K}^2_{X,x}$ of $\mathcal{K}^2_X$ at $x$ coincides with the first deficiency module $K^1(\mathcal{O}_{X,x}) = \mathrm{Ext}^{r-1}_{\mathcal{O}_{X,x}}(\mathcal{O}_{X,x},\mathcal{O}_{\mathbb{P}^r,x})$ of the local ring $\mathcal{O}_{X,x}$ of $X$ at $x$. In particular
- $\mathrm{length}_{\mathcal{O}_{X,x}}(\mathcal{K}^2_{X,x}) = \mathrm{e}_x(X)$ for all closed points $x \in X;$
- $\mathrm{Supp}(\mathcal{K}^2_X) = \mathrm{Sing}(X);$
- $\mathrm{length}_{\mathcal{O}_X}(\mathcal{K}^2_X) = \mathrm{e}(X).$
- If the equivalent conditions of Proposition \[proposition:soc.eq\] hold, then $K^2(A) \cong S/J(d-r)$, where $J \subset S$ is a saturated graded ideal such that:
- $I+L \subset J;$
- ${\operatorname{reg}}(J) = d-r+1;$
- $J$ is minimally generated by $L$ and $d-r+2$ forms of degree $d-r+1.$
- If the equivalent conditions of Proposition \[proposition:soc.eq\] do not hold, then the $S$-module $K^2(A)$ is minimally generated by one element of degree $r-d$ and some additional elements of degrees $> r-d.$
(a): Let $C := B/D.$ As seen in the proof of Theorem \[theorem 2.9\] (d) we have $$H^2(A) \cong H^2(D) \cong H^1(C), \quad \widetilde{D_{A_+}(C)} \cong \widetilde{C} \cong \mathcal{F} \mbox{ and } D_{A_+}(C) \cong \Gamma_{*}(X,\mathcal{F}).$$ Applying the functor $\bullet^\vee$ to the first isomorphism, we get $K^2(A) \cong K^1(C) = K^1(B/A)$. As $C$ is a Cohen-Macaulay module of dimension $1$, so is its canonical module $K^1(C)$.\
To prove claim (1), keep in mind that $K^2(A)$ is a Cohen-Macaulay module of dimension $1$, so that indeed $$\begin{aligned}
{\operatorname{reg}}(K^2(A)) &= \mathrm{end}\big(H_1(K^2(A))\big) +1 = \sup\{j \in \mathbb{Z} \mid \mathrm{dim}_{\Bbbk}(K^2(A)_{j-1}) < \mathrm{dim}_{\Bbbk}(K^2(A)_j )\}+1 \\
& = \sup\{j \in \mathbb{Z} \mid \mathrm{dim}_{\Bbbk}(H^2(A)_{-j+1}) < \mathrm{dim}_{\Bbbk}(H^2(A)_{-j})\}+1 = -1+1 = 0.\end{aligned}$$ To prove claim (2), we apply the Matlis duality functor $\bullet^\vee$ to the short exact sequence $0\rightarrow C
\rightarrow D_{A_+}(C) \rightarrow H^2(A) \rightarrow 0$ and get an exact sequence of graded $A$-modules $0 \longrightarrow K^2(A)
\longrightarrow D_{A_+}(C)^\vee \longrightarrow C^\vee
\longrightarrow 0$. As $\widetilde{C^\vee} = 0$, we obtain $\mathcal{K}^2_X \cong \widetilde{\Gamma_{*}(X,\mathcal{F})^{\vee}}.$\
(b): The standard isomorphism $\widetilde{\mathrm{Ext}^{r-1}_S(A,S(-r-1))}_x \cong \mathrm{Ext}^{r-1}_{\mathcal{O}_{X,x}}(\mathcal{O}_{X,x},\mathcal{O}_{\mathbb{P}^r,x})$ implies the requested isomorphism $\mathcal{K}^2_{X,x} \cong K^1(\mathcal{O}_{X,x}).$ As taking local Matlis duals preserve lengths, local duality implies that $$\mathrm{e}_x(X) = \mathrm{length}\big(H^1_{\mathfrak{m}_{X,x}}(\mathcal{O}_{X,x})\big) = \mathrm{length}\big(K^1(\mathcal{O}_{X,x})\big),$$ and this implies claim (1). Now, claims (2) and (3) are immediate.\
(c): As $H^2(A) \cong H^2_{*}(\mathbb{P}^r,\mathcal{I}_X )$ our hypotheses imply that $\mathrm{Soc}(H^2(A)) = H^2(A)_{d-r}\cong
\Bbbk,$ and hence the standard isomorphism $\mathrm{Soc}\big(H^2(A)\big)^\vee \cong H^2(A)^\vee \otimes_A
\Bbbk$ shows that $K^2(A)$ is generated by one single element of degree $r-d$. Therefore $K^2(A) \cong S/J(d-r)$ for some homogeneous ideal $J \subset S$. By statement (a), this ideal $J$ is saturated. If we apply cohomology to the exact sequence $0\rightarrow S/I\cap L
\rightarrow S/I \oplus S/L \rightarrow S/(I+L) \rightarrow 0$ and keep in mind, that $H^2(S/I\cap L) = H^2_{*}(\mathbb{P}^r,\mathcal{I}_{X \cup \mathbb{F}(X)}) = 0$, we get a monomorphism $H^2(A) \rightarrow H^2(S/(I+L))$. Therefore $(I+L)K^2(A) = 0$ and hence $I+L \subseteq J$. As $K^2(A)$ is of dimension $1$, the inclusion is strict. This proves claim (1).\
According to Statement (a)(1) we have ${\operatorname{reg}}(J) = {\operatorname{reg}}(S/J)+1 = {\operatorname{reg}}(K^2(A)(r-d))+1 = {\operatorname{reg}}(K^2(A)) + d-r+1 = d-r+1$, and this proves claim (2).\
Moreover, in view of condition (v) of Proposition \[proposition:soc.eq\] we have $$\dim_\Bbbk(S/J)_k = \dim_\Bbbk\big(H^2(A)_{r-d+k}\big) = \binom{\min\{k,d-r\}+2}{2} \mbox{ for all } k \in \mathbb{N}_0,$$ so that $$\dim_\Bbbk\big((S/J)_k\big) = \dim_\Bbbk\big((S/L)_k\big) \mbox{ for all } k \leq d-r \mbox{ and }$$ $$\dim_\Bbbk\big((S/J)_{d-r+1} \big) = \dim_\Bbbk\big((S/L)_{d-r+1}\big) - (d-r+2).$$ In view of claims (1) and (2) this proves claim (3).\
(d): If the equivalent conditions of Proposition \[proposition:soc.eq\] do not hold, then $\dim_\Bbbk\big(\mathrm{Soc}(H^2(A))_j\big)$ vanishes for all $j > d-r$, takes the value $1$ for $j=d-r$ and does not vanish for some $j < d-r$ (see Theorem \[4.14” Theorem\] (a)(3)). Now, we get our claim by the isomorphism $\mathrm{Soc}\big(H^2(A)\big)^\vee \cong H^2(A)^\vee \otimes_A \Bbbk.$
\[Remark:def.mod\] Observe that the previous proposition generalizes Theorem 3.6 (e) of [@BS6].
An application in higher dimensions {#an-application-in-higher-dimensions .unnumbered}
-----------------------------------
We now draw a conclusion for higher dimensional varieties of maximal sectional regularity.
\[cor:higherdim\] Let $n \geq 2$ and let $X \subset \mathbb{P}^r$ be an $n$-dimensional variety of maximal sectional regularity of degree $d$ and of type II. Then $X$ is linearly normal and we have $$h^0 (\mathbb{P}^r ,\mathcal{I}_X (2)) \geq {{r-n+1}\choose{2}}-d-1$$ with equality if and only if $X \cup \mathbb{F}(X)$ is arithmetically Cohen-Macaulay. Moreover, if equality is attained, then ${\operatorname{depth}}(X)=n$.
For $n = 2,$ our claim follows by Theorem \[4.14” Theorem\] (a)(2) and (d).\
So let $n > 2$. Note that a general hyperplane section $X' = X \cap \mathbb{H} \subset \mathbb{H}, (\mathbb{H} =\mathbb{P}^{r-1})$ of $X$ is again a variety of maximal sectional regularity and of degree $d$ of type II. So, by induction and on use of the exact sequence $$0 \longrightarrow \mathcal{I}_X (-1) \longrightarrow \mathcal{I}_X \longrightarrow \mathcal{I}_{X'} \longrightarrow 0$$ it first follows that $h^1(\mathbb{P}^r,\mathcal{I}_X (1)) = 0$ and then that $h^0 (\mathbb{P}^r ,\mathcal{I}_X (2)) = h^0 (\mathbb{H} ,\mathcal{I}_{X'}(2)) \geq {{r-n+1}\choose{2}}-d-1.$ So, $X$ is linearly normal and satisfies the requested inequality.\
As $\mathbb{F}(X') = \mathbb{F}(X \cap \mathbb{H})= \mathbb{F}(X) \cap \mathbb{H}$ (see [@BLPS1 Lemma 5.1 (a)]), we have $\big(X \cup \mathbb{F}(X)\big)\cap \mathbb{H} = X' \cup \mathbb{F}(X')$. So, $\big(X \cup \mathbb{F}(X)\big)\cap \mathbb{H}$ is arithmetically Cohen-Macaulay if and only $X' \cup \mathbb{F}(X')$ is. Therefore, again by induction, equality holds if and only if $X \cup \mathbb{F}(X)$ is arithmetically Cohen-Macaulay.\
Finally, by [@BLPS1 Theorem 7.1] we know that $X \cap \mathbb{F} \subset \mathbb{F}$ is a hypersurface. Therefore ${\operatorname{depth}}(X) = n$ if $X \cup \mathbb{F}(X)$ is arithmetically Cohen-Macaulay.
Comparing Betti numbers {#comparing-betti-numbers .unnumbered}
-----------------------
We finish this section with a comparison of the Betti numbers of $X$ and $Y = X \cup \mathbb{F}(X).$
\[prop:BettiNumbers\] Let the notations and hypotheses be as in Convention and Notation \[convention and notation\] and assume that $X$ is not a cone. Set $m:= {\operatorname{reg}}(Y)$. Then the following statements hold:
- For all $i \geq 1$ we have $$\beta_{i,j}(X) = \begin{cases}
\beta_{i,j}(Y) & \mbox{for $1 \leq j \leq m-1$},\\
\beta_{i,j}(Y) = 0 & \mbox{for $m \leq j \leq d-r+1$},\\
\beta_{i,d-r+2}(Y) + \binom{r-2}{i-1} & \mbox{for $j=d-r+2$}.
\end{cases}$$
- $m \leq d-r+2$ if and only if $\beta_{i,d-r+2}(X) = \binom{r-2}{i-1}$ for all $i \geq 1$.
Let $I$ and $L$ respectively denote the homogeneous vanishing ideals of $X$ and $\mathbb{F}(X)$ in $S = \Bbbk[x_0,x_1,\ldots,x_r]$, so that $\beta_{i,j}(X) = \beta_{i,j}(S/I)$ and $\beta_{i,j}(Y) = \beta_{i,j}(S/I\cap L)$ for all $i,j \in \mathbb{N}$.\
(a): The exact sequence $$0 \rightarrow (S/L)(-d+r-3) \rightarrow S/I\cap L \rightarrow S/I \rightarrow 0$$ used in the proof of Theorem \[4.14” Theorem\] induces the following long exact sequence: $${\operatorname{Tor}}^S_i(\Bbbk,S/L)_{i+j-d+r-3} \rightarrow {\operatorname{Tor}}^S_i(\Bbbk,S/I \cap L)_{i+j} \rightarrow {\operatorname{Tor}}^S_i(\Bbbk,S/I)_{i+j}$$ $$\rightarrow{\operatorname{Tor}}^S_{i-1}(\Bbbk,S/L)_{(i-1)+j-d+r-2} \rightarrow {\operatorname{Tor}}^S_{i-1}(\Bbbk,S/I\cap L)_{(i-1)+j+1}$$ For all non-negative integers $k$ we have $$\dim_K\big({\operatorname{Tor}}^S_k(\Bbbk,S/L)_{k+l}\big) = \beta_{k,l}(S/L) = \begin{cases}
0 & \mbox{if $l \neq 0$}\\
\binom{r-2}{k} & \mbox{if $l = 0$}
\end{cases}$$ Therefore, the above long exact sequence makes us end up with isomorphisms $${\operatorname{Tor}}^S_i(\Bbbk,S/I \cap L)_{i+j} \cong {\operatorname{Tor}}^S_i(\Bbbk,S/I)_{i+j} \mbox{ for all } i \geq 1 \mbox{ and all } j \in \{1,2,\ldots, d-r+1\}.$$ As ${\operatorname{reg}}(S/I\cap L) = {\operatorname{reg}}(Y) - 1 = m-1$, we have $\beta_{i,j}(S/I\cap L) = 0 \mbox{ for all } i \geq 1 \mbox{ and all } j \geq m$. So by the above isomorphisms we get the requested values of $\beta_{i,j}(S/I)$ for all $i \geq 1$ and all $j \in\{1,\ldots,d-r+1\}$.
As ${\operatorname{reg}}(S/I\cap L) = {\operatorname{reg}}(Y) -1 \leq d-r+2$ (see Theorem \[4.14” Theorem\](b)(1)), the last module in the above exact sequences vanishes for $j=d-r+2$. So, our previous observation on the Betti numbers $\beta_{k,l}(S/L)$ yields a short exact sequence $$0\rightarrow {\operatorname{Tor}}^S_i(\Bbbk,S/I\cap L)_{i+d-r+2} \rightarrow {\operatorname{Tor}}^S_i(\Bbbk,S/I)_{i+d-r+2} \rightarrow \Bbbk^{\binom{r-2}{i-1}} \rightarrow 0 \mbox{ for all } i \geq 1,$$ which shows that $\beta_{i,d-r+2}(S/I) = \beta_{i,d-r+2}(S/I \cap L) + \binom{r-2}{i-1}$, and this proves our claim.\
(b): As already said above, we have $m \leq d-r+3$ and hence $\beta_{i,j}(Y) = 0$ for all $i \geq 1$ and all $j \geq d-r+3$. Thus $m \leq d-r+2$ if and only if $\beta_{i,d-r+2} (Y)=0$ for all $i \geq 1$. Also by statement (a), the second condition holds if and only if $\beta_{i,d-r+2}(X) = \binom{r-2}{i-1}$ for all $i \geq 1$.
Special extremal secant lines {#special-extremal-secant-lines .unnumbered}
-----------------------------
In case of surfaces of type II, the special extremal secant locus is easily understood. We also shall see that proper $3$-secant lines which meet $X$ only in regular points are already special extremal lines and we shall approximate the singular locus of $X$ by the singular locus of the intersection of $X$ with the extremal $\mathbb{F}(X)$ plane of $X$.
\[prop:extseclinesII\] Let the hypotheses be as in Convention and Notation \[convention and notation\]. Then the following statements hold:
- The image $\psi\big({}^*\Sigma(X)\big)$ of the special extremal locus ${}^*\Sigma(X)$ of $X$ under the Plücker embedding $$\psi:\mathbb{G}(1,\mathbb{P}^r\big) \longrightarrow \mathbb{P}^{\binom{r+1}{2}-1}$$ is a plane.
- Let $\mathbb{L} \in \mathbb{G}(1,\mathbb{P}^r)$ such that $\mathbb{L} \nsubseteq X.$ Then, the following statements are equivalent:
- $\mathbb{L} \in {}^*\Sigma(X);$
- $\mathrm{length}(X \cap \mathbb{L}) \geq 3$ and $X \cap \mathbb{L} \subset \mathrm{Reg}(X).$
- $\mathrm{Sing}(X) = \{x \in \mathrm{Sing}\big(X \cap \mathbb{F}(X)\big) \mid \mathbb{F}(X) \subsetneq \mathrm{T}_x(X)\}$. In particular, each point $x \in \mathrm{Sing}\big(X \cap \mathbb{F}(X)\big)$ not contained in a line $\mathbb{L} \subset X \cap \mathbb{F}(X)$ is a singular point of $X$.
(a): As $X$ is of type II, we know that $\overline{\bigcup_{\mathbb{L} \in {}^*\Sigma(X)} \mathbb{L}} = \mathbb{F}(X) = \mathbb{P}^2 \subset \mathbb{P}^r$ is a plane, so that ${}^*\Sigma(X) = \mathbb{G}(1,\mathbb{F}(X)) = \mathbb{G}(1,\mathbb{P}^2)$. Standard arguments on Plücker embeddings show that $\psi\big(\mathbb{G}(1,\mathbb{P}^2)\big)$ is a plane in $\mathbb{P}^{\binom{r+1}{2}-1}$.\
(b): The implication “(i) $\Leftarrow $ (ii)" follows as $\mathcal{C}_\mathbb{H} := X \cap \mathbb{H}$ is smooth for each $\mathbb{H} \in \mathcal{U}(X)$ and hence can only contain smooth points of $X$.\
So, let $\mathbb{L} \in \Sigma_3(X)$ such that $X \cap \mathbb{L}$ is finite and contained in ${\operatorname{Reg}}(X)$, and assume that $\mathbb{L}_\mathbb{H} \neq \mathbb{L}$ for all $\mathbb{H} \in
\mathcal{U}(X)$. We aim for a contradiction. Let $\pi =
\pi_\Lambda:\widetilde{X} \twoheadrightarrow X$ be the standard normalization of $X$, induced by the linear projection $\pi' =
\pi'_\Lambda : \mathbb{P}^{d+1} \twoheadrightarrow \mathbb{P}^r$, consider the closed preimage $\mathbb{L}' :=
\overline{\pi'^{-1}(\mathbb{L})} \in
\mathbb{G}(d-r+2,\mathbb{P}^{d+1}),$ of $\mathbb{L}$ and observe that $\widetilde{X} \cap \mathbb{L}' = \pi^{-1}(X\cap\mathbb{L})$ is finite. Let $\mathbb{H}' \in \mathbb{G}(d,\mathbb{P}^{d+1})$ be a general hyperplane which contains the space $\mathbb{L}'$. If $\widetilde{X}$ is not a cone, we may conclude by [@BP2 Remark 2.3 (B)], that the intersection $\widetilde{X} \cap \mathbb{H}'
\subset \mathbb{H}'$ is a rational normal curve. If $\widetilde{X}$ is a cone, the fact that $\mathbb{L}$ avoids the singular locus of $X$ implies that $\mathbb{L}'$ does not contain the vertex of $\widetilde{X}$ and we end up again with the conclusion that $\widetilde{X} \cap \mathbb{H}' \subset \mathbb{H}'$ is a rational normal curve. As $\mathbb{H}'$ is general, the hyperplane $\mathbb{H} := \pi'(\mathbb{H}' \setminus \Lambda)
\in \mathbb{G}(r-1,\mathbb{P}^r)$ avoids the finite set $\mathrm{Sing}(\pi)$, and hence $\mathcal{C}_\mathbb{H} = X \cap \mathbb{H} = \pi(\widetilde{X}\cap\mathbb{H}') \subset \mathbb{P}^r$ is an integral curve, whence $\mathbb{H} \in \mathcal{U}(X).$\
By our assumption we have $\mathbb{L} \neq \mathbb{L}_\mathbb{H}$, hence $\mathbb{V} := \langle\mathbb{L}_\mathbb{H},\mathbb{L}\rangle \in \mathbb{G}(s,\mathbb{H})$ with $s \in \{2,3\}$. In particular, the intersection $X \cap \mathbb{V}$ is finite. As $\mathbb{L} \neq \mathbb{L}_\mathbb{H}$, we have $$\mathrm{length}(X \cap (\mathbb{L} \cup \mathbb{L}_\mathbb{H})) \geq \mathrm{length}(X \cap \mathbb{L}) + \mathrm{{\operatorname{length}}}(X \cap \mathbb{L}_\mathbb{H}) - \varepsilon$$ with $\varepsilon = 1$ if $\mathbb{L} \cap \mathbb{L}_\mathbb{H} \subset X$ and $\varepsilon = 0$ otherwise. In the first case, we have $s = 2$, so that always $3-\varepsilon \geq s$. Therefore, we obtain $$\mathrm{{\operatorname{length}}}(X \cap (\mathbb{L} \cup \mathbb{L}_\mathbb{H})) \geq \mathrm{length} (X \cap \mathbb{L}) +
\mathrm{length}(X\cap \mathbb{L}_\mathbb{H}) - \varepsilon \geq 3 + d-r+3 - \varepsilon \geq d-r+s+3.$$ As $\mathbb{L} \cup \mathbb{L}_\mathbb{H} \subset {\rm Reg}(X) \cap \mathbb{V}$ this contradicts Proposition \[prop:loc,properties\] (b).\
(c): Let $x \in \mathrm{Sing}(X)$. Let $\mathbb{H} \in \mathcal{U}(X)$ such that $\mathbb{L}_{\mathbb{H}} \cap \mathrm{Sing}(X) = \emptyset$ and consider the plane $\mathbb{E}_{\mathbb{H}} := \langle x, \mathbb{L}_{\mathbb{H}} \rangle \subset \mathbb{P}^r$. Then, by Proposition \[prop:loc,properties\] (c), we have $\mathrm{dim}(X \cap \mathbb{E}_\mathbb{H}) = 1$. If $x \notin \mathbb{F}(X)$, this would imply the contradiction that the intersection of $X$ with the three-space $\langle x, \mathbb{F}(X) \rangle$ contains infinitely many curves. Therefore $x \in \mathbb{F}(X)$.\
Now, let $\mathbb{L} \subset \mathbb{F}(X)$ be a general line such that $x \in \mathbb{L}$. Then, by Lemma \[4.12” Lemma+\] (a) we have $\mathrm{Sing}(X) \cap \mathbb{L} = \{x\}$ and ${\operatorname{length}}(X \cap
\mathbb{L}) = d-r+3$. Assume, that $\mathrm{mult}_x(\mathbb{L} \cap
X) = 1,$ so that $\mathrm{length}(\mathrm{Reg}(X) \cap \mathbb{L}) =
\mathrm{length}(X \cap \mathbb{L}) - 1 = d-r+2$. By Proposition \[prop:loc,properties\] (b) it follows that $d-r+4 =
d-r+2+2 \leq d-r+1+2 = d-r+3,$ and this contradiction shows that $\mathrm{mult}_x(\mathbb{L} \cap X) > 1$. This first shows that $\mathbb{L}$ is a tangent line to $X$ in $x$, and hence proves that $\mathbb{F}(X) \subseteq \mathrm{T}_x(X)$. As $x \in
\mathrm{Sing}(X)$, the inclusion is strict. As $\mathrm{mult}_x\big(\mathbb{L} \cap (X \cap \mathbb{F}(X)\big) =
\mathrm{mult}_x(\mathbb{L} \cap X) >1$ it also follows that a general line $\mathbb{L} \subset \mathbb{F}(X)$ which runs through $x,$ is tangent to $X \cap \mathbb{F}(X)$ in $x$, so that $x \in
\mathrm{Sing}\big(X\cap \mathbb{F}(X)\big).$ This proves the inclusion ”$\subseteq$” between the two sets in question. As the converse inclusion is obvious, we get the requested equality. The additional claim now follows easily, as $X$ is a union of lines.
The index of normality of $X$
=============================
Index of normality and extremal planes {#index-of-normality-and-extremal-planes .unnumbered}
--------------------------------------
Our next main result is devoted to the study of the relations among the index of normality $N(X)$, the Betti numbers $\beta_{i,j}(X)$ and the nature of the union $X \cup \mathbb{F}(X)$, where $X$ is a surface of maximal sectional regularity which is of type II. We begin with two auxiliary results.
\[4.16”” Lemma\] Let the notations and hypotheses be as in Convention and Notation \[convention and notation\]. Assume that $X$ is not a cone. Then we have the following statements
- ${\rm Soc}(H^1(S/I))(-r-1) \cong {\operatorname{Tor}}^S_r(\Bbbk,S/I)$.
- If ${\operatorname{depth}}(X) = 1$, then $H^1\big(\mathbb{P}^r,\mathcal{I}_X(N(X))\big) \cong {\operatorname{Tor}}^S_r(\Bbbk,S/I)_{N(X)+r+1}.$
- $N(X) \leq d-r$ if and only if $\beta_{r,d-r+2}(X) = 0$.
(a): If ${\operatorname{depth}}(X) > 1$, both of the occurring modules vanish and our claim is obvious. So, we assume that ${\operatorname{depth}}(X) = 1$ and consider the total ring of sections $D := D_{S_+}(S/I) = \bigoplus_{n \in \mathbb{Z}} H^0(\mathbb{P}^r, \mathcal{O}_X(n))$ of $X$, as well as the short exact sequence $$0 \longrightarrow S/I \longrightarrow D \longrightarrow H^1(S/I) \longrightarrow 0.$$ We apply the Koszul functor $K(\underline{x}; \bullet)$ with respect to $\underline{x} := x_0,x_1,\ldots,x_r$ to this sequence and end up in homology with an exact sequence $$H_{r+1}(\underline{x};D) \rightarrow H_{r+1}(\underline{x};H^1(S/I)) \rightarrow H_r(\underline{x}; S/I) \rightarrow H_r(\underline{x}; D).$$ As ${\operatorname{depth}}(D) > 1$ the first and the last module in this sequence vanish, so that $$H_{r+1}(\underline{x};H^1(S/I)) \cong H_r(\underline{x}; S/I).$$ As the Koszul complex $K(\underline{x},S)$ provides a free resolution of $\Bbbk = S/S_+$ and $K(\underline{x}; S/I) \cong K(\underline{x}; S) \otimes_S S/I$ we have $H_r(\underline{x}; S/I) \cong {\operatorname{Tor}}^S_r(\Bbbk,S/I)$. As the sequence $\underline{x}$ has length $r+1$, we have $H_{r+1}(\underline{x}; H^1(S/I)) \cong
{\rm Soc}(H^1(S/I))(-r-1)$. Altogether, we now obtain the requested statement (a).\
(b): As $N(X) = {\rm end}(H^1(S/I))$, we have $$H^1\big(\mathbb{P}^r,\mathcal{I}_X(N(X))\big) \cong H^1(S/I)_{N(X)} = {\rm Soc}(H^1(S/I))_{N(X)}.$$ Now, our claim follows immediately by statement (a).\
(c): If ${\operatorname{depth}}(X) > 1$ we have $N(X) = -\infty$ and $\beta_{r,d-r+2}(X) = 0$, so that our claim is true. We thus may assume that ${\operatorname{depth}}(X) = 1$. As ${\operatorname{reg}}(X) = d-r+3$ we have $N(X) \leq d-r+1$ and ${\operatorname{Tor}}^S_r(\Bbbk,S/I)_{r + l} = 0$ for all $l \geq d-r+3$. Now, we may conclude by statement (b).
Now, we are ready to give the announced main result.
\[4.17” Proposition\] Let the notations and hypotheses be as in Convention and Notation \[convention and notation\]. Assume that $X$ is not a cone. Then
- The following statements are equivalent:
- $N(X) \leq d-r$.
- ${\operatorname{reg}}(X \cup \mathbb{F}(X)) \leq d-r+2$.
- $\beta_{i,d-r+2}(X) = \binom{r-2}{i-1}$ for all $i \geq 1$.
- $\beta_{r,d-r+2}(X) = 0$.
- The following statements are equivalent:
- $\beta_{1,d-r+2}(X) = 1$.
- $I\cap L = (I_{\leq d-r+2})$, where $I$ and $L$ are the homogeneous vanishing ideals of $X$ respectively of $\mathbb{F}(X)$ in $S$.
(a): (i) $\Rightarrow$ (ii): Let $N(X) \leq d-r$ and $I$ and $L
\subset S =\Bbbk[x_0,x_1,\ldots,x_r]$ respectively denote the homogeneous vanishing ideals of $X$ and of $\mathbb{F}(X)$. According to Theorem \[4.14” Theorem\] (b)(1) we have ${\rm
end}(H^1(S/I\cap L) = {\rm end}(H^1(S/I))= N(X) \leq d-r$. So, it follows by Theorem \[4.14” Theorem\] (b)(3) and (6) that ${\operatorname{reg}}(S/I \cap L) \leq d-r+1$, whence ${\operatorname{reg}}(X) \cup \mathbb{F}(X) =
{\operatorname{reg}}(I \cap L)
\leq d-r+2$.\
(ii) $\Rightarrow$ (i): As ${\rm end}(H^1(S/I)) = N(X)$, this is an easy consequence of Theorem \[4.14” Theorem\] (b)(2).\
(ii) $\Rightarrow$ (iii): This follows by Propsition \[prop:BettiNumbers\].\
(iii) $\Rightarrow$ (ii): Assume that statement (iii) holds. Then we have in particular that $\beta_{1,d-r+2}(X) = 1$. Now, we may again conclude by Proposition \[prop:BettiNumbers\].\
(iii) $\Leftrightarrow$ (iv): This is clear by Lemma \[4.16”” Lemma\].\
(b): (i) $\Rightarrow $ (ii): Assume that $\beta_{1,d-r+2}(X) = 1$. Then, it follows by Proposition \[prop:BettiNumbers\] (a) that $(I_{\leq d-r+2}) = I \cap L$.\
(ii)$\Rightarrow$(i): This also follows immediately by Proposition \[prop:BettiNumbers\] (a).
\[remark:normality\] (A) The extremal plane $\mathbb{F}(X)$ of a surface $X$ of type II which satisfies $N(X) \leq d-r$ has some nice properties. We namely can say that the equivalent properties (i),(ii) of Theorem \[4.17” Proposition\] (b) imply the following statements, in which, for all $m \in \mathbb{N}$, we use $$\mathrm{Sec}_m(X) := \bigcup_{\mathbb{L} \in \mathbb{G}(1,\mathbb{P}^r) : \mathrm{length}(X \cap \mathbb{L})\geq m} \mathbb{L}$$ to denote the $m$-*th secant variety* of $X$.
- If $\mathbb{L} \in \Sigma(X)$, then $\mathbb{L} \subset X \cup \mathbb{F}(X),$
- $\mathrm{Sec}_{d-r+3}(X) = X \cup \mathbb{F}(X),$
- $\mathbb{F}^+(X) = \mathbb{F}(X).$
Indeed, assume that the equivalent statements (i) and (ii) of Theorem \[4.17” Proposition\](b) hold. Let $\mathbb{L} \in \mathbb{G}(1,\mathbb{P}^r)$ such that $d-r+3 \leq \mathrm{length}(X \cap \mathbb{L}) < \infty.$ Let $M \subset S$ be the homogeneous vanishing ideal of $\mathbb{L}$. Then, $(I_{d-r+2}) \subset M$. As $I\cap L = (I_{\leq d-r+2})$ it follows that $I \cap L \subset M$. As $\mathbb{L}$ is not contained in $X$, the ideal $I$ is not contained in $M$. It follows that $L \subset M$, and hence that $\mathbb{L} \subset \mathbb{F}(X)$. As $\Sigma(X)$ is the closure of all lines $\mathbb{L}$ as above, this proves claim (1).\
Claims (2) and (3) are immediate by claim (1), as $\mathbb{F}(X)$ is a union of lines and each line $\mathbb{L} \subset \mathbb{P}^r$ with $\mathrm{length}(X \cap \mathbb{L}) > d-r+3$ is contained in $X$.\
(B) Observe that statement (iii) of Theorem \[4.17” Proposition\] (a) implies that $\beta_{1,d-r+2}(X) = 1$. So, the equivalent statements (a) (i)–(iv) imply the equivalent statements (b)(i) and (b)(ii) of this theorem.\
(C) We have seen above, that surfaces $X$ of type II and sub-maximal index of normality behave nicely. We therefore can expect, that in the extremal case $N(X) = -\infty$ – hence in the case where ${\operatorname{depth}}(X) = 2$ – we get even more detailed information on the Betti numbers if $X$ is of “small degree".
Surfaces of degree $r+1$ in $\mathbb{P}^r$ {#surfaces-of-degree-r1-in-mathbbpr .unnumbered}
------------------------------------------
We now briefly revisit the special case of surfaces $X \subset \mathbb{P}^r$ of degree $r+1$.
\[3.3’ Remark\] (s. [@B2], [@BS6]) (A) Assume that $r \geq 5$ and let our surface $X \subset \mathbb{P}^r$ be of degree $r+1$. Then, we can distinguish $9$ cases, which show up by their numerical invariants as presented in the following table. Here $\sigma(X)$ denotes the *sectional genus* of $X$, that is the arithmetic genus of the generic hyperplane section curve $\mathcal{C}_h \quad
(h \in \mathbb{U}(X))$ or equivalently, the sectional genus of the polarized surface $(X, \mathcal{O}_X(1))$ in the sense of Fujita [@Fu]. Moreover $\mathrm{sreg}(X)$ denotes the sectional regularity introduced in Remark and Definition \[remark and definition 0\].
$$\begin{tabular}{| c | c | c | c | c | c | c |}
\hline
$\rm{Case}$ & ${\operatorname{sreg}}(X)$ & ${\operatorname{depth}}(X)$
& $\sigma(X)$ & ${\operatorname{e}}(X)$ & $h^1_A(1)$ & $h^1_A(2)$ \\ \hline
$1$ &$2$ &$3$ &$2$ &$0$ &$0$ &$0$ \\ \hline
$2$ &$3$ &$2$ &$1$ &$0$ &$0$ &$0$ \\ \hline
$3$ &$3$ &$2$ &$1$ &$1$ &$0$ &$0$ \\ \hline
$4$ &$3$ &$1$ &$1$ &$0$ &$1$ &$0$ \\ \hline
$5$ &$3$ &$2$ &$0$ &$2$ &$0$ &$0$ \\ \hline
$6$ &$3$ &$1$ &$0$ &$1$ &$1$ &$\leq 1$ \\ \hline
$7$ &$3$ &$1$ &$0$ &$0$ &$2$ &$\leq 2$ \\ \hline
$8$ &$4$ &$2$ &$0$ &$3$ &$0$ &$0$ \\ \hline
$9$ &$4$ &$1$ &$0$ &$0$ &$2$ &$3$ \\ \hline
\end{tabular}$$\
The case $9$ occurs only if $r = 5$. In [@B2] and [@BS6] we listed indeed two more cases $10$ and $11$, of which we did not know at that time, whether they might occur at all. For these two cases we had ${\operatorname{sreg}}(X) = 4 = d-r+3$ and ${\rm e}(X) \in \{1,2\}$. As these surfaces would be of maximal sectional regularity, this would contradict Theorem \[t1-coh\] (b) and Theorem \[4.14” Theorem\] (a). So, surfaces which fall under the cases $10$ and $11$ cannot occur at all. In the case $9$ we have ${\rm e}(X) = 0$, and hence by Theorem \[4.14” Theorem\] (b)(3) the surface $X$ is of type I in this case.\
(B) The surfaces of types 8 and 9 are of particular interest, as they are the ones of maximal sectional regularity within all the 9 listed types. Observe, that among all surfaces $X$ of degree $r+1$ in $\mathbb{P}^r$, those of type 8 are precisely the ones $X$ which are of maximal sectional regularity and of arithmetic depth $\geq 2$. If $r \geq 6$, the surfaces of type 8 are precisely the ones which are of maximal sectional regularity.\
(C) Observe, that in the cases 5 – 9 we have $\sigma(X) = 0$. This means, that the surfaces which fall under these 5 types are all sectionally rational and have finite non-normal locus. So, by Theorem 4.1 in [@BLPS1], these surfaces are almost non-singular projections of a rational normal surface scroll $\widetilde{X} =
S(a,r+1-a)$ with $0 \leq a \leq \frac{r+1}{2}$, even if they are cones (see [@BS5 Corollary 5.11] for the non-conic case). So, according to Theorem \[2.5’ Theorem\] (b) the surfaces $X$ of types 5 – 9 all satisfy the Eisenbud-Goto inequality ${\operatorname{reg}}(X) \leq
4$, with equality in the cases 8 and 9 (see Theorem \[4.14” Theorem\] (a)(1) and Theorem \[t1-betti\] ). In the cases 1 – 5, the values of $h^i(\mathbb{P}^r, \mathcal{I}_X(n)) =: h^i(S/I)_n
\quad (i=1,2, n\in \mathbb{Z})$ (see [@BS6 Reminder 2.2 (C) and (D)]) show, that ${\operatorname{reg}}(X) = 3$. In the case 6 we may have ${\operatorname{reg}}(X) = 3$ whereas in the case 7, we know even that ${\operatorname{reg}}(X)$ may take both values $3$ and $4$ (see [@BS6 Example 3.5, Examples 3.4 (A),(B) and (C)]). This shows in particular, that there are sectionally rational surfaces $X \subset \mathbb{P}^r$ of degree $r+1$ with finite non-normal locus and ${\operatorname{sreg}}(X) < {\operatorname{reg}}(X)$.
\[3.5’ Corollary\] Assume that the surface $X \subset \mathbb{P}^r_K$ is of degree $r+1$. Then, the following statements are equivalent
- The surface $X$ is of type 8.
- ${\operatorname{e}}(X) = 3$.
- ${\operatorname{sreg}}(X) = 4$ and ${\operatorname{depth}}(X) =2$.
- ${\operatorname{sreg}}(X) = 4$ and $X$ does not fall under the case 9 of Remark \[3.3’ Remark\].
This follows easily on use of the table in Remark \[3.3’ Remark\] (A).
Examples and Problems
=====================
Surfaces of extremal regularity with small extremal secant locus {#surfaces-of-extremal-regularity-with-small-extremal-secant-locus .unnumbered}
-----------------------------------------------------------------
As announced already in the Introduction, we now shall present a construction, which allows to provide examples of surfaces of extremal regularity whose extremal secant variety is of dimension $-1,0,$ or $1$. These surfaces are in particular not of maximal sectional regularity. We already have spelled out the meaning of such examples in relation what is said about varieties of extremal regularity in [@GruLPe].
\[7.1 Construction and Examples\] Let $a,b,d \in \mathbb{N}$ with $a \leq b$, let $r:=a+b+3$, assume that $d > r$ and consider the smooth threefold rational normal scroll of degree $a+b+1 = r-2$ $$Z := S(1,a,b) \subset \mathbb{P}^r.$$ Let $H, F \in {\rm Div}(Z)$ respectively be a hyperplane section and a ruling plane of $Z$, so that each divisor on $Z$ is linearly equivalent to $mH+nF$ for some integers $m,n$. Let $X \subset \mathbb{P}^r$ be an non-degenerate irreducible surface of degree $d$ which is contained in $Z$ as a divisor linearly equivalent to $H+(d-r+2)F$.
\(A) One can easily see that $h^0 (X, \mathcal{O}_X (1)) = h^0 (Z, \mathcal{O}_Z (1))+d-r+1 = d+2$. This means that the linearly normal embedding $\widetilde{X} \subset \mathbb{P}^{d+1}$ of $X$ by means of $\mathcal{O}_X (1)$ is of minimal degree and $X$ is a regular projection of $\widetilde{X}$. Keep in mind, that $\widetilde{X}$ is either a smooth rational normal surface scroll, a cone over a rational normal curve or the Veronese surface in $\mathbb{P}^5$. As $d+1 > 5$, and as a cone does not admit a proper isomorphic linear projection, $\widetilde{X} \subset \mathbb{P}^{d+1}$ is a smooth rational normal surface scroll. This means that $X$ is smooth and sectionally rational. Also ${\operatorname{reg}}(X) = d-r+3$ (cf. [@P Theorem 4.3]) and hence $X$ is a surface of extremal regularity.
\(B) Let $\mathbb{L}$ be a line section of $Z$. Then the intersection number $\mathbb{L}\cdot X$ takes the maximal possible value $d-r+3$. This means that either $\mathbb{L}$ is contained in $X$ or else it is a proper $(d-r+3)$-secant line to $X$. On the other hand, observe that any proper $(d-r+3)$-secant line to $X$ must be contained in $Z$ as a line section, since $Z$ is cut out by quadrics.\
To reformulate this observation, we introduce the locally closed subset $$\Sigma^{\circ}(X) := \{\mathbb{L} \in \mathbb{G}(1,\mathbb{P}^r) \mid d-r+3 \leq \mathrm{length}(X \cap\mathbb{L}) < \infty\} \subset \mathbb{G}(1,\mathbb{P}^r)$$ of proper $(d-r+3)$-secant lines to $X$, so that $\overline{\Sigma^{\circ}(X)} = \Sigma(X)$. The previous observation now may be written in the form $$\Sigma^{\circ}(X) = \{\mathbb{L} \in \mathbb{G}(1,\mathbb{P}^r) \mid \mathbb{L} \mbox{ is a line section of } Z \mbox{ and } \mathbb{L} \nsubseteq X\}.$$
\(C) Suppose that $a \geq 2$ and let $\mathbb{L}$ be the unique line section $S(1)$ of $Z$. If $\mathbb{L}$ is contained in $X$, then we have $\Sigma^{\circ}(X) = \emptyset$ and hence $\Sigma (X) =
\emptyset$. So, in this case $X$ is a surface of maximal regularity, having no proper extremal secant line at all. Next, if $\mathbb{L}$ is not contained in $X$, then we have $\Sigma^{\circ}(X) =
\Sigma(X)=\{ \mathbb{L} \}$, and hence $\dim\big(\Sigma(X)\big) =
0$.
\(D) Suppose next, that $a =1$ and $b \geq 2$. Then, by part (B), we have $$\Sigma^{\circ}(X) = \{\mathbb{L} \in S(1,1) \mid \mathbb{L} \mbox{ is a line section of } Z \mbox{ and } \mathbb{L} \nsubseteq X\}.$$ Since $X \neq S(1,1)$, all but finitely many line sections of $Z$ are proper $(d-r+3)$-secant lines to $X$. This implies that $\dim\big(\Sigma(X)\big) = \dim\big(\Sigma^{\circ}(X)\big) = 1$ and that $\mathbb{F}^{+}(X)$ is exactly equal to $S(1,1)$.
Surfaces of maximal sectional regularity of type I {#surfaces-of-maximal-sectional-regularity-of-type-i .unnumbered}
--------------------------------------------------
We now provide a few examples for surfaces of maximal sectional regularity of type I, focusing on the various Betti tables which may occur. These tables have been computed by means of the Computer Algebra System Singular [@GrPf]. We use the divisorial description of surfaces of type I given in Theorem \[theorem 2.1\]
\[example:t1\] Let $X \subset W := S(1,1,1) \subset \mathbb{P}^{5}$ be a divisor which is linearly equivalent to $H + (d-3)F$. Then $X$ is given by an isomorphic projection of a smooth rational normal scroll $\widetilde{X} \subset \mathbb{P}^{d+1}$ (see [@P Lemma 3.1]).\
(A) Let $d=8$ and assume that $X$ has the parametrization $$\{\big(u^7s:u^7t:vs^7:vs^6t:vst^6:vst^7\big) \mid (s,t), (u,v) \in \Bbbk^2 \setminus\{(0,0)\}\}.$$ Then, $X$ has the following Betti table. $$\begin{tabular}{|c|c|c|c|c|c|c|}\hline
\multicolumn{2}{|c||}{$i$}& $1$&$2$&$3$&$4$&$5$ \\\cline{1-7}
\multicolumn{2}{|c||}{$\beta_{i,1}$}& $3$&$2$&$0$&$0$&$0$ \\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,2}$}& $0$& $0$&$0$& $0$&$0$ \\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,3}$}& $0$& $0$&$0$& $0$&$0$ \\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,4}$}& $0$&$0$&$0$&$0$&$0$ \\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,5}$}& $21$&$70$&$87$&$48$&$10$\\\cline{1-7}
\end{tabular}$$
\(B) Let $d=9$ and assume that $X$ has the parametrization $$\{\big(u^7s:u^7t:vs^8:vs^7t:vst^7:vt^8 \big) \mid (s,t), (u,v) \in
\Bbbk^2 \setminus\{(0,0)\}\}.$$ Then, we get the following Betti table for $X$. $$\begin{tabular}{|c|c|c|c|c|c|c|}\hline
\multicolumn{2}{|c||}{$i$}& $1$&$2$&$3$&$4$&$5$ \\\cline{1-7}
\multicolumn{2}{|c||}{$\beta_{i,1}$}& $3$&$2$&$0$&$0$&$0$ \\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,2}$}& $0$& $0$&$0$& $0$&$0$ \\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,3}$}& $0$& $0$&$0$& $0$&$0$ \\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,3}$}& $0$& $0$&$0$& $0$&$0$ \\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,4}$}& $0$&$0$&$0$&$0$&$0$ \\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,5}$}& $28$&$96$&$123$&$70$&$15$\\\cline{1-7}
\end{tabular}$$
\(C) Let $d=10$ and assume that $X$ has the parametrization $$\{\big(u^9s:u^9t:vs^9:vs^8t:vst^8:vst^9\big) \mid (s,t), (u,v) \in \Bbbk^2 \setminus\{(0,0)\}\}.$$ Then, the Betti table of $X$ is as given below. $$\begin{tabular}{|c|c|c|c|c|c|c|}\hline
\multicolumn{2}{|c||}{$i$}& $1$&$2$&$3$&$4$&$5$ \\\cline{1-7}
\multicolumn{2}{|c||}{$\beta_{i,1}$}& $3$&$2$&$0$&$0$&$0$ \\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,2}$}& $0$& $0$&$0$& $0$&$0$ \\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,3}$}& $0$& $0$&$0$& $0$&$0$ \\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,3}$}& $0$& $0$&$0$& $0$&$0$ \\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,3}$}& $0$& $0$&$0$& $0$&$0$ \\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,4}$}& $0$&$0$&$0$&$0$&$0$ \\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,5}$}& $36$ &$126$&$165$ &$96$&$21$ \\\cline{1-7}
\end{tabular}$$
Surfaces of maximal sectional regularity of type II {#surfaces-of-maximal-sectional-regularity-of-type-ii .unnumbered}
---------------------------------------------------
Next, we aim to present examples which concern surfaces $X \subset \mathbb{P}^r$ $(r \geq 5)$ of maximal sectional regularity of degree $d > r$ and of type II. Set $Y := X \cup \mathbb{F}(X)$ and recall that $\tau(X)$ denotes the pair $({\operatorname{depth}}(X), {\operatorname{depth}}(Y))$. We will construct a few examples of $X$, having all possible $\tau(X)$ listed in Theorem \[4.14” Theorem\](c).\
\[7.2 Construction and Examples\] (A) We assume that the characteristic of the base field $\Bbbk$ is zero. Let $a,b$ be integers such that $3 \leq a \leq b$ and consider the standard smooth rational normal surface scroll $\widetilde{X} := S(a,b) \subset \mathbb{P}^{a+b+1}$. We shall construct surfaces of maximal sectional regularity of type II by projecting $\widetilde{X}$ from appropriate linear subspaces of $\mathbb{P}^{a+b+1}$. The occurring Betti diagrams have been computed by means of the Computer Algebra System Singular [@GrPf].
\(B) Let $\Lambda$ be an $(a-3)$-dimensional subspace of $\langle S(a) \rangle = \mathbb{P}^a$ which avoids $S(a)$ and let $X \subset \mathbb{P}^{b+3}$ be the linear projection of $\widetilde{X}$ from $\Lambda$. Observe that this linear projection maps $\langle S(a) \rangle$ onto a plane $\mathbb{P}^2 = \mathbb{F} \subset \mathbb{P}^{b+3}$. Suppose that this projection maps $S(a)$ birationally onto a plane curve $C_a \subset \mathbb{F}$ of degree $a$. Since $X \subset \mathbb{P}^{b+3}$ is a surface of degree $a+b$, we have ${\operatorname{reg}}(X) \leq a$ by Theorem \[4.14” Theorem\](a). On the other hand, a general line on $\mathbb{F}$ is a proper $a$-secant line to $X$. Therefore ${\rm reg}(X)=a$, $X$ is a surface of maximal sectional regularity of type II and $\mathbb{F}(X) = \mathbb{F}$. Finally, we get $\tau(X) = (2,3)$ by Theorem \[4.14” Theorem\](c).
\(C) Assume that $b \geq 3$. Let $\Lambda$ be a $(b-3)$-dimensional subspace of $\langle S(b) \rangle = \mathbb{P}^b$ which avoids $S(b)$ and let $X \subset \mathbb{P}^{a+3}$ be the linear projection of $\widetilde{X}$ from $\Lambda$. So, this linear projection maps $\langle S(b) \rangle$ onto a plane $\mathbb{P}^2 = \mathbb{F} \subset \mathbb{P}^{a+3}$. From now on, we assume that this projection maps $S(b)$ birationally onto a plane curve $C_b \subset \mathbb{F}$ of degree $b$. Then as in (B), one can see that $X \subset \mathbb{P}^r$ is a surface of maximal sectional regularity of type II and $\mathbb{F}(X) = \mathbb{F}$. If $b \leq a+2$, then we have $\tau(X) = (2,3)$ by Theorem \[4.14” Theorem\](c).
\(D) From now on, we assume that $b \geq a+3$, and we will vary the projection center $\Lambda$. To do so, we first consider the canonical isomorphism $$\kappa: \mathbb{P}^1 \rightarrow S(b), \quad [s:t] \mapsto [0:\ldots:0:s^b:s^{b-1}t:\ldots:st^{b-1}:t^b] \in \mathbb{P}^{a+b+1}.$$ Then, we choose a homogeneous polynomial $f \in \Bbbk[s,t]$ of degree $b$ which is not divisible by $s$ and by $t$. Now, let $$\Lambda_f = \mathbb{P}^{b-3} \subset \langle S(b)\rangle \setminus S(b)$$ be such that the composition map $$\varphi_f := \pi_{\Lambda_f} \circ \kappa: \mathbb{P}^1 \rightarrow C_b \subset \mathbb{F} = \mathbb{P}^2$$ of the linear projection map $$\pi_{\Lambda_f}: \mathbb{P}^{a+b+1}\setminus \Lambda_f \twoheadrightarrow \mathbb{P}^{a+3}$$ with the above map $\kappa$ sends $[s:t]$ to $[s^b : f : t^b]$. Let $$X_f := \pi_{\Lambda_f}(\widetilde{X}) \subset \mathbb{P}^{a+2}$$ denote the image of the scroll $\widetilde{X}$ under the linear projection map $\pi_{\Lambda_f}$ centered at $\Lambda_f$. Then, we may write $$X_f := \{ [us^a:us^{a-1}t:\ldots:ust^{a-1}:ut^a:vs^b:vf(s,t):vt^b ] \mid (s,t), (u,v) \in K^2 \setminus\{(0,0)\}\}.$$ After an appropriate choice of $f$, this latter presentation is accessible to syzygetic computations.
\[7.3 Example\] Let $(a,b)=(3,5)$ and $f := s^4t+s^3t^2+s^2t^3+st^4$. Then $X_f \subset \mathbb{P}^6$ is of degree $d = 8 (= 2r-4)$ and the graded Betti numbers $\beta_{i,j} = \beta_{i,j}(X)$ of $X$ are as presented in the following table.
[|c|c|c|c|c|c|c|c|]{} & $1$&$2$&$3$&$4$&$5$&$6$\
& $6$&$8$&$3$&$0$&$0$&$0$\
& $4$&$12$&$12$&$4$&$0$&$0$\
& $0$&$0$&$0$&$0$&$0$&$0$\
& $1$&$4$&$6$&$4$&$1$&$0$\
By Proposition \[prop:BettiNumbers\] (a) it follows from this graded Betti diagram of $X$, that $$\tau(X)=(2,3).$$
\[7.4 Example\] Let $(a,b)=(3,8)$ and consider $X_{f_i} \subset \mathbb{P}^6$ $(i=1,2,3)$ for the following choices of $f_i$:
1. $f_1 = s^7t+s^6t^2+s^5t^3+s^4t^4+s^3t^5+s^2t^6+st^7$,
2. $f_2 = s^7t+s^6t^2+s^5t^3+s^4t^4+s^3t^5+s^2t^6$, and
3. $f_3 = s^7t+s^6t^2+s^5t^3+s^4t^4$.
Then $X_{f_i} \subset \mathbb{P}^6$ is of degree $d = 11 \quad (= 2r-1 = 3r-7)$ for all $i=1,2,3$. The graded Betti diagrams of $X_{f_1}$, $X_{f_2}$ and $X_{f_3}$ are given respectively in the three tables below.
$$\begin{tabular}{|c|c|c|c|c|c|c|c|}\hline
\multicolumn{2}{|c||}{$i$}& $1$&$2$&$3$&$4$&$5$&$6$ \\\cline{1-8}
\multicolumn{2}{|c||}{$\beta_{i,1}$}& $6$&$8$&$3$&$0$&$0$&$0$ \\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,2}$}& $0$& $0$&$0$& $0$&$0$&$0$ \\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,3}$}& $4$&$12$&$12$&$4$&$0$&$0$\\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,4}$}& $0$&$0$&$0$&$0$&$0$&$0$\\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,5}$}& $1$&$4$&$6$&$4$&$1$&$0$\\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,6}$}& $0$&$0$&$0$&$0$&$0$&$0$\\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,7}$}& $1$&$4$&$6$&$4$&$1$&$0$\\\cline{1-8}
\end{tabular}$$ $$\begin{tabular}{|c|c|c|c|c|c|c|c|}\hline
\multicolumn{2}{|c||}{$i$}& $1$&$2$&$3$&$4$&$5$&$6$ \\\cline{1-8}
\multicolumn{2}{|c||}{$\beta_{i,1}$}& $5$&$5$&$0$&$0$&$0$&$0$ \\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,2}$}& $1$& $0$&$1$& $0$&$0$&$0$ \\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,3}$}& $1$&$9$&$11$&$4$&$0$&$0$\\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,4}$}& $4$&$18$&$32$&$28$&$12$&$2$\\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,5}$}& $0$&$0$&$0$&$0$&$0$&$0$\\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,6}$}& $0$&$0$&$0$&$0$&$0$&$0$\\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,7}$}& $1$&$4$&$6$&$4$&$1$&$0$\\\cline{1-8}
\end{tabular}$$
$$\begin{tabular}{|c|c|c|c|c|c|c|}\hline
\multicolumn{2}{|c||}{$i$}& $1$&$2$&$3$&$4$&$5$ \\\cline{1-7}
\multicolumn{2}{|c||}{$\beta_{i,1}$}& $3$&$2$&$0$&$0$&$0$ \\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,2}$}& $10$& $27$&$24$& $7$&$0$ \\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,3}$}& $0$&$0$&$0$&$0$&$0$\\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,4}$}& $0$&$0$&$0$&$0$&$0$\\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,5}$}& $0$&$0$&$0$&$0$&$0$\\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,6}$}& $0$&$0$&$0$&$0$&$0$\\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,7}$}& $1$&$4$&$6$&$4$&$1$\\\cline{1-7}
\end{tabular}$$ By Proposition \[prop:BettiNumbers\] (a) we can see from these tables that $$\tau (X_{f_1}) = (2,2), \quad \tau(X_{f_2}) = (1,1) \mbox{ and } \tau (X_{f_3}) = (2,3).$$
\[7.5 Example\] Let $(a,b)=(3,9)$ and consider $X_{f_i} \subset \mathbb{P}^6$, $(i=1,2)$ for the two choices
1. $f_1 = s^8t+s^7t^2+s^6t^3+s^5t^4+s^4t^5+s^3t^6+s^2t^7+st^8$ and
2. $f_2 = s^8t+s^7t^2+s^6t^3+s^5t^4+s^4t^5+s^3t^6+s^2t^7$.
Then $X_{f_i} \subset \mathbb{P}^6$ is of degree $d = 12 \quad (= 2r = 3r-6)$ for $i=1,2$. The graded Betti diagrams of $X_{f_1}$ and $X_{f_2}$ are given respectively in the tables below.
$$\begin{tabular}{|c|c|c|c|c|c|c|c|}\hline
\multicolumn{2}{|c||}{$i$}& $1$&$2$&$3$&$4$&$5$&$6$ \\\cline{1-8}
\multicolumn{2}{|c||}{$\beta_{i,1}$}& $6$&$8$&$3$&$0$&$0$&$0$ \\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,2}$}& $0$& $0$&$0$& $0$&$0$&$0$ \\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,3}$}& $2$&$4$&$0$&$0$&$0$&$0$\\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,4}$}& $1$&$4$&$10$&$6$&$1$&$0$\\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,5}$}& $0$&$0$&$0$&$0$&$0$&$0$\\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,6}$}& $1$&$4$&$6$&$4$&$1$&$0$\\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,7}$}& $0$&$0$&$0$&$0$&$0$&$0$\\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,8}$}& $1$&$4$&$6$&$4$&$1$&$0$\\\cline{1-8}
\end{tabular}$$ $$\begin{tabular}{|c|c|c|c|c|c|c|c|}\hline
\multicolumn{2}{|c||}{$i$}& $1$&$2$&$3$&$4$&$5$&$6$ \\\cline{1-8}
\multicolumn{2}{|c||}{$\beta_{i,1}$}& $5$&$5$&$0$&$0$&$0$&$0$ \\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,2}$}& $0$& $0$&$1$& $0$&$0$&$0$ \\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,3}$}& $5$&$15$&$15$&$5$&$0$&$0$\\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,4}$}& $0$&$0$&$0$&$0$&$0$&$0$\\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,5}$}& $5$&$23$&$42$&$38$&$17$&$3$\\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,6}$}& $0$&$0$&$0$&$0$&$0$&$0$\\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,7}$}& $0$&$0$&$0$&$0$&$0$&$0$\\\cline{1-1}
\multicolumn{2}{|c||}{$\beta_{i,8}$}& $1$&$4$&$6$&$4$&$1$&$0$\\\cline{1-8}
\end{tabular}$$
By Proposition \[prop:BettiNumbers\] (a) we can verify that $$\tau(X_{f_1}) = (2,2) \mbox{ and } \tau(X_{f_2}) = (1,1).$$
\[7.6 Problem and Remark\] (A) Let $5 \leq r <d$ and let $X \subset \mathbb{P}^r$ be a non-degenerate surface of degree $d$ which is of maximal sectional regularity. We consider the three conditions
- $N(X) \leq d-r$.
- $\beta_{1,d-r+2}(X) = 1$.
- $\mathbb{F}(X) = \mathbb{P}^2$ or – equivalently – $X$ is of type II.
\(B) By the implication (i) $\Rightarrow$ (iii) given in statement (a) of Theorem \[4.17” Proposition\] we have the implication (i) $\Rightarrow$ (ii) among the above three conditions. By the implication (i) $\Rightarrow$ (ii) given in statement (b) of Theorem \[4.17” Proposition\] we have the implication (ii) $\Rightarrow$ (iii) among the above three conditions.\
We expect, that the converse of both implications holds but could not prove this. So we aim to pose the problem
- *Are the three conditions* (i), (ii) *and* (iii) *of part* (A) *equivalent ?*
Observe, that in view of Remark \[remark:normality\] (A) an affirmative answer to this would also answer affirmatively the question, whether for surfaces of type II, the extended extremal variety and the extremal variety of $X$ coincide (see Notation and Reminder \[4.2” Notation and Reminder\]), hence the question whether
- $\quad \mathbb{F}^{+}(X) = \mathbb{F}(X)$ *for $X$ of type II ?*
Obviously, this latter question would been affirmatively answered if we could answer affirmatively the question
- $\quad {}^*\Sigma(X) = \Sigma(X)$ *for $X$ of type II ?*
Acknowledgement {#acknowledgement .unnumbered}
---------------
The first named author thanks to the Korea University Seoul, to the Mathematisches Forschungsinstitut Oberwolfach, to the Martin-Luther Universität Halle and to the Deutsche Forschungsgemeinschaft for their hospitality and the financial support provided during the preparation of this work.The second named author was supported by Basic Science Research Program through the National Research Foundation of Korea(NRF) funded by the Ministry of Education(2014008404). The third named author was supported by the NRF-DAAD GEnKO Program (NRF-2011-0021014). The fourth named author thanks to the Korea University Seoul, to the Mathematisches Forschungsinstitut Oberwolfach and to the Deutsche Forschungsgemeinschaft for their hospitality respectively financial support offered during the preparation of this work.
[0000000]{}
: [*Graded mapping cone theorem, multisecants and syzygies*]{}, Journal of Algebra 331 (2011) 243 - 262.
: [*A bound on certain local cohomology modules and application to ample divisors*]{}, Nagoya Mathematical Journal 163 (2001) 87-106.
: [*A few remarks on blowing-up and connectedness*]{}, Journal für die reine und angewandte Mathematik 370 (1986) 51 - 60.
: [*Cohomology of certain projective surfaces with low sectional genus and degree*]{}, in: D. Eisenbud (Ed.) “Commutative Algebra, Algebraic Geometry and Computational Methods”, 172 - 200, Springer, 1999.
: [*On the regularity of varieties having an extremal secant line*]{}. Journal für die reine und angewandte Mathematik. 545 (2002), 167 - 181.
: [*On projective varieties of maximal sectional regularity*]{}. Preprint (2014).
: [*On varieties of maximal sectional regularity in codimension two*]{}. In preparation.
: [*On varieties of almost minimal degree III: Tangent spaces and embedding scrolls*]{}. Journal of Pure and Applied Algebra 215 (2011) 2859 - 2872.
: [*Curves of degree $ r + 2$ in ${\mathbb P}^r$: Cohomological, geometric, and homological aspects*]{}. Journal of Algebra 242 (2001), 577 - 623.
: [*On projective curves of maximal regularity*]{}. Mathematische Zeitschrift 244 (2003), 271 - 289.
: [*Projective curves with maximal regularity and applications to syzygies and surfaces*]{}. Manuscripta Mathematica 135 (2011) 469 - 495.
: [*Projective surfaces of degree $r+1$ in Projective $r$-space and almost non-singular projections*]{}. Journal of Pure and Applied Algebra 216 (2012) 2241-2255.
: [*Local cohomology – an algebraic introduction with geometric applications*]{}. Cambridge Studies in Advanced Mathematics Vol. 60, Cambridge University Press, Cambridge, UK, 1998.
: [*Sui multipli di une serie lineare di gruppi di punti appartenente ad une curva algebraic*]{}, Rend. Circ. Mat. Palermo (2) 7 (1893), 89-110.
: [*Bounds for the Castelnuovo-Mumford regularity of modules*]{}. Mathematische Zeitschrift 258 (2008) 69 - 80.
: $3-1-2$ – [A computer algebra system for polynomial computations]{}. (2011).
: [*Commutative algebra with a view toward algebraic geometry*]{}. Graduate Texts in Mathematics 150, Springer, New York, 1994.
: [*Linear free resolutions and minimal multiplicity*]{}. J. Algebra 88 (1984) 89 - 133.
: [*Canonical models of surfaces of general type in positive characteristic*]{}. Publications Mathématiques de l’ I.H.E.S, 67 (1988) 97 - 144.
: [*Classification theories of polarized varieties*]{}, London Mathematical Society Lecture Notes Series 155, Cambridge University Press, 1990.
: [*Some results on the syzygies of finite sets and algebraic curves*]{}. Compositio Mathematica 67 (1988), 301 - 314.
: [*On a theorem of Castelnuovo and the equations defining space curves*]{}. Inventiones Mathematicae 72 (1983), 491 - 506.
: [*Some effects of property $N_p$ on the higher normality and defining equations of nonlinearly normal varieties*]{}. Journal für die reine und angewandte Mathematik 582 (2005) 87 - 105.
: [*A sharp Castelnuovo bound for smooth surfaces*]{}. Duke Mathematical Jornal 55 (1987) 423 - 429.
: [*Lectures on curves on an algebraic surface*]{}. Annals of Mathematics Studies 59, Princeton University Press, Princeton, 1966.
: [*A bound on the Castelnuovo-Mumford regularity for curves*]{}. Mathematische Annalen 322 (2002) 69 - 74.
: [*On syzygies of divisors of rational normal scrolls*]{}. Mathematische Nachrichten 287 (2014), no. 11-12, 1383 - 1393.
: [*A Castelnuovo bound for smooth surfaces*]{}. Inventiones Mathematicae 83 (1986), 321 - 332.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We show that the computation of formation probabilities (FP) in the configuration basis and the full counting statistics (FCS) of observables in the quadratic fermionic Hamiltonians are equivalent to the calculation of emptiness formation probability (EFP) in the Hamiltonian with a defect. In particular, we first show that the FP of finding a particular configuration in the ground state is equivalent to the EFP of the ground state of the quadratic Hamiltonian with a defect. Then, we show that the probability of finding a particular value for any quadratic observable is equivalent to a FP problem and ultimately leads to the calculation of EFP in the ground state of a Hamiltonian with a defect. We provide new exact determinant formulas for the FP in the generic quadratic fermionic Hamiltonians. As applications of our formalism we study the statistics of the number of particles and kinks. Our conclusions can be extended also to the quantum spin chains that can be mapped to the free fermions via Jordan-Wigner (J-W) transformtion. In particular, we provide an exact solution to the problem of the transverse field XY chain with a staggered line defect. We also study the distribution of magnetization and kinks in the transverse field XY chain and show how the dual nature of these quantities manifest itself in the distributions.'
author:
- 'M. N. Najafi'
- 'M. A. Rajabpour'
bibliography:
- 'LineDefectEFPRef.bib'
title: Formation probabilities and statistics of observables as defect problems in the free fermions and the quantum spin chains
---
Introduction {#sec:intro}
============
Consider a quantum many body state written in a configuration basis, then the formation probability of a particular configuration in a subsystem is the probability of finding the configuration if we do the measurement in that particular basis. For example, if we take the ground state of a spinless fermionic system and ask the same question for the subsystem $D$ then there are $2^{|D|}$ possibilities for the configurations, where $|D|$ is the size of the subsystem. Every configuration appears with a particular probability which we call formation probability [@Najafi2016]. The simplest example of the FP is the emptiness formation probability which is about the FP of the configuration without any fermion in the subsystem and has a long history. It was studied in the context of the XXZ spin chain in [@Korepin1994; @Korepin1995; @Kitanine2002a; @Kitanine2002b; @Korepin2002; @Korepin2003; @Cantini2012] and in the context of the XY chain in [@Shiroishi2001; @Franchini2003; @Franchini2005; @Viti2019]. FP has been also studied from the conformal field theory (CFT) point of view in [@Stephan2014; @Rajabpour2015; @Rajabpour2016; @viti2016]. The problem is related to the solution of the free energy of a quantum field theory with a slit which in the context of the CFT studied in [@Dubail2013] and references therein. A relation to the Casimir energy problem is also established in [@Rajabpour2015; @Rajabpour2016]. The FP can be also used to find the Shannon entropy with plethora of applications in the studies of quantum phase transitions, see[@Stephan2009; @AR2013; @Stephan2014PRB; @AR2014; @Luitz2014a; @Luitz2014b; @AR2015; @Alcaraz2016; @NR2014; @Alba2016; @NR2019].
One can also look to the problem of the FP from a different point of view. Consider a configuration basis which is associated with the local on-site observable $\hat{o}_i$ at site $i$ with eigenvalues(eigenvectors) $o_i^j$($|o_i^j\rangle$), $j=1,2,...,d$, where $d$ is the dimension of the local on-site Hilbert space. Then one can define an on-site projection operator as $\pi_i^j=|o_i^j\rangle\langle o_i^j|$. Multiplication of these operators in a subsystem leads to a projection operator $\Pi_{\hat{o}}^{\{j\}}=\prod_i\pi_i$, where the set ${\{j\}}$ fix the configuration by picking a particular $\pi^j$ at every site. Finally the average over, for example, the ground state $|g\rangle$ of the Hamiltonian $H$ gives the FP for the desired configuration, i.e. $P\left({\{j\}}\right)=\langle g|\Pi_{\hat{o}}^{\{j\}}|g\rangle$. Now consider one [*[primitive]{}*]{} configuration with all the sites being in the eigenstate corresponding to, for example, $o^1$ i.e. $|\{1\}\rangle$. In the case of fermions this primitive configuration can be the configuration without fermions. Then one can always find a unitary similarity transformation matrix $T^{\{j\}}$ which $P({\{j\}})=\langle g'|\Pi_{\hat{o}}^{\{1\}}|g'\rangle$, where $|g'\rangle=T^{\{j\}}|g\rangle$. The state $|g'\rangle$ can be considered as the ground state of the Hamiltonian $H'=T^{\{j\}}H\left(T^{\{j\}}\right)^{-1}$ which is basically the same Hamiltonian as $H$ but with a line defect. This simple argument shows that formally one can look to the problem of generic FP as the problem of the probability of a primitive configuration in the deformed Hamiltonian. Many body systems with line defects have been studied for decades, for some earlier studies see for example[@PFEUTY1979245; @Bariev1979; @McCoy1980; @Kadanov1981; @Brown1982; @Cabrera1987; @Abraham1989; @Hinrichsen1989]. For studies related to boundary CFT and integrable quantum field theories see [@Turban1985; @Guimaraes1986; @Henkel1987; @Henkel1988; @Grimm1989; @Zhang1996; @Oshikawa1997] and [@Mussardo1994].
In a different approach to the above one could also define the generating function $M(\{\lambda_i\})=\langle e^{\sum_i\lambda_i\hat{o}_i}\rangle$, where the coefficients of the exponentials are the FP’s. This is a non-trivial example of a more widely known concept called full counting statistics which deals with the fluctuations of an observable defined in a subsystem. In most of the FCS problems one is interested in the moments of an observable $\hat{\mathcal{O}}$ defined for the entire subsystem, i.e. $\langle\hat{\mathcal{O}}^n\rangle$. Being a very natural concept FCS has been studied for a long time in different communities. It has been studied in the context of charge fluctuations [@Levitov1993; @Klich2014], Bose gases [@Gritsev2006; @Hofferberth2008; @Bastianello2018; @Piroli2018; @vanNieuwkerk2018], particle number fluctuations [@Song2012; @Vicari2012; @Rachel2012; @Herviou2017; @Herviou2019], quantum spin chains [@Eisler2003; @Demler2007; @Fendley2008; @Ivanov2013; @De-Chiara2016; @NR2017; @Collura2017; @Groha2018] and out of equilibrium quantum systems [@Essler2018; @vanNieuwkerk2018; @Collura2019].
It is not difficult to see that the FCS of an observable defined in a subsystem can be formulated as a formation probability problem by just considering the basis that diagonalizes the observable $\hat{\mathcal{O}}$. In this basis one can consider every eigenstate of the observable as a [*[configuration]{}*]{} and then the probability to find a particular value for the observable is just the FP for that [*[configuration]{}*]{}. This connection seems too formal to be useful, however, in this article we show that this point of view is extremely useful when one deals with quadratic observables in the study of FCS in the quadratic fermion Hamiltonians and the corresponding spin chains. The main advantage with respect to the more standard generating function point of view is that here one does not need to do the inverse Laplace transformation which is normally impossible to do analytically. Even numerical calculation of such kind of inverse transformation is usually extremely difficult because of the presence of a square root of a determinant in the final result of the generating function. Apart from its conceptual appeal our approach provides an explicit formula for the probabilities which is its main advantage with respect to the generating function method.
The paper is organized as follows: In the next section we first introduce the concept of FP and EFP in the quadratic Hamiltonians. Then we provide new determinant formulas for the FP and show explicitly how the FP is an EFP for the Hamiltonian with a defect, a line defect in the case of one dimensional systems. In section III we study the statistics of quadratic observables and show that this problem is related to the FP and ultimately EFP of a Hamiltonian with defect. A couple of examples including the number of particles and kinks will be discussed briefly. In section IV we study the transverse field XY chain with a staggered line defect. We find the exact correlation matrix of this Hamiltonian and argue that our result is valid for a generic translational invariant quadratic fermion Hamiltonian with a staggered line defect. In section V we use the general equations introduced in section III to calculate the probability distribution of the particle numbers and kinks in the transverse field XY chain. Finally in section VI we summarize our findings and comment on the future directions.
Formation probability as emptiness formation probability {#sec:2}
========================================================
Consider the following free fermion Hamiltonian with real generic couplings: $$\label{H1}\
\mathcal{H}_{\text{free}}(\textbf{A},\textbf{B})=\textbf{c}^{\dagger}\textbf{A}\textbf{c}+\frac{1}{2}\textbf{c}^{\dagger}\textbf{B}\textbf{c}^{\dagger}+\frac{1}{2}\textbf{c}\textbf{B}^{T}\textbf{c}-\frac{1}{2}{\rm Tr}{\textbf{A}},$$ where $\textbf{A}$ and $\textbf{B}$ are symmetric and anti-symmetric matrices respectively and $\textbf{c} \equiv (c_1,c_2,...,c_{|D|})$ with similar definition for $\textbf{c}^{\dagger}$. We define the correlation matrix of the ground state of the above Hamiltonian as: $$\begin{aligned}
\label{G normal}\
iG_{jk}&=&\langle \bar{\gamma}_j\gamma_k\rangle,\end{aligned}$$ where we defined the Majorana operators $\gamma_k=c_k+c_k^{\dagger}$ and $\bar{\gamma}_j=i(c_j^{\dagger}-c_j)$ and $\left\langle \right\rangle $ is normally the expectation value in the ground state. Note that here we have $\delta_{jk}=\langle \gamma_j\gamma_k\rangle=-\langle \bar{\gamma}_j\bar{\gamma_k}\rangle$. Using the Wick’s theorem which is valid for the eigenstates of the above Hamiltonian one can write all the other correlation functions with respect to these three basic correlation functions. Other interesting quantities such as entanglement entropy and FP’s can be also written as a function of the correlation matrix. Intuitively FP is defined as the probability of finding a particular configuration $C$ in a subsystem of the full system. It was shown in [@Najafi2016] that the result can be written with respect to the correlation matrix as follows: $$\begin{aligned}
\label{Main NR2015}\
P(C)=\det\frac{\mathbf{I}-\mathbf{G}}{2}\text{Min}[\mathbf{F}],\end{aligned}$$ where $\text{Min}\mathbf{F}$ is a particular principal minor of the matrix $\mathbf{F}=\frac{\mathbf{I}+\mathbf{G}}{\mathbf{I}-\mathbf{G}}$ derived after removing the rows and columns of the sites without any fermion in the configuration $C$. When there is no fermion in the configuration $C$ then the corresponding probability is called emptiness formation probability and we have $$\begin{aligned}
\label{emptiness}\
P(\{0\})=\det\frac{\mathbf{I}-\mathbf{G}}{2},\end{aligned}$$ whereas when all the sites are occupied with fermions we have $$\begin{aligned}
\label{fulliness}\
P(\{1\})=\det\frac{\mathbf{I}+\mathbf{G}}{2}.\end{aligned}$$ The above two formulas are very useful equations to do analytical calculations when there is a translational invariance. However, because of the presence of the minor in the equation (\[Main NR2015\]) analytical calculations do not seem to be feasible in more generic cases. One way to overcome this problem can be mapping the problem of FP to the problem of EFP and see the outcome. This procedure can be done as follows: Consider the formation probability, i.e. $P(C)$, of the ground state of the Hamiltonian $\mathcal{H}_{\text{free}}(\textbf{A},\textbf{B})$. This probability is equal to the EFP of the ground state of the Hamiltonian $\mathcal{H}_{\text{free}}(\textbf{A}',\textbf{B}')$ with a defect. In section IV we show for an explicit example how this procedure can be followed. Finally we have: $$\begin{aligned}
\label{formation-emptiness}\
P(C)=\det\frac{\mathbf{I}-\mathbf{G'}}{2},\end{aligned}$$ where $\mathbf{G'}$ is the correlation matrix of the Hamiltonian with defects. Following [@LIEB1961407] one can, in principle, find this correlation matrix numerically for any values of the coupling constants. However, analytical calculations are commonly feasible when we have translational invariance or some extra structure. For example, the Hamiltonian of a translational invariant (periodic) free fermions with time-reversal symmetry can be written as $$\label{Hamiltonian}\
H=\sum_{r=-R}^R\sum_{j\in\Lambda}a_rc_j^{\dagger}c_{j+r}+\frac{b_r}{2}(c_j^{\dagger}c_{j+r}^{\dagger}-c_jc_{j+r})+\text{const}.\hspace{0.33cm}$$ Using the Majorana operators one can also write, $$\begin{aligned}
\label{Hamiltonian-majorana}\
H=\frac{i}{2}\sum_{r=-R}^R\sum_{j\in\Lambda}t_r\bar{\gamma}_j\gamma_{j+r};\end{aligned}$$ where $t_r=-(a_r+b_r)$ and $t_{-r}=-(a_r-b_r)$. It is very useful to put the coupling constants as the coefficients of the following holomorphic function $f(z)=\sum_rt_rz^r$. Then the Hamiltonian can be diagonalized by going to the Fourier space and then Bogoliubov transformation as follows, see for example [@Jones2019]: $$\begin{aligned}
\label{Hamiltonian-diagonalization}\
H=\sum_q|f(q)|\eta_q^{\dagger}\eta_q+\text{const},\end{aligned}$$ where $\eta_q=\frac{1}{2}(1+\frac{f(q)}{|f(q)|})c_q^{\dagger}+\frac{1}{2}(1-\frac{f(q)}{|f(q)|})c_{-q}$ with $f(q):=f(e^{iq})$. Finally one can write the following explicit formula for the correlation matrix of the ground state: $$\begin{aligned}
\label{G matrix translational invariance}\
G_{jk}=\int_{0}^{2\pi}\frac{\text{d}q}{2\pi}\frac{f(q)}{|f(q)|}e^{iq(j-k)}.\end{aligned}$$ The above correlation matrix has a Toeplitz structure which makes it a suitable candidate for analytical calculations. It is possible to extend the above result to also excited states without much difficulty. When there is no translational invariance as it is the case for $\mathcal{H}_{\text{free}}(\textbf{A}',\textbf{B}')$ following the above procedure is not simple. An explicit calculations will be presented later for the transverse field XY chain with the staggered magnetization. In this work we will follow another path which is going to be one of our main results.
The basic idea of the second method is based on writing the Eq. (\[Main NR2015\]) as the Eq. (\[formation-emptiness\]). We show here that there are at least two different ways to do this. Consider a generic principal minor of a generic matrix $\mathbf{M}$. The basic idea is to write the principal minor as follows: $$\begin{aligned}
\label{main relations0}\
\text{Min}[\mathbf{M}]&=&\det\frac{\mathbf{M}(\mathbf{I}-\mathbf{I}_c)+\mathbf{I}+\mathbf{I}_c}{2};\\
\text{Min}[\mathbf{M}]&=&\det\frac{(\mathbf{I}-\mathbf{I}_c)\mathbf{M}+\mathbf{I}+\mathbf{I}_c}{2};\end{aligned}$$ where $\mathbf{I}$ is an identity matrix and $\mathbf{I}_c$ is a diagonal matrix made out of $\pm1$ which clearly depends on which columns and rows are getting removed. We set its diagonal element to minus one when we have a fermion and one when there is no fermion at the corresponding site. Now it is easy to show that we have $$\begin{aligned}
\label{main relations a}\
P(C)&=&\det\frac{\mathbf{I}-\mathbf{G}\mathbf{I}_c}{2},\\
\label{main relations b}
P(C)&=&\det\frac{\mathbf{I}-\mathbf{I}_c\mathbf{G}}{2}.\end{aligned}$$ The above equations means that the formation probability $P(C)$ is actually the EFP for a defect Hamiltonian with the correlation matrix $\mathbf{G}\mathbf{I}_c$ or $\mathbf{I}_c\mathbf{G}$. As we mentioned before none of these correlation matrices are necessarily the actual correlation matrix of the defect Hamiltonian $\mathcal{H}_{\text{free}}(\textbf{A}',\textbf{B}')$ introduced above. In fact we will show explicitly later that for the staggered Ising chain the correlation matrix has quite a different form. This should not be surprising because one can extract the minor of a matrix using quite different methods and although they all end up to the same number for $P(C)$, they have been derived from different matrices. However, clearly finding one is enough to get the others by proper manipulation of the rows and columns of the correlation matrix. For example, if $\mathbf{G}'$ is the actual correlation matrix then there is a similarity transformation $\mathbf{S}$ which we have $\mathbf{G}'=\mathbf{S}^{-1}\mathbf{G}\mathbf{I}_c\mathbf{S}$. For generic correlation matrices finding the $\mathbf{S}$ matrix is not necessarily an easy problem. It is worth mentioning that using the correlation matrices $\mathbf{G}\mathbf{I}_c$ or $\mathbf{I}_c\mathbf{G}$ leads to the same set of FP’s as the $\mathbf{G}$ matrix. In principle, the Eqs. (\[main relations a\]) and (\[main relations b\]) probably can be useful for analytical calculations when the $\mathbf{G}$ matrix is a Toeplitz matrix and the configuration $C$ has a pattern. In these cases the $\mathbf{G}\mathbf{I}_c$ has always a block Toeplitz structure. As an explicit example consider the ground state of the Hamiltonian (\[Hamiltonian\]) and let us focus on the probability of the configuration $C=(s_1,s_2,...,s_l)$, where $s_j=-1$ or $+1$ depending on the presence or the lack of a fermion at site $j$. Then we can write: $$\begin{aligned}
\label{defect G matrix }\
(\mathbf{G}\mathbf{I}_c)_{jk}=\text{sgn}_r(j,k)\int_{0}^{2\pi}\frac{\text{d}q}{2\pi}\frac{f(q)}{|f(q)|}e^{iq(j-k)},\\
(\mathbf{I}_c\mathbf{G})_{jk}=\text{sgn}_l(j,k)\int_{0}^{2\pi}\frac{\text{d}q}{2\pi}\frac{f(q)}{|f(q)|}e^{iq(j-k)},\end{aligned}$$ where the matrices $\textbf{sgn}_r$ and $\textbf{sgn}_l$ are the sign matrices and for example for a configuration with four sites have the following forms: $$\begin{aligned}
\label{sing matrices}\
\begin{split}
&\textbf{sgn}_l=\left(
\begin{array}{cccc}
s_1 & s_1 & s_1 & s_1 \\
s_2 & s_2 & s_2 & s_2 \\
s_3 & s_3 & s_3 & s_3 \\
s_4 & s_4 & s_4 & s_4
\end{array}
\right),\\
&\textbf{sgn}_r=\left(
\begin{array}{cccc}
s_1 & s_2 & s_3 & s_4 \\
s_1 & s_2 & s_3 & s_4 \\
s_1 & s_2 & s_3 & s_4 \\
s_1 & s_2 & s_3 & s_4
\end{array}
\right).
\end{split}\end{aligned}$$ The generalization for bigger sizes is strightforward. We note that when the configuration has a crystal structure the above matrices have block Toeplitz forms.
Statistics of a generic quadratic observable as a formation probability {#sec:3}
=======================================================================
In this section we argue that the problem of finding the statistics of a generic quadratic observable is essentially a FP problem and consequently the formulas derived in the previous section have much more applications than at first might appear. Consider the following quadratic observable $$\begin{aligned}
\label{Observable}\
\mathcal{O}_{D}=\textbf{c}^{\dagger}\textbf{M}\textbf{c}+\frac{1}{2}\textbf{c}^{\dagger}\textbf{N}\textbf{c}^{\dagger}+
\frac{1}{2}\textbf{c}\textbf{N}^{T}\textbf{c}-\frac{1}{2}{\rm Tr}{\textbf{M}},\end{aligned}$$ where $\textbf{M}$ and $\textbf{N}$ are symmetric and anti-symmetric matrices respectively. It is much more convenient to write the above observable in the following form: $$\begin{aligned}
\label{Observable2}\
\mathcal{O}_{D}=\frac{1}{2}(\textbf{c}^{\dagger}\,\,\textbf{c})\begin{pmatrix}
\textbf{M} & \textbf{N}\\
-\textbf{N} & -\textbf{M}\\
\end{pmatrix}
\begin{pmatrix}
\textbf{c}\\
\textbf{c}^{\dagger}\\
\end{pmatrix},\end{aligned}$$ Note that the above observable can have support just in a subsystem $D$ of the full system. With the statistics of this observable we mean if the full system is in the ground state or any other state what is the probability of finding a particular value if we measure the above quantity. We prove here that this is a FP problem. To show this we first diagonalize the observable $\mathcal{O}_{D}$ with the standard method of Ref. [@LIEB1961407], see Appendix A. The idea is based on a canonical transformation $$\begin{aligned}
\label{mat_U1}\
\begin{pmatrix}
\textbf{c} \\
\textbf{c}^{\dagger} \\
\end{pmatrix}=
\textbf{U}^{\dagger}
\begin{pmatrix}
\boldsymbol{\delta} \\
\boldsymbol{\delta}^{\dagger} \\
\end{pmatrix},\end{aligned}$$ which leads to $$\begin{aligned}
\label{Hdiag1}\
\mathcal{O}_{D}=\sum_{k}|\lambda_{k}|(\delta_{k}^{\dagger}\delta_{k}-\frac{1}{2}).\end{aligned}$$ The eigenvalues and the eigenvectors of the observable can be derived as usual by applying the modes on the [*[ground state]{}*]{} properties. The next step is to write the Hamiltonian with respect to the $\delta_{k}^{\dagger}$ and $\delta_{k}$ as follows: $$\label{Hamiltonian New Form}\
\mathcal{H}_{\text{free}}(\textbf{A},\textbf{B})=\frac{1}{2}(\boldsymbol{\delta}^{\dagger}\,\,\boldsymbol{\delta})\textbf{U}\begin{pmatrix}
\textbf{A} & \textbf{B}\\
-\textbf{B} & -\textbf{A}\\
\end{pmatrix}\textbf{U}^{\dagger}
\begin{pmatrix}
\boldsymbol{\delta}\\
\boldsymbol{\delta}^{\dagger}\\
\end{pmatrix},$$ The above equation can now be used to make the main argument. In the new basis if one calculates the FP with the new matrices the result is exactly equal to the probability of finding a particular value for the corresponding observable. For example, the EFP for the above Hamiltonian is exactly equal to the probability of finding the minimum value for the corresponding observable $\mathcal{O}_{D}$. To find the exact formula one first needs to calculate the correlation matrix $$\begin{aligned}
\label{G updated}\
iG'_{jk}&=&\langle \bar{\alpha}_j\alpha_k\rangle,\end{aligned}$$ where we defined the new Majorana operators $\alpha_k=\delta_k+\delta_k^{\dagger}$ and $\bar{\alpha}_j=i(\delta_j^{\dagger}-\delta_j)$. Having found the above correlation matrix the rest of the calculation is exactly as described in the previous section.
When the observable is defined for a subsystem one can also use the procedure that was outlined in [@NR2019] which is based on the reduced density matrix. When the system is in the ground state the reduced density matrix can be written as: $$\begin{aligned}
\label{Eq:reducedDM}
\rho_{D}&=&\det\frac{1}{2}(\mathbb{I}-\textbf{G})e^{\mathcal{H}},\\
\mathcal{H}&=&\frac{1}{2}(\textbf{c}^{\dagger}\,\,\textbf{c})\begin{pmatrix}
\textbf{M} & \textbf{N}\\
-\textbf{N} & -\textbf{M}\\
\end{pmatrix}
\begin{pmatrix}
\textbf{c}\\
\textbf{c}^{\dagger}\\
\end{pmatrix}+\frac{1}{2}{\rm Tr}\ln{(\textbf{F}_{s})},\end{aligned}$$ where $\mathcal{H}$ is the entanglement Hamiltonian and $$\begin{aligned}
\label{mat_T3}\
\begin{pmatrix}
\textbf{M} & \textbf{N}\\
\textbf{-N} & \textbf{-M}\\
\end{pmatrix}=
\ln \begin{pmatrix}
\textbf{F}_{s}-\textbf{F}_{a}\textbf{F}_{s}^{-1}\textbf{F}_{a} & \textbf{F}_{a}\textbf{F}_{s}^{-1}\\
-\textbf{F}_{s}^{-1}\textbf{F}_{a}& \textbf{F}_{s}^{-1}\\
\end{pmatrix},\end{aligned}$$ where $\textbf{F}_{a}=\frac{\textbf{F}-\textbf{F}^T}{2}$ and $\textbf{F}_{s}=\frac{\textbf{F}+\textbf{F}^T}{2}$ and as before $\mathbf{F}=\frac{\mathbf{I}+\mathbf{G}}{\mathbf{I}-\mathbf{G}}$. Note that the $\mathbf{G}$ matrix here is calculated for the original creation and annihilation operators appearing in the Hamiltonian. The idea is again based on writing the entanglement Hamiltonian in the basis that the observable is diagonal, i. e. $\delta$ basis. Now, we introduce the fermionic coherent states, i. e. $|\boldsymbol{\gamma}\rangle= |\gamma_1,\gamma_2,...,\gamma_{|D|}\rangle= e^{-\sum_{k=1}^{|D|}\gamma_{_{k}}\delta_k^{\dagger}}|0\rangle,$ where $\gamma_k$’s are Grassmann numbers with the following properties: $\gamma_n\gamma_m+\gamma_m\gamma_n=0$ and $\gamma_n^2=\gamma_m^2=0$. Here, $|D|$ is the number of sites in the region $D$. Then, one can write [@NR2019]: $$\label{rhod8}
\langle \boldsymbol{\gamma} |\rho_{D}|\boldsymbol{\gamma}'\rangle=
\det\frac{1}{2}(\mathbb{I}-\textbf{G})\left[\frac{{\rm \det}(\textbf{F}_{s})}{{\rm \det}(\tilde{\textbf{F}}_{s})}\right]^{\frac{1}{2}} \,\,e^{\frac{1}{2}(\bar{\boldsymbol{\gamma}}-\boldsymbol{\gamma'})\tilde{\textbf{F}}(\bar{\boldsymbol{\gamma}}+\boldsymbol{\gamma'})}\hspace{0.5cm}$$ where we defined $$\begin{aligned}
\label{BB_rel3}
\tilde{\textbf{F}}_{s}=e^{\tilde{\textbf{Y}}},\hspace{1cm}
\tilde{\textbf{F}}=\tilde{\textbf{X}}+e^{\tilde{\textbf{Y}}},\end{aligned}$$ with $$\label{matrixes}\
\tilde{\textbf{X}}=\tilde{\textbf{T}}_{12}(\tilde{\textbf{T}}_{22})^{-1},\hspace{0.15cm}
\tilde{\textbf{Z}}=(\tilde{\textbf{T}}_{22}^{-1})\tilde{\textbf{T}}_{21},\hspace{0.25cm}
e^{-\tilde{\textbf{Y}}}=\tilde{\textbf{T}}_{22}^{T},$$ where $$\begin{aligned}
\label{Ttilde}
\begin{split}
\tilde{\textbf{T}}&=\begin{pmatrix}
\tilde{\textbf{T}}_{11} & \tilde{\textbf{T}}_{12}\\
\tilde{\textbf{T}}_{21} & \tilde{\textbf{T}}_{22}\\
\end{pmatrix}\\
&=\textbf{U}\begin{pmatrix}
\textbf{F}_{s}-\textbf{F}_{a}\textbf{F}_{s}^{-1}\textbf{F}_{a} & \textbf{F}_{a}\textbf{F}_{s}^{-1}\\
-\textbf{F}_{s}^{-1}\textbf{F}_{a}& \textbf{F}_{s}^{-1}\\
\end{pmatrix}\textbf{U}^{\dagger}.
\end{split}\end{aligned}$$ The equation (\[rhod8\]) can be used to calculate all the desired probabilities using the same method that was developed in Ref.[@Najafi2016]. First of all, it is easy to see that to find the probability of finding the observable in its minimum value, one needs to put all the $\gamma$’s equal to zero, then we have $$\begin{aligned}
\label{probability of min}
p(o_{\text{min}})=\det\frac{1}{2}(\mathbb{I}-\textbf{G})\left[ \frac{{\rm \det}(\textbf{F}_{s})}{{\rm \det}(\tilde{\textbf{F}}_{s})}\right]^{\frac{1}{2}}.\end{aligned}$$ To find the probability of other values, say $o$, one needs to know the corresponding modes $\lambda_k$’s in which generate the desired eigenvalue of the observable and then, perform a Grassmann integral over the corresponding $\gamma_k$’s and put the other $\gamma$’s equal to zero. The result is $$\begin{aligned}
\label{generic-probability}
p(o)=\det\frac{1}{2}(\mathbb{I}-\textbf{G})\left[\frac{{\rm \det}(\textbf{F}_{s})}{{\rm \det}(\tilde{\textbf{F}}_{s})}\right]^{\frac{1}{2}}\sum_g \text{Min} \left[\tilde{\textbf{F}}\right],\end{aligned}$$ where $\text{Min} \left[ \tilde{\textbf{F}}\right]$ is the corresponding principal minor of the matrix $\tilde{\textbf{F}}$ and the sum takes care of the degeneracies. Note that one can again use the equation (\[main relations0\]) to get rid of the minor in the above equation. The above equation is very convinient to get explicit results for the probabilities without going through the generating function formalism.
Statistics of the number of particles {#sec:stat-number}
-------------------------------------
Statistics of the number of particles in a subsystem is the simplest possible example one can imagine because the observable itself is already diagonal. For earlier detailed studies regarding fluctuations of the particles from the generating function point of view see [@Song2012; @Vicari2012; @Rachel2012]. It is simple to see that the probability of finding no particle in the subsystem of size $l$ is exactly the EFP. Then the probability of finding one particle is just about summing over all the FP’s with just one fermion. In other words we need to first calculate the sum of the minors of rank one of the matrix $\mathbf{F}$. For generic case we need to find the sum of the minors of a particular rank of the matrix. There is a standard method to calculate these numbers which is called Faddeev-LeVerrier algorithm (see Wikipedia). The probability of having $k$ particles can be explicitly written as: $$\begin{aligned}
\label{probability of numbers}
P(k)=\det\left[ \frac{1}{2}(1-\boldsymbol{G})\right](-1)^kc_{l-k},\end{aligned}$$ where the exact form of the coefficients can be written as: $$\begin{aligned}
\label{coefficiants Bell polynomial}
\begin{split}
c_{l-k}=(-1)^{l-k}\frac{1}{k!}B_k(&\text{tr}\mathbf{F},-\text{tr}\mathbf{F}^2,2!\text{tr}\mathbf{F}^3,...\\
&...,(-1)^{k-1}(k-1)!\text{tr}\mathbf{F}^k),
\end{split}\end{aligned}$$ where $B_k$ is the complete exponential Bell polynomial (see Wikipedia). The complete exponential Bell polynomial can be written as: $$\label{complete Bell polynomial}\
B_n(x_1,x_2,...,x_{n-k+1})=\sum_{k=1}^nB_{n,k}(x_1,x_2,...,x_{n-k+1}),$$ where the partial exponential Bell polynomial $B_{n,k}$ is given by the equation.
$$\begin{aligned}
\label{Bell polynomial0}
B_{n,k}(x_1,x_2,...,x_{n-k+1})=\sum\frac{n!}{j_1!j_2!...j_{n-k+1}!}\left(\frac{x_1}{1!}\right)^{j_1}\left(\frac{x_2}{2!}\right)^{j_2}...\left(\frac{x_{n-k+1}}{(n-k+1)!}\right)^{j_{n-k+1}},\end{aligned}$$
where the sum is over all the non-negative $j_1,j_2,...,j_{n-k+1}$ in a way that we have $j_1+j_2+...+j_{n-k+1}=k$ and $j_1+2j_2+...+(n-k+1)j_{n-k+1}=n$.
The first few terms of the coefficients can be written as: $$\begin{aligned}
\label{coefficiants Bell polynomial examples}
c_{l}&=&1,\\
c_{l-1}&=&-\text{tr}\mathbf{F},\\
c_{l-2}&=&\frac{1}{2!}((\text{tr}\mathbf{F})^2-\text{tr}\mathbf{F}^2),\\
c_{l-3}&=&-\frac{1}{3!}((\text{tr}\mathbf{F})^3-3\text{tr}\mathbf{F}^2\text{tr}\mathbf{F}+2\text{tr}\mathbf{F}^3).\end{aligned}$$ Note that we also have $c_0=(-1)^l\det \mathbf{F}$. It is worth mentioning that using the properties of the Bell’s polynomial one can also write the following recursion relation for the probabilities: $$\begin{aligned}
\label{probabilities recursion}
P(k)=-\frac{1}{k}\sum_{j=1}^k(-1)^j\text{tr}\mathbf{F}^jP(k-j).\end{aligned}$$ The above formulas indicate that the problem of finding statistics of the number of particles boils down to the calculation of the trace of different powers of the matrix $\mathbf{F}$.
Statistics of the kinks {#SEC:kinks}
-----------------------
In this section we provide an example to show how the formalism of this section should be applied for a non-trivial observable. We would like to study the statistics of the following quantity: $$\begin{aligned}
\label{Kink}
\textbf{K}=\frac{l-1}{2}+\frac{1}{2}\sum_{j=1}^{l-1}(c_j^{\dagger}-c_j)(c_{j+1}^{\dagger}+c_{j+1})\end{aligned}$$ The reason that we call this quantity kink statistics comes from the spin representation of this quantity after Jordan-Wigner transformation, i.e. $c_j^{\dagger}=\prod_{l=1}^{j-1}\sigma_l^z\sigma_j^{+}$, which leads to $\textbf{K}=\frac{1}{2}\sum_{j=1}^{l-1}(1-\sigma_j^x\sigma_{j+1}^x)$. In this example we have: $$\label{Kink M and N}
M_{ij}=\frac{1}{2}\delta_{i,j+1}+\frac{1}{2}\delta_{i+1,j},\hspace{0.4cm}N_{ij}=\frac{1}{2}\delta_{i+1,j}-\frac{1}{2}\delta_{i,j+1}$$ Using the method of Appendix \[app1\] one can diagonalize $\textbf{K}$. The $\textbf{C}$ matrix is simply $C_{ij}=\delta_{ij}-\delta_{i,1}\delta_{j,1}$, and its eigenvectors $\psi_i(j)\equiv\psi_{ij}$ are chosen to be $\delta_{i,j}$ ($i$ being the label of the eigenvector). Showing the eigenvectors by $\kappa_i$, and using the Eq. \[phi and psi 1a\] for $\kappa_i\ne 0$, we easily show that $\phi_{i}(j)\equiv\phi_{ij}=\delta_{i,j+1}$. Using these eigenvectors the follwoing forms for $\textbf{g}$ and $\textbf{h}$ can be obtained: $$\begin{split}
g_{i,j}&=\frac{1}{2}\delta_{i,j}-\frac{1}{2}\delta_{i,j+1}+\frac{1}{2}\delta_{i,1}\delta_{j,l}\\
h_{i,j}&=\frac{1}{2}\delta_{i,j}+\frac{1}{2}\delta_{i,j+1}-\frac{1}{2}\delta_{i,1}\delta_{j,l}
\end{split}$$ and the corresponding $\textbf{U}$ is: $$\begin{aligned}
\label{mat_U2}\
\textbf{U}= \begin{pmatrix}
\textbf{g} & \textbf{h}\\
\textbf{h} & \textbf{g}\\
\end{pmatrix}.\end{aligned}$$ The diagonal form of $\textbf{K}$ is then: $$\textbf{K}=\sum_k \kappa(k)\eta_k^{\dagger}\eta_k$$ where $$\begin{split}
\kappa(k)=\left\lbrace \begin{matrix}
+1 & \ \ \text{if} &\ \ 1<k\leq l\\
0 & \ \ \text{if} &\ \ k=1
\end{matrix}\right.
\end{split}$$ Having the $\textbf{U}$ matrix now one can write the desired entanglement Hamiltonian in the basis of $\eta$ and then use the equation (\[generic-probability\]) to calculate the probability of having particular number of kinks in the ground state of the quantum spin chain. We note that here we can have $n=0,1,...,l-1$ number of kinks with the degeneracies 2${l-1}\choose{n}$.
Generating function and moments
-------------------------------
In this section following the same lines of thinking as above we give new formulas regarding the generating function and moments of an arbitrary quadratic observables. Compared to the results in [@NR2017] the new formulas have simpler forms. The generating function for an arbitrary operator $\mathcal{O}_{D}$ is defined as: $$\begin{aligned}
\label{generating function Definition}
%
M(z)=\text{tr}\left[ \rho_De^{z\mathcal{O}_{D}}\right].\end{aligned}$$ The trace can be calculated explicitly in the $\delta$ representation and after some manipulations the final result is $$\label{generating function 2}
M(z)=\det\frac{1}{2}(\mathbb{I}-\textbf{G})\left[{\rm \det}(\textbf{F}_{s})\right]^{\frac{1}{2}}
\det\left[\mathbb{I}+e^{z\Lambda}\tilde{\textbf{T}}\right]^{\frac{1}{2}};$$ where $\boldsymbol{\Lambda}=\begin{pmatrix}
|\boldsymbol{\lambda}| & 0\\
0 & -|\boldsymbol{\lambda}|\\
\end{pmatrix}$ is the matrix of the eigenvalues of the observable $\mathcal{O}_{D}$. After simple expansion we have $$\begin{aligned}
\label{generating function 3}
M(z)=\det\left[\mathbb{I}+\sum_{n=1}^{\infty}\frac{z^n}{n!}\tilde{\boldsymbol{\tau}}_n\right]^{\frac{1}{2}};\\
%M(0)&=&\det\frac{1}{2}(\mathbb{I}-\textbf{G})[{\rm \det}(\textbf{F}_{s})]^{\frac{1}{2}}
%\det[\mathbb{I}+\tilde{\textbf{T}}]^{\frac{1}{2}}=1\end{aligned}$$ where $\tilde{\boldsymbol{\tau}}_n=\boldsymbol{\Lambda}^n\frac{\tilde{\textbf{T}}}{\mathbb{I}+\tilde{\textbf{T}}}$. To calculate the moments we need to use the following formula[@Withers2010] $$\begin{aligned}
\label{T22}\
\det\left[\mathbb{I}+\sum_{n=1}^{\infty}\frac{z^n}{n!}\tilde{\boldsymbol{\tau}}_n\right]^{\frac{1}{2}}=1+\sum_{k=1}^{\infty}t_{k}\frac{z^{k}}{k!},\end{aligned}$$ where $$\begin{aligned}
\label{T222}\
t_{k}=\sum_{j=1}^{k}f_{j}B_{kj}(g);\end{aligned}$$ and $f_{j}=\frac{1}{2^j}$ and $B_{kj}(g)$ is the partial exponential Bell polynomial defined as $$\label{Bell polynomial}
\frac{1}{j!}\Big{(}\sum_{k=1}^{\infty}\frac{g_k\epsilon^k}{k!}\Big{)}^j=\sum_{k=j}^{\infty}B_{kj}(g_1,...,g_{k-j+1})\frac{\epsilon^k}{k!},\ j=0,1,2,...,$$ where $\epsilon$ is just a parameter. Here, we list the first few terms,
$$\begin{aligned}
\label{E with respect to g}
t_0&=&1,\\
t_1&=&f_1g_1, \\
t_2&=&f_1g_2+f_2g_1^2,\\
t_3&=&f_1g_3+f_2(3g_1g_2)+f_3g_1^3,\\
t_4&=&f_1g_4+f_2(4g_1g_3+3g_2^2)+f_3(6g_1^2g_2)+f_4g_1^4,
%t_5&=&f_1g_5+f_2(5g_1g_4+10g_2g_3)+f_3(10g_1^2g_3+15g_1g_2^2)+f_4(10g_1^3g_2)+f_5g_1^5,\end{aligned}$$
where $g=(g_1,g_2,...)$ with $$\begin{aligned}
\label{g}
g_k=\sum_{j=1}^k(-1)^{j-1}(j-1)!{\rm tr} B_{kj}(\boldsymbol{\tilde{\tau}}),\end{aligned}$$ where $\boldsymbol{\tilde{\tau}}=(\boldsymbol{\tilde{\tau}}_1,\boldsymbol{\tilde{\tau}}_2,...)$. The ${\rm tr} B_{kj}(\boldsymbol{\tilde{\tau}})$ can be calculated first by calculating $ B_{kj}(g)$ and then symmetrization of all the terms $G_1G_2..G_r\to \frac{1}{r!}\sum_r G_{\pi_1}G_{\pi_2}...G_{\pi_r}$, where the $G_i$’s are any sequence of the $\{g_k\}$ and the sum is over all permutations. After having a symmetrized form for $ B_{kj}(g)$, we can now replace $\{g_k\}$ with $\{\boldsymbol{\tilde{\tau}}_k\}$ and derive the formulas for ${\rm tr} B_{kj}(\boldsymbol{\tilde{\tau}})$. Here, we list a few of the coefficients,
$$\begin{aligned}
\label{g with respect to tau}
g_1&=&{\rm tr} \boldsymbol{\tilde{\tau}}_1,\\
g_2&=&{\rm tr}\left[\boldsymbol{\tilde{\tau}}_2-\boldsymbol{\tilde{\tau}}_1^2\right],\\
g_3&=&{\rm tr}\left[\boldsymbol{\tilde{\tau}}_3-3\boldsymbol{\tilde{\tau}}_1\boldsymbol{\tilde{\tau}}_2+2\boldsymbol{\tilde{\tau}}_1^3\right],\\
g_4&=&{\rm tr}\left[\boldsymbol{\tilde{\tau}}_4-4\boldsymbol{\tilde{\tau}}_1\boldsymbol{\tilde{\tau}}_3-3\boldsymbol{\tilde{\tau}}_2^2+12\boldsymbol{\tilde{\tau}}_1^2
\boldsymbol{\tilde{\tau}}_2-6\boldsymbol{\tilde{\tau}}_1^4\right],\end{aligned}$$
Finally we have $$\begin{aligned}
\label{T222}\
E_m=\langle\mathcal{O}_{D}^m\rangle=t_m.\end{aligned}$$ The first two moments can be written as follows: $$\begin{aligned}
\label{explicit two moments}\
E_1&=&\frac{1}{2}{\rm tr} \boldsymbol{\tilde{\tau}}_1,\\
E_2&=&\frac{1}{2}{\rm tr}\left[\boldsymbol{\tilde{\tau}}_2-\boldsymbol{\tilde{\tau}}_1^2\right]+\frac{1}{4}{\rm tr^2} \boldsymbol{\tilde{\tau}}_1\end{aligned}$$ It is also possible to get simple formulas for the fluctuations around the average: $$\begin{aligned}
\label{shifted ones}\
\tilde{E}_{m}=\langle (\mathcal{O}_{D}-\langle\mathcal{O}_{D}\rangle)^{m}\rangle.\end{aligned}$$ The first few terms are: $$\begin{aligned}
\label{Hl1}
\tilde{E}_1&=& 0, \nonumber\\
\tilde{E}_2&=&\frac{g_2}{2}, \nonumber\\
\tilde{E}_3&=&\frac{g_3}{2}, \nonumber\\
\tilde{E}_4-3\tilde{E}_2^2&=&\frac{g_4}{2}, \nonumber\\
\tilde{E}_5-10\tilde{E}_2\tilde{E}_3&=&\frac{g_5}{2}. \nonumber\\\end{aligned}$$ The most general case can be written as, $$\label{ E tilde general}
\sum_{j=1}^m(-1)^{j-1}(j-1)!B_{mj}(\tilde{E}_1,\tilde{E}_2,...,\tilde{E}_{m-j+1})=\frac{g_m}{2}.$$ Finally one can also write the comulants defined as: $$\begin{aligned}
\label{ comulant}
\kappa_m=\frac{d^m}{dz^m}\log M(z)|_{z=0}\end{aligned}$$ with respect to the $E_m$ and $\tilde{E}_m$ as follows (see Wikipedia): $$\label{ comulants}
\kappa_m=\sum_{j=1}^m(-1)^{j-1}\left\lbrace \begin{matrix}
(j-1)!B_{mj}(0,\tilde{E}_2,...,\tilde{E}_{m-j+1})\\
(j-1)!B_{mj}(E_1,E_2,...,E_{m-j+1})
\end{matrix}\right.$$ where $m>0$ for the first branch and $m>1$ for the second one. The above formulas have relatively simpler form than the ones presented in [@NR2017]. However, we note that here we assume that the observable should be first diagonalized and then the moments should be calculated. In the approach of [@NR2017] no diagonalization is needed.
Transverse field XY chain with a staggered line defect {#sec:4}
======================================================
In this section we would like to provide an exact determinant formula for the formation probability of the Neel sub-configuration, i.e. $\left| \downarrow\uparrow\downarrow\uparrow...\right\rangle$, in the XY chain by a direct method, that is the configuration in which the spins are in the Neel state for the sites lying in the interval $[1,n]$. In principle the formulas (\[main relations a\]) and (\[main relations b\]) are explicit examples which lead to block Toeplitz matrices. However, the direct method has this advantage that one can get an exact formula for the correlation matrix of the line defect problem which can be useful for its own sake. In this section we first explicitly show that the FP of the Neel configuration is the EFP for the staggered XY Hamiltonian. Then we solve the line defect Hamiltonian by using the J-W transformation and find the exact correlation matrix and EFP for the ground state. The Hamiltonian of the transverse field $XY$ chain is: $$H_{XY}=-J\sum_{l=1}^{L} \left( \frac{1+\gamma}{4}\sigma_l^x\sigma_{l+1}^x+\frac{1-\gamma}{4}\sigma_l^y\sigma_{l+1}^y\right) -\frac{h}{2}\sum_{l=1}^L\sigma_l^z,$$ where $\sigma_l^i$ ($i=x,y,z$) are the Pauli matrices, $J$ is the exchange parameter and $h$ is the magnetic field. The FP of the Neel state for the interval of length $n$ (which we take it always even) is readily found to be $$\begin{split}
&P_{\uparrow\downarrow}(n)=\\
&\left\langle g\right| \left( \frac{1-\sigma_1^z}{2}\right) \left( \frac{1+\sigma_2^z}{2}\right) \left( \frac{1-\sigma_3^z}{2}\right) ...\left( \frac{1+(-1)^n\sigma_n^z}{2}\right)\left| g\right\rangle\\
&=\left\langle g'\right| \prod_{j=1}^n\frac{1-\sigma_j^z}{2}\left| g'\right\rangle,
\end{split}$$ where $\left| g\right\rangle$ is the ground state of the $XY$ chain, $\left| g'\right\rangle\equiv P_x^{(n)} \left| g\right\rangle$, and $P_x^{(n)}$ is the projection operator defined by $\prod_{j=1}^{\frac{n}{2}}\sigma_{2j}^x$, satisfying the relation $\left( P_x^n\right)^2=1$. One can easily check that $\left| g'\right\rangle$ is the ground state of $H'_{XY}\equiv P_x^{(n)}H_{XY}P_x^{(n)}$ with the same ground state energy as $H_{XY}$. Therefore, $P_{\uparrow\downarrow}(n)$ is the EFP (represented by $P(n)$ defined as the probability that all spins are down) of the ground state of $H'_{XY}$. After applying $P_x^{(n)}$, we find that the explicit form of $H'_{XY}$ is: $$\begin{split}
H'_{XY}=& -J\sum_{l=1}^{L} \left( \frac{1+\gamma}{4}\sigma_l^x\sigma_{l+1}^x+\frac{1-\gamma}{4}f_n(l) \sigma_l^y\sigma_{l+1}^y\right)\\
& -\frac{1}{2}\sum_{l=1}^Lh_n(l)\sigma_l^z,
\end{split}
\label{Eq:modifiedXY}$$ where $f_n(l)=-1$ and $h_n(l)=(-1)^{l+1}h$ for the case $l\leq n$, and $f_n(l)=+1$ and $h_n(l)=+h$ for the case $l>n$. To work with fermionic Hamiltonian corresponding to the spin chain, we use J-W transformation defined by $$\begin{split}
&c_l^{\dagger}\equiv \prod_{j<l}\sigma_j^z\sigma_l^+,\\
&c_l\equiv \prod_{j<l}\sigma_j^z\sigma_l^-,
\end{split}
\label{Eq:JW}$$ where $\sigma_l^+\equiv\frac{1}{2}\left(\sigma^x+i\sigma^y\right)$, and $\sigma_l^-\equiv\frac{1}{2}\left(\sigma^x-i\sigma^y\right)$. The transformed Hamiltonian then becomes:
$$\begin{split}
H'_{XY}=&\frac{1}{2}\sum_{l=1}^{L-1}\left[ J_n(l) \left( c_l^{\dagger}c_{l+1}-c_lc_{l+1}^{\dagger} \right) +J\gamma_n(l) \left( c_l^{\dagger}c_{l+1}^{\dagger}-c_lc_{l+1}\right)\right]-\frac{NJ}{2} \left( c_L^{\dagger}c_1+\gamma c_L^{\dagger}c_1^{\dagger}+H.C.\right) -\sum_{l=1}^Lh_n(l)c_l^{\dagger}c_l+const.,
\end{split}$$
where $const.=\frac{1}{2}\sum_{l=1}^Lh_n(l)$, and $$\begin{split}
&J_n(l)\equiv \left\lbrace \begin{matrix}
J & \text{if} \ l>n\\
\gamma J & \text{if} \ l\leq n
\end{matrix}\right. \ , \ \gamma_n(l)\equiv \left\lbrace \begin{matrix}
\gamma & \text{if} \ l>n\\
1 & \text{if} \ l\leq n
\end{matrix}\right. .
\end{split}
\label{Eq:gamma-J}$$ Note that this Hamiltonian is identical to the free fermionic Hamiltonian corresponding to the $XY$ model (without staggered interval) outside the staggered interval as expected. In the above equations $H.C.$ is the Hermitian conjugate terms and $N$ is the eigenvalue of $\hat{N}\equiv\prod_{j\leq L}\left(2c_j^{\dagger}c_j-1\right)$ (note that $c_{L+1}=-Nc_1$). In the $\sigma^z$ basis, if the number of down spins is odd (or equivalently the odd number of fermionic *vacancies*), then $N\equiv -1$ (corresponding to the periodic boundary conditions in the fermionic representation, i.e. *Ramond (R) sector*), and in the other case (even number of down spins) $N\equiv +1$ (corresponding to the antiperiodic boundary conditions, i.e. *Neveu-Schwartz (NS) sector*).\
For the $R$-sector ($N=-1$) the fermionic Hamiltonian becomes periodic as follows (ignoring the constant term):
$$\begin{split}
H'_{XY}=&\frac{1}{2}\sum_{l=1}^L\left[ J_n(l) \left( c_l^{\dagger}c_{l+1}-c_lc_{l+1}^{\dagger} \right) +J\gamma_n(l) \left( c_l^{\dagger}c_{l+1}^{\dagger}-c_lc_{l+1}\right)\right]-\sum_{l=1}^Lh_n(l)c_l^{\dagger}c_l+\frac{1}{2}\sum_{l=1}^Lh_n(l),
\label{Eq:effectiveH1}
\end{split}$$
whereas for the $N=+1$ case, the periodicity is destroyed. To retrieve the periodicity, we can use the transformation $\bar{c}_l=\exp\left[ i\frac{\pi(N+1)}{2L}l\right]c_l $, which results to $\bar{c}_{L+1}=\bar{c}_1$. The cost of this operation is that the allowed momentums become half-integer multiplications of $\frac{2\pi}{L}$. In the followings we denote the fermionic operators for both cases by $c$ and $c^{\dagger}$, keeping in mind that for the $R$-sector the momentums should be integer multiplications of $\frac{2\pi}{L}$, whereas for the *NS*-sector they should be half-integers.\
To proceed in finding the staggered spin probability, it is first useful to represent the same for the EFP for the ordinary transverse field $XY$ chain which is well-studied in the literature [@Shiroishi2001; @Franchini2003; @Franchini2005; @Viti2019]. It can be found using the equation (\[emptiness\]). To make contact with the notation of [@Franchini2005] it is useful to define the matrix $\mathbf{S}=\frac{\mathbb{I}-\mathbf{G}^T}{2}$, where $S_{ij}(n)=s^{(1)}_{ij}+is^{(2)}_{ij}$, $s^{(1)}_{ij}\equiv\left\langle c_ic^{\dagger}_j\right\rangle$, and $s^{(2)}_{ij}\equiv i\left\langle c_ic_j\right\rangle$. Then the EFP of the $XY$ model (i.e. with no staggered interval involved) is shown to be: $$P(n)|_{H_{XY}}=\left|\text{Det}(\textbf{S}(n)) \right|.$$ Additionaly for this case, using the exact forms of the correlation functions one readily finds: $$\begin{split}
\textbf{S}^{\text{free}}_{j,k}(n)&=\frac{1}{2}\delta_{ij}+\int_0^{2\pi}\frac{\text{d}q}{2\pi}\sigma(q)e^{iq(j-k)},\\
\sigma(q)&=\frac{1}{2}\left(\frac{\cos q-h-i\gamma\sin q}{\sqrt{(\cos q-h)^2+\gamma^2\sin^2q}} \right),
\end{split}
\label{Eq:sigma0}$$ from which we see that $$G_{jk}=-\int_0^{2\pi}\frac{\text{d}q}{2\pi}\sigma(q)e^{-iq(j-k)}.$$ The matrix $\textbf{S}(n)$ for the ordinary $XY$ model is a Toeplitz matrix, for which the EFP as the determinant of $\textbf{S}(n)$ can be found in the thermodynamic limit using the Fisher-Hardwig technique [@Franchini2005]. Also using the fact that $a_1=a_{-1}=\frac{J}{2}$, $b_1=\frac{\gamma}{2}$, and $a_0=-h$ in Eq. (\[Hamiltonian\]), one readily finds $$\frac{f(q)}{\left| f(q) \right|}=-\frac{J\cos q-h+i\gamma\sin q }{\Lambda(q)}=- \sigma(q)^*$$ which (noting that the $G$ matrix is real) is compatible with the Eq. (\[G matrix translational invariance\]). For the present case (with the ground state $\left| g'\right\rangle$) we note that $P_{\uparrow\downarrow}(n)|_{H_{XY}}= P(n)|_{H'_{XY}}$, and therefore one should find a way to diagonalize $H'_{XY}$. To this end, we use the recipe due to Lieb *et. al.* [@LIEB1961407], according to which, after writing the Hamiltonian in the form $ H'_{XY}=\sum_{i,j}\left[c_i^{\dagger}A_{i,j}c_j +\frac{1}{2}\left(c_i^{\dagger}B_{i,j}c_j^{\dagger}+ h.c.\right) \right]$, one finds the energy spectrum by diagonalizing $\textbf{C}=(\textbf{A}-\textbf{B})(\textbf{A}+\textbf{B})$. The details of this calculation can be found in the Appendices (\[SEC:diagonalization\]) and (\[SEC:solution\]). The matrix $\textbf{C}$ for $H'_{XY}$ has the following form: $$\begin{split}
&C_{ii}=h^2+\frac{1}{2}J^2(1+\gamma^2),\\
&C_{i,i+1}=\left\lbrace \begin{matrix}
(-1)^i hJ & \text{if} \ \ i\leq n\\
-hJ & \text{if} \ \ i> n
\end{matrix}\right. \\
&C_{i,i+2}=\left\lbrace \begin{matrix}
-\frac{1}{4}J^2(1-\gamma^2) & \text{if} \ \ i\leq n\\
\frac{1}{4}J^2(1-\gamma^2) & \text{if} \ \ i> n
\end{matrix}\right.
\end{split}$$ also $C_{i,j}=-hJ$ if $ i=1,j=L$ or $i=L,j=1$ and $C_{i,j}=0$ for other cases. Following Ref. [@LIEB1961407] one can find the eigenvalues and the eigenvectors of $C$, the eigenvalues of which is represented by $\Lambda_k^2$, being the square of the energy spectrum of the system without the line defect, see Appendices (\[SEC:diagonalization\]) and (\[SEC:solution\]) for detailes. Since $\textbf{A}$ ($\textbf{B}$) is symmetric (antisymmetric), $(\textbf{A}-\textbf{B})(\textbf{A}+\textbf{B})$ and $(\textbf{A}+\textbf{B})(\textbf{A}-\textbf{B})$ are symmetric and their eigenvalues are real and the eigenvectors $\psi$ and $\phi$ (see Eqs. (\[phi and psi 2a\]) and (\[phi and psi 2b\])) can be chosen to be orthogonal. To find such solutions, we use the trial function: $$\begin{split}
\psi_{kj}\equiv &\frac{1}{2}\left[ 1+(-1)^j\right] \left(a_1\sin km_j+a_2\cos km_j \right)+\\
&\frac{1}{2}\left[ 1-(-1)^j\right] \left(a_3\sin km_j+a_4\cos km_j \right),
\end{split}
\label{Eq:pairfunction}$$ where $k$ labels the eigenfunctions, $j$ is the number of the sites in the real space, $m_j\equiv \left[\frac{j+1}{2} \right] $, and $a_1,a_2,a_3$, $a_4$ are the coefficients which have to be fixed using the Eq. (\[phi and psi 2a\]). This pairing mechanism facilitates the calculations (note that $\sin$ and $\cos$ are exact solutions for $m_j\gg m_0\equiv\frac{n}{2}$, and $m_j\ll m_0$). The strategy is as follows: We find two kinds of solutions: one for deep inside the staggered interval (DISI), and the other for deep outside the staggered interval (DOSI) with different coefficients. Then we should glue them by fulfilling the requirements at $j=1$ and $j=n$, i.e. where the staggered interval begins and ends respectively. This has been done in Appendix (\[SEC:solution\]) by details. There we show that the solutions at DISI and DOSI are the same, up to a phase shift $k_s\rightarrow k_s-\pi$ where $k_s=\frac{4\pi}{L}s$ is the twice of the real momentum $q_s$ which is $q_s\equiv\frac{2\pi}{L}s$ for $N=-1$ (R sector) and $q_s\equiv\frac{2\pi}{L}(s+\frac{1}{2})$ for $N=+1$ (NS sector) and $s$ runs over $-\frac{L}{2},-\frac{L}{2}+1...,\frac{L}{2}-1$. After applying the boundary conditions, one reach finally to the following function which diagonalizes $\textbf{C}$ (see Appendix (\[SEC:solution\])): $$\begin{split}
&\psi_{sj}=\sqrt{\frac{2}{L}}\left\lbrace\begin{matrix}
-(-1)^{m_j}\cos q_s\left[j-n\right] & j\leq n\\
\cos q_s\left[j-n\right] & j>n
\end{matrix}\right.
\end{split}$$ with the eigenvalues $\Lambda_q^2=(J\cos q_s-h)^2+\gamma^2\sin^2 q_s$. The other independent solution is obtained by replacing $\cos$ by $\sin$. We consider the above solution for $q_s\leq 0$, and the $\sin$ solution for $q_s>0$. To continue we should find the other solution ($\phi_{sj}$) which can easily be obtained using Eq. (\[phi and psi 1b\]), for $\Lambda_q\neq 0$:
$$\begin{split}
\phi_{XY}(j\leq n)=&-\sqrt{\frac{2}{L}}\Lambda_{q_s}^{-1}\left[(-)^{m_j+a_j}h\cos q_s(j-n)\right. \\
&\left. +(-)^{m_{j-1}}\left(\frac{-1+\gamma}{2}\right)\cos q_s(j-n-1)+(-)^{m_{j+1}}\left(\frac{1+\gamma}{2}\right) \cos q_s(j-n+1) \right], \\
\phi_{XY}(j= n+1)=&-\sqrt{\frac{2}{L}}\Lambda_{q_s}^{-1}\left[h\cos q_s(j-n)\right. \\
&\left. (-)^{m_{j-1}}\left( \frac{-1+\gamma}{2}\right) \cos q_s(j-n-1)-\frac{1+\gamma}{2}\cos q_s(j-n+1) \right], \\
\phi_{XY}(j> n+1)=&-\sqrt{\frac{2}{L}}\Lambda_{q_s}^{-1}\left[h\cos q_s(j-n)\right. \\
&\left. +\frac{-1+\gamma}{2}\cos q_s(j-n-1)-\frac{1+\gamma}{2}\cos q_s(j-n+1) \right],
\end{split}$$
where $a_j$ (not to be confused with the coefficients $a_1$, $a_2$, $a_3$ and $a_4$) is $1$ if $j$ belongs to the first sublattice (odd $j$s) and $0$ for the other sublattice (even $j$s). This solution is reserved for $s\leq 0$, and for $q_s> 0$ one should replace $\cos$ by $\sin$. For $\Lambda_k=0$ the solution is $\phi(j)=\pm \psi(j)$. Having $\psi$ and $\phi$ solutions in hand, one can directly calculate $g_{si}\equiv \frac{1}{2}\left( \psi_{si}+\phi_{si}\right) $ and $h_{si}\equiv \frac{1}{2}\left(\psi_{si}-\phi_{si}\right)$ to diagonalize $H$, see relation \[Hdiag\]. For example $\left\langle c_ic_j\right\rangle=\sum_sg_{si}h_{sj}$ where $s$ is integer (half-integer) for $N=-1$ $(N=+1)$. Working out with $f_{AB}^{i,j}=\sum_{s=-\frac{L}{2}}^{L/2-1}A_{si}B_{sj}$, in which $A,B=\psi, \phi$, one can easily show that always (irrespective to the amount of $i$ and $j$ being inside or outside the staggered interval) $f_{\psi\psi}^{i,j}=f_{\phi\phi}^{i,j}=\delta_{ij}$, and $\delta_{ij}$ is the Kronocker delta. These functions help us to calculate the important correlation functions:\
$$\begin{split}
&\left\langle c_ic_j^{\dagger} \right\rangle =\frac{1}{4}\left[ f_{\psi\psi}+f_{\psi\phi}+f_{\phi\psi}+f_{\phi\phi}\right],\\
&\left\langle c_ic_j \right\rangle =\frac{1}{4}\left[f_{\psi\psi}-f_{\psi\phi}+f_{\phi\psi}-f_{\phi\phi}\right], \\
&\left\langle c_i^{\dagger}c_j^{\dagger} \right\rangle=\frac{1}{4}\left[f_{\psi\psi}+f_{\psi\phi}-f_{\phi\psi}-f_{\phi\phi}\right], \\
&\left\langle c_i^{\dagger}c_j \right\rangle =\frac{1}{4}\left[f_{\psi\psi}-f_{\psi\phi}-f_{\phi\psi}+f_{\phi\phi}\right],
\end{split}
\label{Eq:correlations0}$$
We now calculate the correlation functions explicitly. In what fallows, we consider the case in which $i,j\leq n$, and extension to the other cases is straightforward. To calculate the correlation functions, we need the following identity that has been proved in Appendix (\[SEC:corr\_func\]): $$\begin{split}
\frac{1}{4}\left( f_{\psi\phi}\pm f_{\phi\psi}\right) =\chi_{ij}^{\pm}\sigma_1(i,j)+\chi_{ij}^{\mp}\sigma_2(i,j),
\end{split}$$ where $\chi_{ij}^{+}=\left[\frac{1+(-)^{a_i-a_j}}{2} \right](-)^{m_{j+1}-m_i}$ and $\chi_{ij}^-=\left[\frac{1-(-)^{a_i-a_j}}{2} \right](-)^{m_{j+1}-m_i}$, and also $$\begin{split}
\sigma_1(j,k)&=\frac{1}{2L}\sum_s\cos q_s(k-j)\left( \frac{-h+\cos q_s}{\Lambda_s}\right)\\
&=\frac{1}{2L}\sum_s\left( \frac{-h+\cos q_s}{\Lambda_s}\right)e^{-iq_s(k-j)},\\
\sigma_2(j,k)&=\frac{1}{2L}\sum_s\sin q_s(k-j)\left( \frac{\gamma\sin q_s}{\Lambda_s}\right)\\
&=\frac{i}{2L}\sum_s\left( \frac{\gamma\sin q_s}{\Lambda_s}\right)e^{-iq_s(k-j)},
\end{split}
\label{Eq:sigma}$$ where we have used the symmetry considerations to add extra zero contributions, and for the summation, $s$ is integer for the $R$-sector, and half-integer for the $NS$-sector. Therefore, if $i$ and $j$ belong to the same sublattice then we have: $$\begin{split}
&\Sigma_1(i,j)\equiv\frac{1}{4}\left( f_{\psi\phi}+ f_{\phi\psi}\right)=(-)^{m_{j+1}-m_i}\sigma_1(j-i), \\
&\Sigma_2(i,j)\equiv\frac{1}{4}\left( f_{\psi\phi}- f_{\phi\psi}\right)=(-)^{m_{j+1}-m_i}\sigma_2(j-i),
\end{split}$$ and $$\begin{split}
&\Sigma_1(i,j)=(-)^{m_{j+1}-m_i}\sigma_2(j-i), \\
&\Sigma_2(i,j)=(-)^{m_{j+1}-m_i}\sigma_1(j-i),
\end{split}$$ if they belong to the different sublattices. Therefore, at this stage we can find the explicit form of the correlation functions, which are $$\begin{split}
\left\langle c_ic_j^{\dagger} \right\rangle &=\frac{1}{2}\delta_{ij}+\chi_{ij}^+\sigma_1(i,j)+\chi_{ij}^-\sigma_2(i,j),\\
\left\langle c_ic_j \right\rangle &=\chi_{ij}^+\sigma_2(i,j)+\chi_{ij}^-\sigma_1(i,j).
\end{split}$$ Also note that $\left\langle c_i^{\dagger}c_j^{\dagger} \right\rangle =
-\left\langle c_ic_j \right\rangle$ and $\left\langle c_i^{\dagger}c_j \right\rangle =
\delta_{ij}-\left\langle c_ic_j^{\dagger} \right\rangle$, that can be readily checked.
Finally we turn to the calculation of the formation probability of the Neel configuration which is the EFP for $H'_{XY}$. It can be determined by calculating $|\text{Det}(\textbf{S}_n)|$, the elements of which is $s_{ij}=\left\langle c_ic_j^{\dagger} \right\rangle - \left\langle c_ic_j \right\rangle $ as outlined above. Using the above correlation functions, one simply obtains: $$s_{jk}=\frac{1}{2}\delta_{jk}+(\chi_{jk}^+-\chi_{jk}^-)\sigma(j,k),$$ where $$\begin{split}
\sigma(j,k)&\equiv \sigma_1(j,k)-\sigma_2(j,k)\\
&=\frac{1}{2L}\sum_{s=-L/2}^{L/2-1}\left( \frac{\cos q_s-h-i\gamma\sin q_s}{\Lambda_s}\right)e^{iq_s(j-k)}.
\end{split}$$ This matrix, in the $L\rightarrow \infty$ limit becomes $$\begin{split}
&\sigma(j,k)=\frac{1}{2}\int \frac{\text{d}q}{2\pi} \sigma(q)e^{iq(j-k)},
\end{split}$$ where $\sigma(q)$ defined in the Eq. (\[Eq:sigma0\]). $s_{jk}$ is compatible with the result for the free case, i.e. Eq. (\[Eq:sigma0\]), except that here a sign matrix ($\text{sgn}(j,k)\equiv\chi_{jk}^+-\chi_{jk}^-$) is multiplied. The closed form of this sign matrix is: $$\begin{split}
\text{sgn}(j,k)=&\cos\pi(k-j)\left\lbrace -(-)^{\frac{j+k}{2}}\left| \cos \frac{\pi}{2}(k-j)\right|\right. \\
&\left. +(-)^{\frac{j-k-1}{2}}\left| \sin \frac{\pi}{2}(k-j)\right| \right\rbrace,
\end{split}$$ For example, the explicit form for the above sign matrix for $n=8$ is
$$\begin{split}
\textbf{sgn} =\left[ \begin{matrix}
+ & - & - & + & + & - & - & + \\
+ & - & - & + & + & - & - & + \\
- & + & + & - & - & + & + & -\\
- & + & + & - & - & + & + & -\\
+ & - & - & + & + & - & - & + \\
+ & - & - & + & + & - & - & + \\
- & + & + & - & - & + & + & -\\
- & + & + & - & - & + & + & -
\end{matrix} \right].
\end{split}$$
One can easily check that the above sign matrix is different from the ones suggested in (\[sing matrices\]). As we already discussed in section II there are different sign matrices that lead to the same FP’s but they come from different correlation matrices. It is worth mentioning that since the Hamiltonian of the transverse field XY chain that we considered here was with PBC we ended up to have R and NS sectors for the fermionic counterparts. Finding which one is the actual ground state of the spin system is a non-trivial problem. Since our line defect problem has the same eigenvalues the problem is similar to the clean case and we refer to [@DePasquale2009] for systematic study of the clean case. Note that the calculated sign matrix is not only correct for the ground state of R and NS sectors of the XY chain but also for the generic one dimensional translational invariant free fermions.
Probability distribution of particle numbers and kinks in the transverse field XY chain {#sec:5}
=======================================================================================
In this section we provide a couple of examples to show how the explicit formulas that we provided in the previous sections can be applied to calculate the probability distribution of quadratic observables. In both cases the model that we take is the ground state of the transverse field XY chain which has a rich phase diagram with three critical lines at $h=1$, $\gamma\neq0$; $h=-1$, $\gamma\neq0$ and $\gamma=0$, $-1<h<1$. Here we concentrate mostly on the non-negative transverse field part of the phase diagram and study probability distribution of particle numbers and kinks in the transverse field XY chain.
Probability distribution of magnetization {#sec:5A}
-----------------------------------------
The first example is the distribution of the magnetization in the $\sigma^z$ direction in an interval of size $l$. This is equivalent to the distribution of the number of particles that we have studied in the section \[sec:stat-number\]. The generating function of this quantity in the thermodynamic limit has been already studied in [@Ivanov2013], however, it does not seem to be straightforward to do the inverse Laplace transform analytically or numerically in the most generic cases. Using the formulas of the section \[SEC:kinks\] with appropriate $\textbf{G}$ and $\textbf{F}$ matrices we can easily calculate this distribution numerically for arbitrary parameters of the Hamiltonian. The results are depicted in the Figure \[fig:1\] which show clear change of bahavior when we cross the critical line $h=1$. The emergent oscillations in the region $h\geq1$ are similar to the ones that have already been seen in the study of the EFP in [@Franchini2005] and attributed to the competition between the energy cost of flipping a spin (controled by $h$) and the superconducting terms which create and distroy fermions in pairs (controlled by $\gamma$). Although not shown here, similar oscillations also appear in the region $-1\geq h$ with peaks shifted to the left part of the graph. Since on the line $\gamma=0$ we have a $U(1)$ symmetry the number of particles is fixed and consequently the distribution is just a Dirac delta function. For small $h$ the effect of increasing $\gamma$ is just broadening the distribution by increasing the variance.
![Probability distribution of the magnetization in the $\sigma^z$ direction in terms of $\gamma$ and $h$ for the XY-chain with $L=200$ and $l=30$.[]{data-label="fig:1"}](1.pdf){width="55.00000%"}
![Probability distribution of the kinks in the $\sigma^x$ direction for the XY-chain ($L=200$ and $l=18$) using the exact diagonalization outlined in SEC \[sec:3\]. The graph shows the dependence on $\gamma$ and $h$.[]{data-label="fig:2"}](2.pdf){width="55.00000%"}
![Illustration of the Kramers-Wannier(KW) duality. The figure shows the probability distribution of the magnetization ($p_{particles}$) and the kinks ($p_{kink}$) for $\gamma=1$ and $h=2$ and $\tilde{h}=\frac{1}{h}=0.5$ for $L=200$ and $l=18$.[]{data-label="fig:3"}](3.pdf){width="50.00000%"}
Probability distribution of kinks {#sec:5B}
---------------------------------
As a second example of our formalism we discuss the distribution of kinks in the ground state of the XY-chain. Using the method that was provided in the section \[SEC:kinks\], we calculated with exact numerical calculations the probability of having different number of kinks in the ground state and presented the results in the Figure \[fig:2\]. Similar to the particle number distribution here too we have clear change of behavior around the critical line. However, the oscillations are now appearing in the regions $-1\geq h\geq 1$. This is not surprising because the $\sigma^z_j$ and $\sigma^x_j\sigma^x_{j+1}$ have dual behavior. In the case of the transverse field Ising chain this duality is exact and it is called Kramers-Wannier(KW) duality which connects the Hamiltonian with the magnetic field $h$ to the one with the magnetic field $\tilde{h}\equiv\frac{1}{h}$. To see the effect of this duality on the distribution we plotted in the Figure \[fig:3\] the particle number and kink distribution for two different dual magnetic fields. To make the comparision easy we mirrored and shifted (by one unit) one of the distributions. The two distributions are perfectly matching which is a nice way to see the effect of the KW duality on the FCS.
Conclusions
===========
In this paper we studied formation probabilities and full counting statistics of the quadratic observables in the free fermions and the corresponding spin chains in the most generic possible way. We first showed that the problem of FP of the ground state of a generic free fermion can be translated into an emptiness formation probability of a free fermionic Hamiltonian with defects. In one dimension the defect is a line but in higher dimensions it can have different forms. Using the same line of thinking we then provided new determinant formulas for the FP’s with respect to the correlation matrix of the ground state of the Hamiltonian. In the second part of the paper we studied FCS of a generic quadratic observable in the ground state of a generic free fermion Hamiltonian. We showed that the probability of finding a particular value for the observable is exactly a FP problem for a Hamiltonian written in the basis that diagonalizes the observable. We showed how this can be done for a full system and also for the subsystem in the most generic case. Two simple cases, i.e. fluctuations of particles and kinks were discussed to show how one should implement the presented ideas. Finally, in the last section we solved the problem of the transverse field XY chain with an staggered magnetic line defect. We found exact correlation functions in and outside of the staggered region and provided a determinant formula for the FP of the staggered configuration in the ground state of the XY chain. Throughout the paper we tried to keep the discussion as general as possible except when we were presenting explicit examples to show how the procedure can be followed. Clearly one can take a particular model such as the XY chain and apply the presented methods. The numerical method to calculate these quantities are quite simple, however, to further push the analytical calculations in specific cases one normally needs to hire such methods as the generalized Fisher-Hartwig theorem which is beyond the scope of this paper and we hope to come back to them in future works.
**Acknowledgements.** MAR acknowledges the support from CNPQ. We thank P. Calabrese for early discussions.
Diagonalization of the Free Fermions {#SEC:diagonalization}
====================================
\[app1\] In this subsection, we summarize the result of Ref.[@LIEB1961407]. Consider a generic quadratic observable with real couplings $$\label{Hsub}\
\textbf{O}=\sum_{ij}\left[c_{i}^{\dagger}M_{ij}c_{j}+\frac{1}{2}c_{i}^{\dagger}N_{ij}c_{j}^{\dagger}+\frac{1}{2}c_{i}N_{ji}c_{j}\right]-\frac{1}{2}{\rm Tr}{\textbf{M}},$$ where $c_{i}^{\dagger}$ and $c_{i}$ are fermionic creation and annihilation operators, and $i$ and $j$ run over the interval $\left\lbrace 1,2,...,L\right\rbrace $. The Hermitian observable requires $\textbf{M}$ and $\textbf{N}$ to be symmetric and antisymmetric matrices respectively. To diagonalize the operator we use the following canonical transformation $$\begin{aligned}
\label{mat_U1}\
\begin{pmatrix}
\textbf{c} \\
\textbf{c}^{\dagger} \\
\end{pmatrix}=
\textbf{U}^{\dagger}
\begin{pmatrix}
\boldsymbol{\eta} \\
\boldsymbol{\eta}^{\dagger} \\
\end{pmatrix},\end{aligned}$$ with $$\begin{aligned}
\label{mat_U2}\
\textbf{U}= \begin{pmatrix}
\textbf{g} & \textbf{h}\\
\textbf{h}^{*} & \textbf{g}^{*}\\
\end{pmatrix}.\end{aligned}$$ which results to where $g$ and $h$ are $L\times L$ matrices, and the diagonal form of $\textbf{O}$ is $$\begin{aligned}
\label{Hdiag}\
\textbf{O}=\sum_{k}|\lambda_{k}|(\eta_{k}^{\dagger}\eta_{k}-\frac{1}{2}).\end{aligned}$$ By requiring that $[\eta_k,\textbf{O}]=|\lambda_{k}|\eta_k$, it is found that: $$\begin{split}
& \eta_kg_{ki}=\sum_j \left(g_{kj}M_{ji}-h_{kj}N_{ji} \right)\\
& \eta_kh_{ki}=\sum_j \left(g_{kj}N_{ji}-h_{kj}M_{ji} \right)
\end{split}
\label{Eq:g and h}$$ By defining new matrices $\boldsymbol{\psi}$ and $\boldsymbol{\phi}$ as follows $$\begin{aligned}
\label{gandh}\
\textbf{g}&=&\frac{1}{2}(\boldsymbol{\psi}+\boldsymbol{\phi}),\\
\textbf{h}&=&\frac{1}{2}(\boldsymbol{\psi}-\boldsymbol{\phi}),\end{aligned}$$ then Eq. \[Eq:g and h\] results to $$\begin{aligned}
\label{phi and psi 1a}\
\Psi_k(\textbf{M}-\textbf{N})&=&|\lambda_k|\Phi_k\\
\Phi_k(\textbf{M}+\textbf{N})&=&|\lambda_k|\Psi_k
\label{phi and psi 1b}\end{aligned}$$ or equivalently $$\begin{aligned}
\label{phi and psi 2a}\
\Psi_k(\textbf{M}-\textbf{N})(\textbf{M}+\textbf{N})&=&|\lambda_k|^2\Psi_k\\
\Phi_k(\textbf{M}+\textbf{N})(\textbf{M}-\textbf{N})&=&|\lambda_k|^2\Phi_k
\label{phi and psi 2b}.\end{aligned}$$ where $(\Phi_k)_i=\phi_{ki}$, and $(\Psi_k)_i=\psi_{ki}$. Therefore $\Psi_k$ and $\lambda_k$ can be calculated by solving the eigenvalue equation (\[phi and psi 2a\]), and for $\lambda_k\neq0$, $\Phi_k$ can be determined using (\[phi and psi 1a\]). For $\lambda_k=0$, one should solve Eq. \[phi and psi 1b\] directly to obtain $\Phi_k$.\
Having obtained $\textbf{h}$ and $\textbf{g}$, one can calculate the correlation matrix $\textbf{G}$ for the full system defined as $$\begin{aligned}
\label{Gmat2}
G_{ij}&=&\langle (c_{i}^{\dagger}-c_{i})(c_{j}^{\dagger}+c_{j})\rangle\end{aligned}$$ In terms of $\textbf{h}$ and $\textbf{g}$, $\textbf{G}$ can be also calculated as follows: $$\begin{aligned}
\label{Gmat7}
\textbf{G}= (\textbf{h}^{\dagger}-\textbf{g}^{\dagger})(\textbf{g}+\textbf{h}) .\end{aligned}$$ In the following sections we use the above construction to diagonalize the $XY$ Hamiltonian.
Diagonalization of the staggered $XY$ model {#SEC:solution}
===========================================
In this section we present the details of diagonalization of the (modified) $XY$ Hamiltonian. The observable of interest here is the formation probability of the staggered pattern. The general scheme is to apply a projection transformation in such a way that this probability becomes an EFP in a modified $XY$ Hamiltonian.\
The formation probability for the staggered configuration $\left| \downarrow\uparrow\downarrow\uparrow...\right\rangle$ at zero temperature is:
$$\begin{split}
P_{\text{stag}}(n)&=\left\langle 0\right| \left( \frac{1-\sigma_1^z}{2}\right) \left( \frac{1+\sigma_2^z}{2}\right) \left( \frac{1-\sigma_3^z}{2}\right) ...\left( \frac{1+(-1)^n\sigma_n^z}{2}\right) \left| 0\right\rangle \\
&=\left\langle 0\right|\left( \frac{1-\sigma_1^z}{2}\right) \sigma_2^x\left( \frac{1-\sigma_2^z}{2}\right)\sigma_2^x \left( \frac{1-\sigma_3^z}{2}\right) ...(\sigma_n^x)^{n+1}\left( \frac{1-\sigma_n^z}{2}\right)(\sigma_n^x)^{n+1}\left| 0\right\rangle\\
&=\left( \left\langle 0\right| \prod_{j=1}^{\text{int}(\frac{n}{2})}\sigma_{2j}^x\right) \prod_{j=1}^n\left( \frac{1-\sigma_j^z}{2}\right)\left( \prod_{j=1}^{\text{int}(\frac{n}{2})}\sigma_{2j}^x\left| 0\right\rangle\right)
\end{split}$$
To go to the Fermionic section, we use the Jordan-Wigner (J-W) transformation (Eq. \[Eq:JW\]). After applying $P_x=\prod_{j=1}^{\text{int}(\frac{n}{2})}\sigma_{2j}^x$, and also J-W transformation we obtain the Eq. \[Eq:modifiedXY\], from which we obtain: $$\begin{split}
&H'_{XY}=\\
&\frac{1}{2}\sum_{l=1}^{L-1}\left[ J_n(l) \left( c_l^{\dagger}c_{l+1}-c_lc_{l+1}^{\dagger} \right) +J\gamma_n(l) \left( c_l^{\dagger}c_{l+1}^{\dagger}-c_lc_{l+1}\right)\right]\\
&-\frac{NJ}{2} \left( c_L^{\dagger}c_1+\gamma c_L^{\dagger}c_1^{\dagger}+H.C.\right)-\sum_{l=1}^Lh_n(l)c_l^{\dagger}c_l+cnst.
\end{split}$$ where the constants were defined in Eq. \[Eq:gamma-J\]. Also note that $c_{L+1}=-Nc_1$, from which we see that for $N=-1$ one obtains: $$\begin{split}
&H'_{XY}=\\
&\frac{1}{2}\sum_{l=1}^L\left[ J_n(l) \left( c_l^{\dagger}c_{l+1}-c_lc_{l+1}^{\dagger} \right) +J\gamma_n(l) \left( c_l^{\dagger}c_{l+1}^{\dagger}-c_lc_{l+1}\right)\right]\\
& -\sum_{l=1}^Lh_n(l)c_l^{\dagger}c_l+const.
\label{Eq:effectiveH}
\end{split}$$ If we write the modified $XY$ Hamiltonian in the following form: $$H'_{XY}=\sum_{i,j}\left[c_i^{\dagger}A_{i,j}c_j +\frac{1}{2}\left(c_i^{\dagger}B_{i,j}c_j^{\dagger}+ h.c.\right) \right]
\label{Eq:Ising_AB}$$ then we have: $$\begin{split}
&A_{ij}(i\leq n \ \text{and}\ j\leq n)=\left\lbrace \begin{matrix}
(-1)^ih & \text{if} \ \ i=j\\
\frac{1}{2}\gamma J & \text{if} \ \ i=j\pm 1
\end{matrix}\right. \\
&A_{ij}(i> n \ \text{or}\ j> n)=\left\lbrace \begin{matrix}
-h & \text{if} \ \ i=j\\
\frac{1}{2} J & \text{if} \ \ i=j\pm 1\\
\frac{1}{2} J & \text{if} \ \ i=1,j=L \ \text{or} \ \ i=L,j=1\\
0 & \text{otherwise}
\end{matrix}\right.
\end{split}
\label{Eq:A}$$ Also $$\begin{split}
&B_{ij}(i\leq n \ \text{and}\ j\leq n)=\left\lbrace \begin{matrix}
\frac{1}{2} J & \text{if} \ \ i=j+1\\
-\frac{1}{2} J & \text{if} \ \ i=j-1
\end{matrix}\right. \\
&B_{ij}(i> n \ \text{or}\ j> n)=\left\lbrace \begin{matrix}
\frac{1}{2} \gamma J & \text{if} \ \ i=j+1\\
-\frac{1}{2} \gamma J & \text{if} \ \ i=j-1\\
-\frac{1}{2} \gamma J & \text{if} \ \ i=1,j=L\\
\frac{1}{2} \gamma J & \text{if} \ \ i=L,j=1\\
0 & \text{otherwise}
\end{matrix}\right.
\end{split}
\label{Eq:B}$$ Therefore one obtains the following form for $\textbf{C}\equiv(\textbf{A}-\textbf{B})(\textbf{A}+\textbf{B})$: $$\begin{split}
&C_{ii}=h^2+\frac{1}{2}J^2(1+\gamma^2)\\
&C_{i,i+1}=\left\lbrace \begin{matrix}
(-1)^i hJ & \text{if} \ \ i\leq n\\
-hJ & \text{if} \ \ i> n
\end{matrix}\right. \\
&C_{i,i+2}=\left\lbrace \begin{matrix}
-\frac{1}{4}J^2(1-\gamma^2) & \text{if} \ \ i\leq n\\
\frac{1}{4}J^2(1-\gamma^2) & \text{if} \ \ i> n
\end{matrix}\right.
\end{split}$$ Note that $\textbf{C}$ is symmetric, and also $C_{L-1,1}=C_{L,2}=C_{1,L-1}=C_{2,L}=\frac{1}{4}J^2(1-\gamma^2)$, and also $C_{1,L}=C_{L,1}=-hJ$. All other components of $\textbf{C}$ are zero. We use the Eq. \[Eq:pairfunction\] to diagonolize this matrix and obtain $\Lambda_k$’s and also $\Psi$. We analyze two cases separately: $m\ll \frac{n}{2}$ (deep inside the staggered interval, or DISI region), and $m\gg \frac{n}{2}$ (deep outside the staggered interval, or DOSI region).\
\
**DISI** case:\
\
For the solution, we consider the trial function Eq. \[Eq:pairfunction\] with constants $a_1$, $a_2$, $a_3$ and $a_4$ coefficients to be determined. This function can be re-written in the following form: $$\begin{split}
\psi(m)=\left\lbrace \begin{matrix}
a_1\sin km +a_2\cos km & \ \ \text{for the odd sublattice} \\
a_3\sin km +a_4\cos km & \ \ \text{for the even sublattice}
\end{matrix}\right.
\end{split}$$ In this case, applying Eq. \[phi and psi 2a\], we end up with two set of equations (due to the bipartite nature of the lattice) to be solved (for DISI):
$$\begin{split}
(1)\ \ &\sin km\left(-\frac{1}{2}a_1J^2(1-\gamma^2)\cos k+a_3hJ(\cos k-1)+a_4hJ\sin k +a_1(h^2+\frac{1}{2}J^2(1+\gamma^2)-\Lambda_k^2)\right) \\
&+\cos km\left(-\frac{1}{2}a_2J^2(1-\gamma^2)\cos k-a_3hJ\sin k+a_4hJ(\cos k-1) +a_2(h^2+\frac{1}{2}J^2(1+\gamma^2)-\Lambda_k^2)\right)=0\\
(2)\ \ & \ \ \sin km\left(-\frac{1}{2}a_3J^2(1-\gamma^2)\cos k+a_1hJ(\cos k-1)-a_2hJ\sin k +a_3(h^2+\frac{1}{2}J^2(1+\gamma^2)-\Lambda_k^2)\right) \\
&+\cos km\left(-\frac{1}{2}a_4J^2(1-\gamma^2)\cos k+a_1hJ\sin k+a_2hJ(\cos k-1) +a_4(h^2+\frac{1}{2}J^2(1+\gamma^2)-\Lambda_k^2)\right)=0
\end{split}$$
Each component (the coefficients of $\sin km$ and $\cos km$) should be separately set to zero. Therefore, we obtain: $$\begin{split}
\left\lbrace \begin{matrix}
a_1\zeta_k +a_3hJ(\cos k-1)+a_4hJ\sin k=0\\
a_2\zeta_k -a_3hJ\sin k +a_4hJ(\cos k-1)=0\\
a_3\zeta_k +a_1hJ(\cos k-1)-a_2hJ\sin k=0\\
a_4\zeta_k +a_1hJ\sin k +a_2hJ(\cos k-1)=0
\end{matrix}\right.
\end{split}$$ where $\zeta_k\equiv -\frac{1}{2}J^2(1-\gamma^2)\cos k+h^2+\frac{1}{2}J^2(1+\gamma^2)-\Lambda_k^2$. In the matrix form, we have:
$$\begin{split}
\left[ \begin{matrix}
\zeta_k & 0 & hJ(\cos k -1) & hJ\sin k\\
0 & \zeta_k & -hJ\sin k & hJ(\cos k-1)\\
hJ(\cos k-1) & -hJ\sin k & \zeta_k & 0 \\
hJ\sin k & hJ(\cos k-1) & 0 & \zeta_k
\end{matrix}\right] \left[ \begin{matrix}
a_1 \\
a_2 \\
a_3 \\
a_4
\end{matrix}\right]=0
\end{split}$$
By setting the determinant to zero, we find that the eigenvalues should be of the following form: $$\Lambda_k^2=\frac{1}{2}\left[2h^2+J^2(1+\gamma^2)-J^2(1-\gamma^2)\cos k\pm 4hJ\sin \frac{k}{2} \right]
\label{Eq:deep1-energy}$$ and the corresponding eigenvectors are: $$\begin{split}
&\eta_1^-=\left( \begin{matrix}
-\cos \frac{k}{2} \\
\sin \frac{k}{2} \\
0 \\
1
\end{matrix}\right) \ , \ \eta_2^-=\left( \begin{matrix}
\sin \frac{k}{2} \\
\cos \frac{k}{2} \\
1 \\
0
\end{matrix}\right) \\
&\eta_1^+=\left( \begin{matrix}
\cos \frac{k}{2} \\
-\sin \frac{k}{2} \\
0 \\
1
\end{matrix}\right) \ , \ \eta_2^+=\left( \begin{matrix}
-\sin \frac{k}{2} \\
-\cos \frac{k}{2} \\
1 \\
0
\end{matrix}\right)
\end{split}$$ where the minus (plus) sign refers to minus (plus) sign in the eigenvalues.\
\
**DOSI**\
\
Now let us work out with DOSI following the same steps as the DISI case. Let us consider the coefficients to be $b_1$, $b_2$, $b_3$ and $b_4$. The equation governing $\Psi$ results to the following linear equations:
$$\begin{split}
(1)\ \ &\sin km\left(\frac{1}{2}b_1J^2(1-\gamma^2)\cos k-b_3hJ(\cos k+1)-b_4hJ\sin k +b_1(h^2+\frac{1}{2}J^2(1+\gamma^2)-\Lambda_k^2)\right) \\
&+\cos km\left(\frac{1}{2}b_2J^2(1-\gamma^2)\cos k+b_3hJ\sin k-b_4hJ(\cos k+1) +b_2(h^2+\frac{1}{2}J^2(1+\gamma^2)-\Lambda_k^2)\right)=0\\
(2)\ \ &\sin km\left(\frac{1}{2}b_3J^2(1-\gamma^2)\cos k-b_1hJ(\cos k+1)+b_2hJ\sin k +b_3(h^2+\frac{1}{2}J^2(1+\gamma^2)-\Lambda_k^2)\right) \\
&+\cos km\left(\frac{1}{2}b_4J^2(1-\gamma^2)\cos k-b_1hJ\sin k-b_2hJ(\cos k+1) +b_4(h^2+\frac{1}{2}J^2(1+\gamma^2)-\Lambda_k^2)\right)=0
\end{split}$$
resulting to: $$\begin{split}
\left\lbrace \begin{matrix}
b_1\zeta'_k -b_3hJ(\cos k+1)-b_4hJ\sin k=0\\
b_2\zeta'_k +b_3hJ\sin k -b_4hJ(\cos k+1)=0\\
b_3\zeta'_k +b_1hJ(\cos k+1)+b_2hJ\sin k=0\\
b_4\zeta'_k -b_1hJ\sin k -b_2hJ(\cos k+1)=0
\end{matrix}\right. ,
\end{split}$$ or, in the matrix form:
$$\begin{split}
\left[ \begin{matrix}
\zeta'_k & 0 & -hJ(\cos k +1) & -hJ\sin k\\
0 & \zeta'_k & hJ\sin k & -hJ(\cos k+1)\\
-hJ(\cos k+1) & hJ\sin k & \zeta'_k & 0 \\
-hJ\sin k & -hJ(\cos k+1) & 0 & \zeta'_k
\end{matrix}\right] \left[ \begin{matrix}
b_1 \\
b_2 \\
b_3 \\
b_4
\end{matrix}\right]=0
\end{split}$$
where $\zeta'_k\equiv \frac{1}{2}J^2(1-\gamma^2)\cos k+h^2+\frac{1}{2}J^2(1+\gamma^2)-\Lambda_k^2$. The corresponding eigenvalues are: $$\Lambda_k^2=\frac{1}{2}\left[2h^2+J^2(1+\gamma^2)+J^2(1-\gamma^2)\cos k\pm 4hJ\cos \frac{k}{2} \right]
\label{Eq:deep2-energy}$$ This form is just like the eigenvalues found for the DISI, with different sign for $J^2(1-\gamma^2)\cos k$. The corresponding eigenvectors are: $$\begin{split}
&\eta_1^-=\left( \begin{matrix}
\sin \frac{k}{2} \\
\cos \frac{k}{2} \\
0 \\
1
\end{matrix}\right) \ , \ \eta_2^-=\left( \begin{matrix}
\cos \frac{k}{2} \\
-\sin \frac{k}{2} \\
1 \\
0
\end{matrix}\right) \\
&\eta_1^+=\left( \begin{matrix}
-\sin \frac{k}{2} \\
-\cos \frac{k}{2} \\
0 \\
1
\end{matrix}\right) \ , \ \eta_2^+=\left( \begin{matrix}
-\cos \frac{k}{2} \\
\sin \frac{k}{2} \\
1 \\
0
\end{matrix}\right)
\end{split}$$ where again the the minus (plus) sign referes to the minus (plus) sign in the eigenvalues. Since we pair the sites in the direct space, the size of the first Brillouin zone is doubled. If we use the natural change $k=2q$, then we obtain: $$\begin{split}
\Lambda_q^2&=\frac{1}{2}\left[2h^2+J^2(1+\gamma^2)+J^2(1-\gamma^2)\cos 2q\pm 4hJ\cos q \right]\\
& = (J\cos q\pm h)^2+J^2\gamma^2\sin^2q
\end{split}$$ that is exactly the spectrum of the single particle energies of the Fermions.\
Summarizing, for DOSI, (using the above $\eta$’s) the full eigenvector is readily calculated to be: $$\begin{split}
&\psi_1^-=\left\lbrace \begin{matrix}
\cos k(m-\frac{1}{2}) \\
\cos km
\end{matrix}\right. \ , \ \psi_2^-=\left\lbrace \begin{matrix}
\sin k(m-\frac{1}{2}) \\
\sin km
\end{matrix}\right. \\
&\psi_1^+=\left\lbrace \begin{matrix}
-\cos k(m-\frac{1}{2}) \\
\cos km
\end{matrix}\right. \ , \ \psi_2^+=\left\lbrace \begin{matrix}
-\sin k(m-\frac{1}{2}) \\
\sin km
\end{matrix}\right.
\end{split}$$ whereas for DISI, $$\begin{split}
&\psi_1^-=\left\lbrace \begin{matrix}
-\sin k(m-\frac{1}{2}) \\
\cos km
\end{matrix}\right. \ , \ \psi_2^-=\left\lbrace \begin{matrix}
\cos k(m-\frac{1}{2}) \\
\sin km
\end{matrix}\right. \\
&\psi_1^+=\left\lbrace \begin{matrix}
\sin k(m-\frac{1}{2}) \\
\cos km
\end{matrix}\right. \ , \ \psi_2^+=\left\lbrace \begin{matrix}
-\cos k(m-\frac{1}{2}) \\
\sin km
\end{matrix}\right.
\end{split}$$ Interestingly we see that the phase shift $k\rightarrow \pi-k$ relates the eigenvalues in DISI to the eigenvalues in DOSI. Since we need that these eigenvalues be equal for the case of line defect, we should apply this phase shift for one of the regions, e.g. DISI. Under this action, $\cos k(m-\frac{1}{2})\rightarrow -(-1)^m\sin k(m-\frac{1}{2})$, $\sin k(m-\frac{1}{2})\rightarrow -(-1)^m\cos k(m-\frac{1}{2})$, $\cos km\rightarrow (-1)^m\cos km$ and $\sin km\rightarrow -(-1)^m\sin km$. Therefore, for the DISI: $$\begin{split}
&\psi_1^-\rightarrow\left\lbrace \begin{matrix}
-(-1)^{m+1}\cos k(m-\frac{1}{2}) \\
(-1)^m\cos km
\end{matrix}\right. = (-1)^m\times\left\lbrace \begin{matrix}
\cos k(m-\frac{1}{2}) \\
\cos km
\end{matrix}\right.\\
&\psi_2^-\rightarrow\left\lbrace \begin{matrix}
(-1)^{m+1}\sin k(m-\frac{1}{2}) \\
(-1)^{m+1}\sin km
\end{matrix}\right. \rightarrow (-1)^m\times\left\lbrace \begin{matrix}
\sin k(m-\frac{1}{2}) \\
\sin km
\end{matrix}\right.\\
&\psi_1^+\rightarrow\left\lbrace \begin{matrix}
(-1)^{m+1}\cos k(m-\frac{1}{2}) \\
(-1)^m\cos km
\end{matrix}\right.=(-1)^m\times\left\lbrace \begin{matrix}
-\cos k(m-\frac{1}{2}) \\
\cos km
\end{matrix}\right. \\
&\psi_2^+\rightarrow\left\lbrace \begin{matrix}
-(-1)^{m+1}\sin k(m-\frac{1}{2}) \\
(-1)^{m+1}\sin km
\end{matrix}\right.=(-1)^m\times\left\lbrace \begin{matrix}
\sin k(m-\frac{1}{2}) \\
-\sin km
\end{matrix}\right.
\end{split}$$ showing that $\psi_{DISR}=(-1)^m\psi_{DOSR}$.\
Now let us consider the chain as a whole, for which the boundary conditions at $j=n$ should be worked out. To this end, we mix the two solutions obtained above. Based on the above findings, the following trial function is considered ($m_0\equiv \left[\frac{n}{2} \right] $): $$\begin{split}
&\psi(m\leq m_0)=(-1)^m\times\left[
\begin{matrix}
a\cos k(m-\frac{1}{2})+b\sin k(m-\frac{1}{2})\\
a \cos km +b\sin km
\end{matrix}\right. \\
&\psi(m> m_0)=\times\left[
\begin{matrix}
c\cos k(m-\frac{1}{2})+d\sin k(m-\frac{1}{2})\\
c \cos km +d\sin km
\end{matrix}\right.
\end{split}
\label{Eq:trial_wave}$$ With undetermined constants $a$, $b$, $c$, and $d$, to be found by applying the boundary conditions. To facilitate this calculation let us consider $\gamma=1$ (pure Ising model). We have seen that the solution for the Ising model is also valid for generic $\gamma$. For $\gamma=1$ we have four independent equations at the boundaries:
$$\begin{split}
&(a\cos km_0+b\sin km_0)-(c\cos k(m_0+1)+d\sin k(m_0+1))+2\cos\frac{k}{2}(c\cos k(m_0+\frac{1}{2})+d\sin k(m_0+\frac{1}{2}))=0\\
&-(c\cos\frac{kL}{2}+d\sin\frac{kL}{2})+(a\cos k+b\sin k)+2\cos\frac{k}{2}(-a\cos\frac{k}{2}-b\sin\frac{k}{2})=0\\
& a\cos\frac{k}{2}+b\sin\frac{k}{2}-(c\cos\frac{k(L-1)}{2}+d\sin\frac{k(L-1)}{2})+2\cos\frac{k}{2}(c\cos\frac{kL}{2}+d\sin\frac{kL}{2})=0\\
&-(a\cos k(m_0-\frac{1}{2})+b\sin k(m_0-\frac{1}{2}))+(c\cos k(m_0+\frac{1}{2})+d\sin k(m_0+\frac{1}{2}))+2\cos \frac{k}{2}(a\cos km_0+b\sin km_0)=0
\end{split}$$
These equations can be written in the matrix form: $$\begin{split}
\left[ \begin{matrix}
a_{11} & a_{12} & a_{13} & a_{14} \\
a_{21} & a_{22} & a_{23} & a_{24} \\
a_{31} & a_{32} & a_{33} & a_{34} \\
a_{41} & a_{42} & a_{43} & a_{44}
\end{matrix}\right]\left[ \begin{matrix}
a \\
b \\
c \\
d
\end{matrix}\right] =0
\end{split}$$ with the elements: $$\begin{split}
&a_{11}=\cos km_0,a_{12}=\sin km_0,a_{13}=\cos k m_0,a_{14}=\sin km_0, \\
&a_{21}=-1,\ a_{22}=0,\ a_{23}=-\cos\frac{kL}{2},\ a_{24}=-\sin\frac{kL}{2},\\
&a_{31}=\cos\frac{k}{2},\ a_{32}=\sin\frac{k}{2}, a_{33}=\cos\frac{k(L+1)}{2},\\
&a_{34}=\sin\frac{k(L+1)}{2},a_{41}=\cos k(m_0+\frac{1}{2}),a_{42}=\sin k(m_0+\frac{1}{2})\\
&a_{43}=\cos k(m_0+\frac{1}{2}),\ a_{44}=\sin k(m_0+\frac{1}{2})
\end{split}$$ The determinant of this matrix is $-4\sin^2\frac{k}{2}\sin^2\frac{kL}{2}$, whose zeros take place at $k_s=\frac{4\pi s}{L}$ in accordance with the free case (the case with no staggered interval involved). This is expected since the single particle energy spectrum of the original XY model should not change under the action of the unitary transformation $P_x$. For $\sin k_sm_0 \neq 0$ the solution for $a$, $b$, $c$ and $d$ is: $$a=-\cot k_sm_0, \ \ \ b=-1, \ \ \ c=\cot\left(k_sm_0\right), \ \ \ d=1$$ This results to $$\begin{split}
&\psi(m\leq m_0)=-(-)^m\psi_0\times\left\lbrace
\begin{matrix}
\cos k_s\left[m-m_0-\frac{1}{2}\right]\\
\\
\cos k_s\left[m-m_0\right]
\end{matrix}\right. \\
&\psi(m> m_0)=\psi_0\left\lbrace
\begin{matrix}
\cos k_s\left[m-m_0-\frac{1}{2}\right]\\
\\
\cos k_s\left[m-m_0\right]
\end{matrix}\right.
\end{split}$$ where $\psi_0$ is the normalization factor, which is shown to be $\sqrt{\frac{2}{L}}$. In terms of $j=2m_j$ and $q_s=\frac{k_s}{2}$ we find that: $$\begin{split}
\psi(j)=\sqrt{\frac{2}{L}}\left\lbrace \begin{matrix}
-(-1)^{m_j}
\cos q_s\left[j-n\right] & j\leq n\\
\cos q_s\left[j-n\right] & j> n
\end{matrix}\right.
\end{split}
\label{Eq:psi}$$ Also the other choice for $a$, $b$, $c$ and $d$ is $$\begin{split}
& a=\tan k_sm_0, \ \ \ b=-1, \ \ \ c=-\tan k_sm_0,\ \ \ d=1
\end{split}$$ which is equivalent to $\cos \leftrightarrow \sin$. Althuogh the Eq. \[Eq:psi\] was obtained for $\gamma=1$, it is a general result for the staggered line defect, and is valid for generic $\gamma$.
Correlation functions {#SEC:corr_func}
=====================
Here we present the details of calculation of the correlation functions. In the previous Appendix we showed that the eigenvector of the matrix $\textbf{C}$ for the $XY$ model in the general form is: $$\psi_{XY}=\sqrt{\frac{2}{L}}\left\lbrace \begin{matrix}
-(-)^{m_j}\cos q_s (j-n) & j\leq n\\
\cos q_s (j-n) & j> n
\end{matrix} \right.$$ Therefore using the relation \[phi and psi 1b\] we obtain the general form of $\phi_{si}$ for $\Lambda_k \ne 0$ (generic $\gamma$):
$$\begin{split}
\phi_{XY}(j\leq n)=&-\Lambda_{q_s}^{-1}\sqrt{\frac{2}{L}}\left[(-)^{m_j+a_j}h\cos q_s(j-n)\right. \\
&\left. +\frac{-1+\gamma}{2}(-)^{m_{j-1}}\cos q_s(j-n-1)+\frac{1+\gamma}{2}(-)^{m_{j+1}}\cos q_s(j-n+1) \right] \\
\phi_{XY}(j= n+1)=&-\Lambda_{q_s}^{-1}\sqrt{\frac{2}{L}}\left[h\cos q_s(j-n)\right. \\
&\left. \frac{-1+\gamma}{2}(-)^{m_{j-1}}\cos q_s(j-n-1)-\frac{1+\gamma}{2}\cos q_s(j-n+1) \right] \\
\phi_{XY}(j> n+1)=&-\Lambda_{q_s}^{-1}\sqrt{\frac{2}{L}}\left[h\cos q_s(j-n)\right. \\
&\left. +\frac{-1+\gamma}{2}\cos q_s(j-n-1)-\frac{1+\gamma}{2}\cos q_s(j-n+1) \right]
\end{split}$$
We should $\sin \leftrightarrow \cos$ as we go from positive $q_s$’s to negative ones. Let us work with $f_{AB}^{i,j}=\sum_{s=-\frac{L}{2}}^{L/2-1}A_{si}B_{sj}$, in which $A,B=\psi, \phi$. The importance of these functions can be understood noting that: $$\begin{split}
&\left\langle c_ic_j^{\dagger} \right\rangle =\frac{1}{4}\left[ f_{\psi\psi}+f_{\psi\phi}+f_{\phi\psi}+f_{\phi\phi}\right],\\
&\left\langle c_ic_j \right\rangle =\frac{1}{4}\left[f_{\psi\psi}-f_{\psi\phi}+f_{\phi\psi}-f_{\phi\phi}\right], \\
&\left\langle c_i^{\dagger}c_j^{\dagger} \right\rangle=\frac{1}{4}\left[f_{\psi\psi}+f_{\psi\phi}-f_{\phi\psi}-f_{\phi\phi}\right], \\
&\left\langle c_i^{\dagger}c_j \right\rangle =\frac{1}{4}\left[f_{\psi\psi}-f_{\psi\phi}-f_{\phi\psi}+f_{\phi\phi}\right],
\end{split}
\label{Eq:correlations}$$ One can easily show that always (irrespective to the amount of $i$ and $j$ being inside or outside the staggered interval) $f_{\psi\psi}^{i,j}=f_{\phi\phi}^{i,j}=\delta_{ij}$, and $\delta_{ij}$ is the Kronocker delta.\
Having $\Psi$ and $\Phi$ in hand, one can directly calculate $f_{\psi\psi}$, $f_{\phi\phi}$, $f_{\psi\phi}$, and $f_{\phi\psi}$. We immediately obtain that $f_{\psi\psi}=\delta_{ij}$ as expected. In the following we prove also that $f_{\phi\phi}=\delta_{ij}$. Let us consider $i\leq n$ and $j\leq n$. Then we have:
$$\begin{split}
f_{\phi\phi}=&\sum_s\phi_{si}\phi_{sj}=\frac{1}{L}\sum_s\left\lbrace (-)^{m_i+a_i-m_j-a_j}h^2+\left(\frac{-1+\gamma}{2} \right)^2(-)^{m_{j-1}-m_{i-1}}+ \left(\frac{1+\gamma}{2} \right)^2(-)^{m_{j+1}-m_{i+1}}\right\rbrace \cos q_s(j-i) \\
&+\frac{1}{L}\sum_s\left\lbrace (-)^{m_i+a_i-m_{j-1}}h\left(\frac{-1+\gamma}{2} \right)\cos q_s(j-i-1)+(-)^{m_i+a_i-m_{j+1}}h\left(\frac{1+\gamma}{2} \right)\cos q_s(j-i+1)\right\rbrace \\
&+\frac{1}{L}\sum_s\left\lbrace (-)^{m_{i-1}-m_j-a_j}h\left(\frac{-1+\gamma}{2} \right)\cos q_s(j-i+1)+(-)^{m_{i-1}-m_{j+1}}\left(\frac{-1+\gamma}{2} \right)\left(\frac{1+\gamma}{2} \right)\cos q_s(j-i+2)\right\rbrace \\
&+\frac{1}{L}\sum_s\left\lbrace (-)^{m_{i+1}-m_j-a_j}h\left(\frac{1+\gamma}{2} \right)\cos q_s(j-i-1)+(-)^{m_{i+1}-m_{j-1}}\left(\frac{-1+\gamma}{2} \right)\left(\frac{1+\gamma}{2} \right)\cos q_s(j-i-2)\right\rbrace
\end{split}$$
Noting that $(-)^{m_{i-1}}=(-)^{m_i+a_i}=-(-)^{m_{i+1}}$, we reach to:
$$\begin{split}
f_{\phi\phi}=&(-)^{m_{j+1}-m_{i+1}}\frac{1}{L}\sum_s\Lambda_s^{-2}\left(h^2+\frac{1}{2}(1+\gamma^2) \right)\cos q_s(j-i)\\
& +(-)^{m_{j+1}-m_{i+1}}h\left(\frac{-1+\gamma}{2}-\frac{1+\gamma}{2} \right) \frac{1}{L}\sum_s\Lambda_s^{-2}\cos q_s(j-i-1)\\
& +(-)^{m_{j+1}-m_{i+1}}h\left(\frac{-1+\gamma}{2}-\frac{1+\gamma}{2} \right) \frac{1}{L}\sum_s\Lambda_s^{-2}\cos q_s(j-i+1)\\
& -(-)^{m_{j+1}-m_{i+1}}\left(\frac{\gamma^2-1}{4}\right) \frac{1}{L}\sum_s\Lambda_s^{-2}\left[ \cos q_s(j-i-2)+\cos q_s(j-i+2)\right]\\
&=(-)^{m_{j+1}-m_{i+1}}\frac{1}{L}\sum_s\Lambda_s^{-2}\left[h^2+\frac{1}{2}\gamma^2\left(1-\cos 2q_s \right)+\frac{1}{2}\left(1+\cos 2q_s \right)-2h\cos q_s \right] \cos q_s(j-i)\\
&=(-)^{m_{j+1}-m_{i+1}}\frac{1}{L}\sum_s\Lambda_s^{-2}\left[h^2+\gamma^2\sin^2q_s+\cos^2q_s-2h\cos q_s \right] \cos q_s(j-i)\\
&=(-)^{m_{j+1}-m_{i+1}}\frac{1}{L}\sum_s\cos q_s(j-i)=\delta_{ij}=f_{\psi\psi}\\
\end{split}$$
leading to $\left\lbrace c_i^{\dagger},c_j\right\rbrace=\frac{1}{2}\left(f_{\psi\psi}+f_{\phi\phi}\right)=\delta_{ij} $ as expected. Let us next calculate $f_{\psi\phi}$ and $f_{\phi\psi}$:
$$\begin{split}
f_{\psi\phi}=&(-)^{m_j+a_j-m_i}h\frac{1}{L}\sum_s\Lambda_s^{-1}\cos q_s(j-i)\\
&+(-)^{m_{j-1}-m_i}\left(\frac{-1+\gamma}{2} \right) \frac{1}{L}\sum_s\Lambda_s^{-1}\cos q_s(j-i-1)\\
&+(-)^{m_{j+1}-m_i}\left(\frac{1+\gamma}{2} \right) \frac{1}{L}\sum_s\Lambda_s^{-1}\cos q_s(j-i+1)\\
&=(-)^{m_{j+1}-m_i}\frac{1}{L}\sum_s\Lambda_s^{-1}\left[ -h\cos q_s(j-i)+\left(\frac{1-\gamma}{2} \right)\cos q_s(j-i-1)+\left(\frac{1+\gamma}{2} \right)\cos q_s(j-i+1)\right]\\
&=(-)^{m_{j+1}-m_i}\frac{1}{L}\sum_s\Lambda_s^{-1}\left[-h\cos q_s(j-i)+\cos q_s\cos q_s(j-i)-\gamma\sin q_s\sin q_s(j-i)\right]
\end{split}$$
$$\begin{split}
f_{\phi\psi}=&(-)^{m_i+a_i-m_j}h\frac{1}{L}\sum_s\Lambda_s^{-1}\cos q_s(j-i)\\
&+(-)^{m_{i-1}-m_j}\left(\frac{-1+\gamma}{2} \right) \frac{1}{L}\sum_s\Lambda_s^{-1}\cos q_s(j-i+1)\\
&+(-)^{m_{i+1}-m_j}\left(\frac{1+\gamma}{2} \right) \frac{1}{L}\sum_s\Lambda_s^{-1}\cos q_s(j-i-1)\\
&=(-)^{m_{i+1}-m_j}\frac{1}{L}\sum_s\Lambda_s^{-1}\left[ -h\cos q_s(j-i)+\left(\frac{1-\gamma}{2} \right)\cos q_s(j-i+1)+\left(\frac{1+\gamma}{2} \right)\cos q_s(j-i-1)\right]\\
&=(-)^{m_i-m_j+a_i}\frac{1}{L}\sum_s\Lambda_s^{-1}\left[ h\cos q_s(j-i)-\cos q_s\cos q_s(j-i)-\gamma\sin q_s\sin q_s(j-i)\right]\\
&=(-)^{m_j-m_i+a_j}\frac{1}{L}\sum_s\Lambda_s^{-1}(-)^{a_i-a_j}\left[ h\cos q_s(j-i)-\cos q_s\cos q_s(j-i)-\gamma\sin q_s\sin q_s(j-i)\right]\\
&=(-)^{m_{j+1}-m_i}\frac{1}{L}\sum_s\Lambda_s^{-1}(-)^{a_i-a_j}\left[-h\cos q_s(j-i)+\cos q_s\cos q_s(j-i)+\gamma\sin q_s\sin q_s(j-i)\right]
\end{split}$$
Therefore $$\begin{split}
\frac{1}{4}\left( f_{\psi\phi}\pm f_{\phi\psi}\right) &=\frac{1}{2}(-)^{m_{j+1}-m_i}\sum_s\Lambda_s^{-1}\left\lbrace \left[\frac{1\pm(-)^{a_i-a_j}}{2} \right]\times \right. \\
&\left. \left(-h\cos q_s(j-i)+\cos q_s\cos q_s(j-i)\right) \right.\\
& \left. -\gamma\left[\frac{1\mp (-)^{a_i-a_j}}{2}\right]\sin q_s\sin q_s(j-i)\right\rbrace \\
&=\chi_{ij}^{\pm}\sigma_1(i,j)+\chi_{ij}^{\mp}\sigma_2(i,j)
\end{split}$$ where $\chi_{ij}^{+}=\left[\frac{1+(-)^{a_i-a_j}}{2} \right](-)^{m_{j+1}-m_i}$ and $\chi_{ij}^-=\left[\frac{1-(-)^{a_i-a_j}}{2} \right](-)^{m_{j+1}-m_i}$, and also $$\begin{split}
\sigma_1(j,k)&=\frac{1}{2L}\sum_{s=-L/2}^{L/2-1}\cos q_s(k-j)\left( \frac{-h+\cos q_s}{\Lambda_s}\right)\\
&=\frac{1}{2L}\sum_{s=-L/2}^{L/2-1}\left( \frac{-h+\cos q_s}{\Lambda_s}\right)e^{-iq_s(k-j)}\\
\sigma_2(j,k)&=\frac{1}{2L}\sum_{s=-L/2}^{L/2-1}\sin q_s(k-j)\left( \frac{\gamma\sin q_s}{\Lambda_s}\right)\\
&=\frac{i}{2L}\sum_{s=-L/2}^{L/2-1}\left( \frac{\gamma\sin q_s}{\Lambda_s}\right)e^{-iq_s(k-j)}
\end{split}
\label{Eq:sigmas}$$ where we have used the symmetry considerations to add extra zero contributions. Therefore if $i$ and $j$ belong to the same sublattice, then $\chi_{ij}^+=(-)^{m_{j+1}-m_i}$ and $\chi_{ij}^-=0$, so that: $$\begin{split}
&\Sigma_1(i,j)\equiv\frac{1}{4}\left( f_{\psi\phi}+ f_{\phi\psi}\right)=(-)^{m_{j+1}-m_i}\sigma_1(j-i), \\
&\Sigma_2(i,j)\equiv\frac{1}{4}\left( f_{\psi\phi}- f_{\phi\psi}\right)=(-)^{m_{j+1}-m_i}\sigma_2(j-i)
\end{split}$$ Also if they belong to the different sublattices, then $\chi_{ij}^-=(-)^{m_{j+1}-m_i}$ and $\chi_{ij}^+=0$, and therefore we find that: $$\begin{split}
&\Sigma_1(i,j)=(-)^{m_{j+1}-m_i}\sigma_2(j-i), \\
&\Sigma_2(i,j)=(-)^{m_{j+1}-m_i}\sigma_1(j-i)
\end{split}$$ The correlation functions now can be determined explicitly. Using Eq. \[Eq:correlations\] we find: $$\begin{split}
\left\langle c_ic_j^{\dagger} \right\rangle &=\frac{1}{2}\delta_{ij}+\Sigma_1(i,j)\\
&=\frac{1}{2}\delta_{ij}+\chi_{ij}^+\sigma_1(i,j)+\chi_{ij}^-\sigma_2(i,j)\\
\left\langle c_ic_j \right\rangle &=\Sigma_2(i,j)\\
&=\chi_{ij}^+\sigma_2(i,j)+\chi_{ij}^-\sigma_1(i,j)\\
\left\langle c_i^{\dagger}c_j^{\dagger} \right\rangle &=-\chi^+_{ij}\sigma_2(i,j)-\chi_{ij}^-\sigma_1(i,j)=-\left\langle c_ic_j \right\rangle\\
\left\langle c_i^{\dagger}c_j \right\rangle &=\frac{1}{2}\delta_{ij}-\chi_{ij}^+\sigma_1(i,j)-\chi_{ij}^-\sigma_2(i,j)=\delta_{ij}-\left\langle c_ic_j^{\dagger} \right\rangle
\end{split}$$ Therefore, noting that $s_{ij}=\left\langle c_ic_j^{\dagger} \right\rangle - \left\langle c_ic_j \right\rangle $, we see: $$s_{ij}=\frac{1}{2}\delta_{ij}+(\chi_{ij}^+-\chi_{ij}^-)\sigma(i,j)$$ where we have defined: $$\begin{split}
\sigma(j,k)&\equiv \sigma_1(j,k)-\sigma_2(j,k)\\
&=\frac{1}{2L}\sum_{s=-L/2}^{L/2-1}\left( \frac{\cos q_s-h-i\gamma\sin q_s}{\Lambda_s}\right)e^{iq_s(j-k)}
\end{split}$$ which, in the $L\rightarrow \infty$ limit becomes $$\begin{split}
&\sigma(j,k)=\frac{1}{2}\int \frac{\text{d}q}{2\pi} \sigma(q)e^{iq(j-k)}\\
& \sigma(q)\equiv \frac{\cos q-h-i\gamma\sin q}{\Lambda_q}
\end{split}$$ Noting also that: $$\begin{split}
(-)^{m_{j+1}-m_i}&=\left\lbrace \begin{matrix}
(-)^{\frac{i-j-1}{2}} & \text{if} \ i, \ j\in \ (\text{different sublattices})\\
-(-)^{\frac{j+i}{2}} & \text{if} \ i,j\in \ (\text{same sublattice})
\end{matrix}\right.\\
&=-|\cos \frac{\pi}{2}(j-i)|(-)^{\frac{i+j}{2}}+|\sin \frac{\pi}{2}(j-i)| (-)^{\frac{i-j-1}{2}},
\end{split}$$ one can easily verify that: $$\begin{split}
s_{ij}=\frac{1}{2}\delta_{ij}+\cos\pi(j-i)&\left\lbrace -(-)^{\frac{i+j}{2}}\left| \cos \frac{\pi}{2}(j-i)\right|\right. \\
& \left. +(-)^{\frac{i-j-1}{2}}\left| \sin \frac{\pi}{2}(j-i)\right| \right\rbrace \sigma(j-i)
\end{split}$$
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Let $ L/K $ be a finite separable extension of fields whose Galois closure $ E/K $ has group $ G $. Greither and Pareigis have used Galois descent to show that a Hopf algebra giving a Hopf-Galois structure on $ L/K $ has the form $ E[N]^{G} $ for some group $ N $ such that $ |N|=[L:K] $. We formulate criteria for two such Hopf algebras to be isomorphic as Hopf algebras, and provide a variety of examples. In the case that the Hopf algebras in question are commutative, we also determine criteria for them to be isomorphic as $ K $-algebras. By applying our results, we complete a detailed analysis of the distinct Hopf algebras and $ K $-algebras that appear in the classification of Hopf-Galois structures on a cyclic extension of degree $ p^{n} $, for $ p $ an odd prime number.'
address:
- 'Department of Mathematics, Agnes Scott College, 141 E. College Ave., Decatur, GA 30030 USA'
- 'Department of Mathematics and Statistics, Boston University, 111 Cummington Mall, Boston, MA 02215 USA'
- 'School of Computing and Mathematics, Keele University, Staffordshire, ST5 5BG, UK'
- 'Department of Mathematics and Computer Science, Auburn University at Montgomery, Montgomery, AL, 36124 USA'
author:
- Alan Koch
- Timothy Kohl
- 'Paul J. Truman'
- Robert Underwood
bibliography:
- '../structure.bib'
title: 'Isomorphism problems for Hopf-Galois structures on separable field extensions'
---
\#1[keywords[\#1]{}]{}
[*key words:*]{} Hopf-Galois extension, Greither-Pareigis theory, Galois descent
[*MSC:*]{} 16T05
Introduction {#section_introduction}
============
Let $ L/K $ be a finite extension of fields and $ H $ a $ K $-Hopf algebra. We say that $ L $ is an [*$ H $-Galois extension of $ K $*]{}, or that $ H $ gives a [*Hopf-Galois structure*]{} on $ L/K $, if $ L $ is an $ H $-module algebra and the obvious $ K $-linear map $ L \otimes_{K} H \rightarrow \mathrm{End}_{K}(L) $ is bijective. For example, if $ L/K $ is a Galois extension with group $ G $ then the group algebra $ K[G] $, with action induced from the usual action of $ G $ on $ L $, gives a Hopf-Galois structure on $ L/K $. We say that two Hopf algebras $ H_{1}, H_{2} $ give [*isomorphic*]{} Hopf-Galois structures on a finite extension $ L/K $ if there is an isomorphism of $ K $-Hopf algebras $ \varphi : H_{1} \rightarrow H_{2} $ such that $ h x = \varphi(h) x $ for all $ h \in H_{1} $ and $ x \in L $. Note that it is possible for two distinct Hopf-Galois structures on $ L/K $ to have underlying Hopf algebras which are isomorphic as $ K $-Hopf algebras or as $ K $-algebras; equivalently, one might view this as multiple actions of a single $ K $-Hopf algebra or $ K $-algebra on $ L $. In this paper we study this phenomenon.\
\
If $ L/K $ is purely inseparable, it is known that a single Hopf algebra can act in an infinite number of ways: see e.g. [@Koch2014]. We shall therefore suppose that $ L/K $ is separable. In this case Greither and Pareigis [@GreitherPareigis1987] have classified the (isomorphism classes of) Hopf-Galois structures admitted by $ L/K $. In order to state this classification we require some notation. Let $ E $ be the Galois closure of $ L/K $, $ G = {\mbox{Gal}(E/K)} $, and $ G_{L} = {\mbox{Gal}(E/L)} $. Let $ X $ denote the left coset space $ G/G_{L} $, and define a homomorphism $ \lambda : G \rightarrow {\mbox{Perm}(X)} $ by $ \lambda(\sigma)(\overline{\tau}) = \overline{\sigma \tau} $, where $ \overline{\tau} $ denotes the coset $ \tau G_{L} \in X $. The theorem of Greither and Pareigis asserts that there is a bijection between Hopf-Galois structures on $ L/K $ and subgroups $ N $ of $ {\mbox{Perm}(X)} $ which are regular (that is, having the same size as $ X $ and acting transitively on $ X $) and normalized by $ \lambda(G) $ (that is, stable under the action of $ G $ on $ {\mbox{Perm}(X)} $ defined by $ \sigma \ast \eta = \lambda(\sigma) \eta \lambda(\sigma)^{-1} $). The enumeration of the Hopf-Galois structures admitted by $ L/K $ is therefore equivalent to the enumeration of subgroups of $ {\mbox{Perm}(X)} $ with these properties. If $ |X| $ is large then this is a difficult problem, but Byott’s translation theorem [@Byott1996] provides a useful simplification. Loosely, for each abstract group $ N $ of order $ |X| $, Byott’s theorem relates the number of $ G $-stable regular subgroups of $ {\mbox{Perm}(X)} $ that are isomorphic to $ N $ to the number of subgroups of the holomorph of $ N $ that are isomorphic to $ G $. Since $ {\mbox{Hol}(N)} \cong N \rtimes {\mbox{Aut}(N)} $, this group is much smaller than $ {\mbox{Perm}(X)} $. We give a more precise statement of Byott’s theorem in subsection \[subsection\_Hopf\_algebra\_isomorphisms\_holomorph\] below.\
\
The theorem of Greither and Pareigis also asserts that the Hopf algebra appearing in the Hopf-Galois structure corresponding to the $ G $-stable regular subgroup $ N $ of $ {\mbox{Perm}(X)} $ is $ E[N]^{G} $, the fixed points of the group algebra $ E[N] $ under the simultaneous action of $ G $ on $ E $ as Galois automorphisms and on $ N $ by the action $ \ast $. We will refer to the isomorphism class of $ N $ as the [*type*]{} of Hopf-Galois structure given by $ E[N]^{G} $. The proof that $ E[N]^{G} $ is indeed a $ K $-Hopf algebra, and the proof that it gives a Hopf-Galois structure on $ L/K $, are accomplished via Galois descent. We review this theory briefly since it will feature in many of our arguments.\
\
The field $ E $ is a finite dimensional $ K $-vector space and so, by Morita theory (see, e.g. [@Curtis_Reiner_I §3D]), the functor $ E \otimes_{K} - $ is an equivalence between the category of $ K $-vector spaces and the category of $ R = \mathrm{End}_{K}(E) $-modules, with inverse $ \mathrm{Hom}_{R}(E,-) $. But since $ E $ is a Galois extension of $ K $ with group $ G $, we may identify $ R $ with the skew group ring $ E \circ G $ [@ChaseHarrisonRosenberg1965], and so an $ R $-module is simply an $ E $-vector space with a compatible $ G $-action. In this case the functor $ \mathrm{Hom}_{R}(E,-) $ is naturally isomorphic to the fixed point functor $ (-)^{G} $, and so homomorphisms of $ E $-vector spaces descend to homomorphisms of $ K $-vector spaces if and only if they respect the $ G $-action. Applying this to the structure maps defining an $ E $-algebra (resp. an $ E $-Hopf algebra), we obtain criteria for us to descend to a $ K $-algebra (resp. $ K $-Hopf algebra). In the case of the objects appearing in the theorem of Greither and Pareigis, the group algebra $ E[N] $ is certainly an $ E $-Hopf algebra, and we may verify that the Hopf algebra structure maps do respect the action of $ G $ on $ E[N] $ defined above. We therefore obtain that $ E[N]^{G} $ is indeed a $ K $-Hopf algebra. Since the inverse to the fixed point functor is the base change functor, we also obtain that $ E \otimes_{K} E[N]^{G} = E[N] $.\
\
Now suppose that $ N_{1}, N_{2} $ are $ G $-stable regular subgroups of $ {\mbox{Perm}(X)} $, so that $ H_{1}=E[N_{1}]^{G} $ and $ H_{2}=E[N_{2}]^{G} $ are two Hopf algebras giving Hopf-Galois structures on $ L/K $. In section \[section\_Hopf\_algebra\_isomorphisms\] we seek criteria for $ H_{1} \cong H_{2} $ as $ K $-Hopf algebras. Since $ E \otimes_{K} H_{i} = E[N_{i}] $ for each $ i $, a necessary condition for $ H_{1} \cong H_{2} $ as $ K $-Hopf algebras is that $ E[N_{1}] \cong E[N_{2}] $ as $ E $-Hopf algebras, which occurs if and only if $ N_{1} \cong N_{2} $ as groups. However, this condition is not sufficient: for this, we require an isomorphism $ N_{1} \xrightarrow{\sim} N_{2} $ that respects the action of $ G $ on each of these groups. This idea appears in, for example, [@Childs2013] and [@Crespo2015]. We prove a more general result of this type (Theorem \[theorem\_Hopf\_algebra\_isomorphism\]), and provide a variety of examples. We show that it is possible to detect $ K $-Hopf algebra isomorphisms by studying properties of $ {\mbox{Hol}(N_{1})} $ and $ {\mbox{Hol}(N_{2})} $ (Theorem \[theorem\_Hopf\_algebra\_isomorphism\_holomorph\]), and also determine a criterion for $ F \otimes_{K} H_{1} $ and $ F \otimes_{K} H_{2} $ to be isomorphic as $ F $-Hopf algebras, for $ F $ some extension of $ K $ contained in $ E $ (Theorem \[theorem\_Hopf\_algebra\_isomorphism\_base\_change\]).\
\
In section \[section\_algebra\_isomorphisms\_commutative\] we assume that $ K $ has characteristic zero and that $ H_{1} $ and $ H_{2} $ are commutative (equivalently, that $ N_{1}, N_{2} $ are abelian groups). We determine a criterion, in terms of the dual groups $ \widehat{N_{1}} $ and $ \widehat{N_{2}} $, for $ H_{1} \cong H_{2} $ as $ K $-algebras (Theorem \[theorem\_algebra\_isomorphism\]). If $ N_{1} \cong N_{2} $, we show that it is also possible to detect such isomorphisms by studying $ {\mbox{Hol}(N_{1})} $ and $ {\mbox{Hol}(N_{2})} $ (Theorem \[theorem\_algebra\_isomorphism\_holomorph\]). We show that these results have a particularly simple form in the case that $ N_{1} $ and $ N_{2} $ are both cyclic of order $ n $ and $ K $ contains a primitive $ n^{th} $ root of unity (Theorem \[corollary\_algebra\_isomorphism\_cyclic\]). Finally, in section \[section\_cyclic\_p\_power\_extensions\] we apply the results of the preceding sections to give a detailed analysis of the Hopf-Galois structures admitted by a cyclic extension $ L/K $ of odd prime power degree. We show that the Hopf algebras that appear are pairwise nonisomorphic as Hopf algebras (Theorem \[theorem\_cyclic\_p\_power\_Hopf\_algebra\_isomorphisms\]) and, under the assumption that $ K $ has characteristic zero and contains a primitive $ p^{n} $-root of unity, determine the $ K $-algebra isomorphism classes (Theorem \[theorem\_cyclic\_p\_power\_wedderburn\]).
Hopf Algebra Isomorphisms {#section_Hopf_algebra_isomorphisms}
=========================
In this section we address the question of when two Hopf algebras giving Hopf-Galois structures on a finite separable extension of fields are isomorphic as Hopf algebras. We achieve this by exploiting that fact that, by the theorem of Greither and Pareigis, such a Hopf algebra is the fixed points of a group algebra under the action of a certain group of automorphisms. Rather than focusing specifically on this situation, we make our definitions, and formulate our first theorem, in rather more general terms:
Let $ G $ be a group and let $ (N_{1},\ast_{1}), (N_{2},\ast_{2}) $ be $ G $-sets (where, for $ =1,2 $, $ \ast_{i} $ denotes the action of $ G $ on $ N_{i} $). We say that $ (N_{1},\ast_{1}), (N_{2},\ast_{2}) $ (or just $ N_{1},N_{2} $) are [*isomorphic as $ G $-sets*]{} if there is a $ G $-equivariant bijection $ f : N_{1} \rightarrow N_{2} $. We say that a $ G $-set $ N $ is a [*$ G $-group*]{} if it is a group on which $ G $ acts via automorphisms, and that two $ G $-groups $ N_{1}, N_{2} $ are [*isomorphic as $ G $-groups*]{} if there is a $ G $-equivariant isomorphism $ f : N_{1} \xrightarrow{\sim} N_{2} $.
We note that if $ \varphi: N_{1} \xrightarrow{\sim} N_{2} $ is an isomorphism of $ G $-sets and $ \eta_{1}, \ldots ,\eta_{r} $ are representatives for the orbits of $ G $ in $ N_{1} $ then $ \varphi(\eta_{1}), \ldots ,\varphi(\eta_{r}) $ are representatives for the orbits of $ G $ in $ N_{2} $, and $ \mbox{Stab}(\eta_{i}) = \mbox{Stab}(\varphi(\eta_{i})) $ for each $ i $. We use these definitions to formulate the following general theorem:
\[theorem\_Hopf\_algebra\_isomorphism\] Let $ E/K $ be a Galois extension of fields, let $ G $ be a subgroup of $ {\mbox{Gal}(E/K)} $, and let $ F=E^{G} $. Let $ (N_{1}, \ast_{1}) $ and $ (N_{2}, \ast_{2}) $ be $ G $-groups and, for $ i=1,2 $, let $ G $ act on $ E[N_{i}] $ by acting on $ E $ as Galois automorphisms and on $ N_{i} $ via $ \ast_{i} $. Then $ E[N_{1}]^{G} \cong E[N_{2}]^{G} $ as $ F $-Hopf algebras if and only if $ N_{1} \cong N_{2} $ as $ G $-groups.
Note that $ E/F $ is a Galois extension of fields with group $ G $. If there is a $ G $-equivariant isomorphism of groups $ \varphi : N_{1} \xrightarrow{\sim} N_{2} $, then this extends to a $ G $-equivariant isomorphism of $ E $-Hopf algebras $ \varphi : E[N_{1}] \xrightarrow{\sim} E[N_{2}] $, which (since it is $ G $-equivariant) descends to an isomorphism of $ F $-Hopf algebras $ \varphi : E[N_{1}]^{G} \xrightarrow{\sim} E[N_{2}]^{G} $.\
\
Conversely, suppose that $ \psi : E[N_{1}]^{G} \xrightarrow{\sim} E[N_{2}]^{G} $ is an isomorphism of $ F $-Hopf algebras. Since we have $ E \otimes_{F} E[N_{i}]^{G} $ for each $ i $, it extends to an isomorphism of $ E $-Hopf algebras $ \psi : E[N_{1}] \xrightarrow{\sim} E[N_{2}] $. Restricting to the grouplike elements on both sides we obtain an isomorphism of groups $ \psi : N_{1} \xrightarrow{\sim} N_{2} $. Thus, we need to show that $ \psi $ is $ G $-equivariant. Let $ \eta \in N_{1} $, and let $ \{ a_{i} \} $ be an $ F $-basis of $ E[N_{1}]^{G} $. Then $ \{ a_{i} \} $ is also an $ E $-basis of $ E[N_{1}] $, so there exist unique $ x_{i} \in E $ such that $$\eta = \sum_{i} x_{i} a_{i}.$$ Now for $ g \in G $ we have $$\begin{aligned}
\psi(g \ast_{1} \eta) & = \psi \left(g \ast_{1} \sum_{i} x_{i} a_{i} \right) \\
& = \psi \left( \sum_{i} g(x_{i}) (g \ast_{1} a_{i}) \right) \\
& = \psi \left( \sum_{i} g(x_{i}) a_{i} \right) \\
& = \sum_{i} g(x_{i}) \psi(a_{i}) \\
& = \sum_{i}g( x_{i})(g \ast_{2} \psi(a_{i})) \\
&= g \ast_{2} \left( \sum_{i} x_{i} \psi(a_{i}) \right) \\
&= g \ast_{2} \psi \left( \sum_{i} x_{i} a_{i} \right) \\
& = g \ast_{2} \psi( \eta).\end{aligned}$$ Therefore $ \psi(g \ast_{1} \eta) = g \ast_{2} \psi( \eta) $, and so $ \psi : N_{1} \xrightarrow{\sim} N_{2} $ is $ G $-equivariant.
The following corollary is the principal application of Theorem \[theorem\_Hopf\_algebra\_isomorphism\]. We retain the notation established in section \[section\_introduction\] to describe the theorem of Greither and Pareigis: $ L/K $ is a finite separable extension of fields with Galois closure $ E$, $ G = {\mbox{Gal}(E/K)} $, $ G_{L} = {\mbox{Gal}(E/L)} $, $ X = G/G_{L} $. A Hopf algebra giving a Hopf-Galois structure on $ L/K $ then has the form $ E[N]^{G} $ for some regular subgroup $ N $ of $ {\mbox{Perm}(X)} $ stable under the action of $ G $ on $ {\mbox{Perm}(X)} $ by $ \sigma \ast \eta = \lambda(\sigma) \eta \lambda(\sigma)^{-1} $.
\[corollary\_Hopf\_algeba\_isomorphism\] Let $ E[N_{1}]^{G} $ and $ E[N_{2}]^{G} $ give Hopf-Galois structures on $ L/K $. Then $ E[N_{1}]^{G} \cong E[N_{2}]^{G} $ as $ K $-Hopf algebras if and only if $ N_{1} \cong N_{2} $ as $ G $-groups.
We illustrate the applicability of the theory with a variety of examples.
[**(The classical and canonical nonclassical Hopf-Galois structures)**]{} \[example\_Hopf\_isomorphism\_classical\_nonclassical\] If $ L/K $ is a Galois extension then in the notation of the theorem of Greither and Pareigis we have $ E=L $, $ G_{L} = \{ 1 \} $, and $ X = G $, and the homomorphism $ \lambda : G \rightarrow {\mbox{Perm}(G)} $ is in fact the left regular embedding of $ G $. In this case examples of $ G $-stable regular subgroups of $ {\mbox{Perm}(G)} $ are $ \lambda(G) $ itself and $ \rho(G) $, the image of $ G $ under the right regular embedding (these subgroups coincide if and only if $ G $ is abelian). The elements of $ \rho(G) $ commute with those of $ \lambda(G) $, so the action of $ G $ on $ \rho(G) $ is trivial, and therefore the Hopf algebra appearing in the Hopf-Galois structure corresponding to $ \rho(G) $ is $ L[\rho(G)]^{G} = L^{G}[\rho(G)] = K[\rho(G)] $, which is isomorphic to the Hopf algebra $ K[G] $. We call this the [*classical Hopf-Galois structure*]{} on $ L/K $. If $ G $ is nonabelian then the subgroup $ \lambda(G) $ corresponds to a different Hopf-Galois structure on $ L/K $, which we call the [*canonical nonclassical Hopf-Galois structure*]{}. The Hopf algebra appearing in this structure is $ H_{\lambda} := L[\lambda(G)]^{G} $. Since $ G $ is nonabelian the action of $ G $ on $ \lambda(G) $ is not trivial (the orbits are the conjugacy classes in $ \lambda(G) $), and so $ \rho(G) \not \cong \lambda(G) $ as $ G $-groups in this case. Therefore by Corollary \[corollary\_Hopf\_algeba\_isomorphism\] $ K[G] \not \cong H_{\lambda} $ as $ K $-Hopf algebras.
[**(Elementary abelian extensions of degree $ p^{2} $)**]{} \[example\_Hopf\_isomorphism\_elementary\_abelian\_p2\] Let $ p > 2 $ be prime, and let $ L/K $ be a Galois extension of fields of degree $ p^{2} $ with elementary abelian Galois group $ G $. In [@Byott1996 Corollary to Theorem 1, part (iii)] it is shown that $ L/K $ admits $ p^{2} $ Hopf-Galois structures. By applying Corollary \[corollary\_Hopf\_algeba\_isomorphism\] we can determine which of these Hopf-Galois structures involve isomorphic Hopf algebras. In [@Byott2002 Theorem 2.5] the regular subgroups of $ {\mbox{Perm}(G)} $ that yield the Hopf-Galois structures are determined as follows: let $ T $ be a subgroup of $ G $ of order $ p $, and fix elements $ s, t \in G $ such that $$T = \langle t \rangle, \hspace{4mm} s^{p}=1_{G}, \hspace{4mm} G = \langle s, t \rangle.$$ Let $ d \in \{0, 1, \ldots ,p-1\} $, and define $ \alpha, \beta \in {\mbox{Perm}(G)} $ in terms of their actions on a typical element $ s^{k}t^{l} \in G $: $$\label{equation_byott_p2_permutations}
\begin{array}{rcl}
\alpha[s^{k} t^{l}] & = & s^{k} t^{l-1} \\
\beta[s^{k} t^{l}] & = & s^{k-1} t^{l+(k-1)d}.
\end{array}$$ It is easily verified that $\alpha^{p}=\beta^{p}=1 $ and $ \alpha\beta =\beta\alpha $, whence $ N_{T,d}= \langle \alpha, \beta \rangle \cong G $. Moreover, one can show that $ N_{T,d} $ is a regular subgroup of $ {\mbox{Perm}(G)} $, and that $$\label{equation_byott_p2_lambda_action}
s \ast \alpha = \alpha, \hspace{4mm} t \ast \alpha = \alpha, \hspace{4mm} s \ast \beta = \alpha^{d} \beta, \hspace{4mm} t \ast \beta=\beta.$$ Thus $ N_{T,d} $ is $ G $-stable, and therefore corresponds to a Hopf-Galois structure on $ L/K $ with Hopf algebra $ H_{T,d} = L[N_{T,d}]^{G} $. If $ d=0 $ then $ N_{T,d}= \rho(G) $ regardless of the choice of $ T $, and so we obtain the classical Hopf-Galois structure. Taking $ 1 \leq d \leq p-1 $ and letting $ T $ vary through the subgroups of $ G $ of order $ p $, we obtain $ p^{2}-1 $ distinct groups $ N_{T,d} \neq \rho(G) $, giving in total $ p^{2} $ Hopf-Galois structures on $ L/K $. These are all the Hopf-Galois structures on $ L/K $.\
\
We claim that two Hopf algebras $ H_{1} = H_{T_{1},d_{1}} $ and $ H_{2} = H_{T_{2},d_{2}} $ are isomorphic as Hopf algebras if and only if $ d_{1}=d_{2}=0 $ or $ d_{1}d_{2} \neq 0 $ and $ T_{1}=T_{2} $. Let $$\begin{array}{ccccl}
N_{1} & = & N_{T_{1},d_{1}} & = & \langle \alpha_{1}, \beta_{1} \rangle \\
N_{2} & = & N_{T_{2},d_{2}} & = & \langle \alpha_{2}, \beta_{2} \rangle,
\end{array}$$ where $ \alpha_{1},\beta_{1} $ and $ \alpha_{2},\beta_{2} $ are defined as in Equations , using $ d_{1}, d_{2} $ as appropriate. We have seen that if $ d_{1}=d_{2}=0 $ then $ H_{1} = H_{2} $. If $ d_{1}d_{2} \neq 0 $ and $ T_{1}=T_{2} $ then there exists $ c \in \{1, \ldots ,p-1\} $ such that $ cd_{2}\equiv d_{1} \pmod{p} $. Using this, define a homomorphism $ \varphi: N_{1} \rightarrow N_{2} $ by $$\varphi(\alpha_{1}) = \alpha_{2}, \hspace{4mm} \varphi(\beta_{1}) = \beta_{2}^{c}.$$ It is clear that $ \varphi $ is an isomorphism, and we claim that it is $ G $-equivariant. We have: $$\begin{array}{l}
\varphi(s \ast \alpha_{1} ) = \varphi( \alpha_{1} ) = \alpha_{2} = s \ast \alpha_{2} = s \ast \varphi(\alpha_{1}), \\
\varphi(t \ast \alpha_{1} ) = \varphi( \alpha_{1} ) = \alpha_{2} = t \ast \alpha_{2} = t \ast \varphi(\alpha_{1}), \\
\varphi(s \ast \beta_{1}) = \varphi(\alpha_{1}^{d_{1}}\beta_{1}) = \alpha_{2}^{d_{1}} \beta_{2}^{c} = \alpha_{2}^{d_{2}c} \beta_{2}^{c} = s \ast \beta_{2}^{c} = s \ast \varphi(\beta_{1}), \\
\varphi(t \ast \beta_{1} ) = \varphi( \beta_{1} ) = \beta_{2}^{c} = t \ast \beta_{2}^{c} = t \ast \varphi(\beta_{1}).
\end{array}$$ Thus $ \varphi $ is a $ G $-equivariant isomorphism of $ N_{1} $ onto $ N_{2} $, and so $ H_{1} \cong H_{2} $ as Hopf algebras by Corollary \[corollary\_Hopf\_algeba\_isomorphism\].\
\
For the converse, note that if $ T $ is a subgroup of $ G $ of order $ p $ and $ d \neq 0 $ then by the kernel of the action of $ G $ on $ N_{T,d} $ is precisely $ T $. Therefore, if $ d_{1}d_{2} \neq 0 $ and $ N_{1} \cong N_{2} $ as $ G $-groups then we must have $ T_{1} = T_{2} $.\
\
Thus, the $ p^{2} $ Hopf-Galois structures admitted by $ L/K $ involve exactly $ p+2 $ nonisomorphic Hopf algebras: the group algebra $ K[G] $ with its usual action on $ L $ and, for each of the $ p+1 $ subgroups $ T $ of $ G $ of order $ p $, a Hopf algebra $ H_{T}=H_{T,1} $, acting in $ p-1 $ different ways.
[**(Fixed point free endomorphisms)**]{} \[example\_Hopf\_isomorphism\_fpf\] If $ L/K $ be a finite Galois extension of fields with nonabelian Galois group $ G $ then, as described above, $ L/K $ admits a canonical nonclassical Hopf-Galois structure with Hopf algebra $ H_{\lambda} = L[\lambda(G)]^{G} $. In [@Childs2013], Childs shows how certain endomorphisms of $ G $ can yield further Hopf-Galois structures on $ L/K $, whose Hopf algebras are isomorphic to $ H_{\lambda} $. Specifically, let $ \psi $ be an endomorphism of $ G $ which is abelian (meaning $ \psi(gh) = \psi(hg) $ for all $ g, h\in G $) and fixed point free (meaning $ \psi(g)=g $ if and only if $ g=1_{G} $). From $ \psi $ we may construct a homomorphism $ \alpha_{\psi}: G \rightarrow {\mbox{Perm}(G)} $ defined by $$\alpha_{\psi}(g) = \lambda(g) \rho(\psi(g)).$$ One can show that $ \alpha_{\psi}(G) $ is a $ G $-stable regular subgroup of $ {\mbox{Perm}(G)} $, which therefore corresponds to a Hopf-Galois structure on $ L/K $. It is shown in [@Childs2013 Theorem 5] that the Hopf algebras appearing in the Hopf-Galois structures produced by this construction are all isomorphic to $ H_{\lambda} $ as Hopf algebras. We can reinterpret these ideas via Corollary \[corollary\_Hopf\_algeba\_isomorphism\].\
\
We know that $ \lambda(G) $ and $ \alpha_{\psi}(G) $ are both regular subgroups of $ {\mbox{Perm}(G)} $ normalized by $ \lambda(G) $. Define a map $ \varphi : \lambda(G) \rightarrow \alpha_{\psi}(G) $ by $$\varphi(\lambda(g)) = \alpha_{\psi}(g) = \lambda(g) \rho(\psi(g)).$$ It is clear that $ \varphi $ is an isomorphism of groups. In fact, it is also $ G $-equivariant: if $ h\in G $ then we have $$\begin{aligned}
h \ast \varphi(\lambda(g)) & = \lambda(h) \lambda(g) \rho(\psi(g)) \lambda(h)^{-1} \\
& = \lambda(h) \lambda(g) \lambda(h)^{-1} \rho(\psi(g)) \mbox{ (since $\lambda(G) $ and $ \rho(G) $ commute inside ${\mbox{Perm}(G)} $)} \\
& = \lambda(hg h^{-1}) \rho(\psi(g)) \\
& = \lambda(hg h^{-1}) \rho(\psi(hg h^{-1})) \mbox{ (since $\psi $ abelian implies that $\psi(g) = \psi(hg h^{-1})$).} \\
& = \varphi(\lambda(hg h^{-1})) \\
&= \varphi(h \ast \lambda(g)). \end{aligned}$$ Thus $ \varphi : \lambda(G) \rightarrow \alpha_{\psi}(G) $ is a $ G $-equivariant isomorphism of groups, and so by Corollary \[corollary\_Hopf\_algeba\_isomorphism\] $ L[\alpha_{\psi}(G)]^{G} \cong L[\lambda(G)]^{G} $ as Hopf algebras.
[**(Conjugating regular subgroups by elements of $ \rho(G) $)**]{} \[example\_Hopf\_isomorphism\_conjugation\] Let $ L/K $ be a Galois extension of fields with nonabelian Galois group $ G $, and let $ L[N]^{G} $ give a Hopf-Galois structure on $ L/K $. Since $ G $ is nonabelian, we have $ \lambda(G) \neq \rho(G) $ and so, although $ N $ is normalized by $ \lambda(G) $, it may not be normalized by $ \rho(G) $. Suppose that $ g \in G $ is such that $ N_{g} = \rho(g)N\rho(g)^{-1} \neq N $. Then $ N_{g} $ is a regular subgroup of $ {\mbox{Perm}(G)} $ since $ N $ is, and it is $ G $-stable: if $ h \in G $ then $$\begin{aligned}
h \ast N_{g} & = h \ast\rho(g)N\rho(g)^{-1} \\
& = \lambda(h) \rho(g)N\rho(g)^{-1} \lambda(h)^{-1} \\
& = \rho(g)\lambda(h) N \lambda(h)^{-1}\rho(g)^{-1} \\
& = \rho(g)N\rho(g)^{-1} \\
&= N_{g}. \end{aligned}$$ Therefore $ N_{g} $ corresponds to a Hopf-Galois structure on $ L/K $, with Hopf algebra $ L[N_{g}]^{G} $. The map $ \varphi : N \rightarrow N_{g} $ defined by $$\varphi(\eta) = \rho(g) \eta \rho(g)^{-1}$$ is an isomorphism of groups, and is $ G $-equivariant since if $ h \in G $ then $$\begin{aligned}
h \ast \varphi(\eta) & =h \ast \rho(g) \eta \rho(g)^{-1} \\
&= \lambda(h) \rho(g) \eta \rho(g)^{-1} \lambda(h)^{-1} \\
&= \rho(g)\lambda(h) \eta \lambda(h)^{-1}\rho(g)^{-1} \\
&= \rho(g)(h \ast \eta)\rho(g)^{-1} \\
& = \varphi(h \ast \eta).\end{aligned}$$ Therefore by Corollary \[corollary\_Hopf\_algeba\_isomorphism\] $ L[N_{g}]^{G} \cong L[N]^{G} $ as Hopf algebras.
[**(A specific example of conjugating by elements of $ \rho(G) $)\
**]{} Let $ p,q $ be primes with $ p \equiv 1 \pmod{q} $, and let $ L/K $ be a Galois extension of fields with group isomorphic to the metacyclic group of order $ pq $: $$G = \langle \sigma, \tau \mid \sigma^{p}=\tau^{q}=1, \; \tau \sigma = \sigma^{g} \tau \rangle,$$ where $ g $ is a fixed positive integer whose order modulo $ p $ is $ q $. By [@Byott2004 Theorem 6.2], $ L/K $ admits precisely $ 2+p(2q-3) $ Hopf-Galois structures: the classical structure, the canonical nonclassical structure, $ 2p(q-2) $ further structures of metacyclic type, and $ p $ structures of cyclic type. We can use Example \[example\_Hopf\_isomorphism\_conjugation\] to show that the Hopf algebras appearing in the Hopf-Galois structures of cyclic type are all isomorphic as Hopf algebras. The regular subgroups $ N_{c} $ ($c=0, \ldots ,p-1$) of $ {\mbox{Perm}(G)} $ corresponding to these Hopf-Galois structures are described explicitly in [@Byott2004 Lemma 4.1], each in terms of two generators. Using these descriptions, we can verify that $ N_{c} = \langle \eta_{c} \rangle $, where $$\eta_{c}[\sigma^{u} \tau^{v}] = \sigma^{u+1-cg^{v}} \tau^{v+1}.$$ That is: $$\eta_{c} = \lambda(\sigma) \rho(\sigma^{-c}\tau)^{-1}.$$ Using the fact that $ \lambda(G) $ and $ \rho(G) $ commute inside $ {\mbox{Perm}(G)} $ we have in particular, for $ c=0 $, $$\begin{aligned}
\rho(\sigma^{i}) \eta_{0} \rho(\sigma^{-i}) & = & \lambda(\sigma) \rho(\sigma^{i} \tau \sigma^{-i} )^{-1} \\
& = & \lambda(\sigma) \rho(\sigma^{i(1-g)}\tau)^{-1} \\
& = & \eta_{i(g-1)},\end{aligned}$$ where the subscript should be interpreted modulo $ p $. Since $ g $ has order $ q $ modulo $ p $ we certainly have $ g-1 \neq 0 \pmod{p} $ and so, given $ c= 0, \ldots ,p-1 $ there exists $ i $ such that $$\rho(\sigma^{i}) \eta_{0} \rho(\sigma^{-i}) = \eta_{c}.$$ Therefore the groups $ N_{c} $ are all isomorphic as $ G $-groups, and so the Hopf algebras $ H_{c} $ are all isomorphic as $ K $-Hopf algebras.
In the case that $ q=2 $ (so that $ G \cong D_{p} $), this result is established in [@TARP_PROMS2017 Proposition 4.3], by methods different to those employed above.
Hopf algebra isomorphisms after base change {#subsection_Hopf-algebra_isomorphism_base_change}
-------------------------------------------
We return to the situation addressed by the theorem of Greither and Pareigis, as described in section \[section\_introduction\]: $ L/K $ is a finite separable extension of fields whose Galois closure $ E $ has group $ G $, so that a Hopf algebra giving a Hopf-Galois structure on $ L/K $ has the form $ E[N]^{G} $ for some $ G $-stable regular subgroup of $ {\mbox{Perm}(X)} $. If $ E[N_{1}]^{G} $ and $ E[N_{2}]^{G} $ are two such Hopf algebras then, by Galois descent, we have $ E \otimes_{K} E[N_{i}]^{G} = E[N_{i}] $ for each $ i $, so certainly $ E \otimes_{K} E[N_{1}]^{G} $ and $ E \otimes_{K} E[N_{2}]^{G} $ are isomorphic as $ E $-Hopf algebras. However, there may exist intermediate fields $ K^{\prime} $ such that $ K^{\prime} \otimes_{K} E[N_{1}]^{G} $ and $ K^{\prime} \otimes_{K} E[N_{2}]^{G} $ are isomorphic as $ K^{\prime} $-Hopf algebras. In this section we identify the fields with this property that are also Galois over $ K $.
\[theorem\_Hopf\_algebra\_isomorphism\_base\_change\] Let $ E[N_{1}]^{G} $ and $ E[N_{2}]^{G} $ give Hopf-Galois structures on $ L/K $, let $ G^{\prime} $ be a normal subgroup of $ G $, and let $ K^{\prime} = E^{G^{\prime}} $. Then $$K^{\prime} \otimes E[N_{1}]^{G} \cong K^{\prime} \otimes_{K} E[N_{2}]^{G} \mbox{ as $ K^{\prime} $-Hopf algebras}$$ if and only if $ (N_{1},\ast) \cong (N_{2},\ast) $ as $ G^{\prime} $-groups.
For $ i=1,2 $ we have by Galois descent that $ E[N_{i}]^{G^{\prime}} $ is a $ K^{\prime} $-Hopf algebra, and one may show that action of $ G $ on $ E[N_{i}] $ induces an action of the group $ G/G^{\prime} $ on $ E[N_{i}]^{G^{\prime}} $ and that $$E[N_{i}]^{G} = \left( E[N_{i}]^{G^{\prime}} \right)^{G/G^{\prime}}.$$ Now since $ K^{\prime}/K $ is Galois with group $ G/G^{\prime} $, we have $$K^{\prime} \otimes_{K} \left( E[N_{i}]^{G^{\prime}} \right)^{G/G^{\prime}} = E[N_{i}]^{G^{\prime}},$$ again by Galois descent. Therefore $$K^{\prime} \otimes_{K} E[N_{i}]^{G} = E[N_{i}]^{G^{\prime}}$$ for $ i=1,2 $, and so $ K^{\prime} \otimes_{K} E[N_{1}]^{G} \cong K^{\prime} \otimes_{K} E[N_{2}]^{G} $ as $ K^{\prime} $-Hopf algebras if and only if $ E[N_{1}]^{G^{\prime}} \cong E[N_{2}]^{G^{\prime}} $ as $ K^{\prime} $-Hopf algebras. By Theorem \[theorem\_Hopf\_algebra\_isomorphism\] this occurs if and only if $ (N_{1},\ast) \cong (N_{2},\ast) $ as $ G^{\prime} $-groups.
[**(The smallest extension of scalars giving a group algebra)**]{} \[example\_extension\_of\_scalars\_group\_algebra\] Let $ E[N]^{G} $ be a Hopf algebra giving a Hopf-Galois structure on $ L/K $, and let $ G^{\prime} $ denote the kernel of the action of $ G $ on $ N $. Then $ E[N]^{G^{\prime}} = E^{G^{\prime}}[N] = K^{\prime}[N] $, a group algebra with coefficients drawn from the field $ E^{G^{\prime}}=K^{\prime} $. By Theorem \[theorem\_Hopf\_algebra\_isomorphism\_base\_change\] we have $ K^{\prime} \otimes_{K} E[N]^{G} =K^{\prime}[N] $. In fact $ K^{\prime} $ is minimal amongst the subfields $ F $ of $ E $ such that $ F \otimes_{K} E[N]^{G} $ is isomorphic as a Hopf algebra to a group algebra (see [@GreitherPareigis1987 Corollary 3.2]).
[**(Elementary abelian extensions of degree $ p^{2} $ revisited)**]{} Let $ p $ be an odd prime, and let $ L/K $ be an elementary abelian extension of degree $ p^{2} $ with group $ G $. In Example \[example\_Hopf\_isomorphism\_elementary\_abelian\_p2\] we determined which of the Hopf algebras appearing in the classification of Hopf-Galois structures on $ L/K $ are isomorphic as $ K $-Hopf algebras. Here we can show that, given any two Hopf algebras $ H_{1},H_{2} $ giving nonclassical Hopf-Galois structures on the extension, there exists a subfield $ K^{\prime} $ of $ L/K $ of degree $ p $ over $ K $ such that $ K^{\prime} \otimes_{K} H_{1} \cong K^{\prime} \otimes_{K} H_{2} $ as $ K^{\prime} $-Hopf algebras. Recall from Example \[example\_Hopf\_isomorphism\_elementary\_abelian\_p2\] that for $ i=1,2 $ the Hopf algebra $ H_{i} $ corresponds to a choice of subgroup $ T_{i} $ of degree $ p $ and an integer $ d_{i} \in \{1, \ldots ,p-1\} $ (the possibility $ d_{i}=0 $ is excluded since the Hopf algebras give nonclassical structures). Specifically, we have $ H_{i} = L[N_{i}]^{G} $, where $ N_{i} $ is generated by two permutations $ \alpha_{i}, \beta_{i} $ as described in Equations and the action of $ G $ on $ N_{i} $ is as described in Equation . From this last equation, we see that the kernel of the action of $ G $ on each $ N_{i} $ is precisely $ T_{i} $, which we now write as $ \langle t_{i} \rangle $. Let $ g^{\prime} = t_{1}t_{2} $, $ G^{\prime} = \langle g^{\prime} \rangle $, and $ K^{\prime} = L^{G^{\prime}} $; we claim that $ K^{\prime} \otimes_{K} H_{1} \cong K^{\prime} \otimes_{K} H_{2} $ as $ K^{\prime} $-Hopf algebras. By Theorem \[theorem\_Hopf\_algebra\_isomorphism\_base\_change\], we must show that there is a $ G^{\prime} $-equivariant isomorphism $ \varphi : N_{1} \xrightarrow{\sim} N_{2} $. For each $ i $ we have $ g^{\prime} \ast \alpha_{i} = \alpha_{i} $ for each $ i $, and so we focus our attention on the action of $ g^{\prime} $ on the $ \beta_{i} $. Since each $ T_{i} $ is precisely the kernel of the action of $ G $ on $ N_{i} $ and $ G $ is abelian, by equation we have $$g^{\prime} \ast \beta_{1} = t_{1}t_{2} \ast \beta_{1} = t_{2} \ast \beta_{1} = \alpha_{1}^{u_{1}} \beta_{1} \mbox{ for some $ u_{1} \in \{ 1, \ldots ,p-1 \} $}$$ and $$g^{\prime} \ast \beta_{2}= t_{1}t_{2} \ast \beta_{2} = t_{1} \ast \beta_{2} = \alpha_{2}^{u_{2}} \beta_{2} \mbox{ for some $ u_{2} \in \{ 1, \ldots ,p-1 \} $} .$$ There exists an integer $ c $ such that $ cu_{2} \equiv u_{1} \pmod{p} $; use this to define a homomorphism $ \varphi : N_{1} \rightarrow N_{2} $ by $$\varphi(\alpha_{1}) = \alpha_{2}, \hspace{4mm} \varphi(\beta_{1})=\beta_{2}^{c}.$$ It is clear that $ \varphi $ is a isomorphism of $ N_{1} $ onto $ N_{2} $, and to verify that it is $ G^{\prime} $-equivariant it suffices to consider the action of $ g^{\prime} $ on $ \beta_{1} $. We have: $$\begin{aligned}
\varphi( g^{\prime} \ast \beta_{1}) & = & \varphi( \alpha_{1}^{u_{1}} \beta_{1} ) \\
& = & \alpha_{2}^{u_{1}} \beta_{2}^{c} \\
& = & \alpha_{2}^{cu_{2}} \beta_{2}^{c} \\
& = & g^{\prime} \ast \beta_{2}^{c}.\end{aligned}$$ Therefore $ \varphi $ is a $ G^{\prime} $-equivariant isomorphism of $ N_{1} $ onto $ N_{2} $, and so Theorem \[theorem\_Hopf\_algebra\_isomorphism\_base\_change\] we have $$K^{\prime} \otimes_{K} H_{1} \cong K^{\prime} \otimes_{K} H_{2} \mbox{ as $ K^{\prime} $-Hopf algebras}.$$
Hopf algebra isomorphisms via the holomorph {#subsection_Hopf_algebra_isomorphisms_holomorph}
-------------------------------------------
We retain the notation of subsection \[subsection\_Hopf-algebra\_isomorphism\_base\_change\]. We have seen that a necessary condition for $ E[N_{1}]^{G} \cong E[N_{2}]^{G} $ as Hopf algebras is that the underlying groups are isomorphic; we now use this requirement to change our point of view. Fix an abstract group $ N $ of the same order as $ G $, and study embeddings $ \alpha: N \hookrightarrow {\mbox{Perm}(G)} $ such that $ \alpha(N) $ is $ G $-stable and regular. Byott’s translation theorem [@Byott1996 Proposition 1] relates such embeddings to certain embeddings of $ G $ into the holomorph of $ N $, denoted $ {\mbox{Hol}(N)} $, which is normalizer of $ \lambda(N) $ inside $ {\mbox{Perm}(N)} $. More precisely, there is a bijection between the sets $$\{ \alpha : N \hookrightarrow {\mbox{Perm}(X)} \mid \alpha \mbox{ an injective homomorphism whose image in $ G $-stable and regular} \}$$ and $$\{ \beta : G \hookrightarrow {\mbox{Hol}(N)} \mid \beta \mbox{ an injective homomorphism such that $ \beta(G_{L}) $ is the stabilizer of $ e_{N} $} \}.$$ Since $ {\mbox{Hol}(N)} = \rho(N) \rtimes {\mbox{Aut}(N)} $ [@Ch00 (7.2) Proposition], it is much smaller than $ {\mbox{Perm}(X)} $, and so Byott’s translation theorem is often useful in the enumeration of the Hopf-Galois structures admitted by a given extension. Of course, different embeddings $ \alpha $ can have the same image (and so correspond to the same Hopf-Galois structure), but this can be detected by studying the corresponding $ \beta $: we have $ \alpha_{1}(N)=\alpha_{2}(N) $ if and only if $ \beta_{1}(G) $ and $ \beta_{2}(G) $ are conjugate by an element of $ {\mbox{Aut}(N)} $. In fact, we can also detect when the Hopf-Galois structures corresponding to different embeddings $ \alpha_{1}, \alpha_{2} $ involve isomorphic Hopf algebras by studying properties of the corresponding $ \beta_{1}, \beta_{2} $:
\[theorem\_Hopf\_algebra\_isomorphism\_holomorph\] Let $ \alpha_{1}, \alpha_{2} $ be embeddings of $ N $ into $ {\mbox{Perm}(X)} $ whose images are regular and normalized by $ \lambda(G) $, and let $ \beta_{1}, \beta_{2} $ be the corresponding embeddings of $ G $ into $ {\mbox{Hol}(N)} $. Viewing $ {\mbox{Hol}(N)} $ as $ \rho(N) \rtimes {\mbox{Aut}(N)} $, let $ \overline{\beta_{1}}, \overline{\beta_{2}} $ denote the compositions of $ \beta_{1}, \beta_{2} $ with the projection onto the automorphism component. Then $$E[\alpha_{1}(N)]^{G} \cong E[\alpha_{2}(N)]^{G} \mbox{ as Hopf algebras}$$ if and only if there exists $ \mu \in {\mbox{Aut}(N)} $ such that $$\overline{\beta_{2}}(g) = \mu \overline{\beta_{1}}(g) \mu^{-1} \mbox{ for all } g \in G.$$
By Theorem \[theorem\_Hopf\_algebra\_isomorphism\], we have $ E[\alpha_{1}(N)]^{G} \cong E[\alpha_{2}(N)]^{G} $ as Hopf algebras if and only if $ (\alpha_{1}(N),\ast) \cong (\alpha_{2}(N),\ast) $ as $ G $-groups. For $ i=1,2 $ define an action $ \ast_{i} $ of $ G $ on $ N $ by $ g \ast_{i} \eta = \overline{\beta_{i}}(g) [ \eta ] $; then by [@Ch00 (7.7) Proposition] we have that $ (\alpha_{i}(N), \ast) \cong (N, \ast_{i}) $ as $ G $-groups, and so $ (\alpha_{1}(N),\ast) \cong (\alpha_{2}(N),\ast) $ as $ G $-groups if and only if $ (N, \ast_{1}) \cong (N, \ast_{2}) $ as $ G $-groups. This occurs if and only if there exists $ \mu \in {\mbox{Aut}(N)} $ such that $$\mu(g \ast_{1} \eta) = g \ast_{2} \mu(\eta) \mbox{ for all } g \in G, \eta \in N,$$ that is, if and only if $$\mu \overline{\beta_{1}}(g) = \overline{\beta_{2}}(g) \mu \mbox{ for all } g \in G.$$
As a special case, we have
\[corollary\_Hopf\_algebra\_isomorphism\_Aut\_N\_abelian\] If $ {\mbox{Aut}(N)} $ is abelian (in particular, if $ N $ is cyclic), then $ E[\alpha_{1}(N)]^{G} \cong E[\alpha_{2}(N)]^{G} $ as Hopf algebras if and only if $ \overline{\beta_{1}}(g) = \overline{\beta_{2}}(g) $ for all $ g \in G $.
[**(The classical and canonical nonclassical Hopf-Galois structures revisited)**]{} Recall from Example \[example\_Hopf\_isomorphism\_classical\_nonclassical\] that if $ L/K $ is a Galois extension with nonabelian group $ G $ then $ L/K $ admits at least two Hopf-Galois structures: the classical Hopf-Galois structure, which corresponds to the subgroup $ \rho(G) $ of $ {\mbox{Perm}(G)} $ and has Hopf algebra $ K[G] $, and the canonical nonclassical Hopf-Galois structure, which corresponds to the subgroup $ \lambda(G) $ of $ {\mbox{Perm}(G)} $ and has Hopf algebra $ H_{\lambda} = L[\lambda(G)]^{G} $. In the notation of this subsection, we may take $ N=G $ and view $ \lambda $ and $ \rho $ as embeddings of the abstract group $ G $ into $ {\mbox{Perm}(G)} $ whose images are $ G $-stable and regular. By following the details of the proof of Byott’s translation theorem, we find that the embedding $ G \hookrightarrow {\mbox{Hol}(G)} $ corresponding to $ \rho $ is $ \rho $ itself, and similarly for $ \lambda $. When we view $ {\mbox{Hol}(G)} $ as $ \rho(G) \rtimes {\mbox{Aut}(G)} $, we have $ \rho(G) = \{ (\rho(g),1) \mid g \in G \} $, whereas $ \lambda(G) = \{ (\rho(g^{-1}),c(g)) \mid g \in G \} $, where $ c(g) $ is the inner automorphism of $ G $ arising from conjugation by $ g $. Therefore $ \overline{\rho}(G) $ and $ \overline{\lambda}(G) $ have different orders, and so there cannot exist an automorphism $ \mu \in {\mbox{Aut}(G)} $ with the properties required by Theorem \[theorem\_Hopf\_algebra\_isomorphism\_holomorph\]. Hence we recover the fact that $ K[G] \not \cong H_{\lambda} $ as $ K $-Hopf algebras.
Algebra Isomorphisms for Commutative Structures {#section_algebra_isomorphisms_commutative}
===============================================
In this section, we consider the question of when two Hopf algebras $ E[N_{1}]^{G} $ and $ E[N_{2}]^{G} $ giving Hopf-Galois structures on $ L/K $ are isomorphic as $ K $-algebras. We shall assume that $ E[N_{1}]^{G} $ and $ E[N_{2}]^{G} $ are commutative algebras; this is equivalent to assuming that the underlying groups $ N_{1},N_{2} $ are abelian. Note, however, that we do not assume that these groups are isomorphic. We shall also assume that $ K $ has characteristic zero; a consequence of this is that each $ E[N_{i}]^{G} $ is a semisimple $ K $-algebra and therefore by the Wedderburn-Artin theorem is isomorphic as a $ K $-algebra to a product of extension fields of $ K $. We establish a criterion for $ E[N_{1}]^{G} \cong E[N_{2}]^{G} $ as $ K $-algebras.\
\
The question of explicitly determining the Wedderburn-Artin decomposition of a commutative Hopf algebra of this form has been answered by Boltje and Bley [@BleyBoltje] in terms of the action of the absolute Galois group of $ K $ on the values of characters of the underlying group. We review their approach using a finite extension of $ E $ in place of the separable closure of $ K $. Let $ e $ denote the least common multiple of the exponents of $ N_{1}, N_{2} $, let $ \zeta $ be a primitive $ e^{th} $ root of unity. Then $ K(\zeta) $ is a Galois extension of $ K $, and the compositum of $ E $ and $ K(\zeta) $, say $ \tilde{E} $, is a Galois extension of $ K $. Let $ \Gamma = {\mbox{Gal}(\tilde{E}/K)} $ and $ \Gamma_{E} = {\mbox{Gal}(\tilde{E}/E)} $, so that $ \Gamma / \Gamma_{E} \cong G $. $$\xymatrixcolsep{4pc}
\xymatrixrowsep{2pc}
\xymatrix{
&\tilde{E} \ar@{-}[ddd]_{\Gamma} & \\
E \ar@{-}[d]_{G_{L}} \ar@{-}[ddr]^{G} \ar@{-}[ur]^{\Gamma_{E}}& & \\
L \ar@{-}[dr]& & \ar@{-}[dl] \ar@{-}[uul]K(\zeta)\\
&K& \\
}$$ For $ i=1,2 $ let $ \Gamma $ act on $ \tilde{E}[N_{i}] $ by acting on $ \tilde{E} $ as Galois automorphisms and on $ N_{i} $ by factoring through $ \Gamma / \Gamma_{E} $ (i.e. through $ G $). Then $$\tilde{E}[N_{i}]^{\Gamma} = \left( \tilde{E}[N_{i}]^{\Gamma_{E}} \right)^{\Gamma / \Gamma_{E}} = \left( \tilde{E}^{\Gamma_{E}}[N_{i}] \right)^{\Gamma / \Gamma_{E}} = E[N_{i}]^{G}.$$ (In the second term $ \Gamma_{E} $ acts trivially on $ N_{i} $, so we have $ \tilde{E}[N_{i}]^{\Gamma_{E}} = \tilde{E}^{\Gamma_{E}}[N_{i}] = E[N_{i}] $, the group algebra of $ N_{i} $ with coefficients drawn from $ \tilde{E}^{\Gamma_{E}} = E $.) Now elements of $ \widehat{N_{i}} $, namely the characters of $ N_{i} $, have values which all lie in $ \tilde{E} $, and so $ \tilde{E}[N_{i}] $ is isomorphic to $ \tilde{E}^{n} $ as an $ \tilde{E} $-algebra via orthogonal idempotents, one corresponding to each character. Writing $ \widehat{N_{i}} $ for the dual group of $ N_{i} $, the idempotent corresponding to $ \chi \in \widehat{N_{i}} $ is $$e_{\chi} = \sum_{\eta \in N_{i}} \chi(\eta) \eta \in \tilde{E}[N_{i}].$$ This action of $ \Gamma $ on $ \tilde{E}[N_{i}] $ permutes the orthogonal idempotents, which in turn induces an action of $ \Gamma $ on $ \widehat{N_{i}} $, as follows: if $ \chi \in \widehat{N_{i}} $ and $ \gamma \in \Gamma $ then $$\begin{aligned}
\gamma \ast e_{\chi} & = & \gamma \ast \left( \sum_{\eta \in N_{i}} \chi(\eta) \eta \right) \\
& = & \sum_{\eta \in N_{i}} \gamma(\chi(\eta)) \gamma \ast \eta \\
& = & \sum_{\eta \in N_{i}} \gamma(\chi(\gamma^{-1} \ast \eta)) \eta \mbox{ (reindexing).}\end{aligned}$$ We therefore define $ \gamma \ast \chi $ by $ (\gamma \ast \chi [\eta] = \gamma(\chi[\gamma^{-1} \ast \eta]) $ for all $ \eta \in N_{i} $. Now let $ \chi_{1}, \ldots ,\chi_{r} $ be representatives of the orbits of $ \Gamma $ in $ \widehat{N_{i}} $, and for each $ j=1, \ldots ,r $ let $ S_{j} = \mbox{Stab}(\chi_{j}) $. Then [@BleyBoltje Lemma 2.2] asserts that $$\tilde{E}[N_{i}]^{\Gamma} \cong \prod_{j=1}^{r} \tilde{E}^{S_{j}} \mbox{ as $ K $-algebras.}$$ From this, it is clear that if there exist sets of representatives of the orbits of $ \Gamma $ in $ \widehat{N_{1}} $ and $ \widehat{N_{2}} $ having the same stabilizers then $ \tilde{E}[N_{1}]^{\Gamma} \cong \tilde{E}[N_{2}]^{\Gamma} $ as $ K $-algebras. The following theorem is a refinement of this idea. Note that as in section \[section\_Hopf\_algebra\_isomorphisms\] we formulate the theorem in more general terms than we require: the added flexibility will be useful later.
\[theorem\_algebra\_isomorphism\] Let $ (N_{1}, \ast_{1}) $ and $ (N_{2}, \ast_{2}) $ be abelian $ \Gamma $-groups and, for $ i=1,2 $, let $ \Gamma $ act on $ \tilde{E}[N_{i}] $ by acting on $ \tilde{E} $ as Galois automorphisms and on $ N_{i} $ via $ \ast_{i} $. Then $ \tilde{E}[N_{1}]^{\Gamma} \cong \tilde{E}[N_{2}]^{\Gamma} $ as $ K $-algebras if and only if $ \widehat{N_{1}} \cong \widehat{N_{2}} $ as $ \Gamma $-sets.
Suppose first that $ \psi : \widehat{N_{1}} \rightarrow \widehat{N_{2}} $ is a $ \Gamma $-equivariant bijection. Let $ \chi_{1}, \ldots , \chi_{r} $ be representatives of the orbits of $ \Gamma $ in $ \widehat{N_{1}} $. Then $ \psi(\chi_{1}), \ldots , \psi(\chi_{r}) $ are representatives of the orbits of $ \Gamma $ in $ \widehat{N_{2}} $, and for each $ i = 1, \ldots ,r $ we have $$\begin{aligned}
\gamma \in \mbox{Stab}(\psi(\chi_{i})) & \Leftrightarrow & \gamma \ast \psi(\chi_{i}) = \psi(\chi_{i}) \\
& \Leftrightarrow & \psi(\gamma \ast \chi_{i}) = \psi(\chi_{i}) \mbox{ ($\psi$ is $ \Gamma $-equivariant)}\\
& \Leftrightarrow &\gamma \ast \chi_{i} = \chi_{i} \mbox{ ($\psi$ is a bijection)}\\
& \Leftrightarrow & \gamma \in \mbox{Stab}(\chi_{i}). \end{aligned}$$ Therefore $ \mbox{Stab}(\psi(\chi_{i})) = \mbox{Stab}(\chi_{i}) $ for each $ i $, and so $$H_{1} \cong \prod_{i=1}^{r} \tilde{E}^{\text{Stab}(\chi_{i})} = \prod_{i=1}^{r} \tilde{E}^{\text{Stab}(\psi(\chi_{i}))} \cong H_{2} \mbox{ as $ K $-algebras}.$$ Conversely, suppose that $ \varphi : H_{1} \xrightarrow{\sim} H_{2} $ is an isomorphism of $ K $-algebras. It extends to an isomorphism of $ \tilde{E} $-algebras $ \varphi : \tilde{E}[N_{1}] \cong \tilde{E}[N_{2}] $ which, by the argument employed in the proof of Theorem \[theorem\_Hopf\_algebra\_isomorphism\], is $ \Gamma $-equivariant. Now recall that $ \tilde{E}[N_{1}] \cong \tilde{E}[N_{2}] \cong \tilde{E}^{n} $ as $ \tilde{E} $-algebras via orthogonal idempotents, as described above. Let $ \{ e_{\chi} \mid \chi \in \widehat{N_{1}} \} $ be the orthogonal idempotents of $ \tilde{E}[N_{1}] $, and $ \{ f_{\chi^{\prime}} \mid \chi^{\prime} \in \widehat{N_{2}} \} $ be the orthogonal idempotents of $ \tilde{E}[N_{2}] $. The set $ \{ \varphi(e_{\chi}) \mid \chi \in \widehat{N_{1}} \} $ is the $ \tilde{E} $-basis of orthogonal idempotents of $ \tilde{E}[N_{2}] $, and so for each $ \chi \in \widehat{N_{1}} $ there exists $ \psi(\chi) \in \widehat{N_{2}} $ such that $ \varphi(e_{\chi}) = f_{\psi(\chi)} $. This establishes a bijection $ \psi : \widehat{N_{1}} \rightarrow \widehat{N_{2}} $. Since $ \varphi $ is $ \Gamma $-equivariant, for $ \gamma \in \Gamma $ we have: $$f_{(\gamma \ast\psi)(\chi)} = \gamma \ast f_{\psi(\chi)} = \gamma \ast \varphi( e_{\chi} ) = \varphi(\gamma \ast e_{\chi} ) = \varphi( e_{\gamma \ast \chi} ) = f_{\psi(\gamma \ast \chi)},$$ and so $ (\gamma \ast \psi)(\chi) = \psi(\gamma \ast \chi) $. Therefore the bijection $ \psi : \widehat{N_{1}} \rightarrow \widehat{N_{2}} $ is $ \Gamma $-equivariant.
The following corollary is the principal application of Theorem \[theorem\_algebra\_isomorphism\].
\[corollary\_algebra\_isomorphism\] Let $ E[N_{1}]^{G} $ and $ E[N_{2}]^{G} $ be commutative Hopf algebras giving Hopf-Galois structures on $ L/K $. Then $ E[N_{1}]^{G} \cong E[N_{2}]^{G} $ as $ K $-algebras if and only if $ \widehat{N_{1}} \cong \widehat{N_{2}} $ as $ \Gamma $-sets.
[**(Elementary abelian extensions of degree $ p^{2} $ revisited)** ]{} Let $ p $ be an odd prime and let $ L/K $ be an elementary abelian extension of degree $ p^{2} $ with group $ G $. In Example \[example\_Hopf\_isomorphism\_elementary\_abelian\_p2\] we determined criteria for two Hopf algebras $ H_{1}, H_{2} $ giving Hopf-Galois structures on $ L/K $ to be isomorphic as $ K $-Hopf algebras. Under the additional hypotheses that $ K $ has characteristic zero and contains a primitive $ p^{th} $ root of unity $ \zeta $, we now apply Corollary \[corollary\_algebra\_isomorphism\] to determine criteria for them to be isomorphic as $ K $-algebras. In fact, we show that in this case $ H_{1} \cong H_{2} $ as $ K $-algebras if and only if $ H_{1} \cong H_{2} $ as $ K $-Hopf algebras.\
\
Recall from Example \[example\_Hopf\_isomorphism\_elementary\_abelian\_p2\] a Hopf algebra $ H $ giving a Hopf-Galois structures on $ L/K $ determined by a choice of subgroup $ T $ of degree $ p $ and an integer $ d \in \{0, \ldots ,p-1\} $; the Hopf algebra is then $ L[N]^{G} $, where $ N $ is the subgroup of $ {\mbox{Perm}(G)} $ generated by two permutations $ \alpha, \beta $ as described in Equations , and the action of $ G $ on $ N $ is as described in Equation . The dual group $ \widehat{N} $ is therefore generated by two characters $ \chi, \psi $, defined as follows: $$\chi(\alpha) = \zeta , \hspace{4mm} \chi(\beta) = 1 , \hspace{4mm} \psi(\alpha) = 1 , \hspace{4mm} \psi(\beta) = \zeta.$$ Since $ L/K $ is a Galois extension we have $ E = L $, and the hypothesis that $ K $ contains a primitive $ p^{th} $ root of unity implies that in the notation of this section we have $ \tilde{E}=E=L $ and $ \Gamma = G $. The action of $ G $ on $ N $ described in Equation translates into an action of $ G $ on $ \widehat{N} $ by $$\label{equation_byott_p2_lambda_action_dual}
s \ast \chi = \chi \psi^{d^{-1}}, \hspace{4mm} t \ast \chi = \chi, \hspace{4mm} s \ast \psi = \psi, \hspace{4mm} t \ast \psi=\psi,$$ where $ d^{-1} $ is to be interpreted modulo $ p $. Now let $ H_{1}, H_{2} $ be two such Hopf algebras giving distinct Hopf-Galois structures on $ L/K $, with underlying groups $ N_{1}, N_{2} $ as in Example \[example\_Hopf\_isomorphism\_elementary\_abelian\_p2\]. We claim that $ H_{1} \cong H_{2} $ as $ K $-algebras if and only if $ d_{1}=d_{2}=0 $ or $ d_{1}d_{2} \neq 0 $ and $ T_{1}=T_{2} $; these are identical to the conditions we derived in Example \[example\_Hopf\_isomorphism\_elementary\_abelian\_p2\] for $ H_{1} \cong H_{2} $ as $ K $-Hopf algebras. If these conditions are satisfied then $ H_{1} \cong H_{2} $ as $ K $-Hopf algebras, so certainly as $ K $-algebras. For the converse, note that by if $ d_{i}=0 $ then $ G $ acts trivially on $ \widehat{N_{i}} $, whereas if $ T $ is a subgroup of $ G $ of order $ p $ and $ d \neq 0 $ then the kernel of the action of $ G $ on $ N_{T,d} $ is precisely $ T $. Therefore if $ \widehat{N_{1}} \cong \widehat{N_{2}} $ as $ G $-sets then we must have $ d_{1} = d_{2} = 0 $ or $ d_{1}d_{2} \neq 0 $ and $ T_{1} = T_{2} $.\
\
We remark that if one determines a set of representatives for the orbits of $ \Gamma $ in $ \widehat{N_{i}} $, and the associated stabilizers, then one finds that $ H_{i} \cong K^{p} \times \left(L^{T_{i}}\right)^{p-1} $ as $ K $-algebras; this result was obtained by slightly different methods in [@PJT_CpxCp Proposition 3.4].
Algebra isomorphisms via the holomorph
--------------------------------------
Hitherto in this section we have not assumed that $ N_{1}, N_{2} $ are isomorphic as groups; we now impose this assumption. We can therefore view $ N_{1}, N_{2} $ as the images of a single abstract abelian group $ N $ under two embeddings $ \alpha_{1}, \alpha_{2} : N \hookrightarrow {\mbox{Perm}(X)} $, as in subsection \[subsection\_Hopf\_algebra\_isomorphisms\_holomorph\]. We recall from that subsection that By Byott’s translation theorem these embeddings correspond to embeddings $ \beta_{1}, \beta_{2} : G \hookrightarrow {\mbox{Hol}(N)} = \rho(N) \rtimes {\mbox{Aut}(N)} $, and we write $ \overline{\beta_{1}}, \overline{\beta_{2}} $ for the compositions of $ \beta_{1}, \beta_{2} $ with the projection onto the $ {\mbox{Aut}(N)} $ component. We have seen in Theorem \[theorem\_Hopf\_algebra\_isomorphism\_holomorph\] that it is possible to detect when $ E[\alpha_{1}(N)]^{G} \cong E[\alpha_{2}(N)]^{G} $ as $ K $-Hopf algebras by studying properties of $ \overline{\beta_{1}} , \overline{\beta_{2}} $. We shall show that in our situation these maps also allow us to detect $ K $-algebra isomorphisms.\
\
As in the proof of Theorem \[theorem\_Hopf\_algebra\_isomorphism\_holomorph\], for $ i=1,2 $ define an action $ \ast_{i} $ of $ G $ on $ N $ by $ g \ast_{i} \eta = \overline{\beta_{i}}(g) [ \eta ] $; then by [@Ch00 (7.7)] we have that $ (\alpha_{i}(N), \ast) \cong (N, \ast_{i}) $ as $ G $-groups. We may extend these to actions of $ \Gamma $ by factoring through $ G $, obtaining $ (\alpha_{i}(N), \ast) \cong (N, \ast_{i}) $ as $ \Gamma $-groups. Each of the actions of $ \Gamma $ on $ N $ yields a dual action of $ \Gamma $ on $ \widehat{N} $, which we also denote by $ \ast_{i} $. Similarly, the action of $ \Gamma $ on each $ \alpha_{i}(N) $ yields a dual action $ \ast $ of $ \Gamma $ on each $ \widehat{\alpha_{i}(N)} $. Then we have:
\[lemma\_dual\_equivairant\_isomorphism\] For $ i=1,2 $, we have $ (\widehat{\alpha_{i}(N)}, \ast) \cong (\widehat{N}, \ast_{i}) $ as $ \Gamma $-groups.
For each $ i=1,2 $, the map $ \alpha_{i} : (N,\ast_{i}) \rightarrow (\alpha_{i}(N),\ast) $ is a $ \Gamma $-equivariant isomorphism. For each $ \chi \in \widehat{N} $, the function $ \psi_{\chi} : \alpha_{i}(N) \rightarrow \tilde{E} $ defined by $ \psi_{\chi}[\alpha_{i}(\eta)] = \chi[\eta] $ for all $ \eta \in N $ is actually a character of $ \alpha_{i}(N) $, since for $ \eta,\eta^{\prime} \in N $ we have $$\begin{aligned}
\psi_{\chi}[\alpha_{i}(\eta)\alpha_{i}(\eta^{\prime})] & = & \psi_{\chi}[\alpha_{i}(\eta\eta^{\prime})] \\
& = & \chi[\eta\eta^{\prime}] \\
& = & \chi[\eta]\chi[\eta^{\prime}] \\
& = & \psi_{\chi}[\alpha_{i}(\eta)]\psi_{\chi}[\alpha_{i}(\eta^{\prime})].\end{aligned}$$ We shall show that the map $ \psi : \widehat{N} \rightarrow \widehat{\alpha_{i}(N)} $ defined by $ \psi(\chi) = \psi_{\chi} $ is a $ \Gamma $-equivariant isomorphism. To show that it is an injection, let $ \chi, \chi^{\prime} \in \widehat{N} $. Then $$\begin{aligned}
\psi(\chi) = \psi(\chi^{\prime}) & \Rightarrow & \psi(\chi)[\alpha_{i}(\eta)] = \psi(\chi^{\prime})[\alpha_{i}(\eta)] \mbox{ for all $ \eta \in N $} \\
& \Rightarrow & \chi(\eta) = \chi^{\prime}(\eta) \mbox{ for all $ \eta \in N $}\\
& \Rightarrow & \chi = \chi^{\prime}. \end{aligned}$$ Thus $ \psi $ is an injection and, since $ |\widehat{N}| = |\widehat{\alpha_{i}(N)}| $, therefore a bijection. Next we show that $ \psi $ is a homomorphism. Let $ \chi, \chi^{\prime} $ be as above; then for all $ \eta \in N $ we have $$\begin{aligned}
\psi(\chi \chi^{\prime})[\alpha_{i}(\eta)] & = & (\chi \chi^{\prime})[\eta] \\
& = & \chi[\eta] \chi^{\prime}[\eta] \\
& = & \psi(\chi)[\alpha_{i}(\eta)] \psi(\chi^{\prime})[\alpha_{i}(\eta)] \\
& = & (\psi(\chi)\psi(\chi^{\prime}))[\alpha_{i}(\eta)].\end{aligned}$$ Hence $ \psi(\chi \chi^{\prime}) = \psi(\chi)\psi(\chi^{\prime}) $, and so $ \psi $ is a homomorphism. Finally, we show that $ \psi $ is $ \Gamma $-equivariant. Let $ \chi \in \widehat{N} $ and $ \gamma \in \Gamma $; then for all $ \eta \in N $ we have $$\begin{aligned}
(\gamma \ast \psi(\chi))[\alpha_{i}(\eta)] & = & \gamma(\psi(\chi)[\gamma^{-1} \ast \alpha_{i}(\eta)]) \\
& = & \gamma(\psi(\chi)[\alpha_{i}(\gamma^{-1} \ast_{i} \eta)]) \\
& = & \gamma(\chi[\gamma^{-1} \ast_{i} \eta]) \\
& = & (\gamma \ast_{i} \chi)[\eta] \\
& = & (\psi(\gamma \ast \chi))[\alpha_{i}(\eta)]. \end{aligned}$$ Therefore $ \gamma \ast \psi(\chi) = \psi(\gamma \ast \chi) $, and so $ \psi $ is $ \Gamma $-equivariant.\
\
We have shown that $ \psi : \widehat{N} \rightarrow \widehat{\alpha_{i}(N)} $ is a $ \Gamma $-equivariant isomorphism, and so $ (\widehat{\alpha_{i}(N)}, \ast) \cong (\widehat{N}, \ast_{i}) $ as $ \Gamma $-groups.
\[theorem\_algebra\_isomorphism\_holomorph\] We have $ E[\alpha_{1}(N)]^{G} \cong E[\alpha_{2}(N)]^{G} $ as $ K $-algebras if and only if $ (\widehat{N},\ast_{1}) \cong (\widehat{N},\ast_{2}) $ as $ \Gamma $-sets.
By Theorem \[theorem\_algebra\_isomorphism\] we have $ \tilde{E}[\alpha_{1}(N)]^{\Gamma} \cong \tilde{E}[\alpha_{2}(N)]^{\Gamma} $ as $ K $-algebras if and only if $ (\widehat{\alpha_{1}(N)},\ast) \cong (\widehat{\alpha_{2}(N)},\ast) $ as $ \Gamma $-sets, and by Lemma \[lemma\_dual\_equivairant\_isomorphism\] this occurs if and only if $ (\widehat{N},\ast_{1}) \cong (\widehat{N},\ast_{2}) $ as $ \Gamma $-sets.
The following corollary shows that for certain structures of cyclic type we can dispense with $ \widehat{N} $ and work directly with $ N $:
\[corollary\_algebra\_isomorphism\_cyclic\] Suppose that $ N $ is cyclic and that $ K $ contains a primitive $ |N|^{th} $ root of unity $ \zeta $. Then the following are equivalent:
1. $ E[\alpha_{1}(N)]^{G} \cong E[\alpha_{2}(N)]^{G} $ as $ K $-algebras;
2. $ (\widehat{\alpha_{1}(N)},\ast) \cong (\widehat{\alpha_{2}(N)},\ast) $ as $ \Gamma $-sets.
3. $ (\widehat{N},\ast_{1}) \cong (\widehat{N},\ast_{2}) $ as $ \Gamma $-sets;
4. $ (\alpha_{1}(N),\ast) \cong (\alpha_{2}(N),\ast) $ as $ \Gamma $-sets.
5. $ (N,\ast_{1}) \cong (N,\ast_{2}) $ as $ \Gamma $-sets;
(1), (2), and (3) are equivalent by Theorem \[theorem\_algebra\_isomorphism\] and Theorem \[theorem\_algebra\_isomorphism\_holomorph\], and (4) and (5) are equivalent since for each $ i $ we have $ (\alpha_{i}(N), \ast) \cong (N, \ast_{i}) $ as $ \Gamma $-groups. We show that (3) is equivalent to (5).\
\
Let $ \eta $ be a generator of $ N $, and define $ \chi : N \rightarrow K $ by $ \chi(\eta) = \zeta $; then $ \chi $ is a generator of $ \widehat{N} $. We claim that for $ i=1,2 $, the isomorphism $ f : N \rightarrow \widehat{N} $ defined by $ f(\eta) = \chi $ has the property that $$(\gamma \ast_{i} f)[\eta] = f(\gamma^{-1} \ast_{i} \eta).$$ To see this, note that each $ \gamma \in \Gamma $ acts as an automorphism of $ N $, so there exists an integer $ e(i,\gamma) $ (coprime to $ |N|$) such that $ \gamma \ast_{i} \eta = \eta^{e(i,\gamma)} $. Now using the assumption that $ \zeta \in K $, we have: $$(\gamma \ast_{i} \chi) [\eta] = \chi[\gamma^{-1} \ast_{i} \eta] = \chi[\eta^{e(i,\gamma)^{-1}}] = \chi[\eta]^{e(i,\gamma)^{-1}} = \chi^{e(i,\gamma)^{-1}}[\eta],$$ so in fact $ \gamma \ast_{i} \chi = \chi^{e(i,\gamma)^{-1}} $ (where $ e(i,\gamma)^{-1} $ is computed modulo $ |N| $). It follows that the isomorphism $ f $ has the desired property. We therefore obtain a diagram: $$\xymatrixcolsep{4pc}
\xymatrixrowsep{4pc}
\xymatrix{
(N,\ast_{1}) \ar[d]^{f} \ar@{-->}[r]^{\pi} & (N,\ast_{2}) \ar[d]^{f} \\
(\widehat{N},\ast_{1}) \ar@{-->}[r]^{\widehat{\pi}} & (\widehat{N},\ast_{2}) }$$ If $ \pi $ is a $ \Gamma $-equivariant bijection, then $ f \circ \pi \circ f^{-1} $ is a $ \Gamma $ equivariant bijection, and if $ \widehat{\pi} $ is a $ \Gamma $-equivariant bijection, then $ f^{-1} \circ \psi \circ f $ is a $ \Gamma $-equivariant bijection.
Cyclic Extensions of Prime Power Degree {#section_cyclic_p_power_extensions}
=======================================
Let $ p $ be an odd prime number and $ L/K $ a cyclic extension of degree $ p^{n} $. By a result of Kohl [@Kohl1998 Theorem 3.3] (see also [@Ch00 (9.1)]), there are precisely $ p^{n-1} $ Hopf-Galois structures on $ L/K $, and they all have cyclic type. Explicit $ K $-algebra generators for the Hopf algebras appearing in these Hopf-Galois structures were determined in [@Childs2011 §6.3], requiring intricate manipulations. In this section we apply to results of section \[section\_Hopf\_algebra\_isomorphisms\] and \[section\_algebra\_isomorphisms\_commutative\] to determine which of the Hopf algebras appearing in these Hopf-Galois structures are isomorphic as $ K $-Hopf algebras or $ K $-algebras, and (under certain additional hypotheses) explicitly determine their Wedderburn-Artin decompositions.\
\
Since the Hopf-Galois structures admitted by $ L/K $ all have cyclic type, we can view the corresponding regular subgroups of $ {\mbox{Perm}(G)} $ as images of a single abstract cyclic group $ N = \langle \eta \rangle $ of order $ p^{n} $ under $ p^{n-1} $ different regular embeddings $ \alpha_{s} : N \hookrightarrow {\mbox{Perm}(G)} $. By Byott’s translation, each such $ \alpha_{s} $ corresponds to an embedding $ \beta_{s} : G \rightarrow {\mbox{Hol}(N)} $, and these are described in [@Ch00 (8.6) and (9.1)]: let $ G = \langle \sigma \rangle $, and let $ \delta $ be the $ (p-1)^{st} $ power of some generator of the cyclic group $ {\mbox{Aut}(N)} $. Then the embeddings we seek are of the form $ \beta_{s}: G \hookrightarrow {\mbox{Hol}(N)} $ with $$\beta_{s}(\sigma) = ( \rho(\eta), \delta^{s} ), 0 \leq s < p^{n-1}.$$ For each $ s $, let $ \alpha_{s} : N \hookrightarrow {\mbox{Perm}(G)} $ denote the embedding corresponding to $ \beta_{s} $, and let $ H_{s} = L[\alpha_{s}(N)]^{G} $ denote the corresponding Hopf algebra.
\[theorem\_cyclic\_p\_power\_Hopf\_algebra\_isomorphisms\] Let $ 0 \leq r,s < p^{n-1} $. Then $ H_{r} \cong H_{s} $ as $ K $-Hopf algebras if and only if $ r = s $.
Since $ N $ is cyclic, $ {\mbox{Aut}(N)} $ is abelian, and so by Corollary \[corollary\_Hopf\_algebra\_isomorphism\_Aut\_N\_abelian\] we have $ H_{r} \cong H_{s} $ as $ K $-Hopf algebras if and only if $ \overline{\beta_{r}}(g) = \overline{\beta_{s}}(g) $ for all $ g \in G $. Since in this case $ G $ is generated by $ \sigma $, this occurs if and only if $ \overline{\beta_{r}}(\sigma) = \overline{\beta_{s}}(\sigma) $; that is, if and only if $ \delta^{r} = \delta^{s} $. Hence $ H_{r} \cong H_{s} $ as $ K $-Hopf algebras if and only if $ r = s $.
Therefore the Hopf algebras giving the Hopf-Galois structures on $ L/K $ are pairwise nonisomorphic. We can use Theorem \[theorem\_Hopf\_algebra\_isomorphism\_base\_change\] to determine which of them become isomorphic under various base changes. Let $$K = K_{0} \subset K_{1} \subset \cdots \subset K_{n}=L$$ be the maximal tower of field extensions, and for each $ i=0, \ldots ,n $ let $ G_{i} =\langle \sigma^{p^{i}} \rangle = {\mbox{Gal}(L/K_{i})} $.
\[theorem\_cyclic\_p\_n\_Hopf\_isomorphism\] For $ 0 \leq r,s \leq p^{n-1} $ and $ 0 \leq i \leq n $, we have $ K_{i} \otimes_{K} H_{r} \cong K_{i} \otimes_{K} H_{s} $ as $ K_{i} $-Hopf algebras if and only if $ r \equiv s \pmod {p^{n-1-i}} $.
By Theorem \[theorem\_Hopf\_algebra\_isomorphism\_base\_change\], we have $ K_{i} \otimes_{K} H_{r} \cong K_{i} \otimes_{K} H_{s} $ as $ K_{i} $-Hopf algebras if and only if $ (\alpha_{r}(N),\ast) \cong (\alpha_{s}(N), \ast) $ as $ G_{i} $-groups. By [@Ch00 (7.7)], this is equivalent to $ (N, \ast_{r}) \cong (N, \ast_{s}) $ as $ G_{i} $-groups, so we must show that this occurs if and only if $ r \equiv s \pmod {p^{n-1-i}} $.\
\
Recall that $ \delta \in {\mbox{Aut}(N)} $ has order $ p^{n-1} $, so there exists an element $ d \in (\mathbb{Z}/p^{n}\mathbb{Z})^{\times} $ of order $ p^{n-1} $ such that $ \delta(\eta)=\eta^{d} $. It follows that for $ 0 \leq j \leq p^{n}-1 $ we have $$\sigma^{j} \ast_{r} \eta = \overline{\beta_{r}}(\sigma)^{j} [\eta] = \delta^{rj} \eta = \eta^{d^{rj}},$$ and similarly $ \sigma^{j} \ast_{s} \eta = \eta^{d^{sj}} $. Now let $ \theta $ be an automorphism of $ N $, and write $ \theta(\eta) = \eta^{t} $ for some integer $ t $ coprime to $ p $. Then for $ 0 \leq i \leq n $ we have: $$\begin{array}{cccccc}
&\theta \left( \sigma^{p^{i}} \ast_{r} \eta \right) & = & \theta \left( \eta^{d^{rp^{i}}} \right) & = & \eta^{td^{rp^{i}}} \\
\mbox{ and }&\sigma^{p^{i}} \ast_{s} \theta \left( \eta \right) & = & \sigma^{p^{i}} \ast_{s} \eta^{t} & = & \eta^{td^{sp^{i}}},
\end{array}$$ so $ \theta $ is $ G_{i} $-equivariant if and only if $ d^{rp^{i}} \equiv d^{sp^{i}} \pmod{p^{n}} $. Since $ d $ has order $ p^{n-1} $ in $ (\mathbb{Z}/p^{n}\mathbb{Z})^{\times} $, this occurs if and only if $ rp^{i} \equiv sp^{i} \pmod{p^{n-1}} $, that is, if and only if $ r \equiv s \pmod{p^{n-1-i}} $.
Let $ 0 \leq i \leq n $. Then:
1. The collection $ \{ K_{i} \otimes_{K} H_{0}, \ldots ,K_{i} \otimes_{K} H_{p^{n-1}-1} \} $ can be partitioned into $ p^{n-1-i} $ Hopf algebra isomorphism classes;
2. Each class contains $ p^{i} $ Hopf algebras;
3. $ \{ K_{i} \otimes H_{0}, \ldots ,K_{i} \otimes_{K} H_{p^{n-2-i}} \} $ is a complete set of representatives for the classes;
4. For $ 0 \leq j < p^{n-1} $, the class containing $ K_{i} \otimes_{K} H_{j} $ is $$\{ K_{i} \otimes_{K} H_{j+p^{n-i}m} \mid 0 \leq m < p^{i} \}.$$
Let $ 0 \leq r \leq p^{n}-1 $ and $ 0 \leq i \leq n $. Then $ K_{i} \otimes_{K} H_{r} \cong K_{i}[N] $ as $ K_{i} $-Hopf algebras if and only if $ r \equiv 0 \pmod{n-1-i} $
We have $ H_{0} \cong K[N] $ as $ K $-Hopf algebras, so for each $ i $ we have $ K_{i} \otimes_{K} H_{0} \cong K_{i}[N] $ as $ K_{i} $-Hopf algebras. The result now follows from Theorem \[theorem\_cyclic\_p\_n\_Hopf\_isomorphism\].
We now impose the additional assumption that $ K $ has characteristic zero and contains a primitive $ p^{n} $-root of unity $ \zeta $. In this case we can use Corollary \[corollary\_algebra\_isomorphism\_cyclic\] to determine which of the $ K $-Hopf algebras appearing in the classification of Hopf-Galois structures on $ L/K $ are isomorphic as $ K $-algebras.
\[theorem\_cyclic\_p\_power\_algebra\_isomorphisms\] For $ 0 \leq r,s < p^{n-1} $, we have $ H_{r} \cong H_{s} $ as $ K $-algebras if and only if $ v_{p}(r)=v_{p}(s) $, where $ v_{p} $ denotes the $ p $-adic valuation function.
By Corollary \[corollary\_algebra\_isomorphism\_cyclic\], we have $ H_{r} \cong H_{s} $ if and only if $ (N,\ast_{r}) \cong (N,\ast_{s}) $ as $ G $-sets, so we must show that this occurs if and only if $ v_{p}(r)=v_{p}(s) $.\
\
Suppose first that $ v_{p}(r)=v_{p}(s) $. Then since $ {\mbox{Aut}(N)} $ is cyclic we must have $ \langle \delta^{r} \rangle = \langle \delta^{s} \rangle = \Delta $ say, and we have $ \overline{\beta_{r}}(G) = \overline{\beta_{s}}(G) $. Therefore for each $ \mu \in N $, the orbits of $ \mu $ with respect to $ \ast_{r} $ and $ \ast_{s} $ coincide, and so the stabilizers $ \text{Stab}_{r}(\mu), \text{Stab}_{s}(\mu) $ of $ \mu $ with respect to $ \ast_{r} , \ast_{s}$ have the same order. Since $ G $ is cyclic, this implies that they are equal.\
\
Now let $ \eta_{1}, \ldots, \eta_{k} $ be representatives for the orbits of $ (N,\ast_{r}) $, and define $ \pi : N \rightarrow N $ by setting $ \pi(\eta_{i})=\eta_{i} $ for each $ i $, and insisting that $ \pi( g \ast_{r} \mu ) = g \ast_{s} \pi(\mu) $ for all $ \mu \in N $. It is routine to verify that $ \pi $ is well defined and injective, and so it is a $ G $-equivariant bijection from $ (N,\ast_{r}) $ to $ (N,\ast_{s}) $.\
\
Conversely, suppose that $ v_{p}(r) \neq v_{p}(s) $, and assume without loss of generality that $ v_{p}(r) < v_{p}(s) $. Then since $ {\mbox{Aut}(N)} $ is cyclic we have $ \overline{\beta_{s}}(G) \subsetneq \overline{\beta_{r}}(G) $, and so (for example) the orbit of $ \eta $ with respect to $ \ast_{s} $ is strictly contained in the orbit of $ \eta $ with respect to $ \ast_{r} $. Therefore $ (N,\ast_{r}) $ and $ (N,\ast_{s}) $ cannot be isomorphic $ G $-sets in this case.
\[corollary\_cyclic\_p\_power\_algebra\_isomorphisms\] Precisely $ n $ non-isomorphic $ K $-algebras appear in the classification of Hopf-Galois structures on $ L/K $. For each $ 0 \leq v \leq n-1 $, the $ K $-algebra $ H_{p^{v}} $ has $ \varphi(p^{v}) $ distinct Hopf-Galois actions on $ L/K $.
Finally, retaining the assumption that $ \zeta \in K $, we explicitly compute the Wedderburn-Artin decompositions of these algebras. Recall that, by [@BleyBoltje Lemma 2.5], for $ 0 \leq r < p^{n-1} $ we have $$H_{r} = L[\alpha_{r}(N)]^{G} \cong \prod_{m=1}^{t} L^{ S_{m} } \mbox{ as $ K $-algebras,}$$ where the $ S_{m} $ are the stabilizers of a set of representatives of the orbits of $ G $ in $ \widehat{\alpha_{r}(N)} $. By Corollary \[corollary\_algebra\_isomorphism\_cyclic\], these stabilizers coincide with those of a set of representatives of the orbits of $ G $ in $ N $, with $ G $ acting by $$\sigma^{i} \ast_{r} \eta^{j} = \overline{\beta_{r}}(\sigma^{i})[\eta^{j}] = \delta^{ir}[\eta^{j}] = \eta^{jd^{ir}}.$$ Since $ N $ is cyclic, we may translate this to an action of the additive group $\mathbb{Z}/p^n\mathbb{Z}$ on itself via $$i \cdot_{r} j = jd^{ir},$$ and study the orbits and stabilizers of this action.
\[lemma\_cyclic\_orbits\] Let $ j \in \mathbb{Z}/p^n\mathbb{Z} $ and $ m=\max\{n-1-v_p(j)-v_p(r),0\} $. Then $${\mathcal O}(j)=\{ jd^{is}: 0 \leq i < p^{m} \}, \mbox{ and } \mbox{Stab}(j)= \langle p^{m} \rangle.$$
Note $i\cdot_{r} j = j$ if and only if $$jd^{{ir}} \equiv j \pmod{p^n}$$ i.e., $$d^{{ir}} \equiv 1 \pmod{p^{n-v_p(j)}},$$ which holds if and only if $${ir} \equiv 0 \pmod{p^{n-v_p(j)-1}}.$$ Now if $m=0$ then $v_p(r)=n-1-v_p(j)$, hence $p^{n-1-v_p(j)}\mid r$ and the result is clear. Otherwise, the above congruence holds if and only if $$i\equiv 0 \pmod{p^{n-1-v_p(j)-v_p(r)}}.$$ Thus, $\mbox{Stab}(j)=\langle p^{n-1-v_p(j)-v_p(r)} \rangle $. The orbit computation follows immediately.
The following allows us to count orbit classes in the cases $m>0$.
\[lemma\_cyclic\_orbits\_m>0\] Let $0 < r \leq p^{n-1}$ and pick, if possible, $0 < m \leq n-1-v_p(s)$. Then $\mathbb{Z}/p^n\mathbb{Z}$ has $p^{v_p(r)}(p-1)$ distinct orbits whose stabilizer is $ \langle p^{m} \rangle $.
There are $\varphi(p^{m+v_p(r)+1})= p^{m+v_p(r)}(p-1)$ elements of order $p^{m+v_p(r)+1}$ in $\mathbb{Z}/p^n\mathbb{Z}$. Clearly, $j$ has order $p^{m+v_p(r)+1}$ if and only if $v_p(j)=n-m-v_p(r)-1$. Thus, there are $ p^{m+v_p(r)}(p-1)$ choices of $j$ for which $m=n-1-v_p(j)-v_p(r)$. By Lemma \[lemma\_cyclic\_orbits\] there are $p^m$ choices for $j$ in each orbit. Thus, the number of orbits whose stabilizer is $ \langle p^m \rangle $ is $$\frac{p^{m+v_p(r)}(p-1)}{p^m} = p^{v_p(r)}(p-1).$$
For $m=0$ we have
\[lemma\_cyclic\_orbits\_m=0\] Suppose $v_p(j)\geq n-1-v_p(r)$. Then $${\mathcal O}(j)=\{j\} \mbox{ and } \mbox{Stab}(j)=\mathbb{Z}/p^n\mathbb{Z}.$$
Immediate from Lemma \[lemma\_cyclic\_orbits\] since $m=0$. Note additionally that there are $p^{1+v_p(r)}$ such $j$ since $$\{j: v_p(j)\geq n-1-v_p(r) \}=\{j' p^{n-1-v_p(r)} : 0\leq j^{\prime} < p^{1+v_p(r)} \}.$$
Having computed orbits and stabilizers, we are now able to give Wedderburn-Artin decompositions.
\[theorem\_cyclic\_p\_power\_wedderburn\] There is an isomorphism of $ K $-algebras $$H_{r} \cong K^{p^{1+v_p(r)}} \times \prod_{m=1}^{ n-1-v_p(r)}\left( K_{m} \right)^{p^{v_p(r)}(p-1)}.$$
Suppose first that $r=0$. Then $H_{r}=K[N]$, and since $\zeta\in K$ it follows that $$H_{p^{n-1}}\cong K^{p^n} = K^{p^{1+v_p(p^{n-1})}}.$$ Now let $0 < r < p^{n-1}$. Pick $0 < m \leq n-1-v_p(r)$. From Lemma \[lemma\_cyclic\_orbits\_m>0\], we know that there are $p^{v_p(r)}(p-1)$ distinct orbits whose stabilizer is $ \langle \sigma^{p^m} \rangle $. Since $L^{\langle \sigma^m \rangle }=K_{m}$, the Wedderburn-Artin decomposition contains $p^{v_p(r)}(p-1)$ copies of $K_ {m}$. Therefore, $$H_{r}\cong K^{p^{1+v_p(r)}} \times \prod_{m=1}^{ n-1-v_p(r)}\left( K_{m} \right)^{p^{v_p(r)}(p-1)} \mbox{ as $ K $-algebras.}$$
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The phenomenological linear response theory of non-Markovian Stochastic Resonance (SR) is put forward for stationary two-state renewal processes. In terms of a derivation of a non-Markov regression theorem we evaluate the characteristic SR-quantifiers; i.e. the spectral power amplification (SPA) and the signal-to-noise ratio (SNR), respectively. In clear contrast to Markovian SR, a characteristic benchmark of genuine non-Markovian SR is its distinctive dependence of the SPA and SNR on small (adiabatic) driving frequencies; particularly, the adiabatic SNR becomes strongly suppressed over its Markovian counterpart. This non-Markovian SR theory is elucidated for a fractal gating dynamics of a potassium ion channel possessing an infinite variance of closed sojourn times.'
author:
- Igor Goychuk
- Peter Hänggi
title: 'Non-Markovian Stochastic Resonance '
---
The concept of Stochastic Resonance (SR), originally put forward for the description of the periodicity of the earth’s glacial recurrences [@first], has acquired an immense popularity in the context of weak signal transduction in stochastic nonlinear systems [@review1]. The phenomenon of SR is seemingly rather paradoxical: an optimal dose of either external or internal noise can considerably boost signal transduction. The archetypical situation of SR involves a periodically rocked, continuous state bistable dynamics driven by thermal, white noise [@review1]. The essential features of the perturbed bistable dynamics can be captured by a two-state stochastic process $x(t)$ that switches forth and back between two metastable states $x_{1}$ and $x_{2}$ at random time points $\{ t_i\}$. This two-state random process can be directly extracted from filtered experimental data and subsequently statistically analyzed.
If the sojourn time intervals $\tau_i=t_{i+1}-t_i$ are [*independently*]{} distributed (an assumption being invoked throughout the following), the resulting two-state renewal process is specified by two residence time distributions (RTDs) $\psi_{1,2}(\tau)$ [@cox]. Commonly, one follows the reasoning of McNamara and Wiesenfeld [@mcnamara]; i.e. one approximates the reduced dynamics by a two-state Markovian process with the corresponding RTDs being strictly exponential, $\psi_{1,2}(\tau)=\nu_{1,2}\exp(-\nu_{1,2}\tau)$. Here, $\nu_{1,2}$ are the transition rates which are given by the inverse mean residence times $\langle \tau_{1,2}\rangle :=\int_{0}^{\infty}
\tau\psi_{1,2}(\tau)d\tau$; i.e. $\nu_{1,2}=\langle
\tau_{1,2}\rangle^{-1}$. The input signal $f(t)$ yields time-dependent transition rates $\nu_{1,2}\to\nu_{1,2}(t)$. The probabilities $p_{1,2}(t)$ of the states $x_{1,2}$ obey the Markovian master equation [@review1; @mcnamara]; i.e. $$\begin{aligned}
\label{rate-eq}
\dot p_1(t)& = &-\nu_1(t)p_1(t)+\nu_2(t)p_2(t)\nonumber \\
\dot p_2(t)& = &\;\;\;\nu_1(t)p_1(t)-\nu_2(t)p_2(t).\end{aligned}$$ with the time-dependent rates. Applying a weak periodic signal of the form $f(t)=f_0\cos(\Omega t)$ yields for the asymptotic linear response $\langle \delta x(t)\rangle=
f_0|\tilde\chi(\Omega)|\cos(\Omega t-\varphi(\Omega))$. Here, $\tilde\chi(\Omega)$ is the linear response function in the frequency domain and $\varphi(\Omega)$ denotes the phase shift. For adiabatic, Arrhenius-like transition rates $\nu_{1,2}(t)$ that depend on temperature $T$ and driving signal $f(t)$, the linear response function $\tilde \chi(\Omega)$ is known explicitly [@mcnamara]. The SPA [@JH91; @new], $\eta=|\tilde\chi(\Omega)|^2$, then displays the phenomenon of SR; i.e. the quantity $\eta$ depicts a bell-shaped behavior on the thermal noise strength $T$ [@review1]. This appealing two-state Markovian SR-theory due to McNamara and Wiesenfeld [@mcnamara] enjoys great popularity and wide spread application in SR-research [@review1]. Moreover, this seminal Markovian scheme has recently been generalized in order to unify the various situations of SR – such as periodic or aperiodic SR [@b3] and non-stationary SR – within a unified framework based on information theory [@GH00].
One may encounter, however, an ample number of other physical situations where the observed stochastic two-state dynamics $x(t)$ exhibits strong temporal long range correlations that are manifestly [*non-Markovian*]{} in nature with profoundly non-exponential, experimentally observed RTDs [@boguna; @sullivan; @mercik; @teich]. In principle, any deviation of RTDs from a strictly exponential behavior constitutes a deviation from a Markovian two-state behavior [@boguna], although in practice it can be rather small. A clear-cut, genuine non-Markovian situation emerges when, e.g., one of RTDs possesses a very large, possibly infinite variance ${\rm
var}(\tau_{1,2})=\int_{0}^{\infty} (\tau-\langle
\tau_{1,2}\rangle)^2\psi_{1,2}(\tau)d\tau\to\infty$. As a specific example, this situation occurs for the stochastic dynamics of the conductance fluctuations in biological ion channels for which the RTDs generally assume a non-exponential behavior. The corresponding RTD $\psi(\tau)$ can either be described by a stretched exponential [@sullivan], or possibly also by a power law $\psi(\tau) \propto
1/(b+\tau)^{\beta},\;\beta>0$ [@mercik]. The case with a power law is particularly interesting: In Ref. [@mercik] one finds that the closed time RTD for a large conductance (BK) potassium channel assumes a power law with an exponent $\beta\approx 2.24$ implying that ${\rm
var}(\tau_{closed})=\infty$. As a consequence, the conductance fluctuations are expected to exhibit a characteristic $1/f^{\alpha}$ noise power spectrum $S(f)$ [@teich]. Indeed, this result has been confirmed for the BK ion channel [@siwy], as well as for other types of ion channels [@bezrukov2000].
What are the characteristic signatures of non-Markovian SR in these and several other, manifestly non-Markovian phenomena? To address this challenge we herewith put forward the non-Markovian generalization of the well-known McNamara-Wiesenfeld two-state Markov theory to the case with arbitrary, non-exponential (!) RTDs $\psi_{1,2}(\tau)$ and corresponding survival probabilities $\Phi_{1,2}(\tau)= \int_{\tau}^{\infty}\psi_{1,2} (\tau')d\tau'$, respectively [@cox]. There do exist a few prior studies of non-Markovian SR based on a contraction of a (Markovian) stochastic dynamics onto a non-Markovian process; see, e.g., in [@review1; @new; @melnikov; @lindner]. However, the case of genuine non-Markovian SR with an infinite variance of sojourn times has not been investigated previously. Moreover, in clear contrast to these prior studies [@new; @melnikov; @lindner] we do [*not*]{} presume here any knowledge of the underlying microscopic or mesoscopic dynamics. In practice, such a mesoscopic dynamics is not accessible, or is simply not known. Instead, we pursue with this work a phenomenological scheme of non-Markovian SR which is solely based on the experimentally observed RTDs $\psi_{1,2}(\tau)$ in the absence of an input signal.
[*Propagator for two-state renewal processes.*]{} A first challenge presents the derivation of the propagator $\mathbf \Pi(t|t_0)$ of the unperturbed persistent two-state renewal process $x(t)$. The quantity $\mathbf \Pi(t|t_0)$ relates the probability vector $\vec p(t)=[p_1(t),p_2(t)]^T$ at two different instants of time $t$ and $t_0$, i.e., $\vec p(t)={\mathbf \Pi} (t|t_0)\vec p(t_0)$. One can explicitly find $\mathbf \Pi(t|t_0)$ by considering the various contributions of all possible stochastic paths that lead from $\vec p(t_0)$ to $\vec p(t)$. Let us split up these contributions as follows $$\begin{aligned}
\label{structure}
{\mathbf \Pi} (t|t_0)=\sum_{n=0}^{\infty}{\mathbf \Pi} ^{(n)}(t|t_0),\end{aligned}$$ where the index $n$ denotes the number of corresponding switches that occurred during the stochastic evolution. The contribution with zero alternations is obviously given by $$\begin{aligned}
\label{first}
{\mathbf \Pi} ^{(0)}(t|t_0)=\left [ \begin{array}{cc}
\Phi_1^{(0)}(t-t_0) & 0 \\
0 & \Phi_2^{(0)}(t-t_0)
\end{array} \right ]\;.\end{aligned}$$ Stochastic paths with a single alternation contribute the weight $$\begin{aligned}
{\mathbf \Pi} ^{(1)}(t|t_0)=\int_{t_0}^{t}dt_1{\mathbf P}(t-t_1){\mathbf
F}^{(0)}(t_1-t_0),\end{aligned}$$ where $$\begin{aligned}
{\mathbf P}(t-t_0)=\left [ \begin{array}{cc}
\Phi_1(t-t_0) & 0 \\
0 & \Phi_2(t-t_0)
\end{array} \right ]\end{aligned}$$ and $$\begin{aligned}
\label{second}
{\mathbf F}^{(0)}(t-t_0)=\left [ \begin{array}{cc}
0 & \psi_2^{(0)}(t-t_0) \\
\psi_1^{(0)}(t-t_0) & 0
\end{array} \right ]\;.\end{aligned}$$ Note that for the persistent renewal process to be strictly [*stationary*]{} [@tunaley], the survival probability $\Phi_{1,2}^{(0)}(\tau)$ of the [*first*]{} residence time interval $\tau_0=t_1-t_0$ in Eq. (\[first\]) and the corresponding RTD $\psi_{1,2}^{(0)}(\tau)=-d\Phi_{1,2}^{(0)}(\tau)/d\tau$ in Eq. (\[second\]) must be chosen differently from all subsequent ones. Stationarity requires that [@cox; @boguna; @tunaley], $$\begin{aligned}
\label{pdf2}
\psi_{1,2}^{(0)}(\tau)=\frac{\Phi_{1,2}(\tau)}{\langle
\tau_{1,2}\rangle},\end{aligned}$$ where $\Phi_{1,2}(\tau)=\int_{\tau}^{\infty}\psi_{1,2}(t)dt$ are the given survival probabilities. From (\[pdf2\]) it follows that the mean residence time $\langle \tau_{1,2}\rangle$ must assume finite values, $\langle \tau_{1,2}\rangle\neq \infty$. This imposes a salient restriction. Next, the paths with two switches contribute to Eq. (\[structure\]) as $$\begin{aligned}
{\mathbf \Pi}
^{(2)}(t|t_0)=\int_{t_0}^{t}dt_2\int_{t_0}^{t_2}dt_1{\mathbf
P}(t-t_2) {\mathbf F}(t_2-t_1)\\ \nonumber
\times {\mathbf F}^{(0)}(t_1-t_0),\end{aligned}$$ where $$\begin{aligned}
\label{endstructure}
{\mathbf F}(t-t_0)=\left [ \begin{array}{cc}
0 & \psi_2(t-t_0) \\
\psi_1(t-t_0) & 0
\end{array} \right ]\;,\end{aligned}$$ and, likewise, for all higher $n$. Because ${\mathbf \Pi} (t|t_0)$ depends only on the time difference, $\tau=t-t_0$, the infinite, multiple-integral series (\[structure\])-(\[endstructure\]) can be summed exactly by use of a Laplace-transform. If we denote the Laplace transform for a function $A(\tau)$ by $\tilde A(s):=\int_{0}^{\infty}\exp(-st)A(\tau)d\tau$ we find $$\begin{aligned}
\label{stat}
\tilde {\mathbf \Pi}(s)=\frac{1}{s}\left [ \begin{array}{cc}
1- \frac{\tilde G(s)}{s\langle \tau_1\rangle} &
\frac{\tilde G(s)}{s\langle \tau_2\rangle} \\
\frac{\tilde G(s)}{s\langle \tau_1\rangle} &
1-\frac{\tilde G(s)}{s\langle \tau_2
\rangle}
\end{array} \right ]\;,\end{aligned}$$ where $$\label{aux}
\tilde G(s)=\frac{\left(1-\tilde \psi_1(s)\right)
\left(1-\tilde \psi_2(s)\right)}{\left(1-\tilde \psi_1(s)\tilde
\psi_2(s)\right)},$$ in agreement with the known result in Refs. [@cox; @boguna].
[*Non-Markov Regression Theorem.*]{} From (\[stat\])-(\[aux\]) one finds the stationary probabilities as $\vec p^{st}=\lim_{s\to
0}\left (s \tilde {\mathbf \Pi}(s)\vec p(0) \right)$. These explicitly read $p_{1,2}^{st}=\langle \tau_{1,2}\rangle/[\langle \tau_1\rangle+\langle
\tau_2\rangle]$. The generally non-exponential relaxation of $\langle x(t)\rangle$ to the stationary mean value $x_{st}=x_1p_1^{st}+x_2p_2^{st}$ is described by the unique relaxation function $R(\tau)$, i.e. $$\begin{aligned}
\label{relax}
p_{1,2}(t_0+\tau)=p_{1,2}^{st}+[p_{1,2}(t_0)-p_{1,2}^{st}]\;
R(\tau)
\end{aligned}$$ where $R(\tau)$ obeys the Laplace-transform $$\begin{aligned}
\label{R-function}
\tilde R(s)=\frac{1}{s}-\left(\frac{1}{\langle \tau_1\rangle} +
\frac{1}{\langle \tau_2\rangle} \right) \frac{1}{s^2}
\tilde G(s)\;.\end{aligned}$$
Let us consider next the autocorrelation function $$\begin{aligned}
\label{correlation}
k(\tau)=
\lim_{t\to\infty}\frac{\langle\delta x(t+\tau)\delta x(t)\rangle }
{\langle\delta x^2\rangle_{st}}\end{aligned}$$ of stationary fluctuations, $\delta x(t)=x(t)-x_{st}$. With $\langle \delta x(t+\tau)\delta x(t)\rangle=
\langle x(t+\tau) x(t)\rangle-\langle x\rangle_{st}^2$ as $t\to \infty$, and $$\begin{aligned}
\label{explicit}
\lim_{t\to\infty}\langle x(t+\tau) x(t)\rangle
=\sum_{i=1,2}\sum_{j=1,2}x_ix_j\Pi_{ij}(\tau)p_j^{st}\end{aligned}$$ we find the same result as in Ref. [@strat]; i.e. $$\begin{aligned}
\label{laplace-corr}
\tilde k(s)=\frac{1}{s}-\left(\frac{1}{\langle \tau_1\rangle} +
\frac{1}{\langle \tau_2\rangle}\right)\frac{1}{s^2}\tilde G(s).\end{aligned}$$ Upon comparison of (\[R-function\]) with (\[laplace-corr\]) the following regression theorem holds for these non-Markovian two-state processes, namely $$\label{regression}
R(\tau)=k(\tau)\;.$$ The regression theorem (\[regression\]), which relates the decay of the relaxation function $R(\tau)$ to the decay of stationary fluctuations $k(\tau)$, presents a first main result, yielding the cornerstone for the derivation of linear response theory for non-Markovian SR.
[*Linear Response Theory.*]{} The common linear response approximation $$\begin{aligned}
\label{response2}
\langle \delta x(t)\rangle:=
\langle x(t)\rangle - x_{st}
=\int_{-\infty}^{t}\chi(t-t')f(t')dt',\end{aligned}$$ clearly holds independently of the underlying stochastic dynamics [@HT82]. In (\[response2\]), $\chi(t)$ denotes the linear response function in the time domain. It can be found following an established procedure [@kubo]: (i) apply a small static “force” $f_0$, (ii) let the process $x(t)$ relax to the constrained stationary equilibrium $\langle x(f_0)\rangle$, and (iii) suddenly remove the “force” at $t=t_0$. Then, in accord with (\[response2\]) the response function reads $$\begin{aligned}
\label{response3}
\chi(\tau)=-\frac{1}{f_0}\frac{d}{d\tau}
\langle\delta x(t_0+\tau)\rangle, \;\;\;\;\;\tau>0,\end{aligned}$$ where $\langle\delta x(t_0+\tau)\rangle=x_1p_1(t_0+\tau)+x_2p_2(t_0+\tau)$ is determined by (\[relax\]) with the initial $p_{1,2}(t_0)$ taken as $p_{1,2}(t_0)=\langle \tau_{1,2}(f_0) \rangle/[\langle
\tau_{1}(f_0)\rangle+\langle \tau_{2}(f_0)\rangle]$. Expanding $p_{1,2}(t_0)$ to first order in $f_0$ we obtain $$\begin{aligned}
\label{expand}
\langle\delta x(t_0+\tau)\rangle =
\frac{\langle \delta x^2\rangle_{st}}{\Delta x}
[\beta_2-\beta_1] R(\tau) f_0 +o(f_0),\end{aligned}$$ where $\Delta x=x_2-x_1$ is the fluctuation amplitude and $$\label{msq}
\langle \delta x^2\rangle_{st}=(\Delta x)^2
\frac{\langle \tau_1\rangle\langle \tau_2\rangle}{(\langle \tau_1\rangle+
\langle \tau_2\rangle)^2},$$ is the mean squared amplitude of the stationary fluctuations. Moreover, $\beta_{1,2}:=d\ln \langle \tau_{1,2}(f_0)\rangle/df_0|_{f_0=0}$ in (\[expand\]) denotes the logarithmic derivative of mean residence time with respect to the input-signal strength. Upon combining (\[expand\]) with the regression theorem (\[regression\]) we obtain from (\[response3\]) the [*fluctuation theorem*]{} $$\begin{aligned}
\label{fluctuation1}
\chi(\tau)=-[\beta_2-\beta_1]\frac{\theta(\tau)}{\Delta x}
\frac{d}{d\tau} \langle\delta x(t+\tau)\delta x(t)\rangle_{st} \;.\end{aligned}$$ $\theta(t)$ denotes the unit step function. The non-Markovian fluctuation theorem (\[fluctuation1\]) presents a second main result of this work; in particular, it does not assume thermal equilibrium [@HT82]. If, in addition, the mean residence times commonly obey an Arrhenius-like dependence on temperature $T$ and force $f_0$; i.e. $$\begin{aligned}
\label{Arrhenius}
\langle \tau_{1,2}(f_0)\rangle & = &A_{1,2} \exp\Big (\frac{
\Delta U_{1,2}\mp \Delta x_{1,2} f_0}{k_BT} \Big),\end{aligned}$$ where $\Delta U_{1,2}$ are the heights of activation barriers, $\Delta x_1=z\Delta x$, $\Delta x_2=(1-z)\Delta x$ with $0<z<1$, we recover for the fluctuation theorem in (\[fluctuation1\]) the form which in particular holds true for a classical equilibrium dynamics [@HT82; @kubo]; i.e., $$\begin{aligned}
\label{equilibrium}
\chi(\tau)=-\frac{\theta(\tau)}{k_BT}
\frac{d}{d\tau} \langle\delta
x(\tau)\delta x(0)\rangle_{st}.\end{aligned}$$
[*Spectral Power Amplification.*]{} In presence of an applied periodic signal, see below (\[rate-eq\]), the spectral power amplification (SPA) [@JH91], $\eta(\Omega)=|\tilde\chi(\Omega)|^2$, where $\tilde\chi(\omega)=\int_{-\infty}^{\infty}\chi(t){\rm e}^{i\omega t}dt$, reads by use of the FT in (\[equilibrium\]) upon combining (\[correlation\]), (\[laplace-corr\]), (\[msq\]), (\[Arrhenius\]) as follows
$$\label{res1}
\eta(\Omega,T)=
\frac{(\Delta x/2)^4}{(k_BT)^2}
\frac{\nu^2(T)}{\cosh^4\left[\epsilon(T)/(2k_BT)\right]}
\frac{|\tilde G(i\Omega)|^2}{\Omega^2}.$$
In (\[res1\]), $\nu(T)=\langle \tau_1\rangle^{-1}+\langle \tau_2\rangle^{-1}$ is the sum of effective rates and $\epsilon(T)=\Delta U_2-\Delta U_1+T\Delta S$ denotes the free-energy difference between the metastable states which includes the entropy difference $\Delta S:=S_2-S_1=k_B\ln(A_2/A_1)$. In the Markovian case we obtain $\tilde G(s)=s/(s+\nu)$ and (\[res1\]) equals the known result, see in [@review1].
[*Signal-to-Noise Ratio.*]{} The signal-to-noise ratio (SNR) within linear response theory is given by ${\rm SNR}(\Omega,T):=\pi f_0^2|\tilde\chi(\Omega)|^2/S_N(\Omega)$, where $S_N(\omega)$ is the spectral power of stationary fluctuations [@review1]. In the present case, $S_N(\omega)=2\langle \delta x^2 \rangle_{st}
{\rm Re}\left [\tilde k(i\omega)\right]$ with $\langle \delta x^2
\rangle_{st}$ from (\[msq\]) and $\tilde k(s)$ given in (\[laplace-corr\]). By use of (\[res1\]), we obtain $$\begin{aligned}
\label{res3}
{\rm SNR}(\Omega,T)=\frac{\pi f_0^2(\Delta x/2)^2}{2(k_BT)^2}\frac{\nu(T)}
{\cosh^2\left[\frac{\epsilon(T)}{2k_BT)}\right]}\; N(\Omega),\end{aligned}$$ where the term $N(\Omega)=
|\tilde G(i\Omega)|^2/{\rm Re}[\tilde G(i\Omega)]$ denotes a frequency and temperature dependent non-Markovian correction. For arbitrary continuous $\psi_{1,2}(\tau)$ and the high-frequency signals $\Omega\gg \langle \tau_{1,2}\rangle^{-1}$, the function $N(\Omega)$ approaches unity. Then, Eq. (\[res3\]) reduces to the known Markovian result [@review1], i.e. the Markovian limit of SNR is assumed asymptotically in the high frequency limit. More interesting, however, is the result for small frequency driving. In the zero-frequency limit we obtain $N(0)=2/[{\rm var}(\tau_1)/\langle \tau_1\rangle^2+
{\rm var}(\tau_2)/\langle \tau_2\rangle ^2]$. With ${\rm var}(\tau_{1,2})=\langle \tau^2_{1,2}\rangle
- \langle \tau_{1,2}\rangle^2=
\infty$, $N(0)=0$; i.e. ${\rm SNR}(0,T)=0$ as well. Consequently, ultra-slow signals are difficult to detect within the SNR-measure.
[*Application: fractal ion channel gating.*]{} Let us next illustrate our main findings for the case of a manifestly non-Markovian gating dynamics of ion channels that exhibit a fractal gating kinetics together with a $1/f^{\alpha}$ noise spectrum of fluctuations [@mercik; @siwy]. In this context, $x(t)$ corresponds to the conductance fluctuations and the forcing $f(t)$ is proportional to the time-varying transmembrane voltage. For a locust BK channel the measured unperturbed closed time statistics can be approximated by a Pareto law; i.e. $\psi_1(\tau)=\langle \tau_1\rangle^{-1}(1+\gamma^{-1})/[1+\gamma^{-1}\tau/
\langle \tau_1\rangle]^{2+\gamma}$ with $\gamma\approx 0.24$ and $\langle \tau_1\rangle=0.84$ ms. The open time RTD assumes an exponential form with $\langle \tau_2\rangle=0.79$ ms [@mercik]. For low $\omega$, the noise power reads $S_N(\omega)\propto 1/\omega^{1-\gamma}$. Unfortunately, neither voltage, nor temperature dependence of mean residence times are experimentally available. Thus, we employ here the Arrhenius dependence in (\[Arrhenius\]). Namely, because the temperature dependence of the open-to-closed transitions is typically strong [@hille], we assume a rather high activation barrier; i.e. $\Delta U_2=100$ kJ/mol ($\sim 40\;k_BT_{room}$). The closed-to-open transitions are assumed to be weakly temperature-dependent with $\Delta U_1=10$ kJ/mol. Because $\langle \tau_1 \rangle \sim \langle \tau_2 \rangle$ at $T_{room}$, the difference between $\Delta U_1$ and $\Delta U_2$ is compensated by an entropy difference $\Delta S\sim
-36\;k_B$. The physical reasoning is that the closed time statistics exhibits a power law; i.e. the conformations in the closed state are largely degenerate. This in turn yields a larger entropy as compared to the open state.
For these parameters, the spectral power amplification versus the temperature is depicted for various driving frequencies in Fig. 1a. Furthermore, Fig. 1b corresponds to an overall Markovian modeling with an exponential $\psi_1(\tau)$ possessing the same $\langle \tau_1 \rangle$. We observe a series of striking features in Fig. 1. (i) A distinct SR-maximum occurs in the physiological range of varying temperatures (caused by the entropy effects). (ii) Due to a profound intrinsic asymmetry, the frequency dependence of the spectral amplification $\eta(\Omega,T)$ for the Markov modeling is very weak for small frequencies $\Omega\ll \langle \tau_{1,2}\rangle^{-1}$ [@review1]. In contrast, the non-Markovian SR exhibits a distinct low-frequency dependence (thereby frequency-resolving the three overlapping lines in Fig. 1b). (iii) The evaluation of the SNR yields – in clear contrast to the frequency-independent Markov modeling – a very strong non-Markovian SR frequency suppression of SNR towards smaller frequencies. The SNR-maximum for the top line in Fig. 1a is suppressed by two orders of magnitude as compared to the Markov case (not shown). As a consequence, for a strong non-Markovian situation it is preferable to use low-to-moderate frequency inputs in order to monitor SR.
In conclusion, we have put forward the phenomenological two-state theory of non-Markovian stochastic resonance. This approach carries great potential for many applications in physical and biological systems exhibiting temporal long range correlations, such as they occur in life sciences and geophysical phenomena (e.g. earthquakes), to name but a few. In clear contrast to Markovian-SR, a benchmark of genuine non-Markovian SR is its distinct strong frequency-dependence of corresponding SR quantifiers within the adiabatic driving regime; cf. Fig. 1(a).
[*Acknowledgements.*]{} This work has been supported by the DFG [*Sonderforschungsbereich*]{} 486.
R. Benzi, A.Sutera, and A. Vulpiani, J. Phys. A [**14**]{}, L453 (1981). L. Gammaitoni, P. Hänggi, P. Jung, F. Marchesoni, Rev. Mod. Phys. [**70**]{}, 223 (1998). D. R. Cox, [*Renewal Theory*]{}, (Methuen, London, 1962). B. McNamara and K. Wiesenfeld, Phys. Rev. A [**39**]{}, 4854 (1989). P. Jung and P. Hänggi, Phys. Rev. A [**44**]{}, 8032 (1991). P. Hänggi, P. Jung, C. Zerbe, and F. Moss, J. Stat. Phys. [**70**]{}, 25 (1993). J.J. Collins, C. C. Chow and T. T. Imhoff, Nature (London) [**376**]{}, 236 (1995); Phys. Rev. E [**52**]{}, R3321 (1995). I. Goychuk and P. Hänggi, Phys. Rev. E [**61**]{}, 4272 (2000). M. Boguna, A.M. Berezhkovskii, G.H. Weiss, Physica A [**282**]{}, 475 (2000). L. S. Liebovitch and J. M. Sullivan, Biophys. J. [**52**]{}, 979 (1987). S. Mercik and K. Weron, Phys. Rev. E [**63**]{}, 051910 (2001). S.B. Lowen and M.C. Teich, Phys. Rev. E [**47**]{}, 992 (1993). Z. Siwy and A. Fulinski, Phys. Rev. Lett. [**89**]{}, 158101 (2002). S.M. Bezrukov and M. Winterhalter, Phys. Rev. Lett. [**85**]{}, 202 (2000). V.I. Melnikov, Phys. Rev. E [**48**]{}, 2481 (1993). B. Lindner and L. Schimansky-Geier, Phys. Rev. E [**61**]{}, 6103 (2000). J.K.E. Tunaley, Phys. Rev. Lett. [**33**]{}, 1037 (1974). R. L. Stratonovich, [*Topics in the Theory of Random Noise I*]{} (Gordon and Breach, New York, 1963), p. 176. P. Hänggi and H. Thomas, Phys. Rep. [**88**]{}, 207 (1982). M. Toda, R. Kubo, and N. Saito, [*Statistical Physics II*]{}, (Springer, Berlin, 1991). B. Hille, [*Ionic Channels of Excitable Membranes*]{}, 3-rd ed., (Sinauer Assoc., Sunderland, MA, 2001).
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'F. Mogavero'
- 'J. P. Beaulieu'
bibliography:
- 'bibtex/RotationMicrolens.bib'
date: 'Received ; accepted '
title: |
Microlensing planet detection\
via geosynchronous and low Earth orbit satellites
---
Introduction
============
The fundamental quantity that is routinely measured in a microlensing event is the Einstein timescale $t_E$, $$\label{tE}
t_E = 1.1 \, \text{days} \, {\left( \frac{M}{{M_{\text{Jup}}}} \right)}^{1/2}
{\left( \frac{x(1-x)}{0.25} \right)}^{1/2}
{\left( \frac{D_S}{8 \, \text{kpc}} \right)}^{1/2}
\left( \frac{200 \, \text{km/s}}{V} \right)$$ where $x \equiv D_L/D_S$. It depends on the lens mass $M$, the lens and source distances from Earth $D_L$ and $D_S$, and the lens-source relative velocity $V$. Consequently, the lens mass cannot be inferred without any knowledge of $x$ and $V$. This degeneracy in $t_E$ can be resolved via the equation $M =
\theta_E/\kappa \pi_E$, with $\kappa$ a constant, if the angular Einstein radius $\theta_E = t_E V/D_L$ and the microlens parallax $\pi_E = 1 \, \text{AU} (D_L^{-1} - D_S^{-1})/\theta_E$, $$\label{piE}
\pi_E = 4.3 \left( \frac{1 \, \text{day}}{t_E} \right) \left( \frac{1-x}{0.5} \right)
\left( \frac{200 \, \text{km/s}}{V} \right)$$ can both be measured from the event light curve (@Gould2013, from now on G13). The quantity $\theta_E$ is measured via finite source effects in high-magnification single-lens events and in a large number of the planetary ones. G13 has shown that for planetary events ($t_E \approx 20$ days) with peak amplifications $A_{\text{max}} \gtrsim 20$, a satellite placed in geosynchronous orbit would measure the microlens parallax $\pi_E$, providing lens masses. This result is especially important in the perspective of the Wide Field Infrared Space Telescope (WFIRST) mission [@Barry2011]: a geosynchronous orbit is currently a strong alternative to the second Lagrangian point (L2) [@Spergel2015]. In particular, G13 studies the regime where the satellite orbital period $P$ is much shorter than the shortest timescale of the microlens event, ${t_\text{eff}}\equiv \beta t_E$, with $\beta$ the lens-source impact parameter in units of $\theta_E$. The following scaling law for the relative error on $\pi_E$ is predicted: $$\label{gould}
\frac{{\sigma_{\pi_E}}}{\pi_E} \, \propto \, {t_E}^{1/2} \, \beta \, R^{-1}, \quad
P \ll \beta t_E \, \, \text{and} \, \, \beta \ll 1$$ where $R$ is the satellite orbital radius. The sensitivity to $\pi_E$ increases towards shorter Einstein timescales. However, Eq. does not apply when $P \gtrsim \beta t_E$, as correlations between $\pi_E$ and the other model parameters, such as $\beta$ and $t_E$, start to affect the error on $\pi_E$.\
In this paper, after briefly setting up the Fisher matrix analysis in Sect. \[Fisher\_analysis\], we extend the analysis of G13 to shorter Einstein timescales in Sect. \[geosynchronous\], to cover events raised by Jupiter-mass lenses ($t_E \approx 1$ day). Then, in Sect. \[LEO\] we study the capabilities of a low Earth orbit satellite to detect microlens parallax in case of Earth-mass lenses ($t_E
\approx 0.1$ days).
Fisher matrix analysis {#Fisher_analysis}
======================
The Fisher information matrix provides a way to predict, in a Bayesian framework, constraints on the best fit of a theoretical model to observational data. It allows the covariance matrix of the model parameters to be estimated via their posterior probability distribution [@Sellentin2014]. The Fisher matrix is a fast, analytical alternative to time-consuming methods such as the Markov-Chain Monte Carlo, even though its reliability is not always guaranteed [@Vallisneri2008].
The physical observable of a microlensing event is the light flux $F$, $$\label{flux}
F = F_s \, A + F_b = \overline{F} \, [(1- \nu) \, A + \nu]$$ where $F_s$, $F_b$, and $\overline{F} \equiv F_s + F_b$ are the source, blending, and baseline fluxes, and $\nu \equiv F_b/(F_s + F_b)$ defines the blending ratio. In a point source-point lens (PSPL) model, the amplification factor $A$ is given by $A(u) = (u^2 + 2)/u \sqrt{u^2 + 4}$, where $u$ is the magnitude of the lens-source separation vector $\vec{u}$ as seen by the observer. For an inertial observer, $u = \sqrt{\tau^2 + \beta^2}$, where $\tau \equiv (t-t_0)/t_E$ and $t_0$ is the peak time. The inertial observer model is thus described by five parameters: $\vec{\theta} =
(\overline{F}, \nu, t_0, t_E, \beta)$.
Under the assumption of Gaussian errors, the Fisher matrix elements are given by [@Sellentin2014] $$\label{FisherMatrix}
\mathcal{F}_{ij} = \sum_{k = 1}^N \frac{1}{\sigma_k^2}
\frac{\partial F_k}{\partial \theta_i} \frac{\partial F_k}{\partial \theta_j}, \quad
\sigma^2 = \sigma_0^2 \, [(1- \nu) \, A + \nu]$$ where the index $k$ spans the set of $N$ independent observations, and $\sigma$ is the flux error. We have assumed Poisson noise-limited photometry with $\sigma_0 \simeq (\ln 10/2.5) \overline{F} \sigma_m$, where $\sigma_m$ is the magnitude error at the light curve baseline. In the case of continuous observations, i.e. $f \beta t_E \gg 1$, where $f$ is the number of observations per unit time, the sum appearing in equation can be approximate by an integral over time, $$\label{sum_to_integral}
\sum_{k = 1}^N \longrightarrow f t_E \int_{- \infty}^{+ \infty} d\tau.$$ Since the parallax signal contributing to the matrix $\mathcal{F}$ comes from a few ${t_\text{eff}}$ near $\tau = 0$, the boundaries of integration can be safely extended to infinity, as shown in Eq. , if the observations last more than a few event timescales[^1]. Once the Fisher matrix is computed, the covariance matrix of the model parameters is given by its inverse, $\mathcal{F}^{-1}$ [@Sellentin2014].
Parallax effect for an Earth orbit satellite
--------------------------------------------
Observations from an Earth orbit satellite are affected by the parallax effect of its non-inertial motion, in the same way as for a telescope on Earth [@Hardy1995]. We thus consider an observatory in circular orbit around Earth with radius $R$ and period $P$ ($\omega =
2 \pi/P$ as its angular velocity). If $\lambda$ denotes the latitude of the source star with respect to the plane of this orbit, the projection of the satellite trajectory onto the plane of the sky is an ellipse with semimajor and semiminor axes $R$ and $R \sin(\lambda)$, respectively. Following G13, one then defines ${\epsilon_\|}= \epsilon \equiv
R/1 \, \text{AU}$ and ${\epsilon_\bot}\equiv R \sin(\lambda)/1 \, \text{AU}$. Let $\theta$ denote the direction of the lens-source relative motion in the plane of the sky with respect to the projected major axis of the satellite orbit. The lens-source separation vector seen by the satellite is given by $\vec{u} = (\tau \cos\theta - \beta \sin\theta + {\epsilon_\|}\pi_E \cos(\omega t_E \tau + \varphi), \tau \sin\theta + \beta \cos\theta +
{\epsilon_\bot}\pi_E \sin(\omega t_E \tau + \varphi))$, where $\varphi$ is the satellite orbital phase with respect to the peak time $t_0$ (@Hardy1995, G13). If one introduces the microlens parallax vector $\vec{\pi_E} = ({\pi_\|},
{\pi_\bot}) = \pi_E (\cos\theta, \sin\theta)$, the non-inertial observer model is described by seven parameters: $\vec{\theta} = (\overline{F},
\nu, t_0, t_E, \beta, {\pi_\|}, {\pi_\bot})$. The two additional flux derivatives $\partial F/\partial {\pi_\|}$ and $\partial F/\partial {\pi_\bot}$ appearing in the Fisher matrix are proportional to $\partial u/\partial {\pi_\|}$ and $\partial u/\partial {\pi_\bot}$, respectively, with $$\begin{split}
\label{parallax}
&u \frac{\partial u}{\partial {\pi_\|}} \simeq {\epsilon_\|}\tau \cos(\omega t_E
\tau + \varphi) + {\epsilon_\bot}\beta \sin(\omega t_E \tau + \varphi), \\
&u \frac{\partial u}{\partial {\pi_\bot}} \simeq - {\epsilon_\|}\beta \cos(\omega t_E
\tau + \varphi) + {\epsilon_\bot}\tau \sin(\omega t_E \tau + \varphi).
\end{split}$$ The approximate equalities mean we neglect terms that are linear in $\pi_E$. Moreover, one can assume $u \simeq \sqrt{\tau^2 + \beta^2}$. These approximations are justified as long as $\epsilon \, \pi_E \, \beta^{-1} \ll 1$. In the following analysis, the above inequality is safely verified. G13 also assumes its validity.
Sensitivity to $\pi_E$
----------------------
Once the Gaussian approximation to the posterior probability distribution of ${\pi_\|}$ and ${\pi_\bot}$ is known via the Fisher matrix, one can forecast the sensitivity of microlensing observations to $\pi_E$ through standard error propagation. Since small variations in the parallax parameters are related by $\delta \pi_E \simeq \cos\theta \, \delta {\pi_\|}+ \sin\theta \, \delta {\pi_\bot}$, the 1-$\sigma$ error on $\pi_E$ is given by the equation $$\label{sigmapi}
\sigma_{\pi_E}^2 = \cos^2(\theta) \, {\sigma_{{\pi_\|}}}^2 + \sin^2(\theta) \,
{\sigma_{{\pi_\bot}}}^2 + \sin(2\theta) \, \text{cov}({\pi_\|}, {\pi_\bot}),$$ where $\text{cov}({\pi_\|}, {\pi_\bot})$ is the covariance between the two parallax parameters. In principle, ${\sigma_{\pi_E}}^2$ depends on both $\theta$ and $\varphi$. To present results that are independent of the geometry of the event, we summarise the information contained into Eq. by finding the extrema of ${\sigma_{\pi_E}}^2(\theta, \varphi)$ over $\theta, \varphi \in [0, 2\pi]$. To do this, one first notices that, according to Eq. , ${\sigma_{{\pi_\|}}}^2$, ${\sigma_{{\pi_\bot}}}^2$, and $\text{cov}({\pi_\|}, {\pi_\bot})$ do not depend on $\theta$, but only on $\varphi$. One can thus analytically find the extrema of ${\sigma_{\pi_E}}^2(\theta, \varphi)$ over the $\theta$ range, $$\label{extrema}
{\sigma_{\pi_E, \pm}}^{2}(\varphi)
= \frac{{\sigma_{{\pi_\|}}}^2 + {\sigma_{{\pi_\bot}}}^2}{2} \pm \frac{\sqrt{({\sigma_{{\pi_\|}}}^2
- {\sigma_{{\pi_\bot}}}^2)^2 + 4 \, \text{cov}({\pi_\|}, {\pi_\bot})^2}}{2}$$ where the plus and minus signs stand for the maxima and minima. The values of $\theta$ that correspond to these extrema are given by $\tan(2\theta) = 2 \, \text{cov}({\pi_\|}, {\pi_\bot})/({\sigma_{{\pi_\|}}}^2 -
{\sigma_{{\pi_\bot}}}^2)$ [^2]. Then, one can find the extrema of Eq. over the $\varphi$ range numerically, and they only depend on $t_E$ and $\beta$: $${\sigma_{\pi_E, \, \text{max}}}^2 \equiv \max_{\varphi \in [0, 2 \pi]}
\sigma_{\pi_E, \, +}^2(\varphi), \quad {\sigma_{\pi_E, \, \text{min}}}^2 \equiv
\min_{\varphi \in [0, 2 \pi]} \sigma_{\pi_E, \, -}^2(\varphi).$$ We note that for $P \ll \beta t_E$, the covariance between ${\pi_\|}$ and ${\pi_\bot}$ vanishes, ${\sigma_{{\pi_\|}}}^2$ and ${\sigma_{{\pi_\bot}}}^2$ become independent of $\varphi$, and the extrema of ${\sigma_{\pi_E}}^2(\theta, \varphi)$ are simply given by $\max({\sigma_{{\pi_\|}}}^2, {\sigma_{{\pi_\bot}}}^2)$ and $\min({\sigma_{{\pi_\|}}}^2, {\sigma_{{\pi_\bot}}}^2)$.
Geosynchronous orbit satellite {#geosynchronous}
==============================
![Relative error on $\pi_E$ plotted against $t_E$ for a GSO satellite. We assume 3 min exposures, $\sigma_m = 0.01$, and zero blending. For $\beta = 0.1$ the region ${\sigma_{\pi_E}}/\pi_E
\in [{\sigma_{\pi_E, \, \text{min}}}, \, {\sigma_{\pi_E, \, \text{max}}}]/\pi_E$ is shown, along with the G13 prediction. For $\beta = 0.05, 0.025, 0.01, \,
\text{and} \, 0.005,$ the minimum relative error ${\sigma_{\pi_E, \, \text{min}}}/\pi_E$ is plotted.[]{data-label="figure1"}](figures/figure_1.pdf){width="\hsize"}
We then consider the case of a geosynchronous orbit (GSO) satellite, orbiting in the equatorial plane and targeting the Galaxy bulge, $P = \text{23 h 56 min 4 s}$, $R = 6.6 \, R_{\Earth}$, $\lambda = 30^\degree$. We assume $f = 1 \,
\text{observation}/ 3 \, \text{minutes}$, which is reasonable for a WFIRST-like satellite (G13) and a magnitude error at baseline $\sigma_m = 0.01$. Figure \[figure1\] shows the predicted relative error ${\sigma_{\pi_E}}/\pi_E$ as a function of the Einstein timescale $t_E$, for zero blending ($\nu = 0$). Since, from Eq. , $\pi_E$ depends on $D_L$, $D_S$, and $V$, we consider a typical disk lens at $D_L = 4 \, \text{kpc}$ and $V = 200 \, \text{km/s}$ ($D_S = 8 \, \text{kpc}$ for a source in the bulge of the Galaxy).
For an impact parameter $\beta = 0.1$, which corresponds to a peak amplification $A_\text{max} \approx 10$, we plot the region ${\sigma_{\pi_E, \, \text{min}}}\leq {\sigma_{\pi_E}}\leq {\sigma_{\pi_E, \, \text{max}}}$ to take the variations induced by the parameters $\varphi$ and $\theta$ into account. Figure \[figure1\] clearly shows two different regimes. For $t_E \gg 1 \, \text{day}$, the relative error on $\pi_E$ scales as ${t_E}^{1/2}$, in agreement with Eq. , the analytic prediction of G13. For $t_E \ll 1 \, \text{day}$, the relative error increases by decreasing $t_E$, as a result of the correlations between ${\pi_\|}, {\pi_\bot}$, and the other parameters $\nu, t_0, t_E,$ and $\beta$. In the intermediate region $1 \, \text{day} \leq t_E \leq 3 \, \text{days}$, a maximum sensitivity to $\pi_E$ is attained. In particular, at $t_E \approx 3 \, \text{days}$, the error range is very narrow around ${\sigma_{\pi_E}}/\pi_E \approx 9\%$. By decreasing $t_E$, ${\sigma_{\pi_E, \, \text{max}}}$ increases steeply, while ${\sigma_{\pi_E, \, \text{min}}}$ still decreases to a minimum value of about $7\%$ at $t_E \approx 1 \, \text{day}$. This means that one can end up with a good sensitivity to $\pi_E$ even at $t_E \approx 1 \, \text{day}$, where the maximum relative error is already a steep function of the Einstein timescale.
For smaller impact parameters, $\beta < 0.1$, Fig. \[figure1\] shows the minimum relative error ${\sigma_{\pi_E, \, \text{min}}}/\pi_E$. The region of maximum sensitivity to $\pi_E$ clearly moves to higher Einstein timescales. If one considers the Einstein timescale ${t_E^{\star}}$ which minimizes ${\sigma_{\pi_E, \, \text{min}}}/\pi_E$, it approximately scales according to $$\label{tEstar}
{t_E^{\star}}\, \propto \, P \, \beta^{-1}.$$ Indeed, it separates the two regimes $\beta t_E \ll P$ and $\beta t_E \gg P$. To estimate the minimum relative error for a given $\beta$, one can write that, approximately, ${\sigma_{\pi_E, \, \text{min}}}/\pi_E \, \propto \,
{t_E}^{1/2} \, \beta \, R^{-1} g(\beta t_E \omega)$, where $g(y)$ is a function that tends to $1$ for $y \rightarrow \infty$, to match Eq. . Since Eq. implies $\beta {t_E^{\star}}\omega = \text{const}$, one obtains $$\label{sigmapiE_min}
\frac{{\sigma_{\pi_E, \, \text{min}}}}{\pi_E} \left( {t_E^{\star}}\right) \, \propto \, P^{1/2} \,
\beta^{1/2} \, R^{-1}.$$ Equations and are in good agreement with Fig. \[figure1\]. The maximum sensitivity to $\pi_E$ increases slowly by decreasing the impact parameters $\beta$, while the range of the Einstein timescales that allow for a parallax detection clearly widens.
![Ratio ${\sigma_{\pi_E, \, \text{min}}}(\nu)/{\sigma_{\pi_E, \, \text{min}}}(\nu=0)$ against $\nu$ for a GSO satellite and $\beta = 0.1$, with the same assumptions as in Fig. \[figure1\]. Three Einstein timescales are shown: $t_E =$ 1, 3, and 10. Clearly the error diverges for $\nu \rightarrow 1$. []{data-label="figure3"}](figures/figure_3.pdf){width="\hsize"}
From the above analysis and Eq. , it follows that a GSO satellite is naturally optimized to measure microlens parallax $\pi_E$ in events raised by free-floating objects with masses spanning from a fraction to a few dozens ${M_{\text{Jup}}}$. The peak amplifications for such a measure can be as low as 10, or even somewhat lower for closer lenses. If, additionally, the Einstein angle $\theta_E$ can be measured via finite source effects (starting at $A_{\text{max}} \approx 50$ for a Jupiter-mass lens and a Sun-like source, see Eq. \[z\]), the lens mass can be inferred for this class of events.\
Analogously, since for a bulge source the Einstein radius scales as $R_E = t_E V = 0.12 \, \text{AU} \, {(M/{M_{\text{Jup}}})}^{1/2} {(x(1-x)/0.25)}^{1/2}$, a GSO satellite is naturally optimized to discover planets in tight orbits around low-mass brown dwarfs[^3]. In particular, as a matter of speculation, it could lead to the discovery of miniature planetary systems around planetary-mass brown dwarfs[^4]. In fact, disks have been found in the past fifteen years to surround brown dwarfs with masses in the range $5-15 \, {M_{\text{Jup}}}$ (see, for example, @Luhman2005feb [@Luhman2005dec]). In particular, the disk around OTS 44 has been estimated at roughly $30 \, M_{\Earth}$ [@Joergens2013]. The question of whether planets can form out of such disks naturally arises, but very little is known about these hypothetical planetary systems, even from a theoretical point of view. Gravitational microlensing could thus lead to fundamental discoveries in this field.
Blending
--------
G13 adopts $\sigma^2 = \sigma_0^2 \, A$ in Eq. , meaning that it neglects the blending flux $F_b$ by stating that one is “only concerned with the scaling of the errors when the source is relatively highly magnified”. Actually, Eq. shows that the information contained in the Fisher matrix thanks to the source flux amplification is scaled by a factor $1-\nu$; i.e., the light curve provides less information about the model parameters if the blending factor $\nu$ is bigger. This is illustrated in Fig. \[figure3\], where we plot the ratio ${\sigma_{\pi_E, \, \text{min}}}(\nu)/{\sigma_{\pi_E, \, \text{min}}}(\nu=0)$ against $\nu$ for $\beta = 0.1$ and different values of the Einstein timescale $t_E$. The error roughly doubles at $\nu \approx 0.6$, and the detection of $\pi_E$ becomes quite hopeless for $\nu \gtrsim 0.9$. Blending can thus highly affect the sensitivity to $\pi_E$ and it must always be taken into account in survey planning. It is also important to point out that, even for very low values of the blending factor, $\nu \rightarrow 0$, this parameter still affects the sensitivity to the other ones. Indeed, a small variation $\delta\nu$ produces a change in the total flux $F$ according to $|\delta F|/F \simeq (A-1)\delta\nu/[(1-\nu)A + \nu]$. For $\nu \rightarrow 0$ and $A \gg 1$, one obtains $|\delta F|/F \simeq \delta\nu$, which means that the contribution of $\nu$ to the Fisher matrix is different from zero even for vanishing blending factors. Gould’s assumption in G13 is only valid for $\nu \rightarrow 0$ and $P \ll \beta t_E$, because in this regime the correlations between $\nu$ and the parallax parameters ${\pi_\|}$ and ${\pi_\bot}$ turn out to vanish[^5].
Low Earth orbit satellite {#LEO}
=========================
![Relative error on $\pi_E$ plotted against $t_E$ for a LEO satellite at 550 km above the Earth’s surface. We assume 3 min exposures, $\sigma_m = 0.01$, and zero blending. For $\beta = 0.1$ the region ${\sigma_{\pi_E}}/\pi_E
\in [{\sigma_{\pi_E, \, \text{min}}}, \, {\sigma_{\pi_E, \, \text{max}}}]/\pi_E$ is shown. For $\beta = 0.05, 0.025, 0.01, \,
\text{and} \, 0.005$, the minimum relative error ${\sigma_{\pi_E, \, \text{min}}}/\pi_E$ is plotted.[]{data-label="figure4"}](figures/figure_1_LEO.pdf){width="\hsize"}
Using a space telescope in low Earth orbit (LEO), such as the HST, measuring microlens parallaxes was first proposed by @Honma1999. In G13 Gould argues that this would generally be useless, since the corresponding $\epsilon = R/1 \, \text{AU}$ is too small. However, substituting Kepler’s third law, $P^2 \, \propto \, R^3$, in Eq. leads to the following scaling law: $$\label{sigmapiE_min_R}
\frac{{\sigma_{\pi_E, \, \text{min}}}}{\pi_E} \left( {t_E^{\star}}\right) \, \propto \, \beta^{1/2} \, R^{-1/4}.$$ The maximum sensitivity to $\pi_E$ decreases slowly when one reduces the orbit radius, and what changes between a GSO and a LEO is just a factor of ${(6.6)}^{1/4} \approx 1.6$. Moreover, Eqs. and show that the timescale at which a LEO satellite is most sensitive to $\pi_E$ coincides with the typical Einstein timescale of an Earth-mass object.
These properties are clearly shown in Fig. \[figure4\], where we consider a LEO satellite with HST-like orbital parameters. We assume an orbit radius $R = R_{\Earth} + 550 \, \text{km}$ ($P = 1 \, \text{h} \, 35.5 \, \text{min}$) and an event latitude $\lambda = 60^{\degree}$ for a bulge source, since the Hubble inclination above the equatorial plane is $28.5^{\degree}$. The predicted relative error ${\sigma_{\pi_E}}/\pi_E$ is plotted against the Einstein timescale $t_E$ for $D_L = 4 \, \text{kpc}$, $V = 200 \,
\text{km/s}$, and zero blending ($\nu = 0$). We assume $f = 1 \, \text{observation}/ 3 \, \text{minutes}$ and $\sigma_m = 0.01$ as in Fig. \[figure1\], to allow for a quick comparison with the GSO. The plot shows the same physics as in Fig. \[figure1\], with two different regimes that arise for each choice of $\beta$. At $\beta = 0.1$, the maximum sensitivity to $\pi_E$ corresponds to Einstein timescales of one to several hours. According to Eq. , these $t_E$ are typical of lens masses ranging from $M_{\Earth}$ to the super-Earth/ice giant transition. With a minimum relative error of roughly $10\%$, Fig. \[figure4\] suggests that a detection of $\pi_E$ for these lenses should be possible for $\beta \lesssim 0.1-0.2$. Since finite source effects easily arise for these kinds of lenses (see Sect. \[finite\_source\]) and can provide a measure of $\theta_E$, a LEO survey satellite could discover populations of free-floating objects ranging from terrestrial planets to super-Earths and ice giants. This would be a fantastic prospective for microlensing and exoplanet science in general. Clearly, a LEO survey would be affected by blending limitations, similar to the GSO. Moreover, the effects of Earth umbra have to be taken into account, because it reduces the fraction of the orbital period available to follow the source star (see Sect. \[earth\_umbra\]).
For $\beta < 0.1$, Eqs. and are consistent with Fig. \[figure4\], similar to what is obtained for the GSO. In particular, we point out that with peak amplifications of several dozen, a LEO satellite should also be able to measure microlens parallax for $t_E \approx 1 \, \text{day}$ events, especially if ground-based observations are also available.
Finite source effects {#finite_source}
---------------------
Finite source effects arise when the impact parameter $\beta$ is comparable to $\rho = R_S x/R_E$, the projection of the source star radius $R_S$ onto the lens plane, measured in units of the Einstein radius $R_E$. Taking a Sun-like source as reference, the parameter $z \equiv \beta/\rho$ is given by $$\label{z}
z = 0.5
\left( \frac{\beta}{0.1} \right)
\left( \frac{t_E}{0.1 \, \text{days}} \right)
\left( \frac{R_{\Sun}}{R_S} \right)
\left( \frac{1/2}{x} \right)
\left( \frac{V}{200 \, \text{km/s}} \right).$$ Equation shows that finite source effects are clearly measurable for $\beta = 0.1$ and $t_E = {t_E^{\star}}\approx 0.1 \, \text{days}$. Moreover, they are always detectable for $\beta < 0.1$ and $t_E = {t_E^{\star}}$ since Eq. implies that $\beta {t_E^{\star}}= \text{const}$.
The Fisher analysis presented in Sect. \[Fisher\_analysis\] considers a point source. Can this assumption invalidate the error predictions of Fig. \[figure4\]? The peak amplification influences ${\sigma_{\pi_E}}$ via the relation $A_{\text{max}} \simeq \beta^{-1}$. Therefore, finite source effects strongly modify the error predicted by a PSPL model if the maximum amplification of a finite source model, $A_{\text{max}}' = \sqrt{4 + \rho^2}/\rho$ [@Witt1994], is much smaller than $\beta^{-1}$, that is, $A_{\text{max}}'/A_{\text{max}}
\ll 1$. Because $A_{\text{max}}'/A_{\text{max}} \simeq 2 \, z$, finite source effects strongly affect the error predictions only for $z \ll 1/2$ [@Witt1994], which corresponds to $t_E \ll 0.1\, \text{days}$ at $\beta = 0.1$ for source stars not much bigger than the Sun. As Fig. \[figure4\] shows, these timescales are, however, of little importance since the error is already too big. For lower values of the impact parameter, $\beta < 0.1$, and $t_E \ll {t_E^{\star}}$, the predictions are a priori strongly affected by finite source effects. However, since Fig. \[figure4\] shows that for $t_E \ll {t_E^{\star}}$, the sensitivity to $\pi_E$ rapidly saturates when one reduces $\beta$, we do not expect important deviations even for these timescales.
In the case of giant-like source stars, $R_S \gtrsim 10
\, R_\Sun$, finite source effects are dominant since peak amplifications can be much lower than what is predicted by a PSPL model. Basically, one can think in terms of an effective impact parameter $\beta' = \beta \, (2z)^{-1}$. For $R_S \gtrsim 10 \, R_\Sun$ and $t_E = 0.1\, \text{days}$, one has $A_{\text{max}}' \simeq 1$ and $\beta' \simeq 1$, and Fig. \[figure4\] suggests that ${\sigma_{\pi_E}}$ would not guarantee a robust parallax detection. At $t_E = 1\, \text{day}$, one gets $\beta' \gtrsim 0.1$, and the parallax sensitivity is still marginal. Therefore, the finite source effects of giant-like stars seem to exclude robust measurements of $\pi_E$ for Earth-mass lenses. However, larger sources generally yield smaller photometric errors $\sigma_m$, and a parallax detection could still be feasible for closer lenses, i.e., $D_L \lesssim 500 \, \text{pc}$ or $x \lesssim 0.06$, and $R_S \lesssim
10 \, R_\Sun$ (see Eq. \[z\]).
Earth umbra {#earth_umbra}
-----------
The Earth shadow can represent a significant limitation to the capabilities of a LEO satellite. It reduces the fraction $\eta$ of the orbital period available to observe the microlensed source star. We calculate the dependence of this fraction on the orbital radius $R$ and the event latitude $\lambda$ over the orbital plane, $$\label{umbra}
\eta = 1 - \frac{1}{\pi} \arcsin \left[ \sqrt{1 - \left( \frac{R}{R_{\Earth}} \right)^2
\sin^2 \lambda} \biggm/ \frac{R}{R_{\Earth}} \cos\lambda \right]$$ For $R = R_{\Earth} + 550 \, \text{km}$, $\eta$ has a minimum value of $63\%$ at $\lambda = 0^{\degree}$, and it slowly increases to $79\%$ for $\lambda =
60^{\degree}$. For $\lambda \geq \lambda^{\star} \equiv \arcsin (R_{\Earth}/R)
\approx 67^{\degree}$, the source star is always visible from the satellite. Consequently, even if the analysis of Sect. \[LEO\] is strictly valid for continuous observations ($\eta = 1$), Eq. seems to indicate that Earth umbra does not invalidate its general conclusions when one tracks bulge stars. There is also some room to minimize the impact of the Earth shadow by adjusting the satellite orbital parameters. For an altitude of 1000 km, for example, the angle $\lambda_{\star}$ decreases to $60^{\degree}$. Moreover, increasing satellite inclination above the equatorial plane can substantially reduce the shadow impact.
Conclusions
===========
The present study employs a numerical Fisher matrix analysis to assess the feasibility of measuring microlens parallaxes by means of Earth orbit satellites. It extends the previous analytical analysis of @Gould2013 to shorter Einstein timescales. We predict that, at $A_{\text{max}} \gtrsim 5-10$, a GSO satellite could detect microlens parallaxes for free-floating lenses with masses spanning from a fraction to a few dozen ${M_{\text{Jup}}}$ (providing their mass if $\theta_E$ is also measured via finite source effects). It could also discover planets in tight orbits around very low-mass brown dwarfs. Moreover, at $\beta \lesssim 0.1-0.2$, a LEO satellite could discover free-floating objects ranging from terrestrial planets to super-Earths and ice giants. It could also detect, at $\beta \lesssim 0.05$, microlens parallaxes for Jupiter-mass free-floaters. Limitations to these results can be the strong requirements on the photometry ($0.01$ magnitude error with three-minute exposures), the effects of blending, and in case of a LEO satellite, the Earth umbra. It is useful to note that, even though this study adopts a single-lens model, we extrapolated its results to binary-lens events. Indeed, as far as a parallax effect is concerned, the fundamental quantity is the projected separation between the source star and the lens centre of mass. In planetary events this almost coincides with the host. What really differentiates the two kinds of events is that finite source effects are needed to measure $\theta_E$ and estimate the lens mass: they are not routinely measured in single-lens events, while they are for binary ones.
F. M. is grateful to N. Cornuault, L. Pittau, and C. Ranc for the frequent and fruitful discussions. The authors acknowledge the support of PERSU Sorbonne Université and the Programme National de Planétologie.
[^1]: Substitution causes the element $\mathcal{F}_{\overline{F}
\, \overline{F}}$ to diverge, meaning that an arbitrary precision on $\overline{F}$ can be attained by observing enough in the baseline of the light curve. The parameter $\overline{F}$ can thus be safely considered as a constant and can be removed from the Fisher analysis.
[^2]: The extrema ${\sigma_{\pi_E, \pm}}(\varphi)$ correspond to a lens-source relative motion aligned with one of the principal axes of the bivariate Gaussian distribution which approximates the posterior probability distribution of ${\pi_\|}$ and ${\pi_\bot}$ in the Fisher matrix analysis, when one marginalises over the remaining parameters.
[^3]: @Han2013 report the discovery of such a system. However, the mass ratio of this binary appears too high to envisage that the companion formed in a protoplanetary disk around the host.
[^4]: These could also be interpreted as free-floating planet-moon systems [@Bennett2014].
[^5]: We note that, if for a given event $\nu = 0$ holds exactly, the parameter $\nu$ must be removed from the corresponding Fisher analysis, and the present predictions provide an upper bound to ${\sigma_{\pi_E}}$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A group $G$ is a [$vGBS{}$]{} group if it admits a decomposition as a finite graph of groups with all edge and vertex groups finitely generated and free abelian. We construct the JSJ decomposition of a [$vGBS{}$]{} group over abelian groups. We prove that this decomposition is explicitly computable, and may be obtained by local changes on the initial graph of groups.'
author:
- Benjamin Beeker
bibliography:
- 'abelianjsj.bib'
title: Abelian JSJ decomposition of graphs of free abelian groups
---
Introduction
============
The theory of JSJ decomposition starts with the work of Jaco-Shalen and Johansson on orientable irreducible closed $3$-manifolds giving a canonical family of 2-dimensional tori. Kropholler first introduced the notion into group theory giving a JSJ decomposition for some Poincaré duality groups [@Krop]. Then Sela gave a construction for torsion-free hyperbolic groups [@Sela]. This notion has been more generally developed by Rips and Sela [@RiSe], Dunwoody and Sageev [@DunSa2], Fujiwara and Papasoglu [@FuPa] for various classes of groups. In [@GL3a], Guirardel and Levitt generalize the object by introducing the definition of JSJ deformation space, proving the existence of this space for finitely presented groups.
In this paper we consider the following class of groups.
Let $\Gamma$ be a finite graph of groups with all vertex groups finitely generated free abelian. Let $G$ be the fundamental group of $\Gamma$. If the rank of all edge and vertex groups is equal to a fixed integer $n$, we call such a group a ${$GBS{}$}_n$ group, standing for Generalized Baumslag-Solitar groups of rank $n$. When the rank is variable, we call such a group a [$vGBS{}$]{} group. The goal of this paper is to describe the JSJ decomposition of $G$ over abelian groups and to give a way to construct it.
A decomposition of a group $G$ is a graph of groups with fundamental group $G$. To define what a JSJ decomposition is, we need the notion of universally elliptic subgroups. Given a group $G$ and a decomposition $\Gamma$ of $G$, a subgroup $H$ is *elliptic* if the group $H$ is conjugate to a subgroup of a vertex group of $\Gamma$. Given a class of subgroups $\mathcal A$ of $G$, a subgroup $H\subset G$ is *universally elliptic*, if $H$ is elliptic in every decomposition of $G$ as a graph of groups with edge groups in $\mathcal A$. A decomposition is universally elliptic if all edge groups are universally elliptic.
A decomposition $\Gamma$ *dominates* another decomposition $\Gamma'$, if every elliptic group of $\Gamma$ is elliptic in $\Gamma'$. A decomposition is a JSJ decomposition if it is universally elliptic, and it dominates every other universally elliptic decomposition. Then given a vertex $v$ in a JSJ decomposition, either $G_v$, the vertex group of $v$, is universally elliptic, we say $v$ is *rigid*, or $G_v$ is not then we say $v$ is *flexible*.
For instance, looking at the JSJ decomposition of a torsion-free hyperbolic group over cyclic groups described by Sela in [@Sela] and Bowditch [@bow], the flexible vertex groups are exactly the surface groups.
The defining decomposition $\Gamma$ of a [$vGBS{}$]{} group $G$ is a good approximation of the JSJ. For example, for a ${$GBS{}$}_n$ group, the graph of groups $\Gamma$ is a JSJ decomposition whenever the associated Bass-Serre tree $T$ (which is locally finite in this case) is not a line ([@for],[@GL3a]) or equivalently whenever $G$ is not polycyclic.
In the general case three kinds of local changes may be needed to obtain the JSJ from $\Gamma$.
- When the quotient of a vertex group $G_v$ of the decomposition $\Gamma$ by the group $\tilde G_v$ generated by all adjacent edge groups is virtually cyclic, then the vertex is blown into a loop (see figure \[figurecasexpansion\] for the simplest example).
(0,-0.92)(12.104218,0.92) (1.0395312,0.1)(3.239531,0.1) (8.039532,0.1)(10.239531,0.1) (10.839532,0.1)[0.6]{} (2.09375,-0.195)[${\mathbb Z }^n$]{} (1.46375,0.405)[${\mathbb Z }^{n+2}$]{} (3.4737499,0.405)[${\mathbb Z }^{p}$]{} (3.2637498,-0.155)[$v$]{} (8.46375,0.405)[${\mathbb Z }^{n+2}$]{} (9.093749,-0.195)[${\mathbb Z }^n$]{} (11.693749,0.205)[${\mathbb Z }^n$]{} (10.093749,0.405)[${\mathbb Z }^n$]{} (5.3737504,0.005)[$\Longrightarrow$]{} (8.0,0.1)(7.2,-0.9) (8.0,0.1)(7.0,0.3) (8.0,0.1)(7.0,0.9) (1.0,0.1)(0.2,-0.9) (1.0,0.1)(0.0,0.3) (1.0,0.1)(0.0,0.9) (5.405625,0.485)[if $p=n+1$]{}
To be more specific, in figure \[figurecasexpansion\], if $p=n$, then the edge group has finite index in the vertex group of $v$ and $v$ is rigid. If $p=n+1$, there exists a unique non-trivial splitting of $v$ which leaves the ${\mathbb Z }^n$ part elliptic. The vertex must be blown into a loop as in the figure, and the new vertex is then rigid. If $p\geq n+2$ then $v$ is flexible.
- Conversely some loops must be collapsed.
A loop $l$ based at a vertex $v$ is called a $1-1$ loop if both inclusion maps of its edge group into $G_v$ are bijections. The fundamental group of the subgraph of groups composed of $v$ and $l$ is a semi-direct product ${\mathbb Z }^n\rtimes_\varphi {\mathbb Z }$. Some of the $1-1$ loops are collapsed, depending on $\varphi$ and on the other edges adjacent to $v$.
For example, in figure \[figurecas1-1\], the loop must be collapsed if and only if $k<n$ : if $k<n$, there are many decompositions of ${\mathbb Z }^{n+1}={\mathbb Z }^n\rtimes_{id} {\mathbb Z }$ as an HNN extension leaving ${\mathbb Z }^k$ in the vertex group. If $k=n$, there is exactly one, and its edge group is universally elliptic.
(0,-1.5229688)(11.734219,1.5229688) (8.0,0.12453125)(10.139532,0.1440624) (1.0395312,0.1440624)(3.239531,0.1440624) (3.8395312,0.1440624)[0.6]{} (8.993751,-0.15093762)[${\mathbb Z }^k$]{} (8.36375,0.44906238)[${\mathbb Z }^{m}$]{} (10.81375,0.44906238)[${\mathbb Z }^{n+1}$]{} (1.46375,0.44906238)[${\mathbb Z }^{m}$]{} (2.09375,-0.15093762)[${\mathbb Z }^k$]{} (4.80375,0.2090624)[${\mathbb Z }^{n}$]{} (2.9837499,0.51109374)[${\mathbb Z }^{n}$]{} (1.0,0.12453125)(0.2,-0.67546874) (1.0,0.12453125)(0.4,0.12453125) (1.0,0.12453125)(0.0,0.9245312) (8.0,0.12453125)(7.4,-0.67546874) (8.0,0.12453125)(7.6,0.12453125) (8.0,0.12453125)(7.2,0.9245312) (6.3028126,0.52953125)[$k<n$]{} (6.3028126,0.02953125)[$\Longrightarrow$]{} (4.62,-0.13546875)(4.6030965,-0.97302973)(3.2,-0.97546875)(3.2,-0.07790777) (4.6141744,0.6045312)(4.4678555,1.2845312)(3.2322712,1.2645313)(3.1986957,0.38453126) (4.7002497,-0.15262192)[0.07975015]{}[0.0]{}[180.0]{} (4.687334,0.61769086)[0.07315958]{}[179.04517]{}[19.885164]{} (3.9114063,-1.0004688)[$id$]{} (3.9114063,1.3295312)[$id$]{}
- A similar phenomenon occurs for edges which are not loops and whose group has index $2$ in each adjacent vertex group (see figure \[figurecas2-2\]). We call an edge of this type a $2-2$ edge.
(1,-1.548125)(12.882812,1.548125) (8.301406,0.2739062)[$\Longrightarrow$]{} (5.965938,0.8646874)[${\mathbb Z }^2=\langle c,d\rangle$]{} (5.2367187,0.32968745)(7.036719,0.7296874) (5.2367187,0.32968745)(6.4367185,-1.0703127) (5.2367187,0.32968745)(6.836718,-0.27031255) (2.76,0.348125)(5.2367187,0.32968745) (5.2567186,0.32968745) (10.4167185,0.18968745)(12.216719,0.58968747) (10.4167185,0.18968745)(11.616718,-1.2103126) (10.4167185,0.18968745)(12.016719,-0.41031256) (10.4167185,0.18968745)(9.216719,1.1896875) (10.4167185,0.18968745)(8.616718,-0.6103125) (10.4167185,0.18968745) (3.3459377,0.8846874)[${\mathbb Z }^2=\langle a,b\rangle$]{} (2.8367188,0.32968745)(1.6367185,1.3296874) (2.8367188,0.32968745)(1.0357811,-0.47031257) (2.8367188,0.32968745) (4.2523437,-0.06531255)[$a^2=c^2$]{} (10.7223425,0.73468745)[$\langle a,c\rangle\times\langle b\rangle$]{} (1.3923438,-0.6853125)[$\langle b^4\rangle$]{} (1.4323436,0.8546875)[$\langle b^3\rangle$]{} (9.532343,1.3546875)[$\langle b^3\rangle$]{} (8.992344,-0.9053124)[$\langle b^4\rangle$]{} (6.772344,-1.2453125)[$\langle d^2\rangle$]{} (7.132344,-0.4253125)[$\langle d^7\rangle$]{} (6.732344,0.3346875)[$\langle d^3\rangle$]{} (12.032344,-1.3253125)[$\langle d^2\rangle$]{} (12.512343,-0.5453125)[$\langle d^7\rangle$]{} (12.372344,0.2146875)[$\langle d^3\rangle$]{} (4.282344,-0.46531254)[$b=d$]{}
Given an edge $e$ of $\Gamma$, call $\Gamma_e$ the subgraph consisting of the single edge $e$ (and its vertices) and $\Pi_e$ its fundamental group.
To obtain the JSJ decomposition of a [$vGBS{}$]{} group over abelian groups, we show that it suffices to expand vertex groups $G_{v }$ such that $G_{v } / \tilde G_{v }$ is virtually cyclic as in figure \[figurecasexpansion\], and collapse edges which are not universally elliptic. These edges have polycyclic groups, and so are $1-1$ loops and $2-2$ edges as in figures \[figurecas1-1\] and \[figurecas2-2\].
\[theoremintro\] Let $G=\pi_1(\Lambda)$ be a [$vGBS{}$]{} group. For ${v }$ a vertex, let $\tilde G_{v }$ be the subgroup of $G_v$ generated by groups of edges adjacent to ${v }$. A JSJ decomposition of $G$ over abelian groups can be obtained by
- expanding the groups $G_{v }$ such that $G_{v } / \tilde G_{v }$ is virtually cyclic,
- collapsing $2-2$ edges and $1-1$ loops $e$ which are not universally elliptic.
As for hyperbolic groups, in a JSJ of a [$vGBS{}$]{} group, the rigid vertices are also easily identifiable. A vertex is rigid if and only if its vertex group is abelian and virtually generated by the adjacent edge groups.
In order to make theorem \[theoremintro\] more explicit, we shall now describe the edges which are not universally elliptic.
We first describe the edges $e$ for which $\Pi_e$ has a trivial JSJ decomposition. This is the same as giving the JSJ decomposition of polycyclic ${$GBS{}$}_n$ groups. As (reduced) decompositions of polycyclic groups must have a line (or a point) as Bass-Serre tree, these decompositions have at most one edge, which may be of two types.
The first are the $2-2$ edges, that is, non-loop edges whose group is of index $2$ in the groups of the two adjacent vertices. Then $\Pi_e$ is isomorphic to a direct product of a ${\mathbb Z }^n$ by either the Klein bottle group $\langle a, b | a^2=b^2 \rangle$, or the twisted Klein bottle group $\langle a, b | a^2b=ba^2, ab^2=b^2a\rangle$.
The second are $1-1$ loops (whose inclusion maps into the vertex group are bijections), and their groups can be seen as semi-direct product ${\mathbb Z }^n \rtimes_\varphi{\mathbb Z }$:
\[1-1intro\]
1. If $e$ is an edge of type $2-2$ then $\Pi_e$ has a trivial JSJ decomposition over abelian groups.
2. Let us take $\varphi$ in $\text{Aut}({\mathbb Z }^n)\simeq GL_n({\mathbb Z })$.
The group $G_\varphi={\mathbb Z }^n\rtimes_{\varphi}{\mathbb Z }$ has a trivial JSJ decomposition if and only if $\varphi$ can be written in a well-chosen basis $(x,h_1,\dots, h_{n-1})$ in one of the following ways:
1. $\left(
\begin{array}{cc}
1 & 0 \\
p & M
\end{array}
\right)$ with $M$ an $(n-1)\times(n-1)$ matrix of finite order and $p$ in ${\mathbb Z }^{n-1}$,
2. $\left(
\begin{array}{cc}
-1 & 0 \\
p & Id_{n-1}
\end{array}
\right)$ with $p$ in ${\mathbb Z }^{n-1}$.
In every other case the semi-direct product ${\mathbb Z }^n\rtimes_{\varphi}{\mathbb Z }$ is a JSJ decomposition.
This theorem is proved in section \[section22\] and \[section11\].
By [@GL3a lemma 4.10], we know that the edges $e$ which are not universally elliptic are those for which $\Pi_e$ has a trivial JSJ decomposition relative to adjacent edge groups. We therefore have to understand how $\Pi_e$ is embedded in the whole group. More precisely, we look at the way the adjacent edge groups inject into $\Pi_e$. We obtain the following theorem which makes theorem \[theoremintro\] explicit.
Define a *hyperplane* of ${\mathbb Z }^n$ as the kernel of a morphism from ${\mathbb Z }^n$ to ${\mathbb Z }$.
\[simplesintro\]
Let $G=\pi_1(\Gamma)$ be a [$vGBS{}$]{} group. For ${v }$ a vertex, let $\tilde G_{v }$ be the subgroup of $G_v$ generated by groups of edges adjacent to ${v }$. A JSJ decomposition of $G$ over abelian groups can be obtained by
- expanding the groups $G_{v }$ such that $G_{v } / \tilde G_{v }$ is virtually cyclic,
- collapsing $1-1$ edges $e$ such that $\Pi_e$ may be decomposed as in theorem \[1-1intro\] with all adjacent edge groups included in the hyperplane $\langle h_1, \dots, h_{n-1}\rangle$.
- collapsing $2-2$ edges $e$ with vertices $v$ and $v'$, such that there is a hyperplane $H$ of $G_{e }$ which is also a hyperplane of $G_{v }$ and $G_{{v' }}$ and which contains all groups of adjacent edges, $G_{e }$ excepted.
We prove this proposition in sections \[sectioninteraction\] and \[finalsection\]. We prove in section \[algo\] that the construction of the JSJ is algorithmic.
From theorem \[simplesintro\], we obtain the JSJ decomposition of [$vGBS{}$]{} groups over abelian groups with bounded rank:
Let $G=\pi_1(\Gamma)$ a [$vGBS{}$]{} group, and $n\in {\mathbb N }$. Suppose that $\Gamma$ is a JSJ decomposition over free abelian groups. Then a JSJ decomposition over free abelian groups of rank $\leq n$ may be obtained by collapsing every edge of $\Gamma$ with group of rank $> n$.
Preliminaries
=============
Let $G$ be a finitely generated group. We denote by $T$ the Bass-Serre tree associated to a finite graph of groups decomposition $\Gamma$ of $G$. For a vertex $v$ (resp. an edge $e$) of $\Gamma$, we denote $G_v(\Gamma)$ (resp $G_e(\Gamma)$) its stabilizer. Most of the time $\Gamma$ will be omitted.
For $e$ an (unoriented) edge of $T$ with vertices $v$ and $v'$, we call the *type* of $e$ the couple $m-n$ of potentially infinite numbers with $m\leq n$ such that $G_e$ is of index $m$ in $G_v$ and $n$ in $G_{v'}$. For example, an edge whose stabilizer equals the stabilizer of its vertices, is of type $1-1$. This type only depends on the orbit of the edge, so we define the type of an edge of the graph of groups as the type of an edge in $T$ representing it. We say that an edge $e$ of $T$ is a loop if its vertices are in the same orbit under the action of $G$. A graph of groups is *reduced* if all edges of type $1-n$ are loops in the graph. Every decomposition may be reduced by contracting successively the edges of type $1-n$ which are not loops.
Given a Bass-Serre tree $T$, an element $g\in G$ is *elliptic* if it fixes a vertex. Otherwise $g$ is said to be *hyperbolic*. The *characteristic space* of $g$ is the minimal subtree of $T$ containing the vertices $v$ such that the distance between $v$ and $g\cdot v$ is minimal (seeing $T$ as a metric space with all edges of lenght $1$). When $g$ is elliptic, it is the set of all fixed edges and vertices. When $g$ is hyperbolic this is the only line on which $g$ acts by a translation. In this case we call it the *axis* of $g$. A subgroup $H$ of $G$ is *elliptic* if it fixes a vertex or equivalently when $H$ is finitely generated if all of its elements are elliptic. In the case of finitely generated abelian groups, the ellipticity of a generating set implies the ellipticity of the whole group.
From now on and for the rest of the paper, the decompositions we consider will be over free abelian groups, meaning that every edge stabilizer is finitely generated free abelian.
An element or a subgroup of $G$ is *universally elliptic* if it is elliptic in the Bass-Serre trees of all graph of groups decompositions. An edge is universally elliptic if it carry a universally elliptic group. A decomposition is *universally elliptic* if all edge are universally elliptic. A graph of group decomposition $\Gamma$ dominates an other decomposition $\Gamma'$ if every elliptic subgroups of $\Gamma$ is elliptic in $\Gamma'$. A *JSJ decomposition* is a universally elliptic decomposition which dominates every other universally elliptic decomposition ([@GL3a]).
If $e$ is an edge of a Bass Serre tree, with vertices $v_1$ and $v_2$, of type $1-1$ representing a loop $l$ in the graph of groups, let $t$ be a hyperbolic element such that $t\cdot v_1 = v_2$. We call *modulus* of $l$ the linear map $\varphi\in Aut(G_{v_1})$ such that for all $x$ in $G_{v_1}$ we have $\varphi(x)=txt^{-1}$. As $G_v\simeq G_l \simeq Z^n$, we can see $\varphi$ as an element of $GL_n({\mathbb Z })$. Up to conjugacy, the modulus does not depend neither on the choice of $t$ nor on the choice of $e$ representing $l$ but is switched to its inverse if we change the orientation of $l$.
Universally elliptic edges {#sectionuniverselle}
==========================
\[centralisateur\]
Let ${G }=\pi_1({\Gamma })$ be a graph of groups and $T$ its Bass-Serre tree. Let ${x }$ be a hyperbolic element of ${G }$. Then the centralizer $C_{G }({x })$ of ${x }$ in ${G }$ is a semi-direct product ${E}\rtimes {H }$ with ${E}$ a subgroup of an edge stabilizer of $T$ and ${H }$ a cyclic subgroup of ${G }$ generated by a hyperbolic element.
When all edge groups of $\Gamma$ are finitely generated free abelian, the centralizer of a hyperbolic element is a polycyclic group ${\mathbb Z }^n\rtimes{\mathbb Z }$.
The group $C_{G }({x })$ acts on the axis of ${x }$ by translation. This action defines a morphism from $C_G({x })$ to ${\mathbb Z }$. The kernel of this morphism fixes the axis pointwise and so belongs to the stabilizer of the axis.
The centralizers of hyperbolic elements have a very specific structure, which is not the case for elliptic ones. This forces most of edge groups of [$vGBS{}$]{} group to be universally elliptic:
\[rigidite1\]
Let ${G }=\pi_1({\Gamma })$ be a [$vGBS{}$]{} group with $\Gamma$ reduced. Let ${e }$ be an edge of $\Gamma$.
1. If ${e }$ is not a loop of $\Gamma$ and is not a $2-2$ edge then $G_{e }$ is universally elliptic.
2. If ${e }$ is a loop of $\Gamma$ but is not $1-1$ then $G_{e }$ is universally elliptic.
Call $\tilde e$ a representative of $e$ in the Bass-Serre tree. An abelian group generated by finitely many elliptic elements is elliptic. We just have to show that each element of the edge group $G_{\tilde e}$ is universally elliptic.
Let ${\tilde v }$ and ${\tilde v' }$ be the endpoints of ${\tilde e }$. Then $G_{{\tilde v }}\ast_{G_{\tilde e }}G_{{\tilde v' }}$ is contained in the centralizer of $G_{\tilde e }$. If $G_{\tilde e }$ is not of index $\leq 2$ in both ${\tilde v }$ and ${\tilde v' }$, the amalgam contains a free group. So the centralizer of $G_{\tilde e }$ cannot be a polycyclic group. By proposition \[centralisateur\], every element of the edge stabilizer $G_{\tilde e }$ is elliptic and $G_{e }$ is universally elliptic.
There remain two cases: when $e$ is a $2-2$ loop, and when $e$ is a $1-k$ loop with $k>1$. In both cases, let $t\in{G }$ be such that $t\cdot {\tilde v } = {\tilde v' }$.
If ${e }$ is a $2-2$ loop, let ${\alpha }$ be a square in $G_{\tilde e }$. Then the centralizer of ${\alpha }$ contains $G_{{\tilde v }}$, $tG_{{\tilde v }}t^{-1}$ and $t^{-1}G_{{\tilde v }}t$ and so the group $tG_{{\tilde v }}t^{-1}\ast_{G_{{\tilde e }}}(G_{{\tilde v }}\ast_{t^{-1}G_{{\tilde e } }t} t^{-1} G_{{\tilde v }} t)$. But this amalgamated product is of $2-\infty$ type, the same argument as before works. And so ${\alpha }$ is universally elliptic. Every element of $G_{\tilde e }$ has a universally elliptic square, and $G_{e }$ is universally elliptic.
If ${e }$ is a $1-k$ loop with $k>1$, then for all $n>0$, the group $G_{\tilde e }$ is a subgroup of $t^nG_{{\tilde v }}t^{-n}$ which is abelian. So its centralizer contains $\bigcup_{n>0} t^nG_{{\tilde v }}t^{-n}$, a strictly increasing union of groups. The centralizer contains a non-finitely generated group and cannot be polycyclic, contradicting proposition \[centralisateur\]. So $G_{e }$ is universally elliptic.
Graphs of groups consisting of an HNN-extension of type $1-1$ or an amalgamation of type $2-2$ are exactly the reduced ones whose Bass-Serre trees are lines. They are the polycyclic [$vGBS{}$]{} groups. The goal of the next two paragraphs is to study the different decompositions of these groups in order to determine their JSJ.
{#section22}
In this part, one shows that, for a given n, there are exactly two [$vGBS{}$]{} groups whose graphs of groups have two vertices with vertex group ${\mathbb Z }^n$ linked by an edge of type $2-2$, and that these two groups can be seen as semi-direct products of ${\mathbb Z }^n$ by ${\mathbb Z }$. So these groups have both an action by translation and a dihedral action on ${\mathbb R }$. We shall prove in next section that they both have a trivial JSJ decomposition.
The first one is the direct product of the Klein bottle group $K=\langle b_0,b_1|\ b_0^2=b_1^2\rangle$ by $E={\mathbb Z }^{n-1}$. One will call $K$ the *untwisted Klein bottle group*.
The second one is a twisted version of the first, it can be described as the product of $F={\mathbb Z }^{n-2}$ and the group $K'$ with presentation $$\langle b_0,b_1| \ b_0^2b_1=b_1b_0^2,\ b_1^2b_0=b_0b_1^2 \rangle.$$ One will call $K'$ the *twisted Klein bottle group*.
Graph of groups decompositions of $K$ and $K'$ are as in figure \[figure2-2\].
(0,-0.34921876)(10.8628125,0.36921874) (1.9809375,-0.32921875)(3.7809374,-0.32921875) (1.9423437,0.17578125)[$\langle b_0\rangle$]{} (3.7923439,0.17578125)[$\langle b_1\rangle$]{}
(6.9809375,-0.32921875)(8.7809374,-0.32921875) (6.9423437,0.17578125)[$\langle b_0,~b_1^2\rangle$]{} (8.7923439,0.17578125)[$\langle b_0^2,~b_1\rangle$]{}
The two sets of groups $K\times E$ and $K'\times F$ will be called *extended (untwisted or twisted) Klein bottle groups*. Like $K$ and $K'$, they have a decomposition in a amalgam of type $2-2$.
\[TFAE\] Let $G$ be a [$vGBS{}$]{} group and $n\in{\mathbb N }$. The following are equivalent:
1. \[un\]The group $G$ is an extended Klein bottle group $K\times{\mathbb Z }^{n-1}$ or $K'\times{\mathbb Z }^{n-2}$.
2. \[deux\]The group $G$ is a semi-direct product $Z^n\rtimes_\varphi Z$ with $\varphi=\left(
\begin{array}{cc}
-1 & 0 \\
p & Id_{n-1}
\end{array}
\right)$ in a suitable basis of ${\mathbb Z }^n$.
3. \[trois\]The group $G$ admits a graph of group decomposition with two vertices carrying groups ${\mathbb Z }^n$ and an edge of type $2-2$.
We shall prove successively \[un\] $\Leftrightarrow$ \[deux\] and \[trois\] $\Leftrightarrow$ \[un\]
\[pre2-2\] The group $K$ can be seen as a semi-direct product ${\mathbb Z }\rtimes_{-Id} {\mathbb Z }$, and $K'$ as ${\mathbb Z }^2\rtimes_\varphi {\mathbb Z }$ with $\varphi=\left(
\begin{array}{cc}
-1 & 0 \\
1 & 1
\end{array}
\right)$.
The proof consists in giving for $K$ and $K'$ a change of presentation, such that the second one is a semi-direct product. In the case of $K$, the change of presentation is well known: $$\begin{array} {crcl}
\psi: &\langle b_0,b_1|b_0^2=b_1^2\rangle &\stackrel{\sim}{\longrightarrow} &\langle t,a_0|tat^{-1}=a^{-1}\rangle \\
&b_0 &\longmapsto &t\\
&b_0b_1^{-1} &\longmapsto &a
\end{array}.$$
In the second case, the presentation $$\mathcal P = \langle b_0,b_1|b_0^2b_1=b_1b_0^2,\ b_1^2b_0=b_0b_1^2\rangle$$ of $K'$ can be changed to $$\mathcal P' = \mathcal \langle t,a_1,a_2|ta_2 t^{-1}=a_2, ta_1 t^{-1} =a_1^{-1}a_2,[a_1,a_2]=1\rangle$$ via the map $$\begin{array} {crcl}
\psi:&\mathcal P &\stackrel{\sim}{\longrightarrow} &\mathcal P'\\
&b_0 &\longmapsto &t\\
&b_0b_1^{-1} &\longmapsto &a_1\\
&b_0^2b_1^{-2} &\longmapsto &a_2
\end{array}.$$
\[2-2\] Each group $K\times E$ and $K'\times F$ can be identified with a semi-direct product $\mathbb Z^n \rtimes_\varphi \mathbb Z$ where $\varphi\in\text{Aut}(Z^n)$ has matrix $
\left(
\begin{array}{cc}
-1 & 0 \\
p & Id_{n-1}
\end{array}
\right)$ in a well-chosen basis $(x,h_1,\dots h_{n-1})$ .
Moreover, with this identification:
- $x=b_0b_1^{-1}$,
- for all $h$ in $\langle h_1,\dots h_{n-1}\rangle$, the element $xh$ is hyperbolic in the decomposition in amalgam of type $2-2$,
- $\langle h_1,\dots h_{n-1}\rangle = E$ in the case of $K\times E$,
- $\langle h_1,\dots h_{n-1}\rangle=\langle F,b_0^2b_1^{-2}\rangle$ in the case of $K'\times F$.
In particular, in the semi-direct product decomposition, seen as an HNN extension, the groups $E$, $F$ and the element $b_0^2b_1^{-2}$ stay elliptic, the element $b_0b_1^{-1}$ is elliptic too. On the opposite $b_0$ and $b_1$ are hyperbolic.
One can notice that, in the case of untwisted Klein bottle groups, the element $b_0^2b_1^{-2}$ is trivial.
The changes of presentation described in the previous proof can be extended to $K\times E$ or $K'\times F$ by noticing that $({\mathbb Z }^p\rtimes_\varphi {\mathbb Z })\times {\mathbb Z }^q \simeq ({\mathbb Z }^p\times {\mathbb Z }^q) \rtimes_{(\varphi\times Id_q)} {\mathbb Z }$. The four points are easy to check.
This proves \[un\] $\Rightarrow$ \[deux\] in proposition \[TFAE\].\
The automorphism $\varphi=
\left(
\begin{array}{cc}
-1 & 0 \\
p & Id_{n-1}
\end{array}
\right)$ is conjugate to $
\left(
\begin{array}{cc}
-1 & 0 \\
0 & Id_{n-1}
\end{array}
\right)$ if all coordinates of $p$ are even, to $\left(
\begin{array}{rcc}
-1 & 0 &0 \\
1 & 1 &0\\
0 & 0 & Id_{n-2}
\end{array}
\right)$ if one is odd. So the semi-direct product ${\mathbb Z }^n\rtimes_\varphi{\mathbb Z }$ is isomorphic to one of $K\times E$ or $K'\times F$. This implies the \[deux\] $\Rightarrow$ \[un\] part.
\[un\] $\Rightarrow$ \[trois\] has already been done.
Let us prove \[trois\] $\Rightarrow$ \[un\].
Let $A$ and $B$ be two copies of ${\mathbb Z }^n$ generated by the sets $\{a_1, \dots, a_n\}$ and $\{b_1,\dots, b_n\}$ respectively. By assumption, ${G }$ has a presentation $$\langle A, B |\ a_1^2=\varphi(a_1^2),\ a_j=\varphi(a_j), 2 \leq j \leq n\rangle$$ where $\varphi$ is an isomorphism between $\langle a_1^2,a_2,\dots, a_n\rangle $ and $\langle b_1^2,b_2,\dots, b_n\rangle $.
If $\varphi^{-1}(b_1^2)$ is a square in $A$, let $c$ be its square root. Let us take the family $\left\{ c,~\varphi^{-1}(b_i), i>1\right\}$ as a basis of $A$. The presentation of ${G }$ is then changed to $$\langle b_1,b_2,\dots, b_n,c \ |\ b_ib_j=b_jb_i,\ c^2=b_1^2,\ cb_j=b_jc\ for\ j\neq 1,\rangle$$ which is a presentation of an extended Klein bottle group.
If $\varphi^{-1}(b_1^2)$ is not a square in $A$, then $\varphi^{-1}(b_1^2)$ may be written $\prod_{i=1}^na_i^{\nu_i}$, and there exists $r>1$ for which $\nu_r$ is odd. We may assume $r=2$. The element $\tilde a_2 = a_2^{1-\nu_2}\varphi^{-1}(b_1^2)$ is such that its image by $\varphi$ is a square $\tilde b_2^2$ in $B$ with $\tilde b_2$ out of the image of $\varphi$. As $a_2$ can be written $a_2=\tilde{a_2}\cdot \prod_{i=1, i\neq 2}^na_i^{-\nu_i}$, the family $\left\{ \tilde a_2,~a_i,~i\neq 2\right\}$ is a basis of $A$. Taking as generators for $B$ the elements $c_2=\tilde b_2$, $c_1=\varphi(a_1^2) $ and $c_i=\varphi(a_i)$ for $i>2$, the group ${G }$ admits as presentation $$\langle a_i,c_i|a_ia_j=a_ja_i, c_ic_j=c_jc_i, a^2_1=c_1, a_2=c_2^2, a_i=c_i, i>2\rangle.$$ And ${G }$ is isomorphic to $ K'\times{\mathbb Z }^{n-2}$.
{#section11}
The groups we consider here are semi-direct products of ${\mathbb Z }^n$ by ${\mathbb Z }$. The goal of the section is to determine their JSJ decompositions.
For $\varphi$ an element of $Gl_n({\mathbb Z })$ we write $G_\varphi$ the group ${\mathbb Z }^n\rtimes_\varphi {\mathbb Z }$. We will call *hyperplane* of ${\mathbb Z }^n$ the kernel of a linear map $f:{\mathbb Z }^n\twoheadrightarrow {\mathbb Z }$.
\[simples\] If $\varphi$ cannot be written (up to conjugation) in one of the following ways:
1. $\left(
\begin{array}{cc}
1 & 0 \\
p & M
\end{array}
\right)$ with $M$ an $(n-1)\times(n-1)$ matrix of finite order and $p$ in ${\mathbb Z }^{n-1}$,
2. $\left(
\begin{array}{cc}
-1 & 0 \\
p & Id_{n-1}
\end{array}
\right)$ with $p$ in ${\mathbb Z }^{n-1}$.
then the group $G_\varphi={\mathbb Z }^n\rtimes_\varphi {\mathbb Z }$ has a unique non-trivial (reduced) graph of groups decomposition.
Its JSJ decomposition is the HNN-extension ${\mathbb Z }^n\ast_\varphi$.
\[1-1\] Let ${x }$ be an element of the ${\mathbb Z }^n$ part of $G_\varphi$. Assume that there exists a graph of groups decomposition $\Lambda$ of $G_{\varphi }$ for which ${x }$ is hyperbolic. Then there exists a hyperplane ${H }\subset {\mathbb Z }^n$ elliptic in the decomposition $\Lambda$, stable under the action of ${\varphi }$, and such that ${\varphi }_{|{H }}$ has finite order. Moreover, the set of elliptic elements of ${{\mathbb Z }^n}$ in $\Lambda$ is exactly ${H }$, and if ${\varphi }({x }){x }^{-1}$ does not belong to ${H }$ then ${\varphi }_{|{H }} = Id$. In particular $\varphi$ can be written in one of the forms described in proposition \[simples\].
The group $G_{{\varphi }}$ is polycyclic, so the Bass-Serre tree $T$ associated to ${\Lambda }$ is a line. The set of elliptic elements of ${{\mathbb Z }^n}$ in ${\Lambda }$ is the kernel of a non trivial homomorphism from ${\mathbb Z }^n$ onto ${\mathbb Z }$. It is therefore a hyperplane ${H }$ in ${{\mathbb Z }^n}$ not containing ${x }$. Elements of $H$ act as the identity on $T$. Moreover ${\varphi }({H }) = t{H } t^{-1}$ is also an elliptic subgroup of ${{\mathbb Z }^n}$ in ${\Lambda }$. We obtain the inclusion ${\varphi }({H }) \subset {H }$, and so ${\varphi }$ stabilizes ${H }$.
If ${\varphi }_{|{H }}=Id$, then the lemma holds. Let us assume ${\varphi }_{|{H }}\neq Id$ and let us show that ${\varphi }_{|{H }}$ has finite order and that ${\varphi }(x)x^{-1}$ belongs to ${H }$.
Write $G_{\varphi } = {{\mathbb Z }^n} \rtimes_{\varphi } \langle t\rangle$. If the stable letter $t$ is elliptic in $\Lambda$, then it commutes with ${H }$ and so ${\varphi }_{|{H }}=Id$. The letter $t$ must therefore be hyperbolic. There exist two non-zero integers $h$ and $k$ such that ${x }^ht^k$ is elliptic. Then ${x }^ht^k$ must commute with ${H }$, but ${x }$ also commutes with ${H }$, so $t^k$ commutes with ${H }$. One has ${\varphi }^k_{|{H }} = Id_{H }$.
There exists an element ${g } \in {H }$ for which $t{x } t^{-1}={\varphi }({x }) = {x }^{p}{g }$ with $p = \pm 1$. But $t$ and $x$ act by translation implying that $p=+1$ and that ${\varphi }(x)x^{-1}$ belongs to ${H }$.
By lemma \[1-1\], when $\varphi$ is not as in 1. and 2. the vertex group of $G_\varphi$ (seen as an HNN extension) is universally elliptic. By [@GL3a lemma 4.6], the JSJ decomposition of $G_\varphi$ is the HNN extension ${\mathbb Z }^n\ast_\varphi$.
Every other group $G_\varphi$ has at least one other action:
\[flexible\] Let $({x },h_1,\dots,h_{n-1})$ be a basis of ${\mathbb Z }^n$ in which $\varphi$ can be written as in 1. or 2. of proposition \[simples\]. Then the semi-direct product $G_{\varphi } = \mathbb Z^n \rtimes_{\varphi } \langle t\rangle $ can be decomposed as a graph of groups in which ${H }=\langle h_1,\dots, h_{n-1}\rangle$ is elliptic and ${x } h$ is hyperbolic for all $h$ in $H$.
The second case has already been done (see proposition \[TFAE\] and corollary \[2-2\]). It remains the first case.
We produce another decomposition of $G_{{\varphi }}$ as a semi-direct product. Call $k$ the order of $M$.
Define $f: G_\varphi \rightarrow {\mathbb Z }$ by $f(t)=1$, $f(x)=-k$, and $f(g)=0$ for all $g\in H$. As $tHt^{-1}=H$ and $txt^{-1}=xh$ with $h$ in $H$, this is well defined.
Let $G_\varphi$ act on the line by translations via $f$. The kernel of the action is generated by $H$ and $xt^k$ which commute. It is therefore an abelian group ${\mathbb Z }^n$.
If $\varphi$ can be written in one of the following ways
1. $\left(
\begin{array}{cc}
1 & 0 \\
p & M
\end{array}
\right)$ with $M$ an $(n-1)\times(n-1)$ matrix of finite order and $p$ in ${\mathbb Z }^{n-1}$,
2. $\left(
\begin{array}{cc}
-1 & 0 \\
p & Id_{n-1}
\end{array}
\right)$ with $p$ in ${\mathbb Z }^{n-1}$
then the group $G_{\varphi}$ has trivial JSJ decomposition.
Let $x$, $H$ and $t$ be as in lemma \[flexible\]. Following lemma \[flexible\], no element of ${x }{H } $ is universally elliptic. In every JSJ decomposition each of these elements is hyperbolic or fixes a unique vertex of the Bass-Serre tree $T$ of this decomposition.
Assume every element of ${x } {H } $ is elliptic. Since ${E}=\langle {x }, {H } \rangle$ is abelian, they all fix the same vertex. So the group ${E}$ is elliptic and fixes a unique vertex $v\in T$. But ${E} $ is normal, so the decomposition is trivial.
Let us now assume that there exists a hyperbolic element ${y } \in {x } {H }$. Either $t$ is elliptic and $t^2$ fixes the whole tree, or $t$ is hyperbolic and acts by a translation. In particular, there exist $p$ in ${\mathbb Z }$ and $q$ in ${\mathbb Z }^*$ such that ${y } ^pt^q$ is elliptic and belongs to an edge stabilizer. This element, hyperbolic in the initial graph of groups, is not universally elliptic, which is a contradiction.
Interactions between subgraphs {#sectioninteraction}
==============================
For $e$ an edge of a graph of group $\Gamma$, call $\Gamma_e$ the subgraph consisting of the single edge $e$ (and its vertices) and $\Pi_e$ its fundamental group.
Until now, we have shown that the non-universally elliptic edges must be $1-1$ loops or $2-2$ edges (Corollary \[rigidite1\]). In this section, we characterize the non-universally elliptic edges $e$. We show that necessarily $\Pi_e$ has trivial JSJ decomposition, and that adjacent edge groups are included in a specific hyperplane of $G_e$.
We first prove that two adjacent edges cannot be both non-universally elliptic.
\[adjacent\]
Let $G=\pi_1(\Gamma)$ be a [$vGBS{}$]{} group and $T$ its Bass-Serre tree. Let ${v }$ be a vertex of $\Gamma$. Let ${e }$ and ${f }$ be loops of type $1-1$ or edges of type $2-2$, distinct and adjacent to ${v }$. Then $G_{ {e }}$ and $G_{ {f }}$ are universally elliptic.
Let $\tilde {e }$, $\tilde {f }$ and $\tilde {v }$ be preimages in $T$ of ${e }$, ${f }$ and ${v }$ respectively such that $\tilde {e }$ and $\tilde {f }$ have $\tilde {v }$ as common vertex. We call $\tilde {w_1 }$ and $\tilde {w_2 }$ the second vertices of $\tilde {e }$ and $\tilde {f }$ respectively. There are three cases to handle.
1. The edges $\tilde {e }$ and $\tilde {f }$ are of type $1-1$.
Let $t$ and $t'$ such that $t\cdot \tilde {v }= \tilde {w_1 }$ and $t'\cdot \tilde {v }= \tilde {w_2 }$ Take ${y }\in G_{\tilde {v }}=G_{\tilde {e }}=G_{\tilde {f }}$ and suppose ${y }$ is not universally elliptic. Let $\Gamma'$ be a graph of groups in which ${y }$ is hyperbolic. As ${y }$, $t{y } t^{-1}$ and $t'{y } t'^{-1}$ commute, the elements $t$ and $t'$ must stabilize the axis of ${y }$. So there exist $p$, $q$, $r$ and $s$ integers such that $q\neq 0$, $s\neq 0$, and ${y }^pt^q$ and ${y }^rt'^s$ are elliptic with characteristic spaces containing the axis of ${y }$. So ${y }^pt^q$ and ${y }^rt'^s$ must commute. Yet their projections in the (topological) fundamental group of the graph $\Gamma$ generate a free group of rank $2$, which is a contradiction.
2. The edges $ \tilde {e }$ and $\tilde {f }$ are of type $2-2$.
We show that every ${y } \in G_{\tilde {e }}$ is universally elliptic. Replacing ${y }$ by ${y }^2$ we may assume that ${y } \in G_{\tilde {f }}$. Let us fix $x_1\in G_{\tilde {w_1 }}\setminus G_{\tilde {e }}$, $x_2 \in G_{\tilde {w_2 }}\setminus G_{\tilde {f }}$ and $z\in G_{\tilde {v }}\setminus (G_{\tilde {e }}\cup G_{\tilde {f }})$. So $x_1z$ and $x_2z$ are in the centralizer of ${y }$, since $x_1$, $x_2$ and $z$ are in it. Yet those are hyperbolic elements with distinct axis crossing precisely in $\tilde {v }$, they generate a free group. Applying proposition \[centralisateur\] ${y }$ is universally elliptic.
3. The edge $\tilde {e }$ is of type $1-1$ and $\tilde {f }$ of type $2-2$.
Let $t$ be as in the first case. Let us fix $x\in G_{\tilde {w_2 }} \setminus G_{\tilde {f }}$ and $z \in G_{\tilde {v }} \setminus G_{\tilde {f }}$. In the same way as last case, the elements $xz$ and $txzt^{-1}$ are in the centralizer of $G_{\tilde {e }}^2$ (the set of squares of $G_{\tilde e}$) and generate a free group.
\[cas1\]
Let ${G }=\pi_1(\Gamma)$ be a [$vGBS{}$]{} group. Let ${e }$ be a $1-1$ loop based at a vertex ${v }$ and $\varphi\in Aut(G_v)$ the modulus of ${e }$. Let $\tilde G_{v } \subset G_{v }$ be the group generated by the groups of adjacent edges, ${e }$ excepted. Then $G_{e }$ is universally elliptic if and only if there is no decomposition $G_v=\langle x \rangle \times H$ such that
1. \[uns\] $\tilde G_{v } \subset H$,
2. \[bis\] $\varphi$ stabilize $H$
3. \[ter\]
- either $\varphi(x)x^{-1}\in H$ and $\varphi$ act on $H$ with finite order,
- or $\varphi(x)x\in H$ and $\varphi_{|H}=Id$.
Assume $G_{e }$ is not universally elliptic, and let $\Gamma'$ be a decomposition in which $G_{e}$ is not elliptic. From lemma \[1-1\], there exist a decomposition $G_v=\langle x\rangle \times H$ satisfying \[bis\] and \[ter\] such that $H$ is exactly the set of all elliptic elements of $G_v$ in the decomposition $\Gamma'$. By corollary \[rigidite1\] and lemma \[adjacent\] , the group $\tilde G_{v }$ is universally elliptic, so $\tilde G_{v }$ is included in $H$.
Conversely, suppose such a decomposition $\langle x\rangle\times H$ exists.
Let $t$ be a stable letter of $e$. Take $(h_1,\dots,h_{n-1})$ as basis of $H$, completing it by $x$ to make a basis $(x,h_1,\dots,h_{n-1})$ of $G_{v }$. Then $t$ act on $G_{v }$ by a linear map $
\left(
\begin{array}{cc}
\varepsilon & 0 \\
p & M
\end{array}
\right)$ with $\varepsilon=\pm1$ , $M$ a finite order matrix and $M=Id$ if $\varepsilon=-1$. We can apply lemma \[flexible\], and so there is a graph of groups $\Lambda$ of $\langle G_{v }, t\rangle$ in which $x$ is hyperbolic and $H$ is elliptic. Call $v'$ an vertex of $\Lambda$ with $H\subset G_{v'}$. We can construct a new graph of group decomposition $\Gamma'$ of $G$ as follows. The underlying graph is obtained by removing $e$ from $\Gamma$ and gluing $\Lambda$ by adding an edge between $v\in\Gamma$ and $v'\in\Lambda$. We define the vertex groups in the following way:
- for every vertex $w$ of $\Gamma'$ coming from a vertex of $\Gamma\setminus \{v\}$, we define $G_w(\Gamma')=G_w(\Gamma)$,
- for every edge $f$ of $\Gamma'$ coming from an edge of $\Gamma$ not adjacent to $v$, we define $G_f(\Gamma')=G_f(\Gamma)$ with the natural inclusions in the adjacent vertices,
- $G_v(\Gamma')=H$,
- for every edge $f$ of $\Gamma'$ coming from an edge of $\Gamma$ adjacent to $v$, we define $G_f(\Gamma')=G_f(\Gamma)$, with the inclusion $G_f(\Gamma') \hookrightarrow G_{v}(\Gamma')$ coming from the inclusions $G_f(\Gamma)\subset H$ coming from the assumption \[uns\],
- for every vertex $w$ of $\Gamma'$ coming from a vertex of $\Lambda$ including $v'$, we define $G_w(\Gamma')=G_w(\Lambda)$,
- for every edge $f$ of $\Gamma'$ coming from an edge of $\Lambda$, we define $G_f(\Gamma')=G_f(\Lambda)$ with the natural inclusions in the adjacent vertices,
- for $f$ the edge between $\Gamma$ and $\Lambda$, we define $G_f=H$, with natural inclusions $G_f=H\xrightarrow{Id}G_{v}(\Gamma')=H$ and $G_f\hookrightarrow G_{v'}(\Gamma')$ coming from $H\subset G_{v'}(\Lambda)$.
Using the isomorphism $G_{v'}\simeq G_{v'}\ast_H H$, we easily check that $\pi_1(\Gamma')=G$ In particular $\Gamma'$ is a decomposition of $G$ in which $x$ is hyperbolic. The element $x$ is not universally elliptic.
The case of a type $2-2$ edge is similar.
\[cas2\]
Let $G = \pi_1(\Gamma)$ be a [$vGBS{}$]{} group. Let ${e }$ be an edge of type $2-2$ with vertices ${v }$ and ${v' }$. We identify $G_e$ with its images into $G_v$ and $G_{v'}$. Then $G_{e }$ is universally elliptic if and only if there is no hyperplane $H$ of $G_{e }$ such that $H$ is also a hyperplane of $G_{v }$ and $G_{{v' }}$, and $H$ contains all groups of adjacent edges, $G_{e }$ excepted.
If $G_e$ is not universally elliptic, a decomposition $\Gamma'$ in which $G_e$ is not elliptic may be obtain from $\Gamma$ by replacing $e$ by a $1-1$ loop whose stabilizer is not universally elliptic (see figure \[decomp\]).
Let $G_1$ and $G_2$ be two [$vGBS{}$]{} groups with defining graphs $\Gamma_1$ and $\Gamma_2$ and $v\in \Gamma_1$, $v'\in \Gamma_2$ two vertices with groups ${\mathbb Z }^n$. We construct a new [$vGBS{}$]{} group with graph of groups $\Gamma$ as the union of $\Gamma_1$ and $\Gamma_2$ and a new edge $e$ between $v$ and $v'$. We define $G_{v}(\Gamma)=G_{v}(\Gamma_1)\times \langle a \rangle$ and $G_{v'}(\Gamma)=G_{v}(\Gamma_2)\times \langle b \rangle$. We then define the edge group $G_e$ to be ${\mathbb Z }^n\times \langle c\rangle$ identifying $c$ with $a^2$ in $G_{v}(\Gamma)$ and with $b^2$ in $G_{v'}(\Gamma)$ and the $Z^n$ part with $G_v(\Gamma_1)$ and $G_{v'}(\Gamma_2)$.
The group $\pi_1(\Gamma)$ has an other decomposition $\Gamma'$, with underlying graph obtained gluing $\Gamma_1$ and $\Gamma_2$ by identifying $v$ and $v'$ together, and adding a loop $l$ over the new vertex. Define $G_v(\Gamma')={\mathbb Z }^n\times\langle d \rangle$. The loop carries the HNN extension with sable letter $t$ define by $tdt^{-1}=d^{-1}$ and $tht^{-1}=h$ for all $h$ in ${\mathbb Z }^n$. To obtain the isomorphism between $\Pi_l$ and $\Pi_e$, it suffices to identify $t$ with $a$, $d$ with $ab^{-1}$, and the ${\mathbb Z }^n$ parts together. In $\Gamma$ the group $\tilde K=\Pi_e=G_{v}(\Gamma)\ast_{G_e} G_{v'}(\Gamma)$ is a extended Klein bottle group. The ${\mathbb Z }^n$ part of $G_{e}(\Gamma)$ plays the role of $H$ in proposition \[cas2\].
(0,-1.62)(10.865468,1.62) (8.78,-0.2)(8.58,1.2)(9.38,1.6)(9.58,1.4)(9.78,1.2)(9.78,0.6)(8.78,-0.2) (4.255,0.115)[$v'$]{} (5.695,-0.285)[$G'$]{} (4.0,-0.2)(5.8,0.2) (4.0,-0.2)(5.2,-1.6) (4.0,-0.2)(5.6,-0.8) (1.8,-0.2)(4.0,-0.2) (3.94,-0.2) (8.465,-0.885)[$v=v'$]{} (10.474999,-0.285)[$G'$]{} (8.78,-0.2)(10.58,0.2) (8.78,-0.2)(9.98,-1.6) (8.78,-0.2)(10.38,-0.8) (7.2749996,-0.285)[$G$]{} (8.78,-0.2)(7.58,0.8) (8.78,-0.2)(6.98,-1.0) (8.78,-0.2) (2.055,0.115)[$v$]{} (0.89500004,-0.085)[$G$]{} (1.8,-0.2)(0.6,0.8) (1.8,-0.2)(0.0,-1.0) (1.8,-0.2)
\
We call $\tilde K$ the group $G_{v }\ast_{G_{e }} G_{{v' }}$. By proposition \[TFAE\] $\tilde K$ is an extended Klein bottle group, so we have $\tilde K=\langle \lambda,\mu\ |\ \mathcal R\rangle \times L$ with $L$ a free abelian subgroup, $\lambda \in G_v \setminus G_e$, $\mu \in G_{v'} \setminus G_e$ and $\mathcal R$ the relators of the -twisted or not- Klein bottle group. We have $G_{e }=\langle \lambda^2,\mu^2\rangle \times L$ with $\lambda^2=\mu^2$ in the case of the untwisted Klein bottle group.
First, assume there exists a hyperplane $H$ of $G_{e }$ which is also a hyperplane of $G_{v }$ and $G_{{v' }}$ and contains all groups of adjacent edges, $G_e$ excepted. Let us prove $G_e$ is not universally elliptic.
We can fix $x_{v }$ and $x_{{v' }}$ two elements such that $G_{v } = \langle x_{v }, H \rangle$ and $G_{{v' }}= \langle x_{{v' }}, H \rangle$. As $G_{e }$ has index 2 in both $G_{v }$ and $G_{{v' }}$, and $H$ is included in $G_{e }$, we must have $G_{e } = \langle x^2_{v }, H \rangle = \langle x^2_{{v' }}, H \rangle$. So there exists $h\in H$ such that $\tilde K=\langle x_{v },x_{{v' }}, H | x^2_{v } = x^{\pm2}_{{v' }}h, [H,x_{v }]=[H,x_{v' }]=1\rangle$. Up to taking the inverse of $x_{v'}$, one may assume $x^2_{v } = x^{2}_{{v' }}h$.
If $h$ is a square of an element $h'$, replacing $x_{{v' }}$ by $x_{{v' }}h'$, the group $\tilde K$ may have the presentation $\tilde K=\langle x_{v }, x_{{v' }} | x_{v }^2=x_{{v' }}^2\rangle \times H$. By lemma \[2-2\], the group $\tilde K$ admit a decomposition in semi-direct product ([*i*. e.]{} a graph of groups with one vertex and one edge) in which $H$ and $x_{v }x_{{v' }}^{-1}$ are elliptic and $x_{v }$ is hyperbolic.
By collapsing $e$, we obtain a new vertex $v'$ carrying the group $\tilde K$. Since the groups of edges adjacent to $G_{v }$ et $G_{v' }$ are included in $H$, we can expand $v'$ in a $1-1$ loop with $H$ and $x_{v }x_{{v' }}^{-1}$ elliptic and $x_{v }$ hyperbolic. So the subgroup $G_{e }$ is not universally elliptic, and the edge ${e }$ may be replace by a loop of type $1-1$ which is not universally elliptic.
If $h$ is not a square, up to modifying $h$ and $x_{v'}$ by a square of $H$ we can assume $h$ is primitive. Let $H'$ such that $H=\langle h \rangle\times H'$. By the Tietze transformation consisting of replacing $h$ by $x_{v'}^{-2}x_v^2$, the relations of commutation $x_vh=hx_v$ and $x_{v'}h=hx_{v'}$ become $x_vx_{{v' }}^2=x_{{v' }}^2x_v$ and $x_{v }^2x_{{v' }}=x_{{v' }}x_{v }^2$. So the group $\tilde K$ admits a presentation $\tilde K=\langle x_{v }, x_{{v' }} | x_{v }^2x_{{v' }}=x_{{v' }}x_{v }^2, x_vx_{{v' }}^2=x_{{v' }}^2x_v\rangle \times H'$. The lemma \[2-2\] assure us to have a decomposition of $\tilde K$ in graph of groups with one vertex and one loop in which $H'$, $x_{v }^2x_{{v' }}^{-2}=h$ and $x_{v }x_{{v' }}^{-1}$ are elliptic and $x_{v }$ is hyperbolic. Then $H=\langle h, H'\rangle$ is elliptic, the same argument permits to extend the construction to the whole group.
In both cases, the $2-2$ edge can be removed and replaced by a $1-1$ loop with the element $x_vx_{v'}^{-1}$ belonging to the edge group. This edge cannot be universally elliptic.
Conversely, if ${e }$ is not universally elliptic, let ${\alpha } \in G_{e }$ be a non universally elliptic element and $\Gamma'$ a graph of groups in which ${\alpha }$ is hyperbolic with axis $\mathcal A$. The set of the elliptics of $G_{e }$ in $\Gamma'$ is exactly a hyperplane $H$. It remains to show that $H$ has the properties we want.
We first claim that $H$ is exactly the set of elements of $\tilde K$ elliptic in both $\Gamma$ and $\Gamma'$.
Let ${\beta }$ be an element of $G_{e }$, as $\lambda$ belongs to $G_v\setminus G_e$ and $\mu$ to $G_{v'}\setminus G_e$, the elements $\lambda {\beta }$ and $\mu\lambda {\beta }\mu^{-1}$ do not commute. However they both commute with $\alpha$ so their characteristic spaces contain the axis $\mathcal A$ in the Bass-Serre tree of $\Gamma'$. Necessarily $\lambda {\beta }$ is hyperbolic: since the edge groups are abelian, if $\lambda {\beta }$ is elliptic in $\Gamma'$, the elements $\lambda {\beta }$ and $\mu\lambda {\beta }\mu^{-1}$ should commute. And so each element of $\lambda G_{e }$ is hyperbolic. The same argument works to show that every element in $\mu G_{e }$ is hyperbolic in $\Gamma'$.
Yet every elliptic element of $\tilde K$ for the decomposition $\Gamma$ is conjugated to an element of $G_{v }$ or $G_{{v' }}$, and so is conjugate to an element of $$G_{e } \cup \lambda G_{e } \cup \mu G_{e }.$$ The elements of $\tilde K$ elliptic in both $\Gamma$ and $\Gamma'$ belongs to $G_e$, hence to $H$. This prove the claim.
By the claim $H$ contains every elliptic element of $G_{v }$ and $G_{{v' }}$ for the decomposition $\Gamma'$. As $H$ is a subgroup of rank $n-1$ in both $G_{v }$ and $G_{{v' }}$, it is a hyperplane in both $G_{v }$ and $G_{{v' }}$. Moreover, by corollary \[rigidite1\] and lemma \[adjacent\], the group of every edge adjacent to ${e }$ is universally elliptic. So each of these edge groups must be include in $H$.
{#finalsection}
Let $G=\pi_1(\Gamma)$ be a [$vGBS{}$]{} group and ${v }$ a vertex of $\Gamma$. Let $\tilde G_{v }\subset G_v$ the group generated by the groups of edges adjacent to ${v }$. If $G_{v } / \tilde G_{v }$ is infinite, then there exists a hyperplane $H$ of $G_{v }$ containing $\tilde G_{v }$. Then one can replace the vertex ${v }$ by an HNN extension with vertex and edge groups equal to $H$ and the modulus of the edge equal to the identity. Call ${v' }$ and ${e }$ the new vertex and the new edge. We have the equality $G_{{v' }}=G_{{e }}=H$, and so ${e }$ is of type $1-1$. We call this construction the *expansion* of $G_{v }$ over $H$ (see figure \[expansion\]).
(1,-0.8)(11.507187,1.8) (1.761875,0.0)(4.161875,0.0) (7.561875,0.0)(9.761875,0.0) (10.561875,0.0)(0.8,0.8) (1.8354688,0.515)[$\langle a, b \rangle$]{} (4.2254686,0.515)[$\langle c, d \rangle$]{} (4.2254686,-0.285)[$v$]{} (7.4354687,0.515)[$\langle a, b \rangle$]{} (9.395469,0.515)[$\langle c \rangle$]{} (9.595469,-0.285)[$v'$]{} (11.695469,-0.085)[$\langle c \rangle$]{} (11.2,0.715)[$e$]{} (3.1454687,-0.485)[$\langle b\rangle =\langle c\rangle$]{} (8.545468,-0.285)[$\langle b\rangle =\langle c\rangle$]{}
\
\[stable\]
Let $G=\pi_1(\Gamma)$ be a [$vGBS{}$]{} group. Let ${v }$ be vertex, if $G_{v }/{\tilde G_{v }}$ is virtually cyclic, then there is a unique expansion of $G_{v }$ and it creates a universally elliptic edge.
Let $H$ be the unique hyperplane of $G_{v }$ containing $\tilde G_{v }$. Let $\Gamma'$ be the graph of groups obtained by expansion of $G_{v }$ over $H$. From lemma \[adjacent\], the adjacent edges are universally elliptic. Moreover, the group $\tilde G_v$ generated by all those groups has finite index in $G_{{v' }}=H=G_e$, so the edge ${e }$ is universally elliptic.
\[ecrase\]
Let $G=\pi_1(\Gamma)$ be a [$vGBS{}$]{} group. Let ${e }$ be a non-universally elliptic edge, and ${v }$ the vertex obtained by collapsing ${e }$. Let $\tilde G_{v }\subset G_v$ be the group generated by groups of edges adjacent to ${v }$. Denote $\langle\langle \tilde G_{v } \rangle\rangle$ the normal closure of $\tilde G_{v }$ in $G_{v }$. Then $G_{v }/\langle\langle \tilde G_{v } \rangle\rangle$ is not virtually cyclic.
By proposition \[cas2\], we may assume that ${e }$ is a loop of type $1-1$. The group $G_{v }$ is a semi-direct product ${\mathbb Z }^n\rtimes_\varphi {\mathbb Z }$. By proposition \[cas1\], the subgroup $\tilde G_{v }$ is included in a hyperplane $H$ of ${\mathbb Z }^n$ stabilized under the action of $\varphi$. In particular $H$ is normal in $G_v$ and $\langle\langle \tilde G_{v } \rangle\rangle \subset H$. So there exists a projection from $G_{v }/\langle\langle \tilde G_{v } \rangle\rangle$ onto $G_{v }/H \simeq {\mathbb Z }\rtimes {\mathbb Z }$ which is virtually ${\mathbb Z }^2$. The group $G_{v }/\langle\langle \tilde G_{v } \rangle\rangle$ is not virtually cyclic.
\[final\]
Let $G=\pi_1(\Gamma)$ a [$vGBS{}$]{} group. For ${v }$ a vertex, let $\tilde G_{v }$ be the group generated by groups of edges adjacent to ${v }$. A JSJ decomposition of $G$ can be obtained from $\Gamma$ by collapsing the edges carrying a non-universally elliptic group and expanding the groups $G_{v }$ such that $G_{v } / \tilde G_{v }$ is virtually cyclic.
By lemmas \[stable\] and \[ecrase\], the two operations of collapsing and expanding do not interact, so the construction is well define.
Let $\Lambda$ be the obtained graph of groups. According to lemmas \[stable\] and \[ecrase\], this graph does not contain neither vertex ${v }$ such that $G_{v } / \tilde G_{v }$ is virtually cyclic nor edges carrying a non-universally elliptic group. In particular $\Lambda$ is universally elliptic.
It remains to prove the maximality of $\Lambda$.
Let $\Lambda'$ be another decomposition in which all edge groups are universally elliptic. We have to show that every elliptic element in $\Lambda$ is elliptic in $\Lambda'$. Then by [@GL3a lemma 3.6] the decomposition $\Lambda$ is a JSJ decomposition of $G$.
Suppose there exists an element $x$, elliptic in $\Lambda$ and hyperbolic in $\Lambda'$. We will show there exists an edge group of $\Lambda'$ containing a non-universally elliptic element, a contradiction.
There are two cases to consider:
1. If $x$ belongs to a vertex group of $\Lambda$ carrying an abelian group.
Let $v$ be a vertex stabilized by $x$ in the Bass-Serre tree of $\Lambda$. Then no power of $x$ belongs to $\tilde G_{v }$ which is universally elliptic. As $G_{v }$ is abelian but not elliptic in $\Lambda'$, it acts non trivially on the axis of $x$ via an map $\varphi:~G_{v } \rightarrow {\mathbb Z }$ with $\tilde G_{v } \subset \ker \varphi$.
We shall find a decomposition $G_v=\langle y \rangle\times H$ such that $\tilde G_{v } \subset H$ and $y\in\ker \varphi$. Then expanding $G_{v }$ over $H$ in $\Lambda$, the element $y$ will be hyperbolic, contradicting the universal ellipticity of $\Lambda'$.
Let $\bar G_{v }$ be the set of elements of $\ker \varphi$ with a power in $\tilde G_{v }$. As $G_{v } / \tilde G_{v }$ is infinite and not virtually cyclic, the set $\ker \varphi \setminus \bar G_{v }$ is non-empty. Take $y$ a primitive element in $\ker \varphi \setminus \bar G_{v }$. Taking any hyperplane $H$ of $G_v$ containing $\tilde G_v$ such that $G_v=\langle y \rangle\times H$, we obtain the decomposition.
2. If $x$ belongs to a vertex group of $\Lambda$ carrying a non-abelian group.
From proposition \[cas2\], we may assume that all non-universally elliptic edges of $\Gamma$ are loops with type $1-1$. Then $x$ belongs to a group $\Pi_l$ for $l$ a non-universally elliptic loop of $\Gamma$. Call $w$ the vertex and $t$ a stable letter of $l$. Let $h$ be an element of $G_{w }$ non universally elliptic. As $\Pi_l$ is polycyclic, it acts on the axis of $x$ in $\Lambda'$. If $h$ is elliptic in $\Lambda'$, then $h^2$ fixes the axis and so $\Lambda'$ is not universally elliptic. If $h$ is hyperbolic , there exists two integers $p$ and $q\neq 0$ such that $h'=h^pt^q$ is elliptic in $\Lambda'$ (the integer $p$ may be equal to $0$ if $t$ is already elliptic in $\Lambda'$). Up to take the square of $h'$, we may assume it fixes the axis pointwise. As $q$ is different from 0, the element $h'$ is hyperbolic in $\Gamma$. There exists an edge of $\Lambda'$ which is not universally elliptic.
Let $G=\pi_1(\Gamma)$ a [$vGBS{}$]{} group, and $n\in {\mathbb N }$. Suppose that $\Gamma$ is a JSJ decomposition over free abelian groups. Then a JSJ decomposition over free abelian groups of rank $\leq n$ may be obtained by collapsing every edge of $\Gamma$ with group of rank $> n$.
We will use the notation $n$-JSJ decomposition and $n$-universally elliptic for decomposition over group of rank $\leq n$ and $\infty$-JSJ decompositions and $\infty$-universally elliptic for decomposition over all free abelian groups.
By [@GL3a Proposition 7.1], we know that an $n$-JSJ decomposition may be obtained by refining an $\infty$-JSJ decomposition, and then collapsing all edges carrying a group of rank $> n$. So it suffices to show that if a vertex of $\Gamma$ can be decomposed over a group of rank $\leq n$, then this decomposition is not $n$-universally elliptic.
Let $v$ be a vertex of $\Gamma$. Its vertex group is a semi-direct product $G_v={\mathbb Z }^p\rtimes {\mathbb Z }$ (the product may be direct). The only non-trivial reduced decompositions of $G_v$ are as $1-1$ loop or a $2-2$ edge with edge group ${\mathbb Z }^p$.
Call $\Lambda$ a decomposition of $G$ refining $\Gamma$ and an $n$-JSJ decomposition $\Delta$, and $\Lambda_v$ the subgraph of $\Lambda$ whose projection on $\Gamma$ is $v$. Assume $\Lambda_v$ has a minimal number of edges over decomposition refining $\Gamma$ and $\Delta$. If $\Lambda_v$ contains a non-reduced edge then after collapsing all edges with groups of rank $>n$ in $\Lambda$ the edge stays unreduced (if an edge having same group as one of its endpoint $v$ is not collapsed then $rank(G_v)\leq n$ and no edge adjacent to $v$ is collapsed). As $\Lambda_v$ has a minimal number of edges, it is reduced.
If $\Lambda_v$ is not a trivial decomposition of $G_v$, as $G_v$ cannot be decomposed over groups of rank $>p$, if $\Lambda$ is $n$-universally elliptic then $\Lambda$ is $\infty$-universally elliptic. This is a contradiction with the fact that $\Gamma$ is a $\infty$-JSJ decomposition.
Contruction of the JSJ decomposition {#algo}
====================================
We now describe an algorithm which gives the abelian JSJ decomposition of a [$vGBS{}$]{} group.
We assume that a [$vGBS{}$]{} group is given under the form of a [$vGBS{}$]{} decomposition, that is, given by the description of the vertex groups, edge groups, and the inclusion maps of the edge groups into the vertex groups. Given an edge $e$ and an endpoint $v$ of $e$, we may decide if $G_e$ is of finite index in $G_v$ and if so, compute this index. In particular, we may detect $1-1$ loops and $2-2$ edges.
\[thalgo\] The construction of a JSJ decomposition of a [$vGBS{}$]{} group is algorithmic.
Let $\Gamma$ be a [$vGBS{}$]{} decomposition and $l$ be a $1-1$ loop with base point $v$ and modulus $\varphi$ of finite order which is not conjugate to $\left(
\begin{array}{cc}
-1 & 0 \\
p & Id_{n-1}
\end{array}
\right)$ with $p\in{\mathbb Z }^{n-1}$. Call $n$ the rank of $G_v$. Call $\tilde G_v$ the subgroup of $G_v$ generated by groups of edges adjacent to $v$ excepting $l$. Call $E(\varphi)\subset G_v\otimes \mathbb C$ the subspace generated by all eigenspaces of $\varphi$ associated to eigenvalues different from $1$. As $\varphi$ is of finite order, thus diagonalizable, the dimension of $E(\varphi)$ is equal to $n$ minus the dimension of $E_1(\varphi)$, the eigenspace associated to $1$. Define $\bar G_v$ as the smallest $\varphi$-invariant subgroup of $G_v$ containing $\tilde G_v$ and $E(\varphi)\cap G_v$.
\[casparticulier\] With the previous notations, the loop $l$ is universally elliptic if and only if $\bar G_v$ is of rank $n$.
First assume $l$ is not universally elliptic. By proposition \[cas1\], as $\varphi$ is not conjugate to $\left(
\begin{array}{cc}
-1 & 0 \\
p & Id
\end{array}
\right)$, in a well-chosen basis $(x,h_1\dots,h_{n-1})$ the modulus $\varphi$ is of the form $\left(
\begin{array}{cc}
1 & 0 \\
p & M
\end{array}
\right)$ with $M$ of finite order and $\tilde G_v$ contained in $H=\langle h_1\dots,h_{n-1}\rangle$. Moreover, as the action of $\varphi$ on $G_v/H$ is trivial, every eigenvector in $G_v\otimes \mathbb C$ with eigenvalue $\neq 1$ must belong to $H\otimes \mathbb C$. Thus the rank of $\bar G_v$ must be at most $n-1$.
Conversely, if the rank of $\bar G_v$ is $<n$, then we may construct a group $L$ of rank $n-1$, containing $\bar G_v$ and stable under the action of $\varphi$, by adding to $\bar G_v$ eigenvectors associated to $1$. This is possible since $\dim(E(\varphi)\oplus E_1(\varphi))=n$. Call $H$ the hyperplane containing $L$, and $x\in G_v$ such that $G_v=\langle x\rangle\times H$. Then the matrix of $\varphi$ in a basis $(x, h_1,\dots, h_{n-1})$ where $H=\langle h_1,\dots, h_{n-1}\rangle$, is of the form $\left(
\begin{array}{cc}
1 & 0 \\
p & M
\end{array}\right)$. By proposition \[cas1\], the loop $l$ is not universally elliptic.
We may now describe the algorithm.
By theorem \[theoremintro\], there are three different algorithms to construct. The first blows up certain vertices into a loop, and the other two decide when a $1-1$ loop and a $2-2$ edge is universally elliptic. Recall that we must collapse non universally elliptic edges.
- The first algorithm is very simple. Given a vertex $v$, it computes the group $\tilde G_v$ generated by adjacent edge groups. If $rank(\tilde G_v)\neq rank(G_v)-1$, then it leaves $v$ unchanged. If the rank of $\tilde G_v$ is equal to the rank of $G_v$ minus one, it finds a primitive element $a$ such that no power is in $\tilde G_v$ (the algorithm of the Smith normal form works for example), and changes $v$ to a new vertex $v'$ and a loop $l$ with $G_l=G_{v'}$ equal to the set of elements of $G_v$ with a power in $\tilde G_v$, and $a$ the stable letter.
- The second algorithm has to decide whether a $2-2$ edge $e$ is universally elliptic or not.
Call $v$ and $v'$ the two endpoints of $e$. By proposition \[cas2\], if the adjacent edge groups are not all contained in $G_e$, then $e$ is universally elliptic. Otherwise, call $\tilde G_e$ the subgroup of $G_e$ generated by all these groups, and $\bar G_e$ the set of elements of $G_e$ with a power in $\tilde G_e$. Then $e$ is universally elliptic if and only if $G_v/\bar G_e$ or $G_{v'}/\bar G_e$ has torsion. This is decidable by looking at the Smith normal form of $\bar G_e$ in $G_v$ and $G_{v'}$.
- The third algorithm decides whether a $1-1$ loop $l$ is universally elliptic or not.
Call $v$ the base point of $l$, call $\varphi$ the modulus of $l$ and $n$ the rank of $G_v$. One first computes the group $\tilde G_v$ generated by adjacent edge groups, $G_l$ excepted. If $l$ is not universally elliptic, necessarily by proposition \[cas1\] the modulus $\varphi$ acts on a hyperplane of $G_v$ with finite order. As we have a minoration for Euler’s phi function $\phi$ given by $\phi(p)\geq \sqrt{\frac{p}{2}}$, the modulus $\varphi$ acts on a hyperplane with finite order if and only if $\varphi^{(2m^2)!}$ acts trivially on a hyperplane. We therefore consider the eigenspace $E$ of $\varphi^{(2m^2)!}$ associated to the eigenvalue $1$.
There are three cases.
- either $E$ has rank less than $n-1$,
- or $E$ has rank $n-1$.
- or $E=G_v$.
In the first case, the loop $l$ is universally elliptic.
In the second case $E$ is a $\varphi$-invariant hyperplane. As the only hyperplane on which $\varphi$ acts with finite order is $E$, it suffices to check if $\tilde G_v$ is included in $E$, and if $\varphi$ acts trivially on $G_v/E$. By proposition \[cas1\], the loop $l$ is not universally elliptic if and only if those two properties hold.
In the third case, we have to decide if there exists a hyperplane $H$ containing $\tilde G_v$ with properties as in proposition \[cas1\].
We first decide if $\varphi$ is conjugate to $\left(
\begin{array}{cc}
-1 & 0 \\
p & Id
\end{array}
\right)$ for some $p$ in ${\mathbb Z }^{n-1}$. This is decidable because matrices conjugate to a matrix of the previous form are exactly the ones with determinant $-1$ and an eigenspace associated to eigenvalue $1$ of dimension $n-1$.
If $\varphi$ is conjugate to $\left(
\begin{array}{cc}
-1 & 0 \\
p & Id
\end{array}
\right)$, then $l$ is not universally elliptic if and only if $\varphi$ act trivially on $\tilde G_v$.
If $\varphi$ is not conjugate to $\left(
\begin{array}{cc}
-1 & 0 \\
p & Id
\end{array}
\right)$, call $P(X)$ the polynomial $\frac{X^{(2m^2)!}-1}{X-1}$, and compute the subgroup $F=\langle \ker(P(\varphi)),\tilde G_v\rangle$. As $rank(\varphi^k(F))$ is first strictly increasing, and then stationary, the group $M=\varphi^n(F)$ is the smallest $\varphi$-invariant subgroup containing $F$. We compute $m$ the rank of $M$. By lemma \[casparticulier\], the edge $l$ is universally elliptic if and only if $m=n$.
<span style="font-variant:small-caps;">Benjamin Beeker</span>\
LMNO,\
Université de Caen BP 5186\
F 14032 Caen Cedex\
France\
[[email protected]]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We study the behaviour of a natural measure defined on the leaves of the genealogical tree of some branching processes, namely self-similar growth-fragmentation processes. Each particle, or cell, is attributed a positive mass that evolves in continuous time according to a positive self-similar Markov process and gives birth to children at negative jumps events. We are interested in the asymptotics of the mass of the ball centered at the root, as its radius decreases to $0$. We obtain the almost sure behaviour of this mass when the Eve cell starts with a strictly positive size. This differs from the situation where the Eve cell grows indefinitely from size 0. In this case, we show that, when properly rescaled, the mass of the ball converges in distribution towards a non-degenerate random variable. We then derive bounds describing the almost sure behaviour of the rescaled mass. Those results are applied to certain random surfaces, exploiting the connection between growth-fragmentations and random planar maps obtained in [@BBCK16]. This allows us to extend a result of Le Gall [@L17] on the volume of a free Brownian disk close to its boundary, to a larger family of stable disks. The upper bound of the mass of a typical ball in the Brownian map is refined, and we obtain a lower bound as well.
**Keywords:** Self-similar growth-fragmentations; intrinsic area; rate of growth.
**AMS MSC 2010:** 60J25; 60G18; 60G57.
author:
- 'François G. Ged[^1]'
bibliography:
- 'biblio.bib'
title: 'Intrinsic area near the origin for self-similar growth-fragmentations and related random surfaces'
---
Introduction
============
Growth-fragmentation processes form a family of continuous time branching processes that have been introduced by Bertoin [@B17]. They model particle systems without interaction where each particle is described by a positive real number that corresponds to its mass (or size), evolving by growing and splitting, with rates that can depend on its current mass. Note that these processes differ from (pure) fragmentation processes [@B01; @B04], for which growth is not allowed. The fragmentations are binary and when a particle splits, its mass is instantaneously randomly distributed among the two resulting fragments. A *self-similar growth-fragmentation* $(\mathbf{X}(t))_{t\geq 0}$ is a growth-fragmentation where every particle evolves according to a driving self-similar Markov process. Self-similarity refers to the property that the evolution of a particle of size $x>0$ is a scaling transformation of that of a particle of unit size, depending on some index $\alpha\in\mathbb{R}$. The law of $\mathbf{X}$ is characterized by $\alpha$ and its *cumulant function* $\kappa:\mathbb{R}{\left(-\infty\,,\infty\right]}$, both depending on the driving process.
In this paper we are interested in the cases where $\alpha<0$ and $\kappa$ has two positive root $\omega_-<\omega_+$. Under some further conditions on $\alpha$ and $\kappa$, the growth-fragmentation yields a natural measure on the genealogical tree of the branching process seen as a metric space, namely the *intrinsic area measure*. Denoting $A(t)$ the intrinsic area of the ball of radius $t$ centered at the root of the tree, one can investigate the regularity of the Stieltjes measure $\mathrm{d}A(t)$. This has been studied in [@G18], in which Theorem 1 shows that if $\alpha>-\omega_-$, then $\mathrm{d}A(t)$ is absolutely continuous whereas it is singular when $\alpha\leq -\omega_-$. In this paper, we consider the absolutely continuous case. A noteworthy fact is that the density is null at $t=0$, meaning that the dissipation of the area occurs as small particles are about to vanish. We study further properties of $A$.
Our first result is Theorem \[Theorem area GF\], that determines the almost sure behaviour of $A(\epsilon)$ as $\epsilon\to 0+$, when $\mathbf{X}$ starts from a single particle of positive size $x$. Namely, there exists a regularly varying function $f:\mathbb{R}_+\to\mathbb{R}_+$ explicitely given in terms of the characteristics of the driving process, such that $A(\epsilon)/f(\epsilon)\to 1$ as $\epsilon\to 0+$.
It is possible to tilt the law of $\mathbf{X}$ such that the Eve cell starts from size 0. Indeed, in [@BBCK16], the authors introduced a new probability measure, under which $\mathbf{X}$ initiates from a distinguished particle that grows indefinitely from the initial size 0, and evolves differently from the others which eventually die out. Let denote $A^+$ the analog of $A$ under this new measure. In this case, Theorem \[Theorem area GF\] does not apply. However, Proposition \[Theorem stationary area from 0\] shows that, $t\mapsto e^{\omega t/\alpha}A^+(e^t)$ defines a stationary process. In particular, $t^{\omega_-/\alpha}A^+(t)$ has a limit in distribution as $t\to 0+$, but has no almost sure limit. We prove in Propositions \[Theorem upper envelope A(t)\] and \[Proposition lower envelope A(t)\] that, almost surely, $A^+(t)$ deviates from $t^{\omega_-/|\alpha|}$ by at most a power of $|\log(t)|$.
One of the motivations of the present work is that these branching processes turn out to be geometrically connected to some random surfaces. The Brownian map is a random surface homeomorphic to the two-dimensional sphere that appears as the Gromov-Hausdorff scaling limit when $n\to\infty$ of uniformly distributed $q$-angulations with $n$ faces of the sphere [@L13; @M13]. Similarly, Brownian disks are random compact metric spaces homeomorphic to the unit disk of $\mathbb{R}^2$, obtained as scaling limits of random planar maps with a boundary [@BM17]. In [@CL14], the Brownian plane, locally isomorphic to the Brownian map, is obtained as scaling limits of the UIPQ and uniform quadrangulations. Moreover it can also be seen as the Gromov-Hausdorff tangent cone of the Brownian map at its root.
In [@BCK15], it was shown that the collection of perimeters of the holes observed when slicing Boltzmann triangulations at all heights converges, when properly rescaled, towards a particular self-similar growth-fragmentation. We also mention Theorems 3 and 23 of Le Gall and Riera [@LR18], which show that, when slicing directly the free Brownian disk, the holes’ perimeters are described by the same growth-fragmentation as in [@BBCK16] (see also [@MS15] Section 4). When $\mathbf{X}$ starts from a single cell of size 0 that grows indefinitely, the geometrical connection corresponds this time to the holes in a sliced discrete approximation of the Brownian plane.
There is actually a broader family of continuum random surfaces, with different scaling exponents, that arise as limiting objects of Boltzmann random planar maps. They are known as *stable maps* and were first obtained by Le Gall and Miermont [@LM11; @LM10 and references therein], by considering very specific distributions for the degree of a typical face, that have infinite variance (see also [@M18] for the same scaling limits under relaxed hypotheses). However, only the Brownian map has been characterized yet, in the sense that for other stable maps, the uniqueness of the scaling limit is not yet proven.
Similarly as in the Brownian case, the authors in [@BBCK16] were able to extend the geometrical connection previously mentioned, to holes’ perimeters in discrete approximations of stable disks and plane (in the so-called *dilute case*) and a specific family of self-similar growth-fragmentations.
There is a natural way of measuring the “size” of these stable surfaces, namely, the so-called *intrinsic volume measure*, which can be constructed as the scaling limit of the number of vertices (or faces) in the approximation by discrete random maps. This measure corresponds to the intrinsic area measure in the related growth-fragmentation; we shall rather use this name since we consider planar objects.
Besides being aesthetic, this connection has already been fruitful in both directions: for instance, it allowed the authors in [@BBCK16] to use results on discrete random planar maps to determine the law of the total intrinsic area of the related growth-fragmentations. On the other hand, it was known that the intrinsic area of the Brownian map cannot be derived as the length of the perimeters of the holes and the height, as for smooth surfaces, since the latter defines a measure that is not locally finite. From the analysis of growth-fragmentations, it has been shown in [@G18] that the intrinsic area of the Brownian map can be written as the integral against the Lebesgue measure of some function of the perimeters of the holes.
Our results in this paper apply to stable surfaces. In particular, Thanks to Theorem \[Theorem area GF\], we retrieve Theorem 3 of Le Gall [@L17], which shows that, in the case of the free Brownian disk with boundary length $x>0$, it holds that $\epsilon^{-2}A(\epsilon)\to x$ almost surely, as $\epsilon\to 0+$ (see ). More generally, we obtain the analogue for other stable disks, with different exponents.
Applying Proposition \[Theorem upper envelope A(t)\] improves the upper bound that was known for the almost sure behaviour of the area of the ball of radius $\epsilon$ around the root of the Brownian map (i.e. $A^+(\epsilon)$): the previous bound was $\epsilon^{4-\delta}$ for $\delta>0$ arbitrary small, we obtain $\epsilon^{4}|\log(\epsilon)|^{1+\delta}$. We moreover obtain the lower bound $\epsilon^4|\log(\epsilon)|^{-q}$ for any $q>6$, thanks to Proposition \[Proposition lower envelope A(t)\]. Again, the analogue holds for stable maps.
The organisation of the paper is the following. In Section 2, we formally introduce the growth-fragmentations setting. This includes definitions of positive self-similar Markov processes and Lamperti’s transformation, the intrinsic area measure of a growth-fragmentation, as well as the two spinal decompositions introduced in [@BBCK16] that are central throughout this work. We also establish an important Markov-branching property of $A$ in Lemma \[Lemma Markov-branching property of A\], which, roughly speaking, reduces the study of $A$ to that of a single self-similar Markov process.
We state our results in Section 3. Theorem \[Theorem area GF\] concerns the almost sure behaviour of $A(\epsilon)$ as $\epsilon\to 0+$, when $\mathbf{X}$ starts from a typical cell with positive initial size. Our second result is Proposition \[Theorem upper envelope A(t)\] which provides an upper bound for the almost sure behaviour of $A^+(\epsilon)$ when the initial cell starts from $0$ and behaves differently from the others, with indefinite growth. A lower bound is then given in Proposition \[Proposition lower envelope A(t)\].
We prove Theorem \[Theorem area GF\] in Section 4. The first subsection looks at the expectation of $A(\epsilon)$. In the second subsection, we introduce a useful martingale to control the fluctuations of $A(\epsilon)$ around its expectation and prove Theorem \[Theorem area GF\].
Section 5 is devoted to the proof of Propositions \[Theorem upper envelope A(t)\] and \[Proposition lower envelope A(t)\], that are respectively upper bound and lower bound for $A^+(\epsilon)$ as $\epsilon\to 0+$. After having established Proposition \[Theorem stationary area from 0\], we prove Proposition \[Theorem upper envelope A(t)\] in the first subsection. We then turn our attention to Proposition \[Proposition lower envelope A(t)\] in the second subsection. The arguments are different and the proof is more involved.
We conclude this paper with Section 6, in which we apply our results to stable surfaces.
Self-similar growth-fragmentations {#Section self-siilar growth-fragmentations}
==================================
#### The law of a typical cell.
For all $x>0$, let $\mathbb{P}_x$ denote the law of a positive self-similar Markov process $X=(X(t))_{t\geq 0}$ starting from $x$ and absorbed at 0. Self-similarity refers to the property that under $\mathbb{P}_x$, the law of $(X(t))_{t\geq 0}$ is the same as that of $(xX(tx^\alpha))_{t\geq 0}$ under $\mathbb{P}_1$, for some real index $\alpha$. We shall always assume here that $\alpha<0$. For all $t\geq 0$, let introduce the time-change $$\begin{aligned}
\label{Equation Lamperti tau_t}
\tau_t:=\int_0^tX(s)^\alpha\mathrm{d}s.
\end{aligned}$$ The well-known Lamperti’s transformation states that there exists a unique Lévy process $\xi=(\xi(t))_{t\geq 0}$ such that $$\begin{aligned}
X(t)=\exp\left(\xi(\tau_t)\right),\quad \forall t\geq 0.
\end{aligned}$$ Equivalently to , one can write $$\begin{aligned}
\tau_t=\inf\left\{s\geq 0:\int_0^s\exp(-\alpha\xi(u))\mathrm{d}u\geq t\right\}.
\end{aligned}$$ Let $\Lambda$ be the Lévy measure of $\xi$ and assume that $\int_{{\left(1\,,\infty\right)}}e^{y}\Lambda(\mathrm{d}y)<\infty$. We thus have, at least for $q\in{\left[0\,,1\right]}$, that $\mathbb{E}(\exp(q\xi(t)))=\exp(t\psi(q))$ for all $t\geq 0$, with $$\begin{aligned}
\label{Equation Laplace exponent psi of xi}
\psi:q\mapsto bq+\frac{\sigma^2}{2}q^2+\int_{\mathbb{R}}\left(e^{qy}-1+q(1-e^y)\right)\Lambda(\mathrm{d}y),\quad q\geq 0,
\end{aligned}$$ where $b,\sigma^2\geq 0$. We point out that this is not the traditional Lévy-Khintchin formula for the Laplace exponent $\psi$ of a Lévy process, in which $(1-e^y)$ above is usually replaced by $y\mathds{1}_{\left\{|y|<1\right\}}$, but is more convenient for our purposes. We assume throughout this paper that $\psi'(0)<0$. It is known that this entails that $\xi$ drifts to $-\infty$ almost surely, and in particular that the absorption time in $0$ of $X$, namely $\int_0^\infty\exp(-\alpha\xi(t))\mathrm{d}t$, is finite almost surely.
#### The cell-system.
We briefly recall Bertoin’s construction of the cell-system as in [@B17]. We use the classical Ulam-Harris-Neveu notation to label the particles of the branching process. Let $\mathbb{U}:=\bigcup_{n\geq 0}\mathbb{N}^n$ and let $\partial\mathbb{U}$ be the set of infinite sequences of natural integers. We shall also denote $\overline{\mathbb{U}}:=\mathbb{U}\cup\partial\mathbb{U}$. If $u\in\mathbb{N}^n$ for some $n\in\mathbb{N}$, we write $u(k)\in\mathbb{N}^k$ its ancestor at generation $k\leq n$ and $|u|=n$ its generation.
The process starts with a single cell that we call the Eve cell, indexed by $\emptyset$ and size at any time $t\geq 0$ denoted by $\chi_{\emptyset}(t)$. It evolves in time, starting from its birthtime $b_\emptyset:=0$ according to $\mathbb{P}_1$ until its absorption time $\zeta_\emptyset$ at 0. In this context, $\zeta_{\emptyset}$ is called the *lifetime* of $\chi_\emptyset$.
Since $\chi_\emptyset$ converges to 0 almost surely, it is possible to rank all its negative jumps in the decreasing order of the absolute values of their sizes. If $\left\{(b_i,\Delta_i);i\geq 1\right\}$ is the collection of times and sizes of these negative jumps ranked in this way, then for each $i\geq 1$, we start at time $b_i$ a new positive self-similar Markov process $\chi_i$ under the law $\mathbb{P}_{|\Delta_i|}$, independently of every other cell. We denote its lifetime $\zeta_i$, that is $\chi_i(t)$ is the size of the cell labelled $i$ if $b_i\leq t<b_i+\zeta_i$, and is sent to some cemetery state $\partial$ otherwise. In this manner, we construct recursively the whole cell-system indexed by $\mathbb{U}$, and we denote $\mathcal{P}_1$ its law. More generally, we can construct such a cell-system with the Eve cell starting from a size $x>0$, its law is then denoted $\mathcal{P}_x$.
#### The growth-fragmentation.
The growth-fragmentation process $(\mathbf{X}(t))_{t\geq 0}$ induced by the cell-system is the process following the collection of particles’ sizes in time, forgetting about their genealogy, that is $$\begin{aligned}
\mathbf{X}(t):=\left\{\!\!\left\{\chi_u(t-b_u):u\in\mathbb{U},b_u\leq t<b_u+\zeta_u\right\}\!\!\right\},
\end{aligned}$$ where the elements are repeated according to their multiplicity.
The law of $\mathbf{X}$ is characterized by $\alpha$ and a particular function $\kappa:\mathbb{R}_+\to\mathbb{R}\cup\{\infty\}$ called the *cumulant function* of $\mathbf{X}$ (see [@S17]), which is defined as $$\begin{aligned}
\kappa(q):=\psi(q)+\int_{{\left(-\infty\,,0\right)}}(1-e^y)^q\Lambda(\mathrm{d}y).
\end{aligned}$$ The importance of $\kappa$ for the study of self-similar growth-fragmentations comes from the fact, taken from [@B17] Lemma 3, that $\mathbb{E}_x\left(\sum_{i\geq 1}\chi_i(0)^{q}\right)=x(1-\kappa(q)/\psi(q))$, when it makes sense. This means that when $\kappa(q)=0$, the initial size of a cell is equal to the expectation of the sum of the sizes raised to the power $q$ of its first generation children at their birthtime. For this reason, we assume throughout this work that the *Cramér hypothesis* holds, that is there exists $\omega_->0$ such that $$\begin{aligned}
\label{Equation Cramer hypothesis}
\kappa(\omega_-)=0,\quad-\infty<\kappa'(\omega_-)<0.
\end{aligned}$$ We suppose that $\kappa$ has a second root $\omega_+>\omega_-$ and is finite in a right neighbourhood of $\omega_+$, which by convexity of $\kappa$ implies $0<\kappa'(\omega_+)<\infty$.
#### Intrinsic area measure.
Thanks to , the process $$\begin{aligned}
\mathcal{M}(n):=\sum_{|u|=n}\chi_u(0)^{\omega_-},\quad n\geq 0
\end{aligned}$$ is a uniformly integrable martingale, see Lemma 2.4 in [@BBCK16]. Its terminal value $\mathcal{M}:=\lim_{n\to\infty}\mathcal{M}(n)$ is called the *intrinsic area* of $\mathbf{X}$ and we point out that, by construction, it does not depend on $\alpha$.
Endowed with the distance $d(\ell,\ell'):=\exp(-\sup\{n\geq 0:\ell(n)=\ell'(n)\})$, $\partial\mathbb{U}$ is a complete metric space. The *intrinsic area measure* $\mathcal{A}$ is then the unique measure on $\partial\mathbb{U}$ such that the mass of the subsets of leaves having ancestor $u\in\mathbb{N}^n$ at generation $n\geq 0$ satisfies $$\begin{aligned}
\mathcal{A}\left(\left\{\ell\in\partial\mathbb{U}:\ell(n)=u\right\}\right)=\lim_{k\to\infty}\sum_{|v|=k}\chi_{uv}(0)^{\omega_-}.
\end{aligned}$$ Note that the total mass of $\mathcal{A}$ is $\mathcal{M}$. Denoting by $\zeta_\ell$ the height of $\ell\in\partial\mathbb{U}$, that is the time at which the lineage of this leaf goes extinct in terms of the growth-fragmentation’s time, we define the area of the ball with radius $t\in\mathbb{R}_+$ centered at the origin by $$\begin{aligned}
A:t\mapsto\mathcal{A}\left(\left\{\ell\in\partial\mathbb{U}:\zeta_\ell\leq t\right\}\right).
\end{aligned}$$ The increasing function $A$ will be the object of interest of the rest of the paper. It satisfies the following Markov-branching type property that we shall use all along this work.
\[Lemma Markov-branching property of A\] For all $t\geq 0$, it holds that $$\begin{aligned}
A(t)=\sum_{s\leq t}|\Delta_- \chi_{\emptyset}(s)|^{\omega_-}A_s\left((t-s)|\Delta_- \chi_\emptyset(s)|^{\alpha}\right),
\end{aligned}$$ where $\Delta_-\chi_\emptyset(s):=\min\left\{0, \chi_\emptyset(s)-\chi_\emptyset(s-)\right\}$ (i.e. we only consider the negative jumps) and the $A_s$’s are i.i.d. copies of $A$ under $\mathcal{P}_1$, independent from $\chi_\emptyset$.
We first introduce a notation. If $\Delta_-\chi_\emptyset(s)\neq 0$ for some $s>0$, i.e. if the Eve cell gives birth at time $s$, we denote $\mathbb{U}_s$ the subtree generated by the newborn cell. We write $$\begin{aligned}
A(t)
&=\mathcal{A}\left(\left\{\ell\in\partial\mathbb{U}:\zeta_\ell\leq t\right\}\right)
=\sum_{s\leq t}\mathds{1}_{\left\{\Delta_-\chi_{\emptyset}(s)\neq 0\right\}}\mathcal{A}\left(\left\{\ell\in\partial\mathbb{U}_s:\zeta_\ell\leq t\right\}\right).
\end{aligned}$$ Thanks to , Lemma 3.2 of [@BBCK16] entails that the areas in the above sum are independent and that for all $s>0$ the conditional distribution of $\mathcal{A}\left(\left\{\ell\in\partial\mathbb{U}_s:\zeta_\ell\leq t\right\}\right)$ given $\chi_{\emptyset}$ is that of $A(t-s)$ under $\mathcal{P}_{|\Delta_-\chi_{\emptyset}(s)|}$, which by self-similarity is identical to that of $|\Delta_-\chi_{\emptyset}(s)|^{\omega_-}A((t-s)|\Delta_-\chi_{\emptyset}(s)|^{\alpha})$ under $\mathcal{P}_1$. (The time shift $-s$ comes from the fact that the root of $\mathbb{U}_s$ is at height $s$ in $\mathbb{U}$.)
We shall sometimes use Lemma \[Lemma Markov-branching property of A\] replacing $\chi_\emptyset(s)$ by $X(s)$, or equivalently $\exp(\xi_{\tau_s})$.
#### Choosing a spine according to the intrinsic area.
A classical tool in the study of branching processes is the so called *spinal decomposition* (see [@S15]). In [@BBCK16], the authors introduced one related to the intrinsic area (we refer to their work for proofs and more details). Let $\mathcal{P}^-_x$ be the joint law of a cell system $(\chi_u:u\in\mathbb{U})$ starting from an Eve cell $\chi_\emptyset$ with initial size $x>0$ and a distinguished leaf $\widehat{\ell}\in\partial\mathbb{U}$, such that the law of $(\chi_u:u\in\mathbb{U})$ under $\mathcal{P}_x^-$ is absolutely continuous with respect to its law under $\mathcal{P}_x$, with density $x^{-\omega_-}\mathcal{M}$, and $\widehat{\ell}$ has conditional law $\mathcal{A}(\cdot)/\mathcal{M}$ given the cell system.
The *spine* is the process following in time the size of the ancestor of $\widehat{\ell}$. Up to modifying the genealogy, we can consider the spine to be the Eve cell of the cell-system, the law of the growth-fragmentation remains unchanged (see [@BBCK16] Section 4.2). Whereas the spine has a tilted distribution, its daughters generate, given the spine, independent growth-fragmentations with laws $(\mathcal{P}_x)_{x>0}$ according to their initial sizes. Its law is that of a positive self-similar Markov process $Y^-=(Y^-(t))_{t\geq 0}$ with same index $\alpha$. Let $\mathbb{P}_x^-$ be the law of $Y^-$ starting from $x>0$. As for $X$, we have that $(Y^-(t))_{t\geq 0}=(\exp(\eta^-(\tau_t)))_{t\geq 0}$, where $\eta^-$ is a Lévy process and $(\tau_t)_{t\geq 0}$ the associated time-change from the Lamperti’s transformation. (It will always be clear from the context whether $\tau_t$ denotes the time-change associated with $Y^-$ or $X$.) Recall that $\Lambda$ is the Lévy measure of $\xi$. The Lévy measure of $\eta^-$, that we call $\Pi^-$, is then given by $$\begin{aligned}
\label{Equation Pi Levy measure of eta^-}
\Pi^-(\mathrm{d}y)=e^{\omega_-y}(\Lambda+\widetilde{\Lambda})(\mathrm{d}y),
\end{aligned}$$ where $\widetilde{\Lambda}$ is the push-forward of $\Lambda$ by the mapping $x\mapsto\log(1-e^x)\mathds{1}_{\{x<0\}}$. The Laplace exponent[^2] of $\eta^-$ is given by $$\begin{aligned}
\label{Equation Laplace exponent phi of eta^-}
\phi_-(q)=\kappa(\omega_-+q),\quad q\in\mathbb{R}.
\end{aligned}$$ Since $\phi_-'(0)<0$ by , we know that $Y^-$ is absorbed at 0 in finite time almost surely. By definition of Lamperti’s time-change, its absorption time $I$ is given by $$\begin{aligned}
\label{Equation definition I}
I=\int_0^\infty\exp\left(-\alpha\eta^-(t)\right)\mathrm{d}t.
\end{aligned}$$ Thus written, $I$ is known as an *exponential functional* of the Lévy process $\eta^-$, and this kind of random variables has been extensively studied (see e.g. [@PS16] and references therein).
One of the interests of this spinal decomposition is that it allows us to study $A$, e.g. through the forthcoming lemma \[Lemma expect A(epsilon) = distr function of I\] that states an explicit and fruitful connection between $I$ and $A$.
#### Conditioning the spine to grow indefinitely.
A second spinal decomposition was also introduced in [@BBCK16]. Under assumption , the process $n\mapsto\sum_{|u|=n}\chi_u(0)^{\omega_+}$ indexed by generation is also a martingale, but now with terminal value $0$ almost surely. One can define the joint law $\mathcal{P}_x^+$ of $(\chi_u:u\in\mathbb{U})$ starting from an Eve cell of initial size $x>0$ and a leaf $\widehat{\ell}$, such that if $\Gamma_n$ is an event measurable with respect to the sigma-field generated by the cells at generations at most $n\geq 0$, then $$\begin{aligned}
\mathcal{P}_x^+(\Gamma_n)
=x^{-\omega_+}\mathcal{E}_x\left(\mathds{1}_{\Gamma_n}\sum_{|u|=n+1}\chi_u(0)^{\omega_+}\right).
\end{aligned}$$ In [@BBCK16], the ancestors of $\widehat{\ell}$ are selected as follows: let $\widehat{\ell}(n+1)\in\mathbb{N}^{n+1}$ be the parent of $\widehat{\ell}$ at generation $n+1$, then its conditional law is $$\begin{aligned}
\mathcal{P}_x^+\left(\left.\widehat{\ell}(n+1)=v\right|(\chi_u)_{|u|\leq n}\right)
=\frac{\chi_v(0)^{\omega_+}}{\sum_{|u|=n+1}\chi_u(0)^{\omega_+}},\qquad\forall v\in\mathbb{N}^{n+1}.
\end{aligned}$$ The law of the spine is again that of a positive self-similar Markov process $(Y^+(t))_{t\geq 0}=(\exp(\eta^+(\tau_t)))_{t\geq 0}$ with index $\alpha$, that we denote by $\mathbb{P}_x^+$ when $Y^+$ starts from $x>0$. Assigning the role of the Eve cell to the spine instead of $\chi_\emptyset$ reorders the genealogy of the cell system, but leaves the law of $\mathbf{X}$ unchanged, see Section 4.2 in [@BBCK16]. We shall henceforth write $\mathcal{P}_x^+$ for the distribution of the reordered cell system where $\chi_\emptyset$ has law $\mathbb{P}_x^+$, and its daughters generate independent growth-fragmentations given the spine, with laws $(\mathcal{P}_y)_{y>0}$ according to their initial sizes.
The Lévy measure $\Pi^+$ of the Lévy process $\eta^+$ is given by $$\begin{aligned}
\label{Equation Pi Levy measure of eta^+}
\Pi^+(\mathrm{d}y)=e^{\omega_+y}(\Lambda+\widetilde{\Lambda})(\mathrm{d}y),
\end{aligned}$$ and the Laplace exponent of $\eta^+$ is $$\begin{aligned}
\label{Equation Laplace exponent phi of eta^+}
\phi_+(q)=\kappa(\omega_++q),\quad q\geq 0.
\end{aligned}$$ Since $\kappa'(\omega_+)>0$, the process $\eta^+$ diverges to $\infty$ almost surely, which is therefore also the case of $Y^+$ (but does not explode in finite time since $\alpha<0$).
It is possible to make sense of $Y^+$ starting from 0 as the limit of $\mathbb{P}_x^+$ as $x\to 0+$. We can thus define $\mathcal{P}_0^+$, as the law of the growth-fragmentation whose Eve cell has law $\mathbb{P}_0^+$, see Corollary 4.4 in [@BBCK16]. Note that in this latter result, the convergence of $\mathcal{P}_x^+$ towards $\mathcal{P}_0^+$ as $x\to 0$ is proved only in the sense of finite dimensional distributions. This does not allow us, for instance, to study $A$ under $\mathcal{P}_0^+$ as the limit of $A$ under $\mathcal{P}_x^+$
To emphasize with which process we are working, we shall write $\mathbb{P}_x^+$ (respectively $\mathbb{P}_x^-$) for the law of $Y^+$ (respectively $Y^-$) started from $x>0$, and $\mathbb{E}_x^+$ (respectively $\mathbb{E}_x^-$) for the induced expectation operator.
Main results {#Sectionmain results}
============
Before stating the results of this paper, we make a last assumption on the parameters. Recall that $\Lambda$ denotes the Lévy measure of $\xi$. We shall assume that its left-tail $\overline{\Lambda}:x\mapsto\Lambda({\left(-\infty\,,-x\right)})$ is regularly varying at $0+$ with index $-\rho$, for some $\rho>0$ that satisfies $$\begin{aligned}
\label{Equation assumption rho}
\max(2\omega_--\omega_+,-\alpha)<\rho<\omega_-.
\end{aligned}$$ This is equivalent to regular variation with index $\omega_--\rho$ of the tail $x\mapsto\Pi^-({\left(-\infty\,,\log(x)\right)})$ as $x\to 0+$. Indeed, suppose holds and recall that by , we have that $$\begin{aligned}
\Pi^-({\left(-\infty\,,\log(x)\right)})
&=\int_{{\left(-\infty\,,\log(x)\right)}}e^{\omega_- y}\Lambda(\mathrm{d}y)+\int_{{\left(\log(1-x)\,,0\right)}}(1-e^{y})^{\omega_-}\Lambda(\mathrm{d}y).
\end{aligned}$$ We see that $$\begin{aligned}
\int_{{\left(-\infty\,,\log(x)\right)}}e^{\omega_- y}\Lambda(\mathrm{d}y)
\leq x^{\omega_-}\overline{\Lambda}(\log(x))
=o(x^{\omega_-}).
\end{aligned}$$ To estimate the second integral, we use and Theorem 1.6.5 in [@BGT87], which then gives us that $$\begin{aligned}
\label{Equation tail Pi^-}
\Pi^-({\left(-\infty\,,\log(x)\right)})
&\underset{x\to 0+}{\sim}\int_{{\left(\log(1-x)\,,0\right)}}(1-e^{y})^{\omega_-}\Lambda(\mathrm{d}y)
\underset{x\to 0+}{\sim}\overline{\Lambda}\left(x\right)x^{\omega_-}\frac{\rho}{\omega_--\rho}.
\end{aligned}$$ The converse also follows from the same theorem, that is if $x\mapsto\Pi^-({\left(-\infty\,,\log(x)\right)})$ has regular variation at $0+$ with index $\omega_--\rho$ then it is also the case for $\overline{\Lambda}$ with index $-\rho$. Indeed, coming back to the first part of and using $1-e^{-y}\sim y$ and $\log(1-y)\sim y$ as $y\to 0+$, we see that $$\begin{aligned}
\Pi^-({\left(-\infty\,,\log(x)\right)})
&\underset{x\to 0+}{\sim}\int_{{\left(-x\,,0\right)}}|y|^{\omega_-}\Lambda(\mathrm{d}y).
\end{aligned}$$ This is enough to conclude using [@BGT87] Theorem 1.6.5 (see the discussion following the proof of the theorem).
Similarly, one can show that regular variation with index $-\rho$ of $\overline{\Lambda}$ at $0+$ is equivalent to regular variation with index $\omega_+-\rho$ of $x\mapsto\Pi^+({\left(-\infty\,,\log(x)\right)})$ as $x\to 0+$.
Recall the definition of $I$ from .
\[Theorem area GF\] Consider a growth-fragmentation $\mathbf{X}$ such that and are satisfied. Then, for every $x>0$, it holds that $$\begin{aligned}
\frac{\epsilon^{-(1+\frac{\omega_-}{|\alpha|})}}{\overline{\Lambda}(\epsilon^{1/|\alpha|})}A(\epsilon)
\underset{\epsilon\to 0}{\longrightarrow}\frac{|\alpha|\rho}{(\omega_--\rho)(\omega_-+|\alpha|-\rho)}\mathbb{E}_1^-(I^{\frac{\omega_--\rho}{\alpha}})x^{\alpha+\rho},
\end{aligned}$$ $\mathcal{P}_x$-almost surely and in $\mathbb{L}^1$, where the expectation is finite.
Since $\alpha<0$, we know from [@BBCK16] Proposition 4.1 that for any $x>0$, $\mathcal{P}_x^+$ is absolutely continuous with respect to the restriction of $\mathcal{P}_x$ to the natural filtration of the growth-fragmentation and the spine up to time $t>0$. Therefore the above convergence also holds $\mathcal{P}_x^+$-a.s.
In Section 6, we shall see that the following interesting result easily follows from self-similarity.
\[Theorem stationary area from 0\] Suppose that holds. Then under $\mathcal{P}_0^+$, the process $$\begin{aligned}
\left(e^{\frac{\omega_-}{\alpha}u}A\left(e^u\right)\right)_{u\in\mathbb{R}}
\end{aligned}$$ is stationary. In particular, the law of $t^{\frac{\omega_-}{\alpha}}A(t)$ does not depend on $t>0$.
The next two propositions provide some bound about the almost sure behaviour of this area and shows that almost surely, $A(t)$ does not deviate from $t^{\omega_-/|\alpha|}$ by more than a power of $|\log(t)|$:
\[Theorem upper envelope A(t)\] Suppose that and $\kappa(\omega_++\omega_-+\alpha)<\infty$ hold. For all $\delta>0$, we have that $$\begin{aligned}
\limsup_{t\to 0+\text{ or }\infty}|\log(t)|^{-1-\delta}t^{\frac{\omega_-}{\alpha}}A(t)=0,\quad\mathcal{P}_0^+\text{-a.s.}
\end{aligned}$$
\[Proposition lower envelope A(t)\] Suppose that and hold. Then, there exists $q_0>0$ such that for all $q>q_0$, we have that $$\begin{aligned}
\liminf_{t\to 0+\text{ or }\infty}|\log(t)|^{q}t^{\frac{\omega_-}{\alpha}}A(t)=\infty,\quad\mathcal{P}_0^+\text{-a.s.}
\end{aligned}$$
The proof of Proposition \[Proposition lower envelope A(t)\] provides an explicit $q_0$, with a rather complicated expression given in Remark \[Remark exponent lower envelope\] in Section 5. We do not claim however that the bounds of the two above Propositions are optimal.
In Section \[Section Application to random maps\], we shall recall the connection between growth-fragmentations and a family of random surfaces, and apply the above results to the latter, refining in particular some results for the Brownian map.
Area near the origin {#Section area near the origin}
====================
We assume in this section that and hold. Our goal is to prove Theorem \[Theorem area GF\]. We first look at the expectation of $A(\epsilon)$ as $\epsilon\to 0+$.
Behaviour of the expectation
----------------------------
Under $\mathcal{P}_1^-$, the spine is chosen according to $\mathcal{A}$. We can hence study $A$ through the lifetime of the spine $I$, thanks to the following relation.
\[Lemma expect A(epsilon) = distr function of I\] For all $t\geq 0$, it holds that $$\begin{aligned}
\mathcal{E}_1(A(t))=\mathbb{P}_1^-(I\leq t).
\end{aligned}$$
This is a straightforward consequence of Lemmas 2.1 and 2.2 in [@G18]. Thanks to Lemma \[Lemma expect A(epsilon) = distr function of I\], we can deduce
\[Lemma asymp expect A(espilon)\] It holds that that $$\begin{aligned}
\mathcal{E}_1(A(\epsilon))\underset{\epsilon\to 0}{\sim}\frac{|\alpha|\rho}{(\omega_--\rho)(\omega_-+|\alpha|-\rho)}\mathbb{E}_1^-(I^{\frac{\omega_--\rho}{\alpha}})\epsilon^{1+\frac{\omega_-}{|\alpha|}}\overline{\Lambda}(\epsilon^{1/|\alpha|}),
\end{aligned}$$ where the expectation on the right-hand side is finite.
Let $\Pi_\alpha^-$ be the Lévy measure of $|\alpha|\eta^-$. The behaviour of $\mathbb{P}_1^-(I\leq t)$ as $t\to 0$ is given in Theorem 7 of [@AR15]. More precisely, provided that $\Pi_\alpha^-({\left(-\infty\,,-(x+y)\right)})/\Pi_\alpha^-({\left(-\infty\,,-x\right)})\sim e^{-\gamma y}$ as $x\to\infty$ with $\mathbb{E}_1^-(\exp(|\alpha|\gamma\eta^-(1)))<1$ (corresponding to $|\alpha|\gamma\in{\left(0\,,\omega_+-\omega_-\right)}$ by ), it holds that $$\begin{aligned}
\mathbb{P}_1^-(I\leq t)
\underset{t\to 0}{\sim}\frac{\mathbb{E}_1^-(I^{-\gamma})}{1+\gamma}t\Pi_\alpha^-({\left(-\infty\,,\log(1/t)\right)}).
\end{aligned}$$ By , we have that $$\begin{aligned}
\Pi_\alpha^-({\left(-\infty\,,-x\right)})
&\underset{x\to \infty}{\sim}\overline{\Lambda}\left(e^{x/\alpha}\right)e^{\omega_-x/\alpha}\frac{\rho}{\omega_--\rho}.
\end{aligned}$$ We thus see that $\Pi_\alpha({\left(-\infty\,,-x-y\right)})/\Pi_\alpha({\left(-\infty\,,-x\right)})\to \exp(-\frac{\omega_--\rho}{|\alpha|} y)$ as $x\to\infty$. It remains to check that $\omega_--\rho<\omega_+-\omega_-$, which is true since $2\omega_--\omega_+<\rho<\omega_-$ by . We therefore get by [@AR15] Theorem 7 that $$\begin{aligned}
\mathbb{P}_1^-(I\leq \epsilon)
&\underset{\epsilon\to 0+}{\sim}\left(1+\frac{\omega_--\rho}{|\alpha|}\right)^{-1}\mathbb{E}_1^-\left(I^{-\frac{\omega_--\rho}{|\alpha|}}\right)\epsilon\Pi_\alpha^-({\left(-\infty\,,\log(\epsilon)\right)})\\
&\underset{\epsilon\to 0+}{\sim}\frac{|\alpha|\rho}{(\omega_--\rho)(\omega_-+|\alpha|-\rho)}\mathbb{E}_1^-\left(I^{-\frac{\omega_--\rho}{|\alpha|}}\right)\epsilon^{1+\frac{\omega_-}{|\alpha|}}\overline{\Lambda}(\epsilon^{1/|\alpha|}),
\end{aligned}$$ as claimed, where the expectation is finite.
We shall need a bit more than the speed of convergence to 0 of the first moment.
\[Lemma speed A\^p\] For $p=p(\epsilon):=1+1/\sqrt{|\log(\epsilon)|}$, for all $0<q<1+\frac{\omega_--\rho}{|\alpha|}$ it holds that $$\begin{aligned}
\mathcal{E}_1\left(A(\epsilon)^p\right)=o\left(\epsilon^q\right),\quad\text{as}\quad\epsilon\to 0.
\end{aligned}$$
Let $n\geq 1$ be arbitrary large. We write $$\begin{aligned}
\mathcal{E}_1\left(A(\epsilon)^p\right)
&\leq\mathcal{E}_1\left(A(\epsilon)\mathds{1}_{\left\{A(\epsilon)\leq \epsilon^{-n}\right\}}\right)\times \epsilon^{-n(p-1)}+\mathcal{E}_1\left(A(\epsilon)^p\mathds{1}_{\left\{A(\epsilon)>\epsilon^{-n}\right\}}\right).
\end{aligned}$$ By Lemma \[Lemma asymp expect A(espilon)\], the first term is of order $\epsilon^{1+\frac{\omega_-}{-\alpha}-n(p-1)}\overline{\Lambda}(\epsilon^{1/|\alpha|})$, with $p-1=1/\sqrt{-\log(\epsilon)}$ going to 0 with $\epsilon$. It remains to bound the second term. Take $q,q'>1$ such that $1/q+1/q'=1$ and $pq<\omega_+/\omega_-$, Hölder’s Inequality then yields that $$\begin{aligned}
\mathcal{E}_1\left(A(\epsilon)^p\mathds{1}_{\left\{A(\epsilon)>\epsilon^{-n}\right\}}\right)
&\leq\mathcal{E}_1\left(A(\epsilon)^{pq}\right)^{1/q}\times\mathcal{P}(A(\epsilon)>\epsilon^{-n})^{1/q'}.
\end{aligned}$$ Since $pq<\omega_+/\omega_-$ we know that $\mathcal{E}_1(A(\epsilon)^{pq})<\infty$ by Lemma 2.3 in [@BBCK16]. Markov’s Inequality shows that the probability on the right-hand side is bounded from above by $$\begin{aligned}
\mathcal{E}_1(A(\epsilon))^{1/q'}\times\epsilon^{n/q'}.
\end{aligned}$$ Since $n$ was chosen arbitrary large, this concludes the proof.
The proof above would work for any $p(\epsilon)$ converging to 1 from above as $\epsilon\to 0$, but will shall use it with this specific choice of $p(\epsilon)$ in order to have a technical result, that we record now for later use.
\[Lemma equiv p(epsilon)\] For $p=p(\epsilon)$ as in Lemma \[Lemma speed A\^p\], it holds that $$\begin{aligned}
\epsilon^{1+\frac{p\omega_-}{|\alpha|}}\overline{\Lambda}(\epsilon^{1/|\alpha|})=o\left(\epsilon^{p(1+\frac{\omega_-}{|\alpha|})}\overline{\Lambda}(\epsilon^{1/|\alpha|})^{p}\right),\quad\text{as }\epsilon\to 0+.
\end{aligned}$$
By $\eqref{Equation assumption rho}$ and Theorem 1.4.1(iii) in [@BGT87], we know that there exists a function $\ell:{\left(0\,,\infty\right)}\to{\left(0\,,\infty\right)}$ with slow variation at $0$ such that $\overline{\Lambda}(\epsilon)\sim\epsilon^{-\rho}\ell(\epsilon)$. We write $$\begin{aligned}
&\log\left(\frac{\epsilon^{1+\frac{p\omega_-}{|\alpha|}}\overline{\Lambda}(\epsilon^{1/|\alpha|})}{\epsilon^{p(1+\frac{\omega_-}{|\alpha|})}\overline{\Lambda}(\epsilon^{1/|\alpha|})^{p}}\right)
=(1-p)\log(\epsilon)+(1-p)\log\left(\overline{\Lambda}(\epsilon^{1/|\alpha|})\right)\\
&\hspace{3.8cm}\underset{\epsilon\to 0+}{\sim}(1-p)\log(\epsilon)+\frac{(1-p)\rho}{\alpha}\log(\epsilon)+(1-p)\log(\ell(\epsilon^{1/|\alpha|})).
\end{aligned}$$ Since $\ell$ has slow variation, the terms with $\log(\epsilon)$ dominates. By definition of $p$ and since $|\alpha|<\rho$ by , we see that $(1-p)(1+\rho/\alpha)\log(\epsilon)=(1+\rho/\alpha)\sqrt{|\log(\epsilon)|}\to -\infty$ as $\epsilon\to 0+$, which entails the claim.
The almost sure behaviour
-------------------------
Throughout this subsection, we shall only work under $\mathcal{P}_1$, since the other cases follow from the self-similarity. Indeed, under $\mathcal{P}_x$, we have $(A(t))_{t\geq 0}\stackrel{d}{=}(x^{\omega_-}A(t x^\alpha))_{t\geq 0}$ under $\mathcal{P}_1$. It means that if Theorem \[Theorem area GF\] is true under $\mathcal{P}_1$, we easily deduce that under $\mathcal{P}_x$, $$\begin{aligned}
\frac{\epsilon^{-(1+\frac{\omega_-}{|\alpha|})}}{\overline{\Lambda}\left(\epsilon^{1/|\alpha|}\right)}A(\epsilon)
&\stackrel{d}{=}x^{\alpha}\frac{\overline{\Lambda}\left((\epsilon x^\alpha)^{1/|\alpha|}\right)}{\overline{\Lambda}\left(\epsilon^{1/|\alpha|}\right)}\cdot\frac{(\epsilon x^{\alpha})^{-(1+\frac{\omega_-}{|\alpha|})}}{\overline{\Lambda}\left((\epsilon x^\alpha)^{1/|\alpha|}\right)}A(\epsilon x^{\alpha}),\quad\text{under }\mathcal{P}_1\\
&\!\!\!\!\underset{\epsilon\to 0+}{\sim}x^{\rho+\alpha}\lim_{\epsilon\to 0+}\frac{(\epsilon x^{\alpha})^{-(1+\frac{\omega_-}{|\alpha|})}}{\overline{\Lambda}\left((\epsilon x^\alpha)^{1/|\alpha|}\right)}A(\epsilon x^{\alpha}),\hspace{1.3cm}\text{under }\mathcal{P}_1,
\end{aligned}$$ where we used the assumption of regular variation of $\overline{\Lambda}$ at $0+$, as written in .
Recall that $\xi$ is the Lévy process associated with $X$ by Lamperti’s transformation, and that under $\mathcal{P}_1$, $\chi_\emptyset$ is distributed as $X$. Let $\sigma:s\mapsto\int_0^s e^{-\alpha\xi(u)}\mathrm{d}u$ be the inverse change time of $(\tau_t)_{t\geq 0}$. Define the compensated process $(M_t)_{t\geq 0}$ as $$\begin{aligned}
M_t
:=&\sum_{s\leq t}(\Delta_- e^{\xi(s)})^{\omega_-}A_s\left((\sigma_t-\sigma_s)(\Delta_-e^{\xi(s)})^\alpha\right)-S_t,\\
\shortintertext{where the $A_s$'s are i.i.d. copies of $A$ under $\mathcal{P}_{1}$, independent from $\xi$, with}
S_t:=&\int_0^t\mathrm{d}se^{\omega_-\xi(s)}\int_{{\left(-\infty\,,0\right)}}\Lambda(\mathrm{d}y)(1-e^y)^{\omega_-}\mathbb{P}_1^-\left(I\leq (1-e^y)^{\alpha}e^{\alpha\xi(s)}(\sigma_t-\sigma_s)|\mathcal{F}_s\right),
\end{aligned}$$ where $I$ is independent from $\xi$ and $(\mathcal{F}_s)_{s\geq 0}$ is the natural filtration of $\xi$.
\[Lemma M martingale and M(tau)=A-S\] The process $(M_t)_{t\geq 0}$ is a martingale. Moreover, identifying $\chi_{\emptyset}$ with $X=\exp(\xi_\tau)$, we have that $M_{\tau_t}=A(t)-S_{\tau_t}$ for all $t\geq 0$.
Thanks to Lemma \[Lemma expect A(epsilon) = distr function of I\] and independence of the $A_s$’s with $\xi$, one sees that $(S_t)_{t\geq 0}$ is the predictable compensator of the series in $(M_t)_{t\geq 0}$, so that the latter is a martingale. We then write $$\begin{aligned}
M_{\tau_{t}}
&=\sum_{s\leq \tau_t}(\Delta_- e^{\xi(s)})^{\omega_-}A_s\left((t-\sigma_s)(\Delta_-e^{\xi(s)})^\alpha\right)-S_{\tau_t}\\
&=\sum_{s\leq t}(\Delta_- e^{\xi(\tau_s)})^{\omega_-}A_{\tau_s}\left((t-s)(\Delta_-e^{\xi(\tau_s)})^\alpha\right)-S_{\tau_t}\\
&=A(t)-S_{\tau_t},
\end{aligned}$$ by Lemma \[Lemma Markov-branching property of A\] (where we reindexed the $A_s$’s), as claimed.
The main idea of the current section is to use Lemma \[Lemma M martingale and M(tau)=A-S\], namely the martingale property of $M$ and its relation to $A$, to show that $A(\epsilon)\sim S_{\tau_\epsilon}$. In this direction, we first establish the following:
\[Lemma predictable compensator as E(A(epsilon))\] $\mathcal{P}_1$-almost surely, it holds that $$\begin{aligned}
S_{\tau_{\epsilon}}&\underset{\epsilon\to 0}{\sim}\frac{|\alpha|\rho}{(\omega_--\rho)(\omega_-+|\alpha|-\rho)}\mathbb{E}_1^-(I^{-\frac{\omega_--\rho}{|\alpha|}})\overline{\Lambda}\left(\epsilon^{1/|\alpha|}\right)\epsilon^{1+\frac{\omega_-}{|\alpha|}}.
\end{aligned}$$
Recall . After a change of variables and replacing $\exp(\xi(\tau_s))$ by $X(s)$, we have that $$\begin{aligned}
S_{\tau_\epsilon}
&=\int_0^\epsilon\mathrm{d}sX(s)^{\omega_-+\alpha}\int_{{\left(-\infty\,,0\right)}}\Lambda(\mathrm{d}y)(1-e^y)^{\omega_-}\mathbb{P}_1^-\left(I\leq (1-e^y)^{\alpha}X(s)^{\alpha}(\epsilon-s)|\mathcal{F}_{\tau_s}\right)
\end{aligned}$$ It is well known that the law of $I$ is absolutely continuous, see e.g. [@BLM08] Theorem 3.9. Hence, since $X(s)\sim 1$ as $s\to 0+$, one sees that $$\begin{aligned}
S_{\tau_\epsilon}
&\underset{\epsilon\to 0+}{\sim}\int_0^\epsilon\mathrm{d}s\int_{{\left(-\infty\,,0\right)}}\Lambda(\mathrm{d}y)(1-e^y)^{\omega_-}\mathbb{P}_1^-\left(I\leq(\epsilon-s)(1-e^y)^\alpha\right),\quad \mathcal{P}_1\text{-almost surely.}
\end{aligned}$$ We express the right-hand side in the form $$\begin{aligned}
&\int_{{\left(-\infty\,,0\right)}}\Lambda(\mathrm{d}y)(1-e^y)^{\omega_-}\int_0^\epsilon\mathrm{d}s\mathbb{P}_1^-(I\leq s(1-e^y)^{\alpha})\\
&\hspace{2cm}=\int_{{\left(-\infty\,,0\right)}}\Lambda(\mathrm{d}y)(1-e^y)^{\omega_-+|\alpha|}\int_0^{\epsilon(1-e^y)^\alpha}\mathrm{d}s\mathbb{P}_1^-(I\leq s)\\
&\hspace{2cm}=C_1(\epsilon)+C_2(\epsilon),\end{aligned}$$ where $C_1(\epsilon)$ is the part of the first integral with domain restricted to $E_1:={\left(-\infty\,,\log\left(1-\epsilon^{\frac{1-\delta}{|\alpha|}}\right)\right)}$ for $\delta>0$ arbitrary small, and $C_2(\epsilon)$ is that restricted to $E_2:={\left[\log\left(1-\epsilon^{\frac{1-\delta}{|\alpha|}}\right)\,,0\right)}$. We first address $C_1(\epsilon)$. Note that for $y\in E_1$, we have that $s\leq\epsilon (1-e^y)^\alpha\leq \epsilon^\delta$. Appealing to lemmas \[Lemma expect A(epsilon) = distr function of I\] and \[Lemma asymp expect A(espilon)\], there exists a constant $C$ such that $$\begin{aligned}
\mathbb{P}_1^-(I\leq s)\leq Cs^{1+\frac{\omega_-}{|\alpha|}}\overline{\Lambda}(s^{1/|\alpha|}),\quad\forall s\leq\epsilon^\delta.\end{aligned}$$ In the following, we shall keep writing $C$ for any positive and finite constant that does not depend on $\epsilon$ and that may change from line to line. We have $$\begin{aligned}
C_1(\epsilon)
&\leq C\int_{E_1}\Lambda(\mathrm{d}y)(1-e^y)^{\omega_-+|\alpha|}\int_0^{\epsilon^{\delta}} s^{1+\frac{\omega_-}{|\alpha|}}\overline{\Lambda}(s^{1/|\alpha|})\mathrm{d}s\\
&\leq C\int_{E_1}\Lambda(\mathrm{d}y)(1-e^y)^{\omega_-+|\alpha|}\times\epsilon^{\delta(2+\frac{\omega_-}{|\alpha|})}\overline{\Lambda}(\epsilon^{\delta/|\alpha|}),\end{aligned}$$ where we used Theorem 1.5.11(ii) of [@BGT87]. Thanks to Theorem 1.6.4 of the same book, we conclude that $$\begin{aligned}
C_1(\epsilon)&=O\left(\epsilon^{(1-\delta)(1+\frac{\omega_-}{|\alpha|})}\overline{\Lambda}(\epsilon^{(1-\delta)/|\alpha|})\times\epsilon^{\delta(2+\frac{\omega_-}{|\alpha|})}\overline{\Lambda}(\epsilon^{\delta/|\alpha|})\right)\\
&=O\left(\epsilon^{\delta}\times\epsilon^{1+\frac{\omega_-}{|\alpha|}}\overline{\Lambda}(\epsilon^{(1-\delta)/|\alpha|})\overline{\Lambda}(\epsilon^{\delta/|\alpha|})\right)\\
&=o\left(\epsilon^{1+\frac{\omega_-}{|\alpha|}}\overline{\Lambda}(\epsilon^{1/|\alpha|})\right),\end{aligned}$$ where the last line follows straightforwardly from the existence of a slowly varying function $\ell:{\left(0\,,\infty\right)}\to{\left(0\,,\infty\right)}$ at the origin such that $\overline{\Lambda}(x)=x^{-\rho}\ell(x)$, see [@BGT87] Theorem 1.4.1(iii).
We now investigate $C_2(\epsilon)$. Let $k$ denote the density of $I$ under $\mathbb{P}_1^-$. After applying Tonelli’s Theorem, we have that $$\begin{aligned}
\label{Equation C_2}
C_2(\epsilon)
&=\int_0^{\epsilon^{\delta}}k(x)\mathrm{d}x\int_{E_2}\Lambda(\mathrm{d}y)(1-e^y)^{\omega_-}(\epsilon-x(1-e^y)^{|\alpha|})+\int_{\epsilon^{\delta}}^{\infty} k(x)xf(\epsilon/x)\mathrm{d}x,\end{aligned}$$ where $$\begin{aligned}
f(u):=\int_{{\left(\log\left(1-u^{\frac{1}{|\alpha|}}\right)\,,0\right)}}\Lambda(\mathrm{d}y)(1-e^y)^{\omega_-}(u-(1-e^y)^{|\alpha|}).\end{aligned}$$ Theorem 1.6.5 of [@BGT87] allows us to bound the first term in the right-hand side of by $$\begin{aligned}
\mathbb{P}_1^-\left(I\leq\epsilon^{\delta}\right)\times\epsilon\int_{E_2}\Lambda(\mathrm{d}y)(1-e^y)^{\omega_-}
&\underset{\epsilon\to 0+}{\sim}C\mathbb{P}_1^-\left(I\leq\epsilon^\delta\right)\epsilon^{-\delta\frac{\omega_--\rho}{|\alpha|}}\epsilon^{1+\frac{\omega_-}{|\alpha|}}\overline{\Lambda}(\epsilon^{(1-\delta)/|\alpha|})\\
&\ \ =o\left(\epsilon^{1+\frac{\omega_-}{|\alpha|}}\overline{\Lambda}(\epsilon^{1/|\alpha|})\right),\end{aligned}$$ with the same argument as for $C_1(\epsilon)$.
We turn our attention to the second term in . First, with the help of Theorem 1.6.5 of [@BGT87], it can be shown that $$\begin{aligned}
f(u)\underset{u\to 0+}{\sim}&\left(\frac{\rho}{\omega_--\rho}-\frac{\rho}{\omega_-+|\alpha|-\rho}\right)u^{1+\frac{\omega_-}{|\alpha|}}\overline{\Lambda}(u^{1/|\alpha|})\\
=\ &\frac{|\alpha|\rho}{(\omega_--\rho)(\omega_-+|\alpha|-\rho)}u^{1+\frac{\omega_-}{|\alpha|}}\overline{\Lambda}(u^{1/|\alpha|})\end{aligned}$$ Note that $\epsilon/x\leq \epsilon^{1-\delta}$ for all $x>\epsilon^\delta$. The estimate above yields $$\begin{aligned}
\int_{\epsilon^\delta}^{\infty} k(x)xf(\epsilon/x)\mathrm{d}x
\underset{\epsilon\to 0}{\sim}\frac{|\alpha|\rho}{(\omega_--\rho)(\omega_-+|\alpha|-\rho)}\epsilon^{1+\frac{\omega_-}{|\alpha|}}\int_{\epsilon^\delta}^{\infty}k(x)x^{\frac{\omega_-}{\alpha}}\overline{\Lambda}((\epsilon/x)^{1/|\alpha|})\mathrm{d}x.\end{aligned}$$ By dominated convergence, we then see that $$\begin{aligned}
\left(\epsilon^{1+\frac{\omega_-}{|\alpha|}}\overline{\Lambda}(\epsilon^{1/|\alpha|})\right)^{-1}\int_{\epsilon^\delta}^{\infty} k(x)xf(\epsilon/x)\mathrm{d}x
\underset{\epsilon\to 0}{\longrightarrow}\frac{|\alpha|\rho}{(\omega_--\rho)(\omega_-+|\alpha|-\rho)}\mathbb{E}_1^-\left(I^{\frac{\omega_--\rho}{\alpha}}\right),\end{aligned}$$ as claimed.
We are ready to provide the proof of Theorem \[Theorem area GF\].
Provided that $\epsilon\mapsto M_\epsilon$ has regular variation at 0, one has that $M_{\tau_\epsilon}\sim M_\epsilon$ as $\epsilon\to 0+$, as a consequence of $\tau_\epsilon\sim\epsilon$. Thus, thanks to Lemmas \[Lemma M martingale and M(tau)=A-S\] and \[Lemma predictable compensator as E(A(epsilon))\], it suffices to show that $$\begin{aligned}
\label{Equation to show}
\lim_{\epsilon\to 0}\left(\epsilon^{1+\frac{\omega_-}{|\alpha|}}\overline{\Lambda}(\epsilon^{1/|\alpha|})\right)^{-1}M_\epsilon=0,\quad\mathcal{P}_1\text{-a.s.}
\end{aligned}$$ Define the stopping time $$\begin{aligned}
T:=\inf\left\{s>0:|\xi(s)|>1\right\}.
\end{aligned}$$ We proceed as follows: we know that $\mathcal{M}$ has finite moment of order $p$ as soon as $p<\omega_+/\omega_-$ by [@BBCK16] Lemma 2.3. Let $p=p(\epsilon)=1+1/\sqrt{|\log(\epsilon)|}$ as in Lemma \[Lemma speed A\^p\] and let $a>0$ be arbitrary small, we have by Markov’s Inequality that $$\begin{aligned}
\mathcal{P}_1\left(\sup_{s\leq\epsilon\wedge T}\frac{|M_s|}{\epsilon^{1+\frac{\omega_-}{|\alpha|}}\overline{\Lambda}(\epsilon^{1/|\alpha|})}\geq a\right)
&\leq \left(a\epsilon^{1+\frac{\omega_-}{|\alpha|}}\overline{\Lambda}(\epsilon^{1/|\alpha|})\right)^{-p}\mathcal{E}_1\left(\sup_{s\leq\epsilon\wedge T}|M_s|^{p}\right)\\
&\leq (6p)^p\left(a\epsilon^{1+\frac{\omega_-}{|\alpha|}}\overline{\Lambda}(\epsilon^{1/|\alpha|})\right)^{-p}\mathcal{E}_1\left([M]_{\epsilon\wedge T}^{p/2}\right),
\end{aligned}$$ thanks to Burkholder-Davis-Gundy Inequality, where $[M]$ denotes the quadratic variation of $M$ (see Theorem 92 p.304 in [@DM2] for the constant $(6p)^p$). In particular, since $M$ is a purely discontinuous martingale, appealing to [@L76] Section 3(c) we get that the previous quantity is bounded from above by $$\begin{aligned}
(6p)^p\left(a\epsilon^{1+\frac{\omega_-}{|\alpha|}}\overline{\Lambda}(\epsilon^{1/|\alpha|})\right)^{-p}\mathcal{E}_1\left(\sum_{s\leq \epsilon\wedge T}(\Delta e^{\xi(s)})^{p\omega_-}A_s\left((\epsilon\wedge T-s)(\Delta e^{\xi(s)})^{\alpha}\right)^p\right),
\end{aligned}$$ Since $T>0$ almost surely by right-continuity of $\xi$, in order to conclude, it is sufficient to show that the expectation above is $o\left(\epsilon^{p(1+\frac{\omega_-}{|\alpha|})}\overline{\Lambda}(\epsilon^{1/|\alpha|})^{p}\right)$, as $\epsilon\to 0$. We write $$\begin{aligned}
&\mathcal{E}_1\left(\sum_{s\leq \epsilon\wedge T}(\Delta e^{\xi(s)})^{p\omega_-}A_s\left((\epsilon\wedge T-s)(\Delta e^{\xi(s)})^{\alpha}\right)^p\right)\\
&\hspace{2cm}=\mathcal{E}_1\left(\sum_{s\leq \epsilon\wedge T}e^{p\omega_-\xi(s-)}(1-e^{\Delta\xi(s)})^{p\omega_-}A_s\left((\epsilon\wedge T-s)e^{\alpha\xi(s-)}(1-e^{\Delta\xi(s)})^{\alpha}\right)^p\right)\\
&\hspace{2cm}\leq e^{p\omega_-}\mathcal{E}_1\left(\sum_{s\leq \epsilon}(1-e^{\Delta\xi(s)})^{p\omega_-}A_s\left(e^{|\alpha|}(\epsilon-s)(1-e^{\Delta\xi(s)})^{\alpha}\right)^p\right).
\end{aligned}$$
(Recall that the $A_s$’s are non-decreasing and non-negative functions.) Applying the compensation formula, the upper bound above becomes, after an implicit change of variables, $$\begin{aligned}
\label{Equation computation}
&e^{p\omega_-}e^{|\alpha|}\int_0^{\epsilon}\mathrm{d}s\int_{{\left(-\infty\,,0\right)}}\Lambda(\mathrm{d}y)(1-e^y)^{p\omega_-}\mathcal{E}_1\left(A\left(e^{|\alpha|}s(1-e^y)^\alpha\right)^p\right)\nonumber\\
&\hspace{1cm}=e^{p\omega_-+|\alpha|}\int_{{\left(-\infty\,,0\right)}}\Lambda(\mathrm{d}y)(1-e^y)^{p\omega_-}\int_0^{\epsilon}\mathrm{d}s\mathcal{E}_1\left(A\left(e^{|\alpha|}s(1-e^y)^\alpha\right)^p\right)\nonumber\\
&\hspace{1cm}=e^{p\omega_-+|\alpha|}\int_{{\left(-\infty\,,0\right)}}\Lambda(\mathrm{d}y)(1-e^y)^{p\omega_-+|\alpha|}\int_0^{\epsilon(1-e^y)^\alpha}\mathrm{d}s\mathcal{E}_1\left(A\left(e^{|\alpha|}s\right)^p\right)\nonumber\\
&\hspace{1cm}=e^{p\omega_-+|\alpha|}\int_0^{\epsilon}\mathrm{d}s\mathcal{E}_1\left(A\left(e^{|\alpha|}s\right)^p\right)\int_{{\left(-\infty\,,0\right)}}\Lambda(\mathrm{d}y)(1-e^y)^{p\omega_-+|\alpha|}\nonumber\\
&\hspace{2cm}+e^{p\omega_-+|\alpha|}\int_{\epsilon}^\infty\mathrm{d}s\mathcal{E}_1\left(A\left(e^{|\alpha|}s\right)^p\right)\int_{{\left(\log(1-(\epsilon/s)^{\frac{1}{|\alpha|}})\,,0\right)}}\Lambda(\mathrm{d}y)(1-e^y)^{p\omega_-+|\alpha|}\nonumber\\
&\hspace{1cm}\leq e^{p\omega_-+|\alpha|}\int_0^{K\epsilon}\mathrm{d}s\mathcal{E}_1\left(A\left(e^{|\alpha|}s\right)^p\right)\int_{{\left(-\infty\,,0\right)}}\Lambda(\mathrm{d}y)(1-e^y)^{p\omega_-+|\alpha|}\nonumber\\
&\hspace{2cm}+e^{p\omega_-+|\alpha|}\int_{K\epsilon}^\infty\mathrm{d}s\mathcal{E}_1\left(A\left(e^{|\alpha|}s\right)^p\right)\int_{{\left(\log(1-(\epsilon/s)^{\frac{1}{|\alpha|}})\,,0\right)}}\Lambda(\mathrm{d}y)(1-e^y)^{p\omega_-+|\alpha|},\end{aligned}$$ where $K$ is a constant intended to be large. The first term is negligeable since by Lemma \[Lemma speed A\^p\], for $\delta>0$ small enough, we have that $$\begin{aligned}
\int_0^{K\epsilon}\mathcal{E}_1\left(A\left(e^{|\alpha|}s\right)^p\right)\mathrm{d}s
=o(\epsilon^{2+\frac{\omega_--\rho}{|\alpha|}-\delta})
=o(\epsilon^{p(1+\frac{\omega_--\rho}{|\alpha|})}).
\end{aligned}$$ (Recall that $p=1+1/\sqrt{|\log(\epsilon)|}$.) To bound the second term, we first note that for all $s>K\epsilon$, we have $\epsilon/s\leq 1/K$, so that Theorem 1.6.5 of [@BGT87] entails the existence of a constant $C$ such that $$\begin{aligned}
\int_{{\left(\log(1-(\epsilon/s)^{\frac{1}{|\alpha|}})\,,0\right)}}\Lambda(\mathrm{d}y)(1-e^y)^{p\omega_-+|\alpha|}
&\leq C\left(\frac{\epsilon}{s}\right)^{\frac{p\omega_-+|\alpha|}{|\alpha|}}\overline{\Lambda}\left((\epsilon/s)^{1/|\alpha|}\right)\\
&\leq C \epsilon^{1+\frac{p\omega_--\rho}{|\alpha|}}\left(\frac{1}{s}\right)^{1+\frac{p\omega_--\rho}{|\alpha|}}\ell\left((\epsilon/s)^{1/|\alpha|}\right),
\end{aligned}$$ where $\ell:{\left(0\,,\infty\right)}\to{\left(0\,,\infty\right)}$ is some slowly varying function at the origin. Finally, we get that the second term in the right-hand side of is bounded from above by $$\begin{aligned}
\epsilon^{1+\frac{p\omega_--\rho}{|\alpha|}}\int_{0}^\infty\mathrm{d}s\mathcal{E}(A(e^{|\alpha|}s)^p)s^{-1-(p\omega_--\rho)/|\alpha|}\ell\left((\epsilon/s)^{1/|\alpha|}\right).
\end{aligned}$$ Since the integral on the right-hand side is finite by Lemma \[Lemma speed A\^p\], Lemma \[Lemma equiv p(epsilon)\] is enough to show that holds. The convergence in $\mathbb{L}^1$ is just a consequence of Lemma \[Lemma asymp expect A(espilon)\] and Scheffé’s Lemma and the proof is complete.
Area near the root starting from 0 {#Section area near the root starting from 0}
==================================
We start this section by proving Proposition \[Theorem stationary area from 0\].
Let $Y^+$ have law $\mathbb{P}_0^+$. We identify $\chi_\emptyset$ with $Y^+$ and similarly to Lemma \[Lemma Markov-branching property of A\], under $\mathcal{P}_0^+$, we can write $$\begin{aligned}
\label{Equation A as a sum under P_0^+}
\left\{t^{\frac{\omega_-}{\alpha}}A(t);t\geq 0\right\}
=\left\{t^{\frac{\omega_-}{\alpha}}\sum_{s\leq t}|\Delta_-Y^+(s)|^{\omega_-}A_s\left((t-s)|\Delta_-Y^+(s)|^\alpha\right);t\geq 0\right\},
\end{aligned}$$ where $(A_s)_{s\in\mathbb{R}_+}$ is a family of i.i.d. copies of $A$ under $\mathcal{P}_1$, mutually independent from $Y^+$. The fact that for all $t\geq 0$, $A(t)<\infty$ $\mathcal{P}_0^+$-a.s. follows from $\sum_{s\leq t}|\Delta_-Y^+(s)|^{\omega_-}<\infty$ $\mathbb{P}_0^+$-a.s. by Lemma 4.3 in [@BBCK16] and that each $A_s(t)$ is bounded for all $t\geq 0$ by its total mass $\mathcal{M}_s$ having expectation 1.
Recall the self-similarity property $(Y^+(t))_{t\geq 0}\stackrel{d}{=}(xY^+(tx^\alpha))_{t\geq 0}$ for any $x>0$. Combining this fact and the above equation yields, fixing $u\in\mathbb{R}$, that $$\begin{aligned}
&\left\{e^{\frac{\omega_-}{\alpha}(u+t)}A(e^{u+t});t\geq 0\right\}\\
&\hspace{1.5cm}=\left\{e^{\frac{\omega_-}{\alpha}(u+t)}\sum_{s\leq e^{u+t}}|\Delta_-Y^+(s)|^{\omega_-}A_{s}\left((e^{u+t}-s)|\Delta_-Y^+(se^{\omega_-u})|^\alpha\right);t\geq 0\right\}\\
&\hspace{1.5cm}=\left\{e^{\frac{\omega_-}{\alpha}(u+t)}\sum_{s\leq e^{t}}|\Delta_-Y^+(se^u)|^{\omega_-}A_{se^u}\left((e^{u+t}-se^u)|\Delta_-Y^+(se^{u})|^\alpha\right);t\geq 0\right\}\\
&\hspace{1.5cm}\stackrel{d}{=}\left\{e^{\frac{\omega_-}{\alpha}t}\sum_{s\leq e^t}|\Delta_-Y^+(s)|^{\omega_-}A_{se^u}\left((e^t-s)|\Delta_-Y^+(s)|^\alpha\right);t\geq 0\right\}\\
&\hspace{1.5cm}=\left\{e^{\frac{\omega_-}{\alpha}t}A(e^t);t\geq 0\right\},
\end{aligned}$$ which shows the claim.
Our goal in the rest of this section is to prove Propositions \[Theorem upper envelope A(t)\] and \[Proposition lower envelope A(t)\].
The upper bound
---------------
In this subsection, we do not assume regular variation of $\overline{\Lambda}$ as in , but only that $\kappa(\omega_++\omega_-+\alpha)<\infty$. We aim at proving Proposition \[Theorem upper envelope A(t)\]. Thanks to Proposition \[Theorem stationary area from 0\], we can restrict ourselves without loss of generality to the case where $t\to 0+$. We shall need the finiteness of $\mathcal{E}_0^+(t^{\omega_-/\alpha}A(t))$ for all $t>0$, in order to apply Markov’s Inequality on some well chosen sequence of events. Thanks to Proposition \[Theorem stationary area from 0\], it suffices to look at $\mathcal{E}_0^+(A(1))$.
\[Lemma E(A(1)) spectrally negative case\] Suppose that $\kappa(\omega_++\omega_-+\alpha)<\infty$. It holds that $\mathcal{E}_0^+(A(1))<\infty$.
Thanks to , under $\mathcal{P}_0^+$, we have that $$\begin{aligned}
A(1)
&=\sum_{s\leq 1}|\Delta_-Y^+(s)|^{\omega_-}A_s\left((1-s)|\Delta_-Y^+(s)|^\alpha\right)
\leq\sum_{s\leq 1}|\Delta_-Y^+(s)|^{\omega_-}\mathcal{M}_s,
\end{aligned}$$ where each $\mathcal{M}_s$ is distributed as $\mathcal{M}$ under $\mathcal{P}_1$ and is independent from $Y^+$. We can see the above upper bound as the stochastic integral of $s\mapsto \mathcal{M}_s$ with respect to the non-decreasing process $t\mapsto\sum_{s\leq t}|\Delta_-Y^+(s)|^{\omega_-}$ until time 1. In particular, we can use the optional projection theorem from [@DM2] Theorem 57 (see Theorem 43 of the same book for the definition of the optional projection), which states that the expectation of the stochastic integral is equal to the same expectation where each $\mathcal{M}_s$ has been replaced by its conditional expectation given the natural filtration of $t\mapsto\sum_{s\leq t}|\Delta_-Y^+(s)|^{\omega_-}$. In particular, for every negative jump time $s>0$, $\mathcal{E}_0^+(\mathcal{M}_s|(Y^+(u))_{u\leq s})=1$. Therefore, $$\begin{aligned}
\mathcal{E}_0^+\left(A(1)\right)
&\leq\mathcal{E}_0^+\left(\sum_{s\leq 1}|\Delta_-Y^+(s)|^{\omega_-}\right).
\end{aligned}$$ The expected value of the above sum is equal to the expectation of its predictable compensator, given in the proof of Lemma 4.3 in [@BBCK16]. We thus obtain that the upper bound is equal to $$\begin{aligned}
&\int_{{\left(-\infty\,,0\right)}}\Pi^+(\mathrm{d}y)(1-e^y)^{\omega_-}e^{\omega_+y}\mathcal{E}_0^+\left(\int_0^1Y^+(s)^{\omega_-+\alpha}\mathrm{d}s\right).
\end{aligned}$$ The first integral is finite by and . After using Tonelli’s Theorem and self-similarity, the other part of the expression becomes $$\begin{aligned}
\int_0^1 s^{\frac{\omega_-}{|\alpha|}-1}\mathrm{d}s\mathcal{E}_0^+\left(Y^+(1)^{\omega_-+\alpha}\right).
\end{aligned}$$ Hence, it only remains to check that the expectation is finite. If $|\alpha|\leq\omega_-$, then this is the case by Proposition 1(ii) in [@BY02], which ensures that the positive moments of $Y^+(1)$ are finite.
Suppose now that $\alpha<-\omega_-$. Theorem 1(iii) in [@BY01] shows that $$\begin{aligned}
\mathcal{E}_0^+\left(Y^+(1)^{\omega_-+\alpha}\right)
&=C\mathbb{E}_1^+\left(I_{\eta^+}^{\omega_-/\alpha}\right),
\end{aligned}$$ where $C$ is an explicit constant and $$\begin{aligned}
\label{Equation definition I_eta^+}
I_{\eta^+}:=\int_0^\infty\exp(\alpha\eta^+(t))\mathrm{d}t
\end{aligned}$$ (recall that $\eta^+(0)=0$ under $\mathbb{P}_1^+$). Proposition 2 in [@BY02] gives the finiteness and an expression for the negative moments of $I_{\eta^+}$, under the assumption that $e^{\eta^+(1)}$ admits positive moments of all order. However, the proof straightforwardly adapts to $\mathbb{E}_1^+(I_{\eta^+}^{\omega_-/\alpha})$ as soon as $\mathbb{E}_1^+(\exp((\omega_-+\alpha)\eta^+(1)))<\infty$, or equivalently $\kappa(\omega_++\omega_-+\alpha)<\infty$.
Let $\delta>0$. For all $n\geq 1$, we have that $$\begin{aligned}
\mathcal{P}_0^+\left(2^{(n+1)\frac{\omega_-}{|\alpha|}}A(2^{-n})>\log(2^{n+1})^{1+\delta}\right)
&=\mathcal{P}_0^+\left(A(1)>2^{\frac{\omega_-}{\alpha}}\log(2^{n+1})^{1+\delta}\right)\\
&\leq\frac{1}{n^{1+\delta}\log(2)^{1+\delta}}\mathcal{E}_0^+(A(1)).
\end{aligned}$$ The latter is finite by Lemma \[Lemma E(A(1)) spectrally negative case\] and therefore summable over $n$. Borel-Cantelli’s Lemma ensures that $\mathcal{P}_0^+$-a.s. for all $n$ sufficiently large, $A(2^{-n})\leq 2^{(n+1)\frac{\omega_-}{\alpha}}\log(2^{n+1})^{1+\delta}$. Since $A$ is non-decreasing, it means that $\mathcal{P}_0^+$-almost surely, for all $n$ large enough and all $t\in{\left[2^{-n-1}\,,2^{-n}\right)}$, it holds that $$\begin{aligned}
A(t)\leq t^{\frac{\omega_-}{|\alpha|}}\log(t)^{1+\delta},
\end{aligned}$$ which concludes the proof.
The lower bound
---------------
Our purpose now is to show Proposition \[Proposition lower envelope A(t)\]. We hence assume that holds. Similarly as for the upper bound, we can study only the asymptotic behaviour as $t\to\infty$ and easily deduce the one as $t\to 0+$, thanks to Proposition \[Theorem stationary area from 0\].
The strategy is to decompose $A(t)$ over the jumps of $Y^+$ as in , then to find two functions such that, $\mathcal{P}_0^+$-almost surely, the motion of $Y^+(s)$ is circumscribed in between them for all $s$ large enough.
### Upper and lower envelopes of the Eve cell
The so-called lower and upper envelopes of positive self-similar Markov processes are described respectively in [@CP06] and [@P09]. We need some preparations to apply the results of these two papers.
Let $U(x)$ be the last passage time of $Y^+$ below $x>0$, that is $U(x):=\sup\left\{t\geq 0:Y^+(t)\leq x\right\}$. Define $\nu:=Y^+(U(x)-)/x$ (note that if $\Lambda({\left(0\,,\infty\right)})=0$ then $\nu=1$ almost surely). When the process admits positive jumps, the law of $\nu$ is given in [@CP06] Lemma 1: it has its support included in ${\left[0\,,1\right]}$ and satisfies $$\begin{aligned}
\label{Equation law of nu}
\mathbb{P}(\nu\leq u)=\mathbb{E}(H(1)^{-1})\int_0^1\mathrm{d}v\int_{{\left(-\frac{\log(u)}{v}\,,\infty\right)}}y\Pi_H(\mathrm{d}y/|\alpha|),\qquad u\in{\left(0\,,1\right)},
\end{aligned}$$ where the subordinator $(H(s))_{s\geq 0}$ is the ascending ladder height process associated with $\eta^+$ and $\Pi_H$ is its Lévy measure (see Chapter VI of [@B96] for background).
The next lemma will be needed to apply the results describing the envelopes of $U$ and $Y^+$. Recall the definition of $I_{\eta^+}$ in .
\[Lemma tails of I\]
1. For all $0<p<(\omega_+-\omega_-)/|\alpha|$, it holds that $$\begin{aligned}
&\mathbb{P}_1(I_{\eta^+}\leq t)=o(t^{p})
\quad\text{and}\quad \mathbb{P}_1(I_{\eta^+}>1/t)=O(t),
\qquad\text{as}\quad t\to 0+.
\end{aligned}$$
2. Let $q^*:=\min\left\{\omega_+-\omega_-,\sup\left\{p\geq 0:\kappa(\omega_++p)<\infty\right\}\right\}$. For all $q<q^*$, it holds that $$\begin{aligned}
\mathbb{P}_1(\nu I_{\eta^+}\leq t)
=o(t^{q/|\alpha|}),\qquad\text{as}\quad t\to 0+,
\end{aligned}$$ where $\nu$ and $I_{\eta^+}$ are independent.
\(i) It essentially follows from [@R12] Lemma 3 that provides finiteness of some positive and negative moments of $I_{\eta^+}$. (Note that Lemma 3 in [@R12] gives $\mathbb{E}_1(I_{\eta^+}^{-1})=|\alpha\kappa'(\omega_+)|$, which is indeed finite under our assumption .)
\(ii) For all $u\in{\left(0\,,1\right)}$, equation entails that $$\begin{aligned}
\label{Equation left tail nu}
\mathbb{P}(\nu\leq u)
&\leq\mathbb{E}(H(1)^{-1})|\alpha|\int_{{\left(-\log(u)/|\alpha|\,,\infty\right)}}y\Pi_H(\mathrm{d}y)\nonumber\\
&=\mathbb{E}(H(1)^{-1})|\alpha|\left(-\frac{\log(u)}{|\alpha|}\overline{\Pi}_H(-\log(u)/|\alpha|)+\int_{-\log(u)/|\alpha|}^{\infty} \overline{\Pi}_H(z)\mathrm{d}z\right),
\end{aligned}$$ by Tonelli’s Theorem. Theorem 7.8 in [@K14] shows that the right tail of $\Pi_H$ is described for all $y>0$ by $$\begin{aligned}
\overline{\Pi}_H(y)
&=\int_{{\left[0\,,\infty\right)}}\widehat{U}(\mathrm{d}z)\overline{\Pi}^+(z+y),
\end{aligned}$$ where $\widehat{U}$ is the renewal measure of the descending ladder height process of $\eta^+$. Note that the integral is finite, see e.g. [@K14] Corollary 5.3. By assumption , there exists $q>0$ such that $q<\sup\left\{p\geq 0:\kappa(\omega_++p)<\infty\right\}$. In particular, $\kappa(\omega_++q)<\infty$, which is equivalent to $$\begin{aligned}
\int_{{\left(1\,,\infty\right)}} e^{qy}\Pi^+(\mathrm{d}y)
=\overline{\Pi}^+(1)+q\int_0^\infty e^{qx}\overline{\Pi}^+(x)\mathrm{d}x
<\infty,
\end{aligned}$$ so in particular $\overline{\Pi}^+(x)=o(e^{-qx})$ as $x\to\infty$. This leads to $$\begin{aligned}
\overline{\Pi}_H(y)
&\leq e^{-qy}\int_{{\left[0\,,\infty\right)}}\widehat{U}(\mathrm{d}z)e^{-qz}=O(e^{-qy})
\end{aligned}$$ Coming back to , we see that $$\begin{aligned}
\mathbb{P}(\nu\leq u)
&=O\left(|\log(u)|u^{q/|\alpha|}+\int_{0}^{u} \overline{\Pi}_H(-\log(y)/|\alpha|)\frac{\mathrm{d}y}{y}\right)
=O\left(|\log(u)|u^{q/|\alpha|}+\int_{0}^{u}y^{q/|\alpha|-1}\mathrm{d}y\right)\\
&=O\left(|\log(u)|u^{q/|\alpha|}\right).
\end{aligned}$$ For $\delta>0$ small enough, the same reasonning holds with $q$ replaced by $q+\delta$, so that $\mathbb{P}(\nu\leq u)=o(u^{q/|\alpha|})$.
To conclude, suppose that $X_1$ and $X_2$ are independent positive random variables such that for $i=1,2$, $\mathbb{P}(X_i\leq t)=o(t^{q_i})$ as $t\to0+$, for some $q_1>q_2>0$. Note that there exists a constant $C$ such that $\mathbb{P}(X_2\leq t)\leq Ct^{q_2}$ for all $t\geq 0$. Then, we write $$\begin{aligned}
\mathbb{P}_1(X_1X_2\leq t)
&\leq\int_{\mathbb{R}_+}\mathbb{P}(X_1\in\mathrm{d}x)\mathbb{P}(X_2\leq t/x)
\leq\int_{\mathbb{R}_+}\mathbb{P}(X_1\in\mathrm{d}x)C(t/x)^{q_2}
=O(t^{q_2}),
\end{aligned}$$ since $q_1>q_2$. The bound for $\mathbb{P}_1(\nu I_{\eta^+}\leq t)$ as $t\to 0+$ thus follows from this fact and part (i).
Let $\underline{p}>1/|\alpha|$ and $\overline{p}>1/q^*$ where $q^*$ is defined in Lemma \[Lemma tails of I\]. Define on ${\left(1\,,\infty\right)}$ the functions $$\begin{aligned}
\label{Equation definition of g^arrow general}
&g^{\uparrow}:t\mapsto
t^{1/|\alpha|}\log(t)^{\overline{p}},
\hspace{1cm} g^{\downarrow}:t\mapsto
t^{1/|\alpha|}\log(t)^{-\underline{p}}.
\end{aligned}$$ Note that $g^{\uparrow}$ and $g^{\downarrow}$ are regularly varying at $\infty$ with index $1/|\alpha|$. Let $\overline{g}(t):=\sup_{s\leq t}g^{\uparrow}(s)$ and similarly, let $\underline{g}(t):=\inf_{s\geq t}g^{\downarrow}(s)$. These functions have the nice property to be strictly increasing and by Theorem 1.5.3. in [@BGT87] we have that $$\begin{aligned}
\label{Equation monotone equivalents}
\overline{g}(t)\underset{t\to \infty}{\sim} g^{\uparrow}(t)\quad\text{and}\quad\underline{g}(t)\underset{t\to \infty}{\sim} g^{\downarrow}(t).
\end{aligned}$$
\[Lemma envelopes Y and U\] $\mathbb{P}_0^+$-almost surely, there exists $s>0$ such that $$\begin{aligned}
\underline{g}(t)
&<Y^+(t)
<\overline{g}(t),\hspace{1.05cm}\forall t\in{\left(s\,,\infty\right)}.
\end{aligned}$$
First, note that $Y^+(t)<x$ implies $U(x)>t$. Let $T(1):=\inf\{t\geq 0:Y^+(t)\geq 1\}$ and note that $T(1)\leq U(1)$. Proceeding as in the proof of Proposition 6 in [@P09], one can show that $$\begin{aligned}
\mathbb{P}(T(1)<t)\leq C\mathbb{P}(\nu I_{\eta^+}\leq t).
\end{aligned}$$ Lemma \[Lemma tails of I\](ii) in this paper and Proposition 4 in [@P09] then show the claim for $\overline{g}$. The statement involving $\underline{g}$ is shown using Theorem 1(i) in [@CP06] and our Lemma \[Lemma tails of I\](i).
### Proof of Proposition \[Proposition lower envelope A(t)\]
For clarity purpose, we prove two Lemmas that will make the proof of Proposition \[Proposition lower envelope A(t)\] straightforward. We need some notation. Define $$\begin{aligned}
\label{Equation definition f}
f(t):=t\log(t)^{\alpha\overline{p}}.
\end{aligned}$$ We note for later use that and implies that $$\begin{aligned}
\label{Equation asymptotic g(f)}
\underline{g}(f(t))&\underset{t\to\infty}{\sim}t^{\frac{1}{|\alpha|}}\log(t)^{-\underline{p}-\overline{p}},
\hspace{1cm}\overline{g}(f(t))\underset{t\to\infty}{\sim}t^{\frac{1}{|\alpha|}}.
\end{aligned}$$ To control the fluctuations of $Y^+$, we work on the event $$\begin{aligned}
E_t:=&\left\{\forall s\geq f(t):\underline{g}(s)<Y^+(s)<\overline{g}(s)\right\}.
\end{aligned}$$ Note that $f$ being eventually increasing, Lemma \[Lemma envelopes Y and U\] ensures that $\mathbb{P}_0^+(\liminf_{t\to\infty}E_t)=1$. Moreover, noting that $f(t)/t\to 0$ as $t\to\infty$, we see that $J_t:={\left[f(t)\,,f(2t)\right]}\subset{\left[0\,,t\right]}$.
\[Lemma lower bound A\] For all $a>0$, $\mathcal{P}_0^+$-almost surely, it holds on the event $E_t$ that $$\begin{aligned}
A(t)\geq t^{\frac{\omega_-}{|\alpha|}}\log(t)^{-\omega_-(a+\underline{p}+\overline{p})}\sum_{s\in J_t}A_s(1/3)\mathds{1}_{\left\{\Delta_-\eta^+(\tau_s)\leq \log(1-\log(t)^{^{-a}})\right\}},
\end{aligned}$$ for all $t$ large enough.
We need the lower envelope of $Y^+$ to ensure that the jumps are large enough, and the upper envelope of $Y^+$ for the area generated from it to be not too small. On the event $E_t$, we have $$\begin{aligned}
A(t)
&=\sum_{s\leq t}|\Delta_-Y^+(s)|^{\omega_-}A_s\left((t-s)|\Delta_-Y^+(s)|^{\alpha}\right)\\
&\geq\sum_{s\in J_t}|\Delta_-Y^+(s)|^{\omega_-}A_s\left((t-s)|\Delta_-Y^+(s)|^{\alpha}\right)\\
&\geq\sum_{s\in J_t}\underline{g}(s)^{\omega_-}(1-e^{\Delta_-\eta^+(\tau_s)})^{\omega_-}A_s\left((t-s)\overline{g}(s)^{\alpha}(1-e^{\Delta_-\eta^+(\tau_s)})^{\alpha}\right)\\
&\geq\sum_{s\in J_t}\underline{g}(f(t))^{\omega_-}(1-e^{\Delta_-\eta^+(\tau_s)})^{\omega_-} A_s\left(\left(t-f(2t)\right)\overline{g}(f(2t))^{\alpha}\left(1-e^{\Delta_-\eta^+(\tau_s)}\right)^{\alpha}\right).
\end{aligned}$$ Note that $f(2t)/t\to 0$ and entails that $t\overline{g}(f(2t))^{\alpha}\to 1/2$ as $t\to\infty$. Hence, taking $t$ large enough, we see for each $s\in J_t$, $$\begin{aligned}
A_s\left(\left(t-f(2t)\right)\overline{g}(f(2t))^{\alpha}\left(1-e^{\Delta_-\eta^+(\tau_s)}\right)^{\alpha}\right)\geq A_s(1/3).
\end{aligned}$$ We write $$\begin{aligned}
A(t)
&\geq\underline{g}(f(t))^{\omega_-}\sum_{s\in J_t}(1-e^{\Delta_-\eta^+(\tau_s)})^{\omega_-}A_s\left(1/3\right)\mathds{1}_{\left\{\Delta_-\eta^+(\tau_s)<\log(1-\log(t)^{-a})\right\}}\\
&\geq\underline{g}(f(t))^{\omega_-}\log(t)^{-a\omega_-}\sum_{s\in J_t}A_s\left(1/3\right)\mathds{1}_{\left\{\Delta_-\eta^+(\tau_s)<\log(1-\log(t)^{-a})\right\}}\\
&\!\!\!\underset{t\to\infty}{\sim}t^{\frac{\omega_-}{|\alpha|}}\log(t)^{-\omega_-(a+\underline{p}+\overline{p})}\sum_{s\in J_t}A_s\left(1/3\right)\mathds{1}_{\left\{\Delta_-\eta^+(\tau_s)<\log(1-\log(t)^{-a})\right\}},
\end{aligned}$$ where we used .
For the lower bound of Lemma \[Lemma lower bound A\] to be useful, we need to make sure that the sum contains at least one non negligeable term. We shall use the following:
\[Lemma bound number of jumps\] For all $a>\max\{0,\frac{|\alpha|}{\rho}(\overline{p}-\underline{p})\}$, $\mathcal{P}_0^+$-almost surely, for all $t$ large enough, the number of elements in $\left\{s\in J_t:\Delta_-\eta^+(\tau_s)\leq\log(1-\log(t)^{-a})\right\}$ is stochastically bounded by a Poisson random variable with parameter $$\begin{aligned}
\frac{1}{2}\log(t)^{|\alpha|(\underline{p}-\overline{p})+a\rho}.
\end{aligned}$$
By definition of the Lamperti time-change $s\mapsto\tau_s$, we have that $$\begin{aligned}
f(2t)-f(t)
=\int_{\tau_{f(t)}}^{\tau_{f(2t)}}e^{|\alpha|\eta^+(s)}\mathrm{d}s,
\end{aligned}$$ so in particular, on the event $E_t$ such that $\eta^+(\tau_s)<\log(\underline{g}(2t))$ for all $s\in J_t$, we get that $$\begin{aligned}
\left(f(2t)-f(t)\right)\underline{g}(2t)^{\alpha}\leq\tau_{f(2t)}-\tau_{f(t)}
\end{aligned}$$ Therefore, changing the variables, the domain of the sum with respect to $u=\tau_s$ becomes at least of length $$\begin{aligned}
\frac{f(2t)-f(t)}{\underline{g}(2t)^{|\alpha|}}
&\underset{t\to\infty}{\sim}\frac{2t|\log(2t)|^{\alpha\overline{p}}-t|\log(t)|^{\alpha\overline{p}}}{2t|\log(2t)|^{\alpha\underline{p}}}
\underset{t\to\infty}{\sim}\frac{1}{2}|\log(t)|^{\alpha(\overline{p}-\underline{p})}
\end{aligned}$$ Hence, $\eta^+$ being a Lévy process, the number of $s\in J_t$ such that $\mathds{1}_{\left\{\Delta_-\eta^+(\tau_s)<\log(1-\log(t)^{-a})\right\}}=1$ is greater than a Poisson random variable with parameter $$\begin{aligned}
&\frac{1}{2}|\log(t)|^{\alpha(\overline{p}-\underline{p})}\Pi^+\left({\left(-\infty\,,\log(1-\log(t)^{-a})\right)}\right)\\
&\hspace{2cm}\underset{t\to\infty}{\sim}\frac{1}{2}|\log(t)|^{\alpha(\overline{p}-\underline{p})}\Pi^+\left({\left(-\infty\,,-\log(t)^{-a}\right)}\right)
\geq\frac{1}{2}|\log(t)|^{\alpha(\overline{p}-\underline{p})+a\rho},
\end{aligned}$$ where we used to write $$\begin{aligned}
\Pi^+\left({\left(-\infty\,,-\log(t)^{-a}\right)}\right)
&\geq\int_{{\left(-\infty\,,-\log(t)^{-a}\right)}}e^{\omega_+y}\Lambda(\mathrm{d}y)
=O\left(\log(t)^{a\rho}\right),
\end{aligned}$$ as $t\to\infty$, by and Theorem 1.6.4 of [@BGT87].
Recall that Lemma \[Lemma envelopes Y and U\], ensures that $\mathbb{P}_0^+(\liminf_{t\to\infty}E_t)=1$. In view of Lemma \[Lemma lower bound A\], it suffices to establish that the sum in its lower bound is almost surely bounded away from 0 for all $t$ large enough.
Choose $a>\max\{0,\frac{|\alpha|}{\rho}(\overline{p}-\underline{p})\}$ and, to ease the notation, let $\delta:=\alpha(\overline{p}-\underline{p})+a\rho$ and note that $\delta>0$. Let $n\geq 1$. Thanks to Lemma \[Lemma bound number of jumps\], we can write $$\begin{aligned}
&\mathcal{P}_0^+\left(\sum_{s\in J_{2^n}}A_{s}\left(1/3\right)\mathds{1}_{\left\{\Delta_-\eta^+(\tau_s)<\log\left(1-\log(2^n)^{-a}\right)\right\}}\leq 1\right)\\
&\hspace{5cm}= O\left(e^{-\log(2^n)^{\delta}}\sum_{m\geq 0}\frac{\log(2^n)^{\delta m}}{m!}\mathcal{P}_1(A(1/3)\leq 1)^m\right)\\
&\hspace{5cm}= \exp\left(-n^{\delta}\log(2)^{\delta}\mathcal{P}_1(A(1)>1/3)\right).
\end{aligned}$$ The latter being summable, Borel-Cantelli’s Lemma shows that $\mathcal{P}_0^+$- almost surely, there exists $n_0\geq 1$ such that for all $n\geq n_0$, it holds that $$\begin{aligned}
\sum_{s\in J_{2^n}}A_{s}\left(1/3\right)\mathds{1}_{\left\{\Delta_-\eta^+(\tau_s)<\log\left(1-\log(2^n)^{-a}\right)\right\}}>1.
\end{aligned}$$ This, the facts that $\bigcup_{n\geq n_0}J_{2^{n}}={\left[f(2^{n_0})\,,\infty\right)}$, that $A$ is non-decreasing and Lemma \[Lemma lower bound A\] together imply that on the event $E_t$, $\mathcal{P}_0^+$-almost surely it holds that $$\begin{aligned}
\label{Equation exponent lower envelope}
A(t)
&\geq t^{\frac{\omega_-}{|\alpha|}}\log(t)^{-\omega_-(a+\underline{p}+\overline{p})},
\end{aligned}$$ which ensures the claim.
\[Remark exponent lower envelope\] Although we do not claim that it is optimal, Equation shows that Proposition \[Proposition lower envelope A(t)\] applies for all $q>\omega_-(1/|\alpha|+1/q^*+\max\{0,\frac{|\alpha|}{\rho}(1/q^*-1/|\alpha|)\})$, where $q^*$ is defined in Lemma \[Lemma tails of I\]. Indeed, we have respectively chosen $\underline{p}$ arbitrary close from above to $1/\alpha$ and $\overline{p}$ arbitrary close from above to $1/q^*$.
Application to random maps {#Section Application to random maps}
==========================
#### A specific family of growth-fragmentations.
We start this section by recalling the connection between growth-fragmentations and random surfaces that has been observed in [@BCK15] for Boltzmann triangulations approximating Brownian disks, and that has been generalised in [@BBCK16] to a broader family of Boltzmann maps approximating stable disks and plane.
More details on what follows can be found in [@BBCK16]. Let $\theta\in{\left(1\,,3/2\right]}$ and for all $q\in{\left(\theta\,,2\theta+1\right)}$, let $$\begin{aligned}
\label{Equation kappa_theta}
\kappa_\theta(q):=\frac{\cos(\pi(q-\theta))}{\sin(\pi(q-2\theta))}\cdot\frac{\Gamma(q-\theta)}{\Gamma(q-2\theta)}.
\end{aligned}$$ Thus defined, $\kappa_\theta$ is the cumulant of a specific self-similar growth-fragmentation $\mathbf{X}_\theta$. Let its index of self-similarity be $\alpha=1-\theta$.
Informally, the collection of cycles’ lengths observed at heights in some discrete random maps with large boundary converges, when properly rescaled, towards $\mathbf{X}_\theta$, where $\theta$ depends on the tail of the distribution of the degree of a typical face (see Theorem 6.8 in [@BBCK16]). The Brownian case corresponds to $\theta=3/2$. This means that we obtain $\mathbf{X}_{3/2}$ under $\mathcal{P}_1$ (respectively under $\mathcal{P}_0^+$) in the scaling limit of the sliced approximation of the free[^3] Brownian disk (respectively plane); As noted in introduction, $\mathbf{X}_{3/2}$ also appears when directly slicing the Brownian disk or the Brownian plane (see [@L17] Theorem 3 and 23).
In the case of $\mathbf{X}_\theta$, $\eta^+$ and $\eta^-$ belong to the class of hypergeometric Lévy processes, see Proposition 5.2 in [@BBCK16] for this fact and [@KP13] for a definition and references on hypergeometric Lévy processes. This yields an explicit expression of the densities of $\Pi^+$ and $\Pi^-$.
\[Lemma integrability Pi\^+\] Let $c_-:=\frac{\Gamma(\theta+1)}{\pi}$ and $c_+:=\frac{\Gamma(\theta+1)}{\pi}\sin(\pi(\theta-1/2))$. The densities of $\Pi^+$ and $\Pi^-$ are given by $$\begin{aligned}
&\Pi^+(\mathrm{d}y)/\mathrm{d}y
=c_-\frac{e^{3y/2}}{(1-e^y)^{\theta+1}}\mathds{1}_{\left\{y<0\right\}}+c_+\frac{e^{3y/2}}{(e^y-1)^{\theta+1}}\mathds{1}_{\left\{y>0\right\}},\\
&\Pi^-(\mathrm{d}y)/\mathrm{d}y
=c_-\frac{e^{y/2}}{(1-e^y)^{\theta+1}}\mathds{1}_{\left\{y<0\right\}}+c_+\frac{e^{y/2}}{(e^y-1)^{\theta+1}}\mathds{1}_{\left\{y>0\right\}}.
\end{aligned}$$
In the proof of Proposition 5.2 in [@BBCK16], it is shown that the density $h$ of the image of $\Pi^+$ by $x\mapsto e^x$ is given by $$\begin{aligned}
h(z)
&=\frac{\Gamma(\theta+1)}{\pi}\frac{z^{1/2}}{(1-z)^{\theta+1}}\mathds{1}_{\left\{0<z<1\right\}}+\frac{\Gamma(\theta+1)\sin(\pi(\theta-1/2))}{\pi}\frac{z^{1/2}}{(z-1)^{\theta+1}}\mathds{1}_{\left\{z>1\right\}}.
\end{aligned}$$ The expression of $\Pi^+(\mathrm{d}y)/\mathrm{d}y$ follows from a straightforward change of variables. One then gets the expression of $\Pi^-(\mathrm{d}y)/\mathrm{d}y$ from [@CKP09] Section 2 by identifying $c_-$ and $c_+$.
From , one sees that is satisfied with $\omega_-=\theta+1/2$ and $\omega_+=\theta+3/2$. Moreover, one can check that the assumption is also satisfied with $\rho=\theta$, by looking at the behaviour of $\Pi^-({\left(-\infty\,,\log(x)\right)})$, as explained after . This means that our results apply in particular to $\mathbf{X}_\theta$. To sum up, the parameters in terms of $\theta$ are $$\begin{aligned}
\label{Equation parameters in terms of theta}
\begin{cases}
\alpha&=1-\theta\\
\omega_-&=\theta+1/2\\
\omega_+&=\theta+3/2\\
\rho&=\theta
\end{cases}
\end{aligned}$$
#### Area of a a small annular in Brownian and stable disks.
Le Gall showed the following in [@L17] Theorem 3: denote $\mathbf{v}$ the intrinsic area measure of the free Brownian disk with boundary size $x>0$, and let $B_\epsilon$ be the annular of width $\epsilon$ from the boundary (the set of points at distance smaller than $\epsilon$ from the boundary). Then it holds that $$\begin{aligned}
\label{Equation Le Gall}
\lim_{\epsilon\to 0+}\epsilon^{-2}\mathbf{v}(B_\epsilon)=x,\quad\text{almost surely.}
\end{aligned}$$ As explained in the beginning of the section, the above convergence can be translated in terms of $\mathbf{X}_{3/2}$ under $\mathcal{P}_x$, where the area of the annular corresponds to $A(\epsilon)$. We check whether we retrieve the same result using our theorem \[Theorem area GF\].
Consider $\mathbf{X}_{\theta}$ for $\theta\in{\left(1\,,3/2\right]}$. The computations before Theorem \[Theorem area GF\] show that $$\begin{aligned}
&\overline{\Lambda}(\epsilon)
=\Lambda({\left(-\infty\,,-\epsilon\right)})
\underset{\epsilon\to 0+}{\sim}\frac{1/2}{\theta}\epsilon^{-\theta-1/2}\Pi^-\left({\left(-\infty\,,\log(\epsilon)\right)}\right),\\
\end{aligned}$$ which thanks to Lemma \[Lemma integrability Pi\^+\] can be written as $\overline{\Lambda}(\epsilon)\sim\frac{c_-}{\theta}\epsilon^{\theta}$. Recalling , Theorem \[Theorem area GF\] thus reads as follows: For all $x>0$, it holds that $$\begin{aligned}
\lim_{\epsilon\to 0+}\epsilon^{-\frac{\theta-1/2}{\theta-1}}A(\epsilon)=\frac{2(\theta-1)}{(\theta-1/2)}\cdot c_-\mathbb{E}_1^-\left(I^{-\frac{1}{2(\theta-1)}}\right)x,\quad\mathcal{P}_x\text{-a.s. and in }\mathbb{L}^1.
\end{aligned}$$ Note that for $\theta=3/2$, we get $$\begin{aligned}
\lim_{\epsilon\to 0+}\epsilon^{-2}A(\epsilon)
&=c_-\mathbb{E}_1^-\left(I^{-1}\right)x
=\frac{\Gamma(5/2)}{2\pi}|\kappa_{3/2}'(2)|x,\end{aligned}$$ where the value of the expectation is given in [@CPY97] Proposition 3.1(iv). (Note that since $I=\int_0^\infty\exp(\eta^-(t)/2)\mathrm{d}t$, we get $\mathbb{E}_1^-(I^{-1})=|\mathbb{E}_1^-(\eta^-(1)/2)|$.) Because $\kappa_{3/2}(2)=0$ and using the explicit expression of $\kappa_{3/2}$ given in , we have that $$\begin{aligned}
\kappa_{3/2}'(2)
&=\lim_{q\to 2}\frac{\Gamma(q-3/2)}{(q-2)\Gamma(q-3)}
=\frac{\Gamma(1/2)}{\text{Res}_{\Gamma}(-1)}
=-\sqrt{\pi},\end{aligned}$$ where we used that the residue of $\Gamma$ at $-1$ is $-1$ and $\Gamma(1/2)=\sqrt{\pi}$. Well known properties of the gamma function entails that $\Gamma(5/2)=3\Gamma(1/2)/4=3\sqrt{\pi}/4$. We then obtain $$\begin{aligned}
\epsilon^{-2}A(\epsilon)\underset{\epsilon\to 0+}{\sim}\epsilon^2\frac{3}{8}x.\end{aligned}$$ Note that we get an additional factor $3/8$ compare to . The cumulant function $\kappa$ in [@LR18] Section 11.1 is equal to $\sqrt{8/3}\cdot\kappa_{3/2}$, which simply corresponds to multiplying the distance in the Brownian map by a constant. This means in particular, considering this $\kappa$ instead of ours, that $\eta^-$ becomes $\sqrt{8/3}\cdot \eta^-$, and the constant $c_-$ of Lemma \[Lemma integrability Pi\^+\] becomes $\sqrt{8/3}\cdot c_-$. The above is therefore consistent with .
#### Area of a small ball in Brownian and stable maps.
If we read Proposition \[Theorem stationary area from 0\] in terms of $\mathbf{X}_\theta$, we get that the law of $t^{(\theta+1/2)/(1-\theta)}A(t)$ is the same for all $t>0$. This coincides with the volume growth exponent of infinite Boltzmann planar maps in the so-called *dilute phase*, meaning that the number of vertices at height $n$ as $n\to\infty$ in such maps grows at a speed $n^{(\theta+1/2)/(1-\theta)}$, see equation (4.3) in Theorem 4.2 of [@BC17].
When $\theta=3/2$, this exponent is equal to $-4$. In [@L07] Lemma 6.2 (see also [@M14] Lemma 4.4.4), it is shown that in this case, for any $\delta>0$, it holds that $$\begin{aligned}
\limsup_{t\to 0+}t^{-4+\delta}A(t)=0,\quad\mathbb{P}_0^+\text{-a.s.}
\end{aligned}$$
The assumption $\kappa(\omega_++\omega_-+\alpha)<\infty$ in Proposition \[Theorem upper envelope A(t)\] reads as $\kappa_{3/2}(5/2)<\infty$, which is indeed the case by . In particular, our propositions \[Theorem upper envelope A(t)\] and \[Proposition lower envelope A(t)\] thus improve the above, replacing $t^{\delta}$ by a power of $|\log(t)|$ and considering the $\liminf$ as well. The exponent $q_0$ on the $\log$ of the lower bound is given in Remark \[Remark exponent lower envelope\], in terms of some constant $q^*$ defined in Lemma \[Lemma tails of I\]. One can check that for any $\theta\in{\left(1\,,3/2\right]}$, one has that $q^*=\min\{1,\theta-1/2\}=\theta-1/2$, and therefore $q_0=(\theta+1/2)\left((\theta-1)^{-1}+(\theta-1/2)^{-1}\right)$. In the Brownian case $\theta=3/2$, we obtain $q_0=6$.
Thus, for $\mathbf{X}_\theta$ with any $\theta\in{\left(1\,,3/2\right]}$, the following holds: for all $\delta>0$ and $q>q_0$ as above, we have that $$\begin{aligned}
&\limsup_{t\to 0\text{ or }\infty}|\log(t)|^{-1-\delta}t^{\frac{\theta+1/2}{1-\theta}}A(t)=0,\quad\mathbb{P}_0^+\text{-a.s.}\\
&\liminf_{t\to 0\text{ or }\infty}|\log(t)|^{q}t^{\frac{\theta+1/2}{1-\theta}}A(t)=\infty,\quad\mathbb{P}_0^+\text{-a.s.}
\end{aligned}$$
[^1]: Insitut für Mathematik, Universität Zürich, Switzerland. E-mail: [email protected]
[^2]: See the argument of [@BBCK16] Lemma 2.1 for the fact that $\phi_-$ defines indeed a Laplace exponent of a Lévy process.
[^3]: *free* refers to the fact that the total intrinsic area of the surface is not fixed but random, with law described in [@BC17] Proposition 4.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Precise understanding of nonlinear evolution of cosmological perturbations during inflation is necessary for the correct interpretation of measurements of non-Gaussian correlations in the cosmic microwave background and the large-scale structure of the universe. The “$\delta N$ formalism” is a popular and powerful technique for computing nonlinear evolution of cosmological perturbations on large scales. In particular, it enables us to compute the curvature perturbation, $\zeta$, on large scales without actually solving perturbed field equations. However, people often wonder why this is the case. In order for this approach to be valid, the perturbed Hamiltonian constraint and matter-field equations on large scales must, with a suitable choice of coordinates, take on the same forms as the corresponding unperturbed equations. We find that this is possible when (1) the unperturbed metric is given by a homogeneous and isotropic Friedmann-Lemaître-Robertson-Walker metric; and (2) on large scales and with a suitable choice of coordinates, one can ignore the shift vector ($g_{0i}$) as well as time dependence of tensor perturbations to $g_{ij}/a^2(t)$ of the perturbed metric. While the first condition has to be assumed [*a priori*]{}, the second condition can be met when (3) the anisotropic stress becomes negligible on large scales. However, in order to explicitly show that the second condition follows from the third condition, one has to use gravitational field equations, and thus this statement may depend on the details of the theory of gravitation. Finally, as the $\delta N$ formalism uses only the Hamiltonian constraint and matter-field equations, it does not [ *a priori*]{} respect the momentum constraint. We show that the error in the momentum constraint only yields a decaying mode solution for $\zeta$, and the error vanishes when the slow-roll conditions are satisfied.'
author:
- 'Naonori S. Sugiyama'
- Eiichiro Komatsu
- Toshifumi Futamase
title: '$\delta N$ formalism'
---
Introduction
============
Given the success of cosmological linear perturbation theory, the focus has shifted to [*nonlinear*]{} evolution of cosmological perturbations. As the magnitude of the primordial curvature perturbation is of order $10^{-5}$, any nonlinearities are expected to be small; however, such nonlinearities can be measured using non-Gaussian correlations of cosmological perturbations (such as temperature and polarization anisotropy of the cosmic microwave background [@Komatsu:2010] and density fluctuations in the large-scale structure of the universe [@Desjacques/Seljak:2010]). For this reason, precise understanding of the nonlinear evolution of cosmological perturbations is of great interest in cosmology.
The so-called “$\delta N$ formalism” [@Starobinsky:1982; @*Starobinsky:1986; @Salopek/Bond:1990; @Sasaki/Stewart:1996; @David/etal:2005; @Lyth/Rodriguez:2005] is a popular technique for computing non-linear evolution of cosmological perturbations on large scales. Here, by “large scales,” we mean the scales greater than the Hubble horizon, in a sense that the comoving wavenumber of perturbations $k$ is much less than the reciprocal of the comoving Hubble length, i.e., $k\ll aH$. In particular, it enables us to compute the curvature perturbation $\zeta$ without actually solving the perturbed field equations. In this paper, we show why this is the case by rederiving the $\delta N$ formalism using the gradient expansion method as applied to Einstein’s field equations and scalar-field equations in the flat gauge. The usual derivation of the $\delta N$ formalism is based on the so-called “separate universe” approach [@Wands/etal:2000], which assumes the existence of a locally homogeneous (but not necessarily isotropic [@Dimopoulos/etal:2009]) region smoothed over some large length scale. We provide a support for this assumption by considering a global region including many such smoothed local regions and show that they behave as if they were locally homogeneous regions which evolve independently from each other. In so doing, we point out a subtlety regarding the momentum constraint, which is not [*a priori*]{} respected by the $\delta N$ formalism.
The organization of this paper is as follows. In Sec. II, we describe our basic setup including the metric, gauge, and scalar-field Lagrangian. In Sec. III, we review the gradient expansion method, which constitutes the basis for the $\delta N$ formalism. In Sec. IV, we rederive the $\delta N$ formalism. In Sec. V, we give the sufficient conditions for the validity of the $\delta N$ formalism and conclude.
Basic Setup
===========
Metric
------
We write the spacetime metric in the Arnowitt-Deser-Misner (ADM) form, which is the standard (3+1)-decomposition of the metric [@ADM:1962] $$ds^2 = - \alpha^2 dt^2 + \gamma_{ij} \left( dx^i + \beta^i dt \right)\left( dx^j + \beta^j dt \right) ,
\label{}$$ where $\gamma_{ij}$ is decomposed as follows: $$\gamma_{ij} \equiv a^2 e^{2\psi} \left( e^h \right)_{ij}.
\label{}$$ Here, $a$ is the scale factor which depends only on time, and $\psi$ is the scalar perturbation to the spatial curvature. The traceless tensor $h_{ij}$ is further decomposed as $$h_{ij} = \partial_i C_j + \partial_j C_i - \frac{2}{3} \delta_{ij} \partial_k C^k + h^{(T)}_{ij},
\label{}$$ where $C_i$ contains both scalar and vector perturbations, whereas $h_{ij}^{(T)}$ represents tensor perturbations.
We decompose the extrinsic curvature $K_{ij}$ into a trace part $K$ and a traceless part $\tilde{A}_{ij}$ as[^1] $$K_{ij} =\frac{\gamma_{ij}}{3}K + a^2e^{2\psi}\tilde{A}_{ij}.$$ Einstein’s field equations written in terms of these variables are summarized in Appendix A.
Flat gauge
----------
In this paper, we shall fix the gauge completely (i.e., leaving no gauge degree of freedom) by imposing the following gauge-fixing condition[^2]: $$\psi = C_i = 0.
\label{gauge-fixing}$$ Therefore, the spatial metric is described only by the scale factor and tensor perturbations as $\gamma_{ij} = a^2 [e^{h^{(T)}}]_{ij}$. This gauge was also used by [@Maldacena:2003] \[see his Eq. (3.2)\].
We shall call this gauge the “flat gauge” throughout this paper. Note that a flat-gauge condition in the literature sometimes does not include $C_i=0$. In such a case a residual gauge degree of freedom would remain. In the flat gauge, the metric is given by $$ds^2 = - \alpha^2 dt^2 + a^2(t) [e^{h^{(T)}}]_{ij} \left( dx^i + \beta^i dt \right)\left( dx^j + \beta^j dt \right).
\label{}$$ The variables in this metric such as $\alpha$, $\beta_i$, and $h^{(T)}_{ij}$ contain [*nonlinear*]{} perturbations. However, we shall assume that the unperturbed metric is still given by a homogeneous and isotropic Friedmann-Lemaître-Robertson-Walker metric: $$ds^2 = - dt^2 + a^2(t) \delta_{ij} dx^idx^j.
\label{background}$$ Therefore, our argument below does not hold if the unperturbed metric is not given by Eq. (\[background\]).
Scalar-field Lagrangian
-----------------------
We shall consider a universe filled with scalar fields: $$\mathcal{L} = - \frac{1}{2} G_{IJ}\partial^{\mu} \varphi^I \partial_{\mu} \varphi^J - V .
\label{}$$ The capital latin indices ($I$, $J$, etc.) denote scalar-field components running from 1 to $n$ where $n$ is the number of scalar fields. Here, $G_{IJ}$ is the metric tensor for scalar-field space. For simplicity, we shall take the canonical kinetic term $G_{IJ} = \delta_{IJ}$ for the moment. We then argue later (in Sec. \[sec:noncanonical\]) that the results are also valid for noncanonical kinetic terms in the first order of the gradient expansion.
With this Lagrangian, the stress-energy tensor and the field equation of scalar fields are given by $$T_{\mu\nu} =G_{IJ} \partial_{\mu}\varphi^I \partial_{\nu} \varphi^J
+ g_{\mu\nu} \left( - G_{IJ} \frac{1}{2} \partial^{\alpha} \varphi^I \partial_{\alpha} \varphi^J - V\right),
\label{}$$ $$\frac{1}{\sqrt{g}} \partial_{\mu} \left(\sqrt{g} g^{\mu\nu} \partial_{\nu} \varphi^I \right) - V_I = 0,
\label{motion}$$ where $V_I \equiv \partial V/ \partial \varphi^I$.
Gradient expansion method
=========================
Ordering in the gradient expansion {#sec:estimation}
----------------------------------
Since we are interested in nonlinear perturbations on superhorizon scales, we shall expand field equations in the number of spatial derivatives: this is called the gradient expansion method [@Salopek/Bond:1990; @David/etal:2005]. In this method, the ratio of the comoving wavenumber and the comoving Hubble scale, $$\epsilon \equiv \frac{k}{aH}$$ is taken to be a small parameter.
Before we proceed, let us emphasize that we assume the validity of perturbative expansion: namely, while we shall deal with nonlinear perturbations, we assume that the $(i+1)$th-order perturbations are smaller than the $i$th-order perturbations. This means that we have [*two*]{} smallness parameters: one is the number of derivatives, $\epsilon$, and the other is a smallness parameter of perturbation theory $\delta$, which corresponds to $\psi$, $\beta^i$, $\varphi^I-\bar{\varphi}^I$, etc.
These two parameters should satisfy the following condition: $$\delta < \epsilon \ll 1.
\label{condition}$$ This is because, if we take the smoothing length to be infinitely large, i.e., $\epsilon\to 0$, then the perturbation must vanish, i.e., $\delta\to 0$. Then, the metric must approach the unperturbed metric given by Eq. (\[background\]) as we take $\epsilon\to 0$. In other words, the amplitude of the perturbations should be limited by the smoothing length we take.
We now estimate the ordering of perturbation variables in terms of the gradient expansion. First, we demand that all physical quantities do not vanish in the lowest order of the gradient expansion: $$\begin{aligned}
&\alpha-1 = \tilde{\gamma}_{ij} - \delta_{ij} = \varphi^I - \bar{\varphi}^I = \mathcal{O}(\epsilon^0,\delta), \notag \\
&\hspace{1cm} \beta^i = {\cal O}(\epsilon^{-1},\delta).
\label{first-estimation}\end{aligned}$$ We do not include $\psi$ here because we work in the flat gauge. We have defined $\tilde{\gamma}_{ij}\equiv
\gamma_{ij}/[a^2(t)e^{2\psi}]$, which is equal to $[e^{h^{(T)}}]_{ij}$ in the flat gauge.
One may wonder why we chose to start with $\beta^i={\cal
O}(\epsilon^{-1})$, which seems to diverge in the limit of $\epsilon\to
0$. However, this is not true. As noted earlier, the existence of the perturbation ($\delta > 0$) guarantees $\epsilon > 0$ and we always have $\beta^i = {\cal O}(\epsilon^{-1},\delta) < 1$ from Eq. (\[condition\]); thus, there is no divergence in the metric. In fact, we recover the standard Friedmann equation in the lowest order approximation. Furthermore, at the end of Sec. \[sec:momentum\], we show that consistency between the Hamiltonian and momentum constraint equations demands $\beta^i={\cal O}(\epsilon^{-1})$.
Note that the shift vector comes with a spatial derivative $\partial_i$ in Einstein’s field equations and scalar-field equations. As $\partial_i\beta^i={\cal O}(\epsilon^0)$, the spatial derivatives are kept in Einstein’s field equations and scalar-field equations for $\alpha-1 = \tilde{\gamma}_{ij} - \delta_{ij} = \varphi^I -
\bar{\varphi}^I = {\cal O}(\epsilon^0)$. In other words, as we keep spatial derivatives in our approach, we are considering some global region in which there are many smoothed local regions. Therefore, we do not [*a priori*]{} demand that these local regions evolve independently of each other, contrary to what is always demanded by a separate universe approach. Specifically, for a separate universe approach, $\beta^i =
{\cal O}(\epsilon)$ is always assumed [*a priori*]{}.
Similarly, when we decompose the quantities $\beta_i$ and $C_i$ into scalar and vector components as $\beta_i = \partial_i \beta^{(S)} + \beta_i^{(V)}$ and $C_i = \partial_i C^{(S)} + C_i^{(V)}$, respectively, the scalar components are of order $\epsilon^{-2}$: $\beta^{(S)} = C^{(S)} = {\cal O}(\epsilon^{-2})$.
In order to see how Eq. (\[first-estimation\]) can be relaxed, we now investigate the nature of solutions for $\beta_i$ and $h^{(T)}_{ij}$.
At the first order in perturbation variables and the lowest order in the gradient expansion, the evolution equation for $\tilde{A}_{ij}$ is given by \[see Eq. (\[full-A-evolusion\])\] $$\dot{\tilde{A}}_{ij} + 3H \tilde{A}_{ij} = {\cal O}(\epsilon,\delta^2),
\label{0th-A}$$ where $H$ is the Hubble expansion rate $H \equiv
\dot{a}/a$. Here, we have ignored the anisotropic stress term on the right-hand side of Eq. (\[full-A-evolusion\]), as it is of the second order in the gradient expansion for scalar fields. This is a stronger-than-necessary condition: Eq. (\[0th-A\]) is still valid if the anisotropic stress of matter fields is of the first order in the gradient expansion.
It follows from Eq. (\[0th-A\]) that the traceless part of the extrinsic curvature $\tilde{A}_{ij}$ has a decaying solution $\tilde{A}_{ij}\propto 1/a^3$ [@Salopek/Bond:1990].[^3] On the other hand, the evolution equation for $\gamma_{ij}$ with $C_i=0$ yields \[see Eq. (\[full-gamma\])\] $$\dot{h}^{(T)}_{ij} = -2 \tilde{A}_{ij} + \frac{1}{a^2}
\left( \partial_i \beta_j + \partial_j \beta_i - \frac{2}{3} \delta_{ij} \delta^{kl} \partial_k \beta_l \right).
\label{0th-gamma}$$ As the scalar, vector, and tensor modes are independent in linear theory, the equations for the shift vector and tensor perturbations are given by $$\begin{aligned}
\ddot{h}_{ij}^{(T)} + 3 H \dot{h}_{ij}^{(T)} = {\cal O}(\epsilon,\delta^2), \nonumber \\
\dot{\beta}^i + 3H \beta^i = {\cal O}(\epsilon^0,\delta^2) .
\label{h_beta}\end{aligned}$$ Therefore, $\dot{h}^{(T)}_{ij}$ and $\beta^i=\beta_i/a^2$ also have decaying solutions scaling as $a^{-3}$. This is a consequence of the fact that the unperturbed metric \[Eq. (\[background\])\] is given by a homogeneous and isotropic Friedmann-Lemaître-Robertson-Walker metric. In other words, this result may not hold for anisotropic models such as Bianchi-type metrics.
At the second order in perturbation variables, as the source terms in Eqs. (\[full-gamma\]) and (\[full-A-evolusion\]) are decaying, the second-order equations for $h^{(T)}_{ij}$ and $\beta_i$ have approximately the same forms as the first-order equations \[Eqs. (\[h\_beta\]) and (\[0th-A\])\], and thus their solutions must also be decaying as $a^{-3}$. Similarly, $n$th-order solutions for $n\ge 3$ are also decaying.
These properties allow us to safely ignore, in the lowest order of the gradient expansion and the $n$th order of perturbation theory, the traceless part of the extrinsic curvature $\tilde{A}_{ij}$, the shift vector $\beta^i$, as well as a time derivative of tensor perturbations $\dot{h}_{ij}^{(T)}$, after the decaying solutions become sufficiently small. This means that these quantities must be higher order in the gradient expansion than naively assumed in Eq. (\[first-estimation\]): $\tilde{A}_{ij} = {\cal O}(\epsilon)$, $\beta^i = {\cal O}(\epsilon^0)$, and $\dot{h}^{(T)}_{ij} = {\cal O}(\epsilon)$.
Now, it turns out that the above argument also applies to the next order of the gradient expansion. At the next order in the gradient expansion $$\begin{aligned}
&&\alpha-1 = \tilde{\gamma}_{ij}-\delta_{ij} = \varphi^I - \bar{\varphi}^I = {\cal O}(\epsilon), \nonumber \\
&& \hspace{1.7cm}\beta^i = {\cal O}(\epsilon^0),\end{aligned}$$ one can show that for scalar fields whose anisotropic stress is of the second order in the gradient expansion, the equations take on the same form as Eqs. (\[0th-A\]) and (\[h\_beta\]): $$\begin{aligned}
\label{aijfull}
\dot{\tilde{A}}_{ij} + 3H \tilde{A}_{ij} &=& {\cal O}(\epsilon^2,\delta^2), \\
\ddot{h}_{ij}^{(T)} + 3H \dot{h}_{ij}^{(T)} &=& {\cal O}(\epsilon^2,\delta^2), \\
\dot{\beta}^i + 3 H \beta^i &=& {\cal O}(\epsilon,\delta^2).\end{aligned}$$ Applying the same argument as above, one finds that $\tilde{A}_{ij}$, $\dot{h}_{ij}^{(T)}$, as well as $\beta^i$ decay for the $n$th order in perturbation theory. We thus find: $\tilde{A}_{ij} = {\cal O}(\epsilon^2)$, $\beta^i = {\cal O}(\epsilon)$, and $\dot{h}^{(T)}_{ij} = {\cal O}(\epsilon^2)$. However, this argument cannot be extended to the second order of the gradient expansion, as Eq. (\[aijfull\]) is valid only when the anisotropic stress term is unimportant. As one can no longer ignore the anisotropic stress of scalar fields at the second order in the gradient expansion, Eq. (\[aijfull\]) is no longer valid in that order.
Therefore, Eq. (\[first-estimation\]) should be revised as $$\begin{aligned}
& \hspace{1cm} \alpha-1 = \varphi^I- \bar{\varphi}^I = {\cal O}(\epsilon^0), \notag \\
& \beta^i = {\cal O}(\epsilon), \ \ \ \dot{h}^{(T)}_{ij} = {\cal O}(\epsilon^2),
\label{estimation}\end{aligned}$$ where we have dropped $\delta$ in ${\cal O}(\dots)$, as the above estimation is valid for all orders of perturbation theory. Note that this result is valid only in the flat gauge given by Eq. (\[gauge-fixing\]). In particular, the condition $C_i=0$ was needed to estimate the gradient-expansion order of $\beta^i$ and $\dot{h}^{(T)}_{ij}$. These results might depend on the details of the theory of gravitation, as we have used Einstein’s field equations to obtain solutions of $\beta^i$ and $\dot{h}^{(T)}_{ij}$. Furthermore, as a perturbative expansion is used in estimating $\beta^i$ and $\dot{h}^{(T)}_{ij}$, the validity of a perturbative description of the metric with the unperturbed metric given by Eq. (\[background\]) has been assumed in the above argument.
Comparison with previous work {#sec:previous}
-----------------------------
How does Eq. (\[estimation\]) compare with the previous work? Our starting point, Eq. (\[first-estimation\]), is different from the assumption made in Lyth, Malik and Sasaki [@David/etal:2005] (also see [@Lyth/Liddle:2009]). They assume that there exists an approximate set of coordinates with which the metric of any local region can be written as a Friedmann-Lemaître-Robertson-Walker metric. This implies that the shift vector $\beta_i$ vanishes and the quantity $\tilde{\gamma}_{ij}$ is time independent in the limit of $\epsilon \to
0$: $\beta_i = {\cal O}(\epsilon)$ and $\dot{\tilde{\gamma}}_{ij}={\cal
O}(\epsilon)$. We do not make this assumption [*a priori*]{}, and thus our argument is more general than that given in [@David/etal:2005]. They then show that by using Einstein’s field equations and ignoring the anisotropic stress term, $\dot{\tilde{\gamma}}_{ij}$ decays in the first order of the gradient expansion, concluding that $\dot{\tilde{\gamma}}_{ij}={\cal O}(\epsilon^2)$.
In [@Weinberg1; @*Weinberg2], Weinberg uses a broken symmetry argument to show $\beta_i = {\cal O}(a^{-2})$ and $\dot{\tilde{\gamma}}_{ij}={\cal
O}(a^{-2})$ for generally covariant theories and with a suitable choice of coordinates, assuming that the unperturbed metric is given by Eq. (\[background\]) and the anisotropic stress term is negligible. He then shows that for Einstein’s field equations in a coordinate system in which $\beta_i=0$ and a certain combination of matter perturbations vanishes, this solution is an attractor. By identifying ${\cal O}(a^{-2})$ with ${\cal O}(\epsilon^2)$ (because each spatial derivative must come with $1/a$), his argument yields $\beta_i = {\cal O}(\epsilon^{2})$ and $\dot{\tilde{\gamma}}_{ij}={\cal
O}(\epsilon^{2})$.
Therefore, our finding agrees with the previous work: if the unperturbed metric is given by Eq. (\[background\]), the anisotropic stress term is negligible on large scales, and if field equations are given by Einstein’s field equations, then $\beta_i = {\cal
O}(\epsilon)$ and $\dot{\tilde{\gamma}}_{ij}={\cal O}(\epsilon^2)$. Note that Weinberg’s estimate for the order of $\beta_i$ is higher by $\epsilon$; however, he does not show $\beta_i = {\cal O}(\epsilon^2)$ explicitly because he chooses coordinates in which $\beta_i=0$.
Gradient expansion of Einstein’s field equations and scalar-field equations in the flat gauge {#sec:noncanonical}
---------------------------------------------------------------------------------------------
Now, we apply the gradient expansion to Einstein’s field equations and scalar-field equations. We shall work with the flat gauge given by Eq. (\[gauge-fixing\]), which gives the gradient-expansion order of perturbation variables given in Eq. (\[estimation\]).
We shall choose the number of $e$-folds $N \equiv \int_{t_\ast}^t
Hdt^{\prime}$ as our time coordinates in the flat gauge. The Hamiltonian constraint \[Eq. (\[full\_hami\])\] and the scalar-field equation \[Eq. (\[motion\])\] in [*both*]{} the lowest order and the next order of the gradient expansion are given by $$3\tilde{H}^2M_p^2 = \rho ,
\label{0th-hami}$$ $$\tilde{H} \partial_N \left( \tilde{H} \varphi^I_N \right) + 3 \tilde{H}^2 \varphi^I_N + V_I = 0 ,
\label{0th-motion-2}$$ where $\tilde{H} \equiv H/\alpha$ is related to a trace of the extrinsic curvature as $\tilde{H}=-K/3$, and the subscript $N$ denotes a partial derivative with respect to $N$. The energy density $\rho$ is given by $\rho = \frac{\tilde{H}^2}{2} G_{IJ} \varphi_N^I \varphi_N^J + V$. Then, $\tilde{H}$ is given by [@Sasaki:1998] $$\tilde{H}^2 = \frac{2V}{6M_p^2 - G_{IJ} \varphi^I_N \varphi^J_N}.
\label{0th-hami-2}$$ All we need to do is to solve Eq. (\[0th-motion-2\]) coupled with Eq. (\[0th-hami-2\]).
On the other hand, the unperturbed equations are $$3H^2M_p^2 = \bar{\rho} ,
\label{back-hami}$$ $$H \partial_N \left( H \bar{\varphi}^I_N \right) + 3H^2 \bar{\varphi}^I_N + V_I(\bar{\varphi}) = 0 ,
\label{back-motion}$$ where $\bar{\rho}$ and $\bar{\varphi}^I$ are the unperturbed energy density and scalar fields, respectively. Apparently, the perturbative equations \[Eqs. (\[0th-hami\]) and (\[0th-motion-2\])\] coincide exactly with the unperturbed equations \[Eqs. (\[back-hami\]) and (\[back-motion\])\]. This result shows that each region smoothed by a superhorizon scale $\epsilon \ll 1$ in the universe evolves independently and behaves like an unperturbed universe providing a support for the assumption made by a separate universe approach.
These results might depend on the details of the theory of gravitation. While the correspondence between the perturbed and unperturbed equations for other theories of gravitation is an interesting problem, in this paper we shall focus on Einstein’s General Relativity. However, these results should not depend on the form of the Lagrangian of scalar fields. This is because the anisotropic stress (i.e., a traceless part of the stress-energy tensor) for scalar fields with arbitrary Lagrangian necessarily comes with two spatial derivatives, and thus it must be ${\cal O}(\epsilon^2)$. As a result, $\tilde{A}_{ij}$ has a decaying solution and a Friedmann-Lemaître-Robertson-Walker universe will be restored on large scales.
One can generalize the above results for the canonical case to noncanonical Lagrangians given by $\mathcal{L} = P(X^{IJ},\varphi^K)$, where $X^{IJ} \equiv -g^{\mu\nu} \partial_{\mu} \varphi^I\partial_{\nu} \varphi^J$. The Hamiltonian constraint is still given by Eq. (\[0th-hami\]) while the scalar-field equation is given by $$\tilde{H} \partial_N \left[ \tilde{H} \partial_N \varphi^J_N \right] P_{IJ}
+ \tilde{H}^2 \varphi^J_N \partial_N P_{IJ} + 3\tilde{H}^2 P_{IJ} \varphi^J_N + \frac{P_I}{2} = 0,$$ where $P_{IJ} \equiv \partial P/\partial X^{IJ}$, $X^{IJ} = \tilde{H}^2\varphi^I_N \varphi^J_N$, and the energy density is defined as $\rho = 2X^{IJ} P_{IJ} - P$. As $P$ and $P_{IJ}$ are functions of $X^{IJ}$, $\varphi^K$, and $\tilde{H}$, one can write $\tilde{H}$ as a function of $X^{IJ}$ and $\varphi^K$ if an explicit form of the Lagrangian $P$ is specified.
What are the implications of these results? As the equations take on the same forms, the functional forms of the solutions for the perturbed equations and those for the unperturbed equations must be the same. Therefore, the perturbed solutions $\varphi^I$ are given by the unperturbed solutions $\bar{\varphi}^I$ [*with perturbed initial conditions computed in the flat gauge, $\varphi^I_{\ast}({\mbox{\boldmath $x$}})$ and $\varphi^I_{N\ast}({\mbox{\boldmath $x$}})$:*]{} $$\varphi^I(N,{\mbox{\boldmath $x$}}) = \bar{\varphi}^I(N,\varphi^J_{\ast}({\mbox{\boldmath $x$}}),\varphi^K_{N\ast}({\mbox{\boldmath $x$}})).
\label{full-solution}$$ [*This is the fundamental result of the gradient expansion as applied to Einstein’s field equations and scalar-field equations in the flat gauge.*]{} Here, the subscript $*$ indicates that the quantity is evaluated at some initial time, where all the relevant fields are sufficiently outside their sound horizon, i.e., $k\ll a(t_*)H(t_*)/c_s^I$, where $c_s^I$ is the speed of sound of propagation of an $I$th scalar-field perturbation.
In order to simplify our notations, from now on we shall use the lower-case alphabet indices, such as $a, b, c \dots$, to denote the numbers of scalar fields and their time derivatives: $$\varphi^a \equiv (\varphi^I,\varphi^J_N),$$ with $a$ running from 1 to $2n$. With this notation, the solution \[Eq. (\[full-solution\])\] is expressed as $\varphi^a(N,{\mbox{\boldmath $x$}}) = \bar{\varphi}^a(N,\varphi_{\ast}^b({\mbox{\boldmath $x$}}))$.
Similarly, we can write the perturbed energy density of multiscalar fields using the unperturbed energy density solution $\rho(N,{\mbox{\boldmath $x$}}) = \bar{\rho}(N,\varphi^a_{\ast}({\mbox{\boldmath $x$}}))$.
The $\delta N$ formalism
========================
We now need to relate perturbed initial scalar fields and their derivatives $\varphi^a_*({\mbox{\boldmath $x$}})$ to the observables. In cosmology, it is now customary to express the observables such as temperature and polarization anisotropies and the large-scale distribution of galaxies in terms of a curvature perturbation in the “uniform-density gauge,” denoted as $\zeta$.
The so-called $\delta N$ formalism [@Starobinsky:1982; @*Starobinsky:1986; @Salopek/Bond:1990; @Sasaki/Stewart:1996; @David/etal:2005; @Lyth/Rodriguez:2005] achieves this by realizing that $\zeta$ is equal to a perturbation to the number of $e$-folds, $N$, arising from perturbed initial scalar fields $\varphi^a_*({\mbox{\boldmath $x$}})$ computed in the flat gauge.
Conservation of $\zeta$ outside the horizon
-------------------------------------------
We define the uniform-density gauge as $\delta \rho = 0$ and $C_i=0$. (Once again, a uniform-density gauge in the literature sometimes does not include $C_i=0$. In such a case the gauge is not completely fixed.) Let us denote a value of $\psi$ in the uniform-density gauge as $\psi|_{\delta \rho = C_i = 0}$[^4], and write $\zeta$ as $$\zeta \equiv \psi|_{\delta \rho = C_i =0}.$$
This quantity is useful for extracting information about the physics of inflation, as it is conserved outside the horizon, provided that the adiabatic condition $p= p[\rho]$ is satisfied [@Wands/etal:2000; @David/etal:2005]. This is easily seen from the energy conservation equation in the lowest order as well as in the next order gradient expansion with the gauge condition $C_i = 0$: $$\dot{\rho} + 3 (H + \dot{\psi}|_{C_i=0})(\rho + p) = 0.
\label{rho}$$ The perturbation to this equation in the uniform-density gauge yields $$\dot{\psi}|_{\delta \rho = C_i = 0} = 0,$$ and thus $\zeta=\psi|_{\delta \rho = C_i = 0}$ becomes a constant, provided that the adiabatic condition is satisfied.
Alternatively, Eq. (\[rho\]) may be integrated with respect to $t$ without imposing $\delta\rho=0$: $$\tilde{\zeta} \equiv \psi + \int_{\bar{\rho}}^{\rho} \frac{d \rho}{3 (\rho + p[\rho])} = \mbox{const.}
\label{conservezeta}$$ One may then identify this quantity $\tilde{\zeta}$ as a generalization of $\zeta$ when $\delta\rho=0$ is not imposed; however, $\tilde{\zeta}$ is not gauge invariant and does not coincide with $\zeta = \psi|_{\delta \rho = C_i = 0}$, unless the adiabatic condition is satisfied.
What about scalar fields? As scalar fields do not satisfy the adiabatic condition in general, $\zeta$ is not conserved in a universe filled with scalar fields. However, as shown by [@Weinberg:2003], $\zeta$ is generally conserved outside the horizon when inflation was driven by a single scalar field. More precisely, $\zeta$ is conserved outside the horizon in a universe dominated by a single scalar field, provided that the slow-roll conditions are satisfied, or that we completely neglect a decaying mode solution without imposing the slow-roll conditions. This implies that the slow-roll conditions correspond effectively to the adiabatic condition and the neglect of a decaying mode solution for a single scalar field.
Relation between $\zeta$ and the difference in the number of $e$-folds
----------------------------------------------------------------------
The relation between the curvature perturbation and the number of $e$-folds is given by the gauge transformation of the spatial metric $\gamma_{ij}$. Under a gauge transformation given by $t\to T=t+\delta T$ and $x^i\to X^i=x^i+\xi^i$, the metric transforms as $g_{ij}(t,{\mbox{\boldmath $x$}})\to \hat{g}_{ij}(T,\mbox{\boldmath $X$})$. Let us write the 3-metric in the original coordinates in terms of the 3-metric in the new coordinates: $$\begin{aligned}
\gamma_{ij}(t,{\mbox{\boldmath $x$}})
= & -\hat{\alpha}^2(T,\mbox{\boldmath $X$}) \frac{\partial\delta
T}{\partial x^i}\frac{\partial \delta T}{\partial x^j} \notag \\
& + \hat{\beta}_k(T,\mbox{\boldmath $X$}) \frac{\partial
X^k}{\partial x^i} \frac{\partial \delta T}{\partial x^j}
+ \hat{\beta}_k(T,\mbox{\boldmath $X$}) \frac{\partial X^k}{\partial
x^j}\frac{\partial \delta T}{\partial x^i} \notag \\
& + \hat{\gamma}_{kl}(T,\mbox{\boldmath $X$})\frac{\partial X^k}{\partial x^i} \frac{\partial X^l}{\partial x^j}.
\label{gauge}\end{aligned}$$ We shall always impose $C_i=0$, which completely fixes the spatial gauge degree of freedom, and thus we can set $\xi^i=0$ without loss of generality.
Let us examine each term in terms of the gradient-expansion order. The first term is of order ${\cal O}(\epsilon^2)$. As shown in Sec. \[sec:estimation\], when $C_i = 0$, the shift vector is of ${\cal O}(\epsilon)$; thus, the second and third terms are of order ${\cal O}(\epsilon^2)$. This means that, up to ${\cal O}(\epsilon^2)$, the 3-metric transforms as $${\gamma}_{ij}|_{C_i = 0}(t,{\mbox{\boldmath $x$}}) = \hat{\gamma}_{ij}|_{C_i =0}(T,{\mbox{\boldmath $x$}}) + {\cal O}(\epsilon^2).
\label{gammatrans}$$
Recalling $\gamma_{ij}=a^2(t)e^{2\psi}(e^h)_{ij}$ and taking the determinant and logarithm of both sides of Eq. (\[gammatrans\]), we find $$\psi|_{C_i = 0}(t,{\mbox{\boldmath $x$}}) = \hat{\psi}|_{C_i =
0}(T,{\mbox{\boldmath $x$}}) + \ln\left( \frac{a(T)}{a(t)} \right) .
\label{psi_C}$$ Thus, $\psi|_{C_i = 0}$ approximately transforms as a scalar quantity having $\ln (a)$ as the unperturbed value. It follows from Eq. (\[psi\_C\]) that the gauge transformation of $\psi|_{C_i = 0}$ from the flat gauge (in which $\psi=0$) into the uniform-density gauge is given by $$\psi|_{\delta \rho = C_i = 0}(T,{\mbox{\boldmath $x$}})
= \ln \left(\frac{a(t)}{a(T)}\right),
\label{1-definition-deltaN}$$ where $T$ denotes time coordinates in the uniform-density gauge.
On the other hand, when we go from the flat gauge to the uniform-density gauge, the number of $e$-folds, $N \equiv \int_{t_*}^t H dt'$, transforms as $N \to \hat{N}$, where $$\hat{N} \equiv \int^{T}_{t_{\ast}} H(t^{\prime}) dt^{\prime} = \ln \left( \frac{a(T)}{a(t_{\ast})} \right) .
\label{Nhat}$$ Here, $t_*$ is an arbitrary initial time.
Comparing Eq. (\[1-definition-deltaN\]) to Eq. (\[Nhat\]), one finds $$\begin{aligned}
\psi|_{\delta \rho = C_i = 0}(T,{\mbox{\boldmath $x$}}) & = \ln \left( \frac{a(t)}{a(t_*)} \right) - \ln \left( \frac{a(T)}{a(t_*)} \right) \notag \\
& = N - \hat{N} \equiv \delta N.\end{aligned}$$ Therefore, $\zeta=\psi|_{\delta \rho = C_i = 0}$ is equal to the difference between the number of $e$-folds computed in the flat gauge and that computed in the uniform-density gauge. The remaining task is to relate $\delta N$ to perturbed initial scalar fields in the flat gauge.
Relation between $\delta N$ and perturbed initial scalar fields in the flat gauge
---------------------------------------------------------------------------------
The most important result that came from the gradient expansion of Einstein’s field equations and scalar-field equations in the flat gauge is that perturbed quantities can be calculated using their unperturbed solutions with perturbed initial scalar-field values and their time derivatives computed in the flat gauge. Therefore, a perturbed energy density in the flat gauge is given by $\rho(N,{\mbox{\boldmath $x$}}) =
\bar{\rho}(N,\varphi_*^a({\mbox{\boldmath $x$}}))$. Here, we choose the number of $e$-folds as time coordinates.
On the other hand, by definition the energy density in the uniform-density gauge (whose time coordinates are denoted as $\hat{N}$) is equal to the unperturbed density. Namely, when we go from the flat gauge to the uniform-density gauge by changing the time coordinates as $N\to \hat{N}= N + \delta N$, the density transforms as $\rho(N,{\mbox{\boldmath $x$}})\to
\hat{\rho}(\hat{N},{\mbox{\boldmath $x$}})=\bar{\rho}(\hat{N})$. Here, $C_i = 0$ is satisfied in both gauges, and thus there is no ambiguity with respect to the spatial gauge degree of freedom. Now, as the energy density is a four scalar, $$\rho(N,{\mbox{\boldmath $x$}}) = \hat{\rho}(\hat{N},{\mbox{\boldmath $x$}})=\bar{\rho}(\hat{N}),
\label{gauge2}$$ which gives $\bar{\rho}(N,\varphi_*^a({\mbox{\boldmath $x$}}))=\bar{\rho}(\hat{N})$. Inverting this result yields $$N=\hat{N}(\bar{\rho},\bar{\varphi}_*^a({\mbox{\boldmath $x$}})),$$ where the functional form of $\hat{N}$ is the same as that of the unperturbed number of $e$-folds. That the unperturbed density $\bar{\rho}$ (not $t$ or $N$) is used as the time coordinates here ensures that the final time slice coincides with the uniform density hypersurface.
With these results, we can finally calculate $\zeta$, $$\begin{aligned}
\zeta&=& N - \hat{N} \notag \\
&=& \hat{N}(\bar{\rho},\varphi_*^a({\mbox{\boldmath $x$}}))-\hat{N}(\bar{\rho}, \bar{\varphi}_*^a) \notag \\
&=&\hat{N}_a \delta \varphi^a_{\ast}({\mbox{\boldmath $x$}})
+ \frac{1}{2}\hat{N}_{ab} \delta \varphi^a_{\ast}({\mbox{\boldmath $x$}}) \delta \varphi^b_{\ast}({\mbox{\boldmath $x$}})
+ \dots ,
\label{eq:deltaN}\end{aligned}$$ where $\delta\varphi_*^a({\mbox{\boldmath $x$}})\equiv \varphi_*^a({\mbox{\boldmath $x$}})-\bar{\varphi}_*^a$ denotes perturbations to initial scalar fields computed in the flat gauge, and $\hat{N}_a$ and $\hat{N}_{ab}$ are defined as $$\hat{N}_a \equiv \frac{\partial \hat{N}[\bar{\rho},\bar{\varphi}_{\ast}^b]}
{\partial \bar{\varphi}_{\ast}^a} , \ \ \
\hat{N}_{ab} \equiv \frac{\partial^2 \hat{N}[\bar{\rho},\bar{\varphi}_{\ast}^c]}
{\partial \bar{\varphi}_{\ast}^a \partial \bar{\varphi}_{\ast}^b} .
\label{}$$ This is the $\delta N$ formalism, which enables us to relate $\zeta$ to the initial scalar-field perturbations (i.e., scalar-field perturbations at the initial time) computed in the flat gauge, once we know derivatives of the number of $e$-folds with respect to the initial values of the unperturbed scalar fields $\varphi_*^I$ and their derivatives $\varphi_{N*}^I$.
Momentum constraint {#sec:momentum}
-------------------
Perhaps a striking thing about the $\delta N$ formalism is that we only had to use the Hamiltonian constraint \[Eq. (\[0th-hami\])\] and the scalar-field equation \[Eq. (\[0th-motion-2\])\] in the gradient expansion. But, should not we also impose the momentum constraint for consistent calculations?
As the momentum constraint comes with a spatial derivative $\partial_i$ we need to consider the momentum constraint in $\mathcal{O}(\epsilon^2)$ in order to derive the correct relationship between physical quantities up to $\mathcal{O}(\epsilon)$, $$\partial_i \tilde{H} = -\frac{\tilde{H}}{2M_p^2} G_{IJ}\varphi_N^I \partial_i \varphi^J + {\cal O}(\epsilon^3) ,
\label{moment}$$ where we have used the fact that $\tilde{A}_{ij} = {\cal O}(\epsilon^2)$.
On the other hand, the Hamiltonian constraint \[Eq. (\[0th-hami\])\] may be differentiated by $\partial_i$ to give $$\partial_i \tilde{H} = -\frac{\tilde{H}}{2M_p^2} G_{IJ}\varphi_N^I \partial_i \varphi^J + B_i ,
\label{B-moment}$$ where $$B_i \equiv \frac{\tilde{H}^3}{2V}G_{IJ}\left( \varphi^I_N \partial_i \varphi^J_N - \varphi_{NN}^I\partial_i \varphi^J \right) .
\label{definition_of_B}$$ Here, we have used the equation of motion for scalar fields given by Eq. (\[0th-motion-2\]), as well as the evolution equation for $K$ given by Eq. (\[full-K-evolution\]), which yields $\tilde{H}_N = - \frac{\tilde{H}}{2M_p^2} G_{IJ} \varphi_N^I \varphi_N^J$ in the gradient expansion in the flat gauge.
Comparing Eqs. (\[B-moment\]) and (\[moment\]), we find that the $\delta N$ formalism, which does not use the momentum constraint but uses only the Hamiltonian constraint, can introduce an error in the momentum constraint by an amount $B_i$. Imposing the momentum constraint gives an additional constraint $B_i= {\cal O}(\epsilon^3)$ for the $\delta N$ formalism.
How important is $B_i$? In order to investigate the behavior of $B_i$, let us take a spatial derivative of the equation of motion for scalar fields $$\begin{aligned}
&\tilde{H}\partial_N(\tilde{H} \partial_i \varphi_N^I) + 3\tilde{H}^2 \partial_i \varphi^I_N \notag \\
&+ \Bigg[ V_{IJ}
- \frac{\tilde{H}}{a^3 M_p^2} \frac{d}{dN}\left(G_{JK} a^3 \tilde{H} \varphi^I_N \varphi^K_N \right) \Bigg]
\partial_i \varphi^J \notag \\
& \hspace{2cm} + H\varphi^I (B_{iN} + 3B_i ) + 2 \varphi_{NN}^I H B_i = 0 .
\label{aa2}\end{aligned}$$ By contracting this equation with $\varphi^I_N$, one finds $$\begin{aligned}
& \partial_N B_i +3B_i = 0 ,\notag \\
& \to B_i = \frac{a_{\ast}^3 B_{i \ast}}{a^3}.
\label{B}\end{aligned}$$ Therefore, $B_i$ has only a decaying solution.[^5] This is good news for the $\delta N$ formalism: while it does not [*a priori*]{} respect the momentum constraint, the error in the momentum constraint rapidly decays away by inflation. That $B_i$ is a decaying mode may be traced back to the fact that the traceless part of the extrinsic curvature $\tilde{A}_{ij}$ is a decaying mode in the gradient expansion in the flat gauge. In other words, it is a consequence of the universe behaving like a Friedmann-Lemaître-Robertson-Walker universe on superhorizon scales, which is guaranteed by Eq. (\[estimation\]).
Remember that ignoring the decaying-mode terms of $\beta^i$ and $h_{ij}^{(T)}$ has led to the $\delta N$ formalism and the separate universe description. This means that the decaying term $B_i$ should also be ignored for consistency and thus should be treated as higher order in $\epsilon$. In fact, the momentum constraint naturally satisfies this condition: $B_i = {\cal O}(\epsilon^3)$. However, as the $\delta N$ formalism does not [*a priori*]{} respect the momentum constraint, it yields the correct growing solutions and incorrect decaying solutions. In other words, the $\delta N$ formalism yields valid solutions only in models in which the decaying-mode terms never affect the curvature perturbation. If one needs to completely remove the decaying-mode contributions from the $\delta N$ formalism, then one should use the $\delta N$ formalism with the initial condition $B_{i*}=0$.
One may wonder why consistency between the momentum constraint and the Hamiltonian constraint gives a relation only among scalar fields $G_{IJ}\left( \varphi^I_N \partial_i \varphi^J_N -
\varphi_{NN}^I\partial_i \varphi^J \right)=0$, rather than a relation between the metric variables and scalar fields. This is because we have ignored the decaying solutions of $\beta^i$ and $\dot{h}^{(T)}_{ij}$. To see this, let us work at the first order in perturbations, and bring $\beta^i$ back into Einstein’s field equations. We find that consistency between the momentum constraint and the Hamiltonian constraint gives $$B_i + \frac{M_p^2 H^2}{V} \partial_i(\partial_j \beta^j) = 0.
\label{linear_B}$$ Indeed, consistency gives a relation between scalar fields (contained in $B_i$) and a metric variable ($\beta^i$)[^6]. This equation also indicates that when we keep the decaying quantities of order $a^{-3}$, the gradient-expansion order of $\beta^i$ is indeed ${\cal
O}(\epsilon^{-1})$ \[see Eq. (\[first-estimation\])\], as Eq. (\[linear\_B\]) gives $\partial_j \beta^j={\cal O}(1)$.
Now is the time to answer the following question: what if we demand $\beta^i = {\cal O}(\epsilon^0)$? In this case, Eq. (\[linear\_B\]) gives $B_i={\cal O}(\epsilon^2)$ or $$G_{IJ} \left( \bar{\varphi}^I_N \delta \varphi^J_N -
\bar{\varphi}_{NN}^I \delta \varphi^J \right) = {\cal O}(\epsilon),
\label{}$$ where $\delta \varphi^I \equiv \varphi^I - \bar{\varphi}^I$ is the perturbation of scalar fields in linear theory. This result indicates that we are not allowed to have a configuration of scalar fields which yields $G_{IJ} \left( \bar{\varphi}^I_N \delta \varphi^J_N -
\bar{\varphi}_{NN}^I \delta \varphi^J \right)={\cal
O}(\epsilon^0)$. On the other hand, if we start with $\beta^i = {\cal O}(\epsilon^{-1})$, then we can show that $G_{IJ} \left( \bar{\varphi}^I_N \delta \varphi^J_N -
\bar{\varphi}_{NN}^I \delta \varphi^J \right)$ becomes negligible as it is a decaying mode. Another way of saying this is that if we start with $\beta^i = {\cal O}(\epsilon^0)$, then $G_{IJ} \left( \bar{\varphi}^I_N \delta \varphi^J_N -
\bar{\varphi}_{NN}^I \delta \varphi^J \right)=0$ gives only one solution for $\delta\varphi^I$, and another solution, which corresponds to a decaying mode, does not exist.[^7] Therefore, demanding that the number of independent solutions (which is two) not reduce, one should start with $\beta^i = {\cal
O}(\epsilon^{-1})$.
Slow-roll conditions and momentum constraint
--------------------------------------------
Interestingly, we can show that $B_i$ vanishes when the slow-roll conditions are satisfied. The slow-roll equations of motion for the canonical scalar fields are $$3M_p^2 \tilde{H}^2 \approx V , \ \ \ 3\tilde{H}^2 \varphi_N^I + V_I \approx 0 .
\label{}$$ This implies that $$\begin{aligned}
\varphi_N^I \approx - M_p^2 \frac{V_I}{V}
, \ \ \ \ \ \varphi_{NN}^I \approx (2 \varepsilon_{IJ} - \eta_{IJ})\varphi_N^J ,
\label{slow-roll}\end{aligned}$$ where the slow-roll parameters $\varepsilon_{ab}$ and $\eta_{ab}$ are defined as $$\varepsilon_{IJ} \equiv \frac{M_p^2}{2}\frac{V_IV_J}{V^2} ,\ \
\eta_{IJ} \equiv M_p^2 \frac{V_{IJ}}{V}.
\label{}$$
Then we find that the following relation is satisfied under the slow-roll condition, $$\begin{aligned}
\varphi^I_N \partial_i \varphi^I_N
\approx \varphi_N^I (2\varepsilon_{IJ} - \eta_{IJ})
\partial_i \varphi^J \approx \varphi_{NN}^I \partial_i \varphi^I,\end{aligned}$$ which yields $B_i \approx 0$. In this sense, the slow-roll conditions are equivalent to the momentum constraint.
What does this imply? This implies that the $\delta N$ formalism happens to respect the momentum constraint if the slow-roll conditions are satisfied at the initial time $t_*$. This may provide a partial explanation as to why the $\delta N$ formalism has been successful in computing $\zeta$ for a wide variety of slow-roll inflation models.
Conclusion
==========
The necessary and sufficient condition for the validity of the $\delta N$ formalism is that, with a suitable choice of coordinates, the perturbed Hamiltonian constraint and matter-field equations on large scales coincide with the corresponding unperturbed equations.
That perturbed solutions in the long-wavelength limit can be obtained from unperturbed solutions was found and investigated by pioneering work in 1998 [@Taruya/Nambu:1998; @*Nambu/Taruya:1998; @Kodama/Hamazaki:1998; @Sasaki:1998]. While their work was restricted to linear theory (and to quasilinear theory [@Sasaki:1998]), we have extended their work to include nonlinear (but still perturbative) perturbations. Such extension is also explored by [@David/etal:2005], who use the so-called “separate universe approach” [@Wands/etal:2000]. As we have described in Sec. \[sec:previous\], our starting point is more general than theirs. In this paper, using the flat gauge $(\psi= C_i = 0)$ and choosing the number of $e$-folds $N$ as our time coordinates, we have shown that the perturbed Hamiltonian constraint and matter-field equations on large scales coincide with the corresponding unperturbed equations, as long as (at least) the following conditions are satisfied:
- The unperturbed metric is given by a homogeneous and isotropic Friedmann-Lemaître-Robertson-Walker metric.
- The final results for the curvature perturbation are not affected by decaying-mode terms, such as the shift vector, a time derivative of tensor perturbations to $g_{ij}/a^2(t)$, or the error in the momentum constraint $B_i$.
- Evolution of scalar-field perturbations outside the horizon can be treated using the lowest order or the next order of the gradient expansion.
In order to show that the shift vector and a time derivative of tensor perturbations are decaying modes, one needs two more conditions.
- Matter fields are given by scalar fields with an arbitrary form of Lagrangian \[whose anisotropic stress is of order ${\cal O}(\epsilon^2)$\] or, in the $n$th-order gradient expansion where $n=0$ or 1, by some fluids with anisotropic stress of order ${\cal
O}(\epsilon^{n+1})$.
- Theory of gravitation determining the physics of inflation is Einstein’s theory, or modified gravitational theories with can be transformed into Einstein’s theory.
The fourth and fifth conditions are sufficient conditions: we had to use Einstein’s field equations to explicitly show that the fourth condition implies that the shift vector and a time derivative of tensor perturbations are decaying modes. It is possible that other theories of gravitation require different conditions for the shift vector and a time derivative of tensor perturbations to be decaying modes.
The discussion in this paper should also apply to vector-field models (see [@Dimastrogiovanni/etal:2010] for a review and references therein), as long as their anisotropic stress is of order ${\cal O}(\epsilon^{n+1})$ in the $n$th-order gradient expansion where $n=0$ or 1.
The third condition is naturally expected in any inflation scenario. However, when the slow-roll conditions are violated, it is known that decaying-mode solutions in the second order of the gradient expansion cannot be neglected in the power spectrum [@Samuel.M.Leach/M.Sasaki/etal:2001]. In such cases, since there are no gauges in which a Friedmann-Lemaître-Robertson-Walker universe can be obtained in the second order of the gradient expansion, we can no longer use the $\delta N$ formalism [@Takamizu/etal:2010; @*Takamizu:2011].
When all of the above conditions are satisfied, one can calculate $\zeta$ using initial scalar-field perturbations computed in the flat gauge and derivatives of the number of $e$-folds with respect to the initial values of the unperturbed scalar fields $\varphi_*^I$ and their derivatives $\varphi_{N*}^I$.
We would like to thank Y. Itoh, D. Lyth, J. Meyers, A. Naruko, T. Tanaka, M. Sasaki, and S. Weinberg for useful discussions and comments. We would also like to thank T. Hamazaki for bringing Refs. [@Kodama/Hamazaki:1998; @Hamazaki:2008] to our attention after we have posted our article to arXiv. This work is supported in part by the GCOE Program “Weaving Science Web beyond Particle-Matter Hierarchy” at Tohoku University, a Grant-in-Aid for Scientific Research from JSPS (Grants. No. 18072001 and No. 20540245 for T. F.), the Core-to-Core Program “International Research Network for Dark Energy,” as well as by NSF Grant No. PHY-0758153. T. F. would like to thank Luc Branchet and the Institute of Astronomical Observatory, Paris, and N. S. S. and E. K. would like to thank M. Sasaki and the Yukawa Institute for Theoretical Physics, for their warm hospitality during the last stage of this work.
Einstein’s field equations
==========================
In this Appendix, we give Einstein’s field equations in terms of the variables of the ADM formalism \[see Eq. (1) for the ADM metric\]. We use latin indices for the 3D spatial components running from 1 to 3, and greek indices for the 4D spacetime components running from 0 to 3.
We decompose the three-space metric tensor as $$\gamma_{ij} \equiv a^2(t)e^{2\psi}\tilde{\gamma}_{ij} ,
\label{}$$ where $a(t)$ is the scale factor. We define $\tilde{\gamma}_{ij}$ such that $\det[\tilde{\gamma}_{ij}] = 1$; thus, $\det[\tilde{\gamma}_{ij}]$ can be written as $\det[(e^{h})_{ij}] = e^{ {\rm Tr}[h]} $ with ${\rm Tr}[h] = 0$. We further decompose $h_{ij}$ as $$h_{ij} = \partial_i C_j + \partial_j C_i - \frac{2}{3} \delta_{ij} \partial_k C^k + h^{(T)}_{ij},
\label{h-deco}$$ where $h^{(T)}_{ij}$ denotes a tensor mode and $C_i$ has a scalar mode and a vector mode.
The extrinsic curvature $K_{ij}$ is defined as $$K_{ij} \equiv - \nabla_i n_j = \frac{1}{2\alpha}\left( D_i \beta_j+ D_j \beta_i - \dot{ \gamma}_{ij} \right) ,
\label{}$$ where $n^{\mu} = \left( 1/\alpha , - \beta^i/\alpha\right)$ is the unit vector normal to the $t$-constant hypersurface, and $\nabla$ and $D$ are the covariant differential operators constructed by using $g_{ \mu \nu}$ and $\gamma_{ij}$, respectively. The dots denote time derivatives with respect to $t$.
It is useful to decompose the extrinsic curvature $K_{ij}$ into a “trace” part $K$ and a “trace free” part $\tilde{A}_{ij}$ as $$K_{ij} =\frac{\gamma_{ij}}{3}K + a^2e^{2\psi}\tilde{A}_{ij} ,
\label{}$$ where the indices of $\tilde{A}_{ij}$ are raised/lowered by $\tilde{\gamma}_{ij}$, and $\tilde{A}^i_{\ i} = 0$ is satisfied.
By using the above notations, we write down Einstein’s field equations. The Hamiltonian constraint is $$R^{(3)} - \tilde{A}_{ij}\tilde{A}^{ij} + \frac{2}{3} K^2 = \frac{2}{M_p^2}T_{\mu\nu}n^{\nu}n^{\mu} .
\label{full_hami}$$ The momentum constraint is $$D_j\tilde{A}^j_{\ i} - \frac{2}{3}\partial_i K = - \frac{1}{M_p^2} T_{i\nu} n^{\nu} .
\label{full-moment}$$ The dynamical equation for $\psi$ is $$(\partial_t - \beta^i\partial_i) \psi + H = \frac{1}{3}(- \alpha K + \partial_i \beta^i) .
\label{full-K}$$ The dynamical equation for $\tilde{\gamma}_{ij}$ is $$(\partial_t - \beta^k\partial_k)\tilde{\gamma}_{ij} = -2\alpha\tilde{A}_{ij} + \tilde{\gamma}_{ik}\partial_j \beta^k
+ \tilde{\gamma}_{jk}\partial_i \beta^k - \frac{2}{3}\tilde{\gamma}_{ij} \partial_k \beta^k ,
\label{full-gamma}$$ which yields the dynamical equation for $K$, $$\begin{aligned}
(\partial_t - \beta^k\partial_k)K = &\alpha\left( \tilde{A}_{ij}\tilde{A}^{ij} + \frac{1}{3}K^2 \right) \notag \\
& - \gamma^{ij}D_iD_j \alpha + \frac{\alpha}{2M_p^2} \left( T_{\mu\nu}n^{\mu}n^{\nu} + \gamma^{ij}T_{ij} \right),
\label{full-K-evolution}\end{aligned}$$ as well as the dynamical equation for $\tilde{A}_{ij}$, $$\begin{aligned}
& (\partial_t - \beta^k\partial_k)\tilde{A}_{ij} \notag \\
= & \frac{1}{a^2e^{2\psi}}\bigg[ \alpha \bigg( R^{(3)}_{ij} - \frac{\gamma_{ij}}{3}R^{(3)}\bigg)
- \bigg( D_iD_j \alpha - \frac{\gamma_{ij}}{3}D_kD^k \alpha \bigg) \bigg] \notag\\
& + \alpha(K \tilde{A}_{ij} - 2\tilde{A}_{ik}\tilde{A}^k_{\ j})
+ \tilde{A}_{ik}\partial_j\beta^k + \tilde{A}_{jk}\partial_i \beta^k - \frac{2}{3}\tilde{A}_{ij}\partial_k \beta^k \notag\\
& \hspace{3cm} - \frac{\alpha}{a^2 e^{2\psi} M_p^2}\bigg( T_{ij} - \frac{\gamma_{ij}}{3}\gamma^{kl} T_{kl} \bigg) .
\label{full-A-evolusion}\end{aligned}$$ Here, $T_{ij} - \frac{\gamma_{ij}}{3}\gamma^{kl} T_{kl}$ is the anisotropic stress.
The three-dimensional Ricci scalar $R^{(3)}$ is constructed from $\gamma_{ij}$, and $M_p^2$ is the reduced Plank mass defined as $1/8\pi G$ where $G$ is the gravitational constant.
[26]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty [****, ()](\doibase 10.1088/0264-9381/27/12/124010), [****, ()](\doibase 10.1088/0264-9381/27/12/124011), [****, ()](\doibase 10.1016/0370-2693(82)90541-X) @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevD.42.3936) [****, ()](\doibase 10.1143/PTP.95.71), [****, ()](\doibase 10.1088/1475-7516/2005/05/004), [****, ()](\doibase 10.1103/PhysRevLett.95.121302), [****, ()](\doibase 10.1103/PhysRevD.62.043527), [****, ()](\doibase 10.1088/1475-7516/2009/05/013), @noop [ ()]{}, , @noop [****, ()]{}, [****, ()](\doibase 10.1103/PhysRevD.78.103513), @noop [**]{} (, , ) [****, ()](\doibase 10.1103/PhysRevD.78.123521), [****, ()](\doibase 10.1103/PhysRevD.79.043504), [****, ()](\doibase 10.1143/PTP.99.763), [****, ()](\doibase 10.1103/PhysRevD.28.679) [****, ()](\doibase 10.1103/PhysRevD.67.123504), [****, ()](\doibase 10.1103/PhysRevD.57.7177), [****, ()](\doibase 10.1016/S0370-2693(98)00378-5), [****, ()](\doibase 10.1088/0264-9381/15/9/021), [****, ()](\doibase 10.1155/2010/752670), [****, ()](\doibase
10.1103/PhysRevD.64.023512), [****, ()](\doibase
10.1088/1475-7516/2010/06/019), [****, ()](\doibase 10.1103/PhysRevD.83.043504),
[^1]: In Salopek and Bond [@Salopek/Bond:1990], a trace-free part of the extrinsic curvature is denoted as $\bar{K}_{ij}$. In our notation, we have $\bar{K}_{ij}=a^2 e^{2\psi} \tilde{A}_{ij}$.
[^2]: Sometimes another gauge condition $\dot{\psi} = 0$ and $\beta_i = 0$ is imposed; however, the degree of freedom of this gauge is not completely fixed, as this gauge condition only gives the relation for the time derivatives of the shifts in temporal and spatial coordinates.
[^3]: As we start with $\beta^i={\cal O}(\epsilon^{-1})$, we need to linearize Eq. (\[full-A-evolusion\]) to obtain Eq. (\[0th-A\]), showing $\tilde{A}_{ij}\propto 1/a^3$. On the other hand, assuming $\beta^i={\cal
O}(\epsilon)$, Hamazaki derives $\tilde{A}_{ij}\tilde{A}^{ij}\propto 1/a^6$ without using perturbation theory (see Eq. (2.54) of [@Hamazaki:2008]).
[^4]: Incidentally, as the gauge is completely fixed for $\psi|_{\delta \rho = C_i =0}$, there is no ambiguity with respect to the residual gauge degree of freedom. Moreover, one can always write perturbation variables (such as $\psi$) after gauge fixing as a combination of perturbation variables before gauge fixing (such as $C_i$ and $\delta \rho$) such that $\psi$ is explicitly gauge-invariant. For example, we have, at the linear order, $$\zeta \equiv \psi|_{\delta \rho = C_i =0}
= \psi - \frac{\partial_i C^i}{3}- \frac{\delta \rho}{\bar{\rho}_N} ,
\label{ex}$$ where the right-hand side of Eq. (\[ex\]) is the well-known form for a gauge-invariant curvature perturbation in the linear order [@Bardeen/Turner/Steinhardt:1983].
[^5]: In linear theory, this quantity is equal to $-a^{-3}\partial_i W$ where $W$ is given by Eq. (2.26) of [@Sasaki:1998], as well as to $-\partial_i \dot{f}$ where $\dot{f}$ is given by Eq. (5.23) of [@Kodama/Hamazaki:1998]. They show $W\propto a^{-3}$ and $\dot{f}\propto a^{-3}$ in linear theory.
[^6]: We thank M. Sasaki for clarifying this point.
[^7]: Assuming $\beta^i={\cal O(\epsilon)}$, Kodama and Hamazaki [@Kodama/Hamazaki:1998] also find this property from the momentum constraint. See their Eq. (5.13) and the argument given below it.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A few years after its discovery as a magnetar, started a new burst-active phase on 2020 April 27, accompanied by a large enhancement of its X-ray persistent emission. Radio single bursts were detected during this activation, strengthening the connection between magnetars and fast radio bursts. We report on the X-ray monitoring of from $\sim$3 days prior to $\sim$3 weeks after its reactivation, using , , and . We detected X-ray pulsations in the and observations, and constrained the spin period derivative to $|\dot{P}| < 6 \times 10^{-11}$ss$^{-1}$ (3$\sigma$ c.l.). The pulse profile showed a variable shape switching between single and double-peaked as a function of time and energy. The pulsed fraction decreased from $\sim34$% to $\sim$11% (5–10keV) over $\sim$10 days. The X-ray spectrum was well fit by an absorbed blackbody model with temperature decreasing from $kT_{\rm BB} \sim$ 1.6 to 0.6–0.7keV, plus a non-thermal component ($\Gamma\sim 1.2$) observed up to $\sim$25keV with . The 0.3–10keV X-ray luminosity increased in less than four days from $\sim4\times 10^{33}$ to about 2.5$\times10^{35}$ and then decreased again to 1.4$\times 10^{34}$ over the following three weeks of the outburst. We also detected several X-ray bursts, with properties typical of short magnetar bursts.'
author:
- 'A. Borghese'
- 'F. Coti Zelati'
- 'N. Rea'
- 'P. Esposito'
- 'G. L. Israel'
- 'S. Mereghetti'
- 'A. Tiengo'
title: 'THE X-RAY REACTIVATION OF THE RADIO BURSTING MAGNETAR '
---
ß
å[1E1547$-$5408]{}
Introduction {#sec:intro}
============
[ccccccccc]{} /XRT (PC) & 00033349044 & 2020-04-23 15:16:16 & 2020-04-23 15:49:27 & 2.0 & 0.012$\pm$0.002 & – & – & 0.04\
/XRT (PC) & 00968211001 & 2020-04-27 19:41:56 & 2020-04-27 20:15:09 & 1.8 & 0.37$\pm$0.01 & 1.6$\pm$0.1 & 0.58$\pm$0.05 & 4.3$\pm$0.3\
/XTI & 3020560101 & 2020-04-28 00:38:31 & 2020-04-28 16:21:20 & 4.7 & 2.88$\pm$0.04 & 1.00$\pm$0.02 & 0.96$\pm$0.02 & 1.83$\pm$0.05\
/XRT (PC) & 00033349045 & 2020-04-28 18:00:36 & 2020-04-28 21:37:41 & 2.9 & 0.077$\pm$0.005 & 0.89$_{-0.06}^{+0.07}$ & 0.59$_{-0.06}^{+0.09}$ & 0.42$\pm$0.04\
/XRT (WT) & 00033349046 & 2020-04-29 13:07:57 & 2020-04-29 13:32:57 & 1.5 & 0.09$\pm$0.01 & 0.77$_{-0.11}^{+0.14}$ & 0.75$_{-0.17}^{+0.31}$ & 0.36$_{-0.06}^{+0.07}$\
/XTI & 3020560102 & 2020-04-29 13:47:17 & 2020-04-29 14:05:20 & 1.1 & 0.88$\pm$0.05 & 0.65$\pm$0.01 & 1.10$\pm$0.05 & 0.33$\pm$0.02\
/XRT (PC) & 00033349047 & 2020-04-29 17:54:22 & 2020-04-29 18:27:38 & 2.0 & 0.072$\pm$0.006 & 0.76$_{-0.07}^{+ 0.08}$ & 0.75$_{-0.12}^{+0.17}$ & 0.32$\pm$0.04\
/XTI & 3655010101 & 2020-04-29 21:31:57 & 2020-04-29 21:48:40 & 0.8 & 0.43$\pm$0.04 & 0.65$\pm$0.01 & 1.00$\pm$0.05 & 0.28$\pm$0.02\
/XTI & 3655010102 & 2020-04-30 00:37:56 & 2020-04-30 07:09:40 & 5.3 & 0.57$\pm$0.02 & 0.65$\pm$0.01 & 0.98$\pm$0.04 & 0.26$\pm$0.01\
/XTI & 3020560103 & 2020-04-30 13:02:45 & 2020-04-30 13:17:20 & 0.8 & 0.56$\pm$0.04 & 0.65$\pm$0.01 & 0.97$\pm$0.05 & 0.26$\pm$0.02\
/XRT (PC) & 00033349048 & 2020-04-30 05:29:05 & 2020-04-30 18:27:53 & 1.9 & 0.054$\pm$0.005 & 0.85$_{-0.09}^{+0.11}$ & 0.54$_{-0.10}^{+0.15}$ & 0.28$_{-0.04}^{+0.05}$\
/XRT (WT) & 00033349049 & 2020-04-30 07:10:24 & 2020-04-30 11:47:56 & 1.5 & 0.05$\pm$0.01 & 0.85$_{-0.14}^{+0.21}$ & 0.57$_{-0.16}^{+0.28}$ & 0.32$_{-0.06}^{+0.08}$\
/XRT (PC) & 00033349050 & 2020-05-01 02:03:14 & 2020-05-01 22:42:20 & 2.1 & 0.056$\pm$0.005 & 0.82$_{-0.08}^{+0.10}$ & 0.57$_{-0.10}^{+0.14}$ & 0.28$_{-0.04}^{+0.05}$\
/XRT (WT) & 00033349051 & 2020-05-01 12:58:08 & 2020-05-01 13:20:56 & 1.4 & 0.05$\pm$0.01 & 0.51$\pm$0.09 & 1.5$_{-0.4}^{+1.0}$ & 0.20$\pm$0.04\
FPMA/B & 80602313002 & 2020-05-02 00:06:09 & 2020-05-02 20:31:09 & 37.1/36.9 & 0.175$\pm$0.003 & 0.63$_{-0.06}^{+0.07}$ & 0.67$_{-0.14}^{+0.28}$ & 0.32$\pm$0.01\
/XRT (WT) & 00033349053 & 2020-05-02 11:50:05 & 2020-05-02 13:28:56 & 0.7 & 0.06$\pm$0.02 & 0.76$_{-0.12}^{+0.15}$ & 0.73$_{-0.19}^{+0.36}$ & 0.31$_{-0.07}^{+0.08}$\
/XRT (PC) & 00033349052 & 2020-05-02 16:33:41 & 2020-05-02 23:02:54 & 1.2 & 0.027$\pm$0.005 & 0.86$_{-0.14}^{+0.21}$ & 0.43$_{-0.11}^{+0.19}$ & 0.19$_{-0.05}^{+0.06}$\
/XRT (WT) & 00033349055 & 2020-05-03 12:55:54 & 2020-05-03 13:23:56 & 1.7 & 0.020$\pm$0.009 & 0.76$_{-0.16}^{+0.21}$ & 0.55$_{-0.17}^{+0.42}$ & 0.17$_{-0.04}^{+0.05}$\
/XRT (PC) & 00033349054 & 2020-05-03 22:23:05 & 2020-05-03 22:48:52 & 1.5 & 0.050$\pm$0.006 & 0.82$_{-0.09}^{+0.11}$ & 0.56$_{-0.10}^{+0.15}$ & 0.27$\pm$0.05\
/XRT (PC) & 00033349056 & 2020-05-04 01:47:23 & 2020-05-04 18:04:51 & 3.4 & 0.040$\pm$0.003 & 0.80$\pm$0.07 & 0.52$_{-0.07}^{+0.11}$ & 0.20$\pm$0.02\
/XRT (WT) & 00033349057 & 2020-05-04 12:40:56 & 2020-05-04 13:07:56 & 1.6 & 0.07$\pm$0.01 & 0.92$_{-0.10}^{+0.12}$ & 0.54$_{-0.10}^{+0.15}$ & 0.42$_{-0.06}^{+0.07}$\
/XRT (PC) & 00033349058 & 2020-05-05 03:17:19 & 2020-05-05 13:01:52 & 1.9 & 0.034$\pm$0.004 & 0.62$_{-0.07}^{+0.09}$ & 0.77$_{-0.17}^{+0.28}$ & 0.13$\pm$0.02\
/XRT (WT) & 00033349059 & 2020-05-05 20:40:09 & 2020-05-05 21:02:56 & 1.4 & 0.05$\pm$0.01 & 0.63$_{-0.08}^{+0.10}$ & 0.88$_{-0.19}^{+0.32}$ & 0.19$\pm$0.04\
/XRT (PC) & 00033349060 & 2020-05-06 06:36:44 & 2020-05-06 08:20:52 & 1.3 & 0.031$\pm$0.005 & 0.62$_{-0.08}^{+0.11}$ & 0.71$_{-0.18}^{+0.31}$ & 0.11$\pm$0.02\
/XRT (PC) & 00033349061 & 2020-05-07 09:30:09 & 2020-05-07 20:56:54 & 3.7 & 0.035$\pm$0.003 & 0.61$\pm$0.05 & 0.86$_{-0.13}^{+0.18}$ & 0.11$\pm$0.02\
/XRT (PC) & 00033349062 & 2020-05-10 04:28:08 & 2020-05-10 22:15:52 & 3.2 & 0.043$\pm$0.004 & 0.75$_{-0.07}^{+0.08}$ & 0.64$_{-0.09}^{+0.13}$ & 0.23$\pm$0.03\
/XRT (WT) & 00033349063 & 2020-05-10 06:01:43 & 2020-05-10 10:56:56 & 3.2 & 0.030$\pm$0.007 & 0.63$_{-0.06}^{+0.07}$ & 0.89$_{-0.15}^{+0.24}$ & 0.20$\pm$0.03\
FPMA/B & 80602313004 & 2020-05-10 23:51:09 & 2020-05-11 20:31:09 & 38.5/38.2 & 0.140$\pm$0.002 & 0.59$\pm$0.05 & 0.67$_{-0.13}^{+0.22}$ & 0.27$\pm$0.01\
/XTI & 3020560104 & 2020-05-11 14:30:54 & 2020-05-11 16:18:40 & 1.3 & 0.36$\pm$0.04 & 0.65$\pm$0.01 & 0.84$_{-0.04}^{+0.05}$ & 0.19$\pm$0.01\
/XRT (WT) & 00033349064 & 2020-05-13 02:22:52 & 2020-05-13 07:29:55 & 1.9 & 0.025$\pm$0.007 & 0.69$\pm$0.10 & 0.61$_{-0.13}^{+0.23}$ & 0.14$\pm$0.03\
/XRT (WT) & 00033349065 & 2020-05-13 09:03:52 & 2020-05-13 10:30:56 & 1.3 & 0.025$\pm$0.007 & 0.69$\pm$0.10 & 0.61$_{-0.13}^{+0.23}$ & 0.14$\pm$0.03\
/XRT (WT) & 00033349066 & 2020-05-15 00:31:07 & 2020-05-15 03:58:39 & 3.5 & 0.059$\pm$0.006 & 0.63$_{-0.06}^{+0.07}$ & 0.83$_{-0.15}^{+0.23}$ & 0.16$\pm$0.03\
Magnetars are isolated X-ray pulsars with spin periods in the 0.3–12s range and large spin-down rates, implying particularly strong surface dipolar magnetic fields of the order of $B\sim10^{14}$–10$^{15}$G [see @kaspi17; @esposito18 for recent reviews]. These objects have a persistent X-ray luminosity of $L_{\mathrm{X}}\sim 10^{31}$–10$^{36}$, which is thought to be powered by the instabilities and decay of their extreme magnetic fields. Among isolated neutron stars, magnetars are the most variable, with an unpredictable bursting activity. They emit short ($<$1s) and bright ($L_{{\rm peak}}$ $\approx$ 10$^{39}$–10$^{41}$) bursts in the X-ray band, either sporadically or clustered in “forests” [e.g., @israel08]. These bursts are often accompanied by an enhancement of the X-ray persistent flux, up to three orders of magnitude above quiescence. Then, the flux usually relaxes back to the pre-outburst level on months/years timescales [@cotizelati18]. Recently, magnetar traits have been observed also in high-$B$ pulsars [e.g., @gavriil08; @archibald16], X-ray pulsars with dipolar fields as low as $6\times10^{12}$G [e.g., @rea10; @rea12a], and the central source of the supernova remnant RCW103 [e.g., @rea16; @dai16; @borghese18]. These findings started to show how magnetar-like emission might be more common within the neutron star population than previously expected.
( hereafter) was discovered in 2014, when the Burst Alert Telescope (BAT) on board of the *Neil Gehrels Swift Observatory* [@gehrels04] triggered on a short burst [@stamatikos14]. A follow-up campaign confirmed the source as a magnetar with spin period $P\sim3.25$s and spin-down rate $\dot{P}\sim1.43 \times 10^{-11}$ß, implying a dipole magnetic field $B\sim4.4 \times 10^{14}$G at the pole and characteristic age $\tau_{\mathrm{c}}\sim3.6$kyr [@israel16]. has been quite active since then, with intense outbursts in February 2015 and May–June 2016 [@younes17] and frequent bursting activity [@lin20a].
reactivated on 2020 April 27–28, emitting a forest of X-ray bursts [e.g., @palmer20; @younes20] accompanied by an increase of the persistent X-ray flux, as typical in magnetar outbursts. More interestingly, two millisecond radio bursts temporally coincident with a double-peaked hard X-ray burst were detected from the direction of the source [@scholz20; @bochenek20; @li20; @mereghetti20; @tavani20], strenghtening the long suspected connection between magnetars and fast radio bursts (FRBs; see @cordes19 [@petroff19] for reviews). However, besides these radio bursts, radio pulsed emission has not been detected so far from the source [e.g., @younes17; @lin20b].
This Letter reports on the results of our monitoring campaign of with , , and , covering the first $\sim$20 days since its reactivation. We describe the observations (§\[sec:obs\]) and report our timing and spectral analysis as well as a search for short bursts (§\[sec:analysis\]). Summary of the results and discussion follow (§\[sec:discuss\]).
{width="2.1\columnwidth"}
Observations and data reduction {#sec:obs}
===============================
We report in Table\[tab:observations\] the log of the observations used in this work. Data reduction was performed using tools in the [heasoft]{} package (version 6.27.2). Photon arrival times were referred to the Solar system barycenter using the source position (RA = 19$^\mathrm{h}$34$^\mathrm{m}$55$\fs$598, Dec = +21$^{\circ}$53$^{\prime}$47$\farcs$79, J2000.0; @israel16) and the JPL planetary ephemeris DE200. In the following, we adopt a distance of 6.6kpc [@zhou20] and quote all uncertainties at 1$\sigma$ confidence level (c.l.).
After the /BAT trigger, was monitored almost daily with the /XRT [@burrows05] either in photon counting (PC; timing resolution of 2.51s) or windowed timing (WT; 1.8ms) modes. The data were reprocessed and analysed with standard prescriptions.
In the first XRT observation performed after the BAT trigger, a dust scattering ring was detected around the source, extending from $\sim$1 to 2 arcmin [@kennea20; @mereghetti20]. This structure was no longer observed in a pointing performed the following day (a detailed study of this structure will be presented in a future paper). We collected the source photons from a 20-pixel circle (1 pixel $ = 2\farcs36$). Background counts were extracted from a region of the same size for WT data and an annulus with radii of 100 and 150 pixels, centered on the source, for the PC observations.
was observed with [@harrison13] twice, on 2020 May 2 and 11. The two focal plane modules FPMA and FPMB observed the source for a total on-source exposure time of 75.6 and 75.1ks, respectively. We used the tool [nupipeline]{} to create cleaned event files and filter out passages through the South Atlantic Anomaly. The source counts were collected within a circular region of radius 100 arcsec, while the background was estimated from a 100-arcsec circle on the same chip of the target. In both pointings, is detected until $\sim$25keV. We ran the script [nuproducts]{} to extract light curves and spectra and generate response files for both FPMs.
[@gendreau12] observed six times for a total on-source exposure time of $\sim$ 14ks. The data were processed via the [nicerdas]{} pipeline, with the tool [nicerl2]{} with standard filtering criteria. The background count rate and spectra were computed from observations of the blank-field regions using [nibackgen3C50]{}.
Analysis and results {#sec:analysis}
====================
Timing analysis {#sec:timing}
---------------
For the timing analysis, we selected events in the 1–5keV energy band for and 3–20keV for . The data sets of observations performed on the same day were merged to increase the source signal-to-noise ratio. We did not include /XRT observations in our timing analysis due to their poor counting statistics.
We calculated a power density spectrum (PDS) for all time series to search for the spin signal, assuming a 3.5$\sigma$ detection threshold for the signal (using the algorithm by @israel96), taking into account all the frequencies in the PDS. Pulsations were significantly detected over a blind search only during the first observation. The signal was then found in the second observation and in the combined pointings IDs. 3655010101$+$3655010102 by looking in the range of periods $P\pm \Delta P$ (at 3$\sigma$; the $\dot{P}$ component can be neglected) around the value measured in the first data set. The period values were then refined by means of a phase-fitting technique. We obtained the following results: $P$ = 3.24733(2)s for the combined data sets (April 29–30), $P$ = 3.247331(3)s for the first observation (May 2) and $P$ = 3.24734(1)s for the second observation (May 11). The above uncertainties and the variable pulse profile (see below) did not allow us to coherently phase the and observations. These period measurements imply an upper limit on the spin period derivative of $|\dot{P}| < 6 \times 10^{-11}$ss$^{-1}$ (3$\sigma$ c.l.), a factor of about four above the value inferred during the 2014 outburst [@israel16].
Fig.\[fig:spec\] shows the pulse profiles at the different epochs and as a function of energy. The profile shape varies considerably in time, changing from quasi-sinusoidal on April 29-30 to double-peaked on May 2 and 11 (the separation between the two peaks is about half rotational cycle). The profile shape is also highly variable with energy in the data sets, the second peak (at phase $\sim$0.6–0.7) being more prominent above 5keV and dominating above 10keV in the first observation.
The background-subtracted pulsed fraction (defined as the semi-amplitude of the sinusoidal functions describing the pulse divided by the source average count rate; Fig.\[fig:spec\]) decreased by a factor of $\approx3$ between May 2 and 11 (in the 3–5 and 5–10keV ranges; see Fig.\[fig:spec\]). No pulsations were detected over the 10–20keV band in the second observation, and we set a 3$\sigma$ upper limit on the pulsed fraction of $\sim$15%.
{width="1.45\columnwidth"} {width="1.45\columnwidth"} {width="1.45\columnwidth"} {width="1.45\columnwidth"} {width="1.45\columnwidth"} {width="1.45\columnwidth"} {width="1.45\columnwidth"} {width="1.45\columnwidth"} {width="1.45\columnwidth"} {width="1.45\columnwidth"} {width="1.45\columnwidth"} {width="1.45\columnwidth"} {width="1.45\columnwidth"} {width="1.45\columnwidth"}
Spectral analysis {#sec:spec}
-----------------
The spectral analysis was performed with the [xspec]{} fitting package. We adopted the [Tbabs]{} model [@wilms00] to describe the photoelectric absorption by the interstellar medium. The and background-subtracted spectra were grouped in at least 50 and 20 counts per bin, respectively. The /XRT spectra were grouped according to a minimum number of counts variable from observation to observation.
We started the spectral analysis by fitting all the /XRT spectra jointly with an absorbed blackbody model. The hydrogen column density was tied across all data sets, yielding $= (1.4 \pm 0.1) \times 10^{22}$. We obtained a reduced chi-square = 1.24 for 132 degrees of freedom (dof). We fit the spectra simultaneously, adopting the same model and fixing to the above-mentioned value. To avoid covariance between the values of the blackbody temperature and normalization, due to the limited energy band adopted for spectra (1–5keV), we tied up the temperature across all data sets in the joint fit except for the first one (= 1.02/486 dof).
The blackbody temperature reached a value of $kT_{\rm BB}=(1.6 \pm 0.1)$keV $\sim$75min after the first BAT trigger on April 27 at 18:26:20 UT [@palmer20]. It decreased to 0.89$_{-0.06}^{+ 0.07}$keV in the following day, and attained values in the range 0.6–0.7keV over the last $\sim$10 days of our monitoring (Table\[tab:observations\]; Fig.\[fig:spec\]). During the first $\sim$20 days of this new active phase, the observed flux dropped from $(4.4 \pm 0.2) \times 10^{-11}$ to $(1.7 \pm 0.3) \times 10^{-12}$ (0.3–10keV; Table\[tab:observations\]; Fig.\[fig:spec\]). These values translate into a luminosity of $(2.5 \pm 0.1) \times 10^{35}$ and $(1.4 \pm 0.2) \times 10^{34}$ (0.3–10keV), respectively. An XRT observation performed on April 23 (only 4 days prior to the outburst onset) found in quiescence with net count rate of $0.012 \pm 0.002$ counts s$^{-1}$ (0.3–10keV), corresponding to an observed flux of $\sim4 \times 10^{-13}$ and a luminosity of $\sim 4 \times 10^{33}$ (assuming an absorbed blackbody spectrum with $kT_{\rm BB} = 0.5$keV, $= 1.4 \times 10^{22}$).
Fig.\[fig:spec\] shows the spectra extracted from nearly simultaneous and /XRT data. The broad-band spectrum is well described by an absorbed blackbody model plus a power-law component accounting for the emission above 10keV ( was again fixed to the above-mentioned value). For the first epoch (May 2), the best-fitting values are: $kT_{\rm BB} = 0.63^{+0.07}_{-0.06}$keV, $R_{\rm BB} = 0.67^{+0.28}_{-0.14}$km, and photon index $\Gamma = 1.16 \pm 0.06$ (= 1.06/146 dof). For the second epoch (May 11), we derived $kT_{\rm BB} = 0.59 \pm 0.05$keV, $R_{\rm BB} = 0.67^{+0.22}_{-0.13}$km and $\Gamma = 1.20 \pm 0.06$ (= 0.9/145 dof). The observed fluxes were (6.8 $\pm$ 0.1) $\times 10^{-12}$ and (5.8 $\pm$ 0.1) $\times 10^{-12}$ (0.3–25keV), chronologically, giving luminosities of (3.79 $\pm$ 0.07) $\times 10^{34}$ and (3.23 $\pm$ 0.07) $\times 10^{34}$. At both epochs, the power-law component accounted for $\sim$85% of the total observed flux and its luminosity varied from (3.14 $\pm$ 0.08) $\times 10^{34}$ to (2.78 $\pm$ 0.07) $\times 10^{34}$ (0.3–25keV).
Burst search and properties
---------------------------
We inspected the light curves of all observations for the presence of short bursts. Our search algorithm estimates the Poisson probability for an event to be a random fluctuation, considering the total number of time bins $N$. We applied this algorithm to light curves binned with different time resolutions (2$^{-4}$, 2$^{-5}$ and 2$^{-6}$s) to be sensitive to bursts of different duration, except for the /XRT PC-mode event files that were binned at the available timing resolution (2.5073s). Bins having a probability smaller than 10$^{-4}(N N_{\rm trials})^{-1}$ are identified as bursts ($N_{\rm trials}$ is the number of different time resolutions adopted for the search). In Table\[tab:bursts\], we report the epochs of the bursts referred to the Solar system barycenter. Fluence and duration are given for the bursts detected in the and data sets. Their light curves are shown in Fig.\[fig:bursts\]. We do not report on the more than 100 short bursts detected in the first observation (Obs. ID 3020560101; @younes20) due to the complex light curve and instrument saturation problems. We extracted the spectra only for those events with at least 30 net counts, that is, two bursts detected in the observations (80602313002 \#4 and 80602313004 \#1 in Table\[tab:bursts\]). We adopted the Cash statistics and fitted the spectra using single-component models (a power-law, a blackbody and a bremsstrahlung). The blackbody and power-law model fits gave a satisfactory description for both events with a goodness probability[^1] of $\sim$55% and $\sim$40%, respectively.
For the blackbody model, we derived a temperature equivalent to (2.9 $\pm$ 0.5)keV for 80602313002 \#4 and (3.9 $\pm$ 0.7)keV for 80602313004 \#1. The corresponding fluxes were (1.0 $\pm$ 0.3) $\times$ 10$^{-8}$ and (1.8 $\pm$ 0.6) $\times$ 10$^{-8}$ in the 3–79keV energy range, converting to luminosity of (5.3 $\pm$ 1.5) $\times$ 10$^{37}$ and (9.3 $\pm$ 3.1) $\times$ 10$^{37}$.
Discussion {#sec:discuss}
==========
Since its discovery in 2014, the magnetar has been a prolific source, showing numerous X-ray outbursts and frequent bursting activity. We presented here the results of an intensive X-ray monitoring campaign of this source over about three weeks since the end of April 2020, when it emitted a forest of X-ray bursts, and two bright radio millisecond bursts with characteristics strongly reminiscent of FRBs [@scholz20; @bochenek20].\
1. [*Spin period and pulse profiles*]{}. We detected the source spin period in a observation on April 29–30, and in both data sets on May 2 and 11. Unfortunately, the spacing between the few detections, and the uncertainties on the periods, prevented us from extracting a phase-connected timing solution. The spin period measurements at the different epochs allowed us to set an upper limit on the period derivative of $|\dot{P}| < 6 \times 10^{-11}$ss$^{-1}$ (at 3$\sigma$ c.l.). This limit is compatible with the spin-down rate of $\dot{P} \sim 1.43 \times 10^{-11}$ß derived by [@israel16] in 2014, using a phase-connected timing analysis. The double-peaked morphology of the pulse profiles is markedly different from the quasi-sinusoidal modulation observed in the observation a few days before and in previous X-ray observations of the source [@israel16]. Timing noise and large pulse profile changes (in time and energy) are common during magnetar outbursts [@rea11], especially following X-ray bursting activity. The magnetar magnetosphere is subject to rapid changes before setting to a new quiescent configuration, which are responsible for the fast profile variations especially in the hard X-rays, where the emission is dominated by nonthermal photons. These changes might also lead to the formation of new bundles and hot spots on the surface, modifying the pulse profile also in the soft X-ray range.\
2. [*Luminosity, spectral evolution, and bursting activity*]{}. About three days before its reactivation, was observed by /XRT at a luminosity of $4\times 10^{33}$. Following the source reactivation, the X-ray luminosity reached a peak value of $\sim$2.5 $\times$ 10$^{35}$ (0.3–10keV), making this event the most powerful outburst so far detected from . The luminosity then dropped by more than one order of magnitude, down to $\sim$1.4 $\times$ 10$^{34}$ about three weeks later. However, this is still a factor $\sim$3 larger than the pre-outburst level. A similar rapid decay pattern was also observed for the strong outburst occurred in May–June 2016 [@younes17] and, overall, is not uncommon for magnetars in outburst [@cotizelati18]. During the entire monitoring, showed a thermal spectrum in the soft X-rays well described by an absorbed blackbody model quickly cooling from a temperature of (1.6$\pm$0.1)keV to $\sim$0.6–0.7keV. This spectral evolution is different from that observed in the previous outbursts, where the spectrum was better described either by a double-blackbody model or a blackbody-plus-power-law model, and the luminosity decay could be ascribed to the evolution of the hot blackbody or the power-law component [@younes17].
Emission was detected up to $\sim$25keV in our observations. The spectral shape was identical at the two epochs, and was adequately modelled by a power law model with index $\Gamma \sim 1.2$ and luminosity $\sim4 \times 10^{34}$ (extrapolated to the 10–50keV energy range). Hard X-ray emission from was seen also in a pointing performed $\sim$ 5 days after the 2015 outburst onset. In that case, the high-energy spectrum could be described by a slightly harder power-law component ($\Gamma$ $\sim$ 0.9) with a lower luminosity, $\sim$1 $\times$ 10$^{34}$ (10–50keV; @younes17).
The bursting activity of during this new outburst is not dissimilar from that previously observed in this and other magnetars. However, such activity is not so prolific in all magnetars, and it is expected to depend on the age of the source and the tangled configuration of its magnetic field [@perna11; @vigano13]. A very rough proxy for it is provided by the quiescent X-ray luminosity, which is predicted to be higher in magnetars with a more tangled and powerful magnetic field in the crust, since they are subject to larger crustal currents and $B$-field crustal dissipation (see Fig.\[fig:discussion\]). A significant anti–correlation between magnetar quiescent luminosities and their luminosity increases in outburst was observed [@pons12; @cotizelati18], suggesting the existence of a limiting luminosity of $\sim 10^{36}$ for magnetar outbursts (regardless of the source quiescent level), which holds also for the case of .\
3. [*Comparison with other magnetars and FRBs*]{}. Comparing the short X-ray bursts and outburst emitted by with those of the other Galactic magnetars, they are perfectly in line with the expectations, and there is nothing in the X-ray emission properties of this magnetar that would make it peculiar in any aspect [@cotizelati18]. However, the detection of radio bursts simultaneous with a bright magnetar-like burst [@scholz20; @bochenek20; @mereghetti20] showed for the first time that magnetar bursts might have bright radio counterparts. This result is particularly interesting in the context of the physical interpretation of FRBs, bright ms-duration transients coming from distant Galaxies. Their brightness temperatures imply a coherent radio emission, inevitably connecting them to pulsars. Several repeating FRBs have been discovered [@spitler14; @spitler16], reinforcing their proposed interpretation in terms of young bursting magnetars in other galaxies (e.g., @popov13 [@margalit20], and references therein).
![Quiescent X-ray luminosity of magnetars as a function of their dipolar magnetic field at the pole. Circles denote radio-loud magnetars, either in the form of bursts (; in bold) or pulsed emission (other sources). Markers are color-coded according to the spin-down power of each source. Values are from the Magnetar Outburst Online Catalogue (<http://magnetars.ice.csic.es/>; @cotizelati18), with updates for , and [@camilo18; @ybk17; @esposito20].[]{data-label="fig:discussion"}](Lx_B_1935_hot.pdf){width="9cm"}
Radio pulsed emission was so far restricted to five magnetars (see Fig.\[fig:discussion\]). Such emission is at variance with the typical radio pulsar emission, and it is always connected to some extent with the magnetar X-ray activation. However, similarly to radio pulsars, all radio-loud magnetars have a large spin-down power compared to their radio-quiet siblings, and quiescent X-ray luminosity below their rotational power (with the exception of ; @rea12b [@cotizelati18] and Fig.\[fig:discussion\]). has a high rotational power, but so far it did not show any radio pulsations [@younes17; @lin20b], while surprisingly emitting radio bursts during the outburst we report here. From the study of the bursting activity of this source, it becomes clear that: 1) not all X-ray magnetar bursts have necessarily a radio counterpart (see also @archibald20), and 2) many radio bursts from magnetars might have been missed due to the lack of large field-of-view instruments in the radio band. Hence, it might be a common characteristic after all. Future detections will shed light on these ms-radio bursts, their connection (or not) with faint radio pulsations (i.e. bright single pulses), their preferred X-ray burst counterparts. Population synthesis studies will allow a comparison between their rates and luminosity distributions and those observed in FRBs.
[cccccc]{}\[htb\] /XRT (PC) & 00968211001 \#1 & 2020-04-27 19:46:52 & – & –\
& \#2 & 19:57:22 & – & –\
& \#3 & 20:03:38 & – & –\
& \#4 & 20:09:12 & – & –\
& \#5 & 20:14:42 & – & –\
& \#6 & 20:15:30 & – & –\
/XRT (WT) & 00033349046 \#1 & 2020-04-29 13:10:41 & – & –\
& \#2 & 13:12:34 & – & –\
& \#3 & 13:19:38 & – & –\
& \#4 & 13:21:59 & – & –\
& \#5 & 13:33:48 & – & –\
/XTI & 3655010101 \#1 & 2020-04-29 21:49:09 & 8 & 15.625\
/XTI & 3655010102 \#1 & 2020-04-30 00:51:27 & 6 & 31.25\
& \#2 & 05:21:38.89 & 7 & 15.625\
& \#3 & 06:56:59 & 11 & 15.625\
/XRT (WT) & 00033349049 \#1 & 2020-04-30 08:53:44 & – & –\
/XRT (WT) & 00033349051 \#1 & 2020-05-01 13:03:55 & – & –\
& 80602313002 \#1 & 2020-05-02 05:43:12 & 11 & 31.25\
& \#2 & 10:19:46 & 13 & 31.25\
& \#3 & 10:27:46 & 58 & 0.125\
/XRT (WT) & 00033349059 \#1 & 2020-05-05 20:50:38 & – & –\
/XRT (WT) & 00033349063 \#1 & 2020-05-10 07:50:37 & – & –\
& 80602313004 \#1 & 2020-05-11 00:33:00 & 59 & 140.868\
& \#2 & 00:47:14 & 6 & 62.5\
& \#3 & 09:46:00 & 6 & 31.25\
& \#4 & 13:20:16 & 26 & 62.608\
& \#5 & 16:26:47 & 6 & 62.5\
& \#6 & 18:22:15 & 5 & 15.625\
& \#7 & 19:38:10 & 4 & 15.625\
/XTI & 3020560104 \#1 & 2020-05-11 14:47:00 & 6 & 15.625\
/XRT (WT) & 00033349064 \#1 & 2020-05-13 02:27:45 & – & –\
& \#2 & 07:10:44 & – & –\
& \#3 & 07:12:16 & – & –\
& \#4 & 07:16:22 & – & –\
& \#5 & 07:27:04 & – & –\
/XRT (WT) & 00033349065 \#1 & 2020-05-13 09:12:39 & – &\
& \#2 & 10:24:18 & – &\
/XRT (WT) & 00033349066 \#1 & 2020-05-15 01:06:18 & – &\
& \#2 & 02:16:50 & – & –\
& \#3 & 02:25:54 & – & –\
& \#4 & 02:36:56 & – & –\
& \#5 & 02:43:10 & – & –\
& \#6 & 02:43:19 & – & –\
& \#7 & 03:53:54 & – & –\
& \#8 & 03:55:00 & – & –\
This research is based on observations with (NASA), (CaltTech/NASA/JPL) and (NASA/ASI/UKSA). We thank the , and teams for promptly scheduling our observations. We made use of public and observations asked by the CHIME/FRB collaboration and the team, that we acknowledge. We acknowledge support from the PHAROS COST Action (CA16214). AB and FCZ are supported by Juan de la Cierva fellowships. AB, FCZ, and NR are supported by grants SGR2017-1383, PGC2018-095512-BI00, and the ERC Consolidator Grant “MAGNESIA" (nr.817661). GLI, SM and AT acknowledge financial support from the Italian MIUR through PRIN grant 2017LJ39LM.
natexlab\#1[\#1]{}\[1\][[\#1](#1)]{} \[1\][doi: [](http://doi.org/#1)]{} \[1\][[](http://ascl.net/#1)]{} \[1\][[](https://arxiv.org/abs/#1)]{}
, R. F., [Kaspi]{}, V. M., [Tendulkar]{}, S. P., & [Scholz]{}, P. 2016, , 829, L21
, R. F., [Scholz]{}, P., [Kaspi]{}, V. M., [Tendulkar]{}, S. P., & [Beardmore]{}, A. P. 2020, , 889, 160
, K. A. 1996, in Astronomical Data Analysis Software and Systems V, Vol. 101, XSPEC: The First Ten Years, ed. G. H. [Jacoby]{} & J. [Barnes]{} (ASP, San Francisco), 17–20
, J. K. 1995, in Astronomical Data Analysis Software and Systems IV., Vol. 77, [FTOOLS: A FITS Data Processing and Analysis Software Package]{}, ed. R. A. [Shaw]{}, H. E. [Payne]{}, & J. J. E. [Hayes]{} (ASP Conf. Ser., San Francisco, CA), 367
, C. D., [Ravi]{}, V., [Belov]{}, K. V., [et al.]{} 2020, arXiv e-prints, arXiv:2005.10828.
, A., [Coti Zelati]{}, F., [Esposito]{}, P., [et al.]{} 2018, , 478, 741,
, D. N., [Hill]{}, J. E., [Nousek]{}, J. A., [et al.]{} 2005, Space Science Reviews, 120, 165
, F., [Scholz]{}, P., [Serylak]{}, M., [et al.]{} 2018, , 856, 180
, J. M., & [Chatterjee]{}, S. 2019, , 57, 417,
, F., [Rea]{}, N., [Pons]{}, J. A., [Campana]{}, S., & [Esposito]{}, P. 2018, , 474, 961
, A., [Evans]{}, P. A., [Burrows]{}, D. N., [et al.]{} 2016, , 463, 2394,
, P., [Rea]{}, N., & [Israel]{}, G. L. 2018, in Timing Neutron Stars: Pulsations, Oscillations and Explosions, ed. T. [Belloni]{}, M. [Mendez]{}, & C. [Zhang]{}, ASSL, Springer, in press (preprint: astro-ph/1803.05716)
, P., [Rea]{}, N., [Borghese]{}, A., [et al.]{} 2020, arXiv e-prints, arXiv:2004.04083.
, F. P., [Gonzalez]{}, M. E., [Gotthelf]{}, E. V., [et al.]{} 2008, Science, 319, 1802
, N., [Chincarini]{}, G., [Giommi]{}, P., [et al.]{} 2004, , 611, 1005
, K. C., [Arzoumanian]{}, Z., & [Okajima]{}, T. 2012, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, Vol. 8443, , 844313
, F. A., [Craig]{}, W. W., [Christensen]{}, F. E., [et al.]{} 2013, , 770, 103
Hunter, J. D. 2007, Computing in Science & Engineering, 9, 90,
, G. L., & [Stella]{}, L. 1996, , 468, 369
, G. L., [Romano]{}, P., [Mangano]{}, V., [et al.]{} 2008, , 685, 1114
, G. L., [Esposito]{}, P., [Rea]{}, N., [et al.]{} 2016, , 457, 3448,
, V. M., & [Beloborodov]{}, A. M. 2017, , 55, 261
, J. A., [Beardmore]{}, A. P., [Page]{}, K. L., & [Palmer]{}, D. M. 2020, The Astronomer’s Telegram, 13679, 1
, C. K., [Lin]{}, L., [Xiong]{}, S. L., [et al.]{} 2020, arXiv e-prints, arXiv:2005.11071.
, L., [G[ö]{}[ğ]{}[ü]{}[s]{}]{}, E., [Roberts]{}, O. J., [et al.]{} 2020, , 893, 156,
, L., [Zhang]{}, C. F., [Wang]{}, P., [et al.]{} 2020, arXiv e-prints, arXiv:2005.11479.
, B., [Beniamini]{}, P., [Sridhar]{}, N., & [Metzger]{}, B. D. 2020, arXiv e-prints, arXiv:2005.05283.
, S., [Savchenko]{}, V., [Ferrigno]{}, C., [et al.]{} 2020, arXiv e-prints, arXiv:2005.06335.
, D. M., & [BAT Team]{}. 2020, GRB Coordinates Network, 27665, 1
, R., & [Pons]{}, J. A. 2011, , 727, L51
, E., [Hessels]{}, J. W. T., & [Lorimer]{}, D. R. 2019, , 27, 4,
, J. A., & [Rea]{}, N. 2012, , 750, L6,
, S. B., & [Postnov]{}, K. A. 2013, arXiv e-prints, arXiv:1307.4924.
, N., [Borghese]{}, A., [Esposito]{}, P., [et al.]{} 2016, , 828, L13
Rea, N., & Esposito, P. 2011, in High-Energy Emission from Pulsars and their Systems. Proceedings of the First Session of the Sant Cugat Forum on Astrophysics, ed. D. F. Torres & N. Rea, Astrophysics and Space Science Proceedings (Springer, Heidelberg), 247–273
, N., [Pons]{}, J. A., [Torres]{}, D. F., & [Turolla]{}, R. 2012, , 748, L12
, N., [Esposito]{}, P., [Turolla]{}, R., [et al.]{} 2010, Science, 330, 944
, N., [Israel]{}, G. L., [Esposito]{}, P., [et al.]{} 2012, , 754, 27,
, A., [Svinkin]{}, D., [Frederiks]{}, D., [et al.]{} 2020, arXiv e-prints, arXiv:2005.11178.
, L. G., [Cordes]{}, J. M., [Hessels]{}, J. W. T., [et al.]{} 2014, , 790, 101,
, L. G., [Scholz]{}, P., [Hessels]{}, J. W. T., [et al.]{} 2016, , 531, 202,
, M., [Malesani]{}, D., [Page]{}, K. L., & [Sakamoto]{}, T. 2014, GRB Coordinates Network, 16520, 1
, M., [Casentini]{}, C., [Ursi]{}, A., [et al.]{} 2020, arXiv e-prints, arXiv:2005.12164.
, [:]{}, [Andersen]{}, B. C., [et al.]{} 2020, arXiv e-prints, arXiv:2005.10324.
van der Walt, S., Colbert, S. C., & Varoquaux, G. 2011, Comput. Sci. Eng., 13, 22
, D., [Rea]{}, N., [Pons]{}, J. A., [et al.]{} 2013, , 434, 123
, J., [Allen]{}, A., & [McCray]{}, R. 2000, , 542, 914
, G., [Baring]{}, M. G., [Kouveliotou]{}, C., [et al.]{} 2017, , 851, 17
, G., [Kouveliotou]{}, C., [Jaodand]{}, A., [et al.]{} 2017, , 847, 85
, G., [Guver]{}, T., [Enoto]{}, T., [et al.]{} 2020, The Astronomer’s Telegram, 13678, 1
, P., [Zhou]{}, X., [Chen]{}, Y., [et al.]{} 2020, arXiv e-prints, arXiv:2005.03517.
[^1]: <https://heasarc.gsfc.nasa.gov/xanadu/xspec/manual/node84.html>
| {
"pile_set_name": "ArXiv"
} |
=-1.0in =0.375in
[**NEUTRINO OSCILLATIONS: SOME THEORETICAL IDEAS[^1]**]{}
Stephen M. Barr
Bartol Research Institute
University of Delaware
Newark, DE 19711
[**INTRODUCTION**]{}
Over the years, and especially since the discovery of the large mixing of $\nu_{\mu}$ seen in atmospheric neutrino experiments, there have been numerous models of neutrino masses proposed in the literature. In the last two years alone, as many as one hundred different models have been published. One of the goals of this talk is to give a helpful classification of these models. Such a classification is possible because in actuality there are only a few basic ideas that underlie the vast majority of published neutrino mixing schemes. After some preliminaries, I give a classification of three-neutrino models, and then in the last part of the talk I discuss in more detail one category of models — those with “lopsided" charged-lepton mass matrices. Finally, I talk about a specific very predictive model based on lopsided mass matrices that I have worked on with Albright and Babu.
[**THE DATA**]{}
There are four indications of neutrino mass that guide recent attemps to build models: (1) the solar neutrino problem, (2) the atmospheric neutrino anomaly, (3) the LSND experiment, and (4) dark matter. Several excellent reviews of the evidence for neutrino mass have appeared recently.$^1$
\(1) The three most promising solutions to the solar neutrino problem are based on neutrino mass. These are the small-angle MSW solution (SMA), the large-angle MSW solution (LMA), and the vacuum oscillation solution (VO). All these solutions involve $\nu_e$ oscillating into some other type of neutrino — in the models we shall consider predominantly $\nu_{\mu}$. In the SMA solution the mixing angle and mass-squared splitting between $\nu_e$ and the neutrino into which it oscillates are roughly $\sin^2 2 \theta \sim 5.5 \times 10^{-3}$ and $\delta m^2
\sim 5.1 \times 10^{-6} eV^2$. For the LMA solution one has $\sin^2 2 \theta \sim 0.79$, and $\delta m^2
\sim 3.6 \times 10^{-5} eV^2$. (The numbers are best-fit values from a recent analysis.$^2$) And for the VO solution $\sin^2 2 \theta \sim 0.93$, and $\delta m^2 \sim 4.4 \times 10^{-10} eV^2$. (Again, these are best-fit values from a recent analysis.$^3$)
\(2) The atmospheric neutrino anomaly strongly implies that $\nu_{\mu}$ is oscillating with nearly maximal angle into either $\nu_{\tau}$ or a sterile neutrino, with the data preferring the former possibility.$^4$ One has $\sin^2 2 \theta \sim 1.0$, and $\delta m^2
\sim 3 \times 10^{-3} eV^2$.
\(3) The LSND result, which would indicate a mixing between $\nu_e$ and $\nu_{\mu}$ with $\delta m^2 \sim 0.1 - 1 eV^2$ is regarded with more skepticism for two reasons. The experimental reason is that KARMEN has failed to corroborate the discovery, though KARMEN has not excluded the entire LSND region. The theoretical reason is that to account for the LSND result and also for both the solar and atmospheric anomalies by neutrino oscillations would require three quite different mass-squared splittings, and that can only be achieved with [*four*]{} species of neutrino. This significantly complicates the problem of model-building. In particular, it is regarded as not very natural, in general, to have a fourth sterile neutrino that is extremely light compared to the weak scale. For these reasons, the classification given in this talk will assume that the LSND results do not need to be explained by neutrino oscillations, and will include only three-neutrino models.
\(4) The fourth possible indication of neutrino mass is the existence of dark matter. If a significant amount of this dark matter is in neutrino mass, it would imply a neutrino mass of order several eVs. In order then to achieve the small $\delta m^2$’s needed to explain the solar and atmospheric anomalies one would have to assume that $\nu_e$, $\nu_{\mu}$ and $\nu_{\tau}$ were nearly degenerate. We shall not focus on such models in our classification, which is primarily devoted to models with “hierarchical" neutrino masses. However, in most models with nearly degenerate masses, the neutrino mass matrix consists of a dominant piece proportional to the identity matrix and a much smaller hierarchical piece. Since the piece proportional to the identity matrix would not by itself give oscillations, such models can be classified together with hierarchical mass models in most instances.
In sum, the models we shall classify are those which assume (a) three flavors of neutrino that oscillate ($\nu_e$, $\nu_{\mu}$, and $\nu_{\tau}$), (b) a hierarchical pattern of neutrino masses, (c) the atmospheric anomaly explained by $\nu_{\mu}$-$\nu_{\tau}$ oscillations with nearly maximal angle, and (d) the solar anomalies explained by $\nu_e$ oscillating primarily with $\nu_{\mu}$ with either small angle (SMA) or large angle (LMA, VO).
[**MAJOR DIVISIONS**]{}
There are several major divisions among models. One is between models in which the neutrino masses arise through the see-saw mechanism,$^5$ and those in which the neutrino masses are generated directly at low energy. In see-saw models, there are both left- and right-handed neutrinos. Consequently, there are five fermion mass matrices to explain: the four Dirac mass matrices, $U$, $D$, $L$, and $N$ of the up quarks, down quarks, charged leptons, and neutrinos, respectively, and the Majorana mass matrix $M_R$ of the right-handed neutrinos. The four Dirac mass matrices are all roughly of the weak scale, while $M_R$ is many orders of magnitude larger than the weak scale. After integrating out the superheavy right-handed neutrinos, the mass matrix of the left-handed neutrinos is given by $M_{\nu} = - N^T M_R^{-1} N$. Typically, in see-saw models, the four Dirac mass matrices are closely related to each other, either by grand unification or by flavor symmetries. That means that in see-saw models neutrino masses and mixings are just one aspect of the larger problem of quark and lepton masses, and are likely to shed great light on that problem, and perhaps even be the key to solving it. On the other hand, in most see-saw models $M_R$ is either not related or is tenuously related to the Dirac mass matrices of the quarks and leptons. The freedom in $M_R$ is the major obstacle to making precise predictions of neutrino masses and mixings in most see-saw schemes.
In non-see-saw schemes, there are no right-handed neutrinos. Consequently, there are only four mass matrices to consider, the Dirac mass matrices of the quarks and charged leptons, $U$, $D$, and $L$, and the Majorana mass matrix of the light left-handed neutrinos $M_{\nu}$. Typically in such schemes $M_{\nu}$ has nothing directly to do with the matrices $U$, $D$, and $L$, but is generated at low-energy by completely different physics.
The three most popular possibilities in recent models for generating $M_{\nu}$ at low energy in a non-see-saw way are (a) triplet Higgs, (b) variants of the Zee model,$^6$ and (c) R-parity violating terms in low-energy supersymmetry. (a) In triplet-Higgs models, $M_{\nu}$ arises from a renormalizable term of the form $\lambda_{ij} \nu_i \nu_j H_T^0$, where $H_T$ is a Higgs field in the $(1,3, +1)$ representation of $SU(3) \times SU(2) \times U(1)$. (b) In the Zee model, the Standard Model is supplemented with a scalar, $h$, in the $(1,1, +1)$ representation and having weak-scale mass. This field can couple to the lepton doublets $L_i$ as $L_i L_j h$ and to the Higgs doublets $\phi_a$ (if there is more than one) as $\phi_a \phi_b h$. Clearly it is not possible to assign a lepton number to $h$ in such a way as to conserve it in both these terms. The resulting lepton-number violation allows one-loop diagrams that generate a Majorana mass for the left-handed neutrinos. (c) In supersymmetry the presence of such R-parity-violating terms in the superpotential as $L_i L_j E^c_k$ and $Q_i D^c_j L_k$, causes lepton-number violation, and allows one-loop diagrams that give neutrino masses.
It is clear that in all of these schemes the couplings that give rise to neutrino masses have nothing to do with the physics that gives mass to the other quarks and leptons. While this allows more freedom to the neutrino masses, it would from one point of view be very disappointing, as it would mean that the observation of neutrino oscillations is almost irrelevant to the burning question of the origin of quark and charged lepton masses.
Another major division among models has to do with the kinds of symmetries that constrain the forms of mass matrices and that, in some models, relate different mass matrices to each other. There are two main approaches: (a) grand unification, and (b) flavor symmetry. Many models use both.
\(a) The simplest grand unified group is $SU(5)$. In minimal $SU(5)$ there is one relation among the Dirac mass matrices, namely $D= L^T$, coming from the fact that the left-handed charged leptons are unified with the right-handed down quarks in a $\overline{{\bf 5}}$, while the right-handed charged leptons and left-handed down quarks are unified in a ${\bf 10}$. In $SU(5)$ there do not have to be right-handed neutrinos, though they may be introduced. In $SO(10)$, which in several ways is a very attractive group for unification, the minimal model gives the relations $N = U \propto D = L$. In realistic models these relations are modified in various ways, for example by the appearance of Clebsch coefficients in certain entries of some of the mass matrices. It is clear that unified symmetries are so powerful that very predictive models are possible. Most of the published models which give sharp predictions for masses and mixings are unified models.
\(b) Flavor symmetries can be either abelian or non-abelian. Non-abelian symmetries are useful for obtaining the equality of certain elements of the mass matrix, as in models where the neutrino masses are nearly degenerate, and in the so-called “flavor democracy" schemes. Abelian symmetries are useful for explaining hierarchical mass matrices through the so-called Froggatt-Nielson mechanism.$^7$ The idea is that different fermion multiplets can differ in charge under a $U(1)$ flavor symmetry that is spontaneously broken by some “flavon" expectation value (or values), $\langle f_i \rangle$. Thus, different elements of the fermion mass matrices would be suppressed by different powers of $\langle f_i \rangle/M \equiv \epsilon_i \ll 1$, where $M$ is the scale of flavor physics. This kind of scheme can explain small mass ratios and mixings in the sense of predicting them to arise at certain orders in the small quantities $\epsilon_i$. A drawback of such models compared to many grand unified models, is that actual numerical predictions, as opposed to order of magnitude estimates, are not possible. On the other hand, models based on flavor symmetry involve less of a theoretical superstructure built on top of the Standard Model than do unified models, and could therefore be considered more economical in a certain sense. Unified models put more in but get more out than flavor symmetry.
[**THE PUZZLE OF LARGE $\nu_{\mu}-\nu_{\tau}$ MIXING**]{}
The most significant new fact about neutrino mixing is the largeness of the mixing between $\nu_{\mu}$ and $\nu_{\tau}$ This comes as somewhat of a surprise from the point of view of both grand unification and flavor symmetry approaches. Since grand unification relates leptons to quarks, one might expect lepton mixing angles to be small like those of the quarks. In particular, the mixing between the second and third family of quarks is given by $V_{cb}$, which is known to be $0.04$. That is to be compared to the nearly maximal mixing of the second and third families of leptons: $U_{\mu 3} \cong 1/\sqrt{2}
\cong 0.7$. It is true that even in the early 1980’s some grand unified models predicted large neutrino mixing angles. (Especially noteworthy is the remarkably prophetic 1982 paper of Harvey, Ramond, and Reiss,$^8$ which explicitly predicted and emphasized that there should be large $\nu_{\mu}- \nu_{\tau}$ mixing. However, in those days the top mass was expected to be light, and Ref. 8 chose it to be 25 GeV. That gave $V_{cb}$ in that model to be about $0.22$. The corresponding lepton mixing was further boosted by a Clebsch of 3. With the actual value of $m_t$ that we now know, the model of Ref. 8 would predict $U_{\mu 3}$ to be 0.12). What makes the largeness of $U_{\mu 3}$ a puzzle in the present situation is the fact that we now know that both $V_{cb}$ and $m_c/m_t$ are exceedingly small.
The same puzzle exists in the context of flavor symmetry. The fact that the quark mixing angles are small suggests that there is a family symmetry that is only weakly broken, while the large mixings of some of the neutrinos suggests that family symmetries are badly broken.
The chief point of interest in looking at any model of neutrino mixing is how it explains the large mixing of $\nu_{\mu}$ and $\nu_{\tau}$. This will be the feature that I will use to organize the classification of models.
[**CLASSIFICATION OF THREE-NEUTRINO MODELS**]{}
Virtually all published models fit somewhere in the simple classification now to be described. The main divisions of this classification are based on how the large $\nu_{\mu}-\nu_{\tau}$ mixing arises. This mixing is described by the element $U_{\mu 3}$ of the so-called MNS matrix (analogous to the CKM matrix for the quarks).
The mixing angles of the neutrinos are the mismatch between the eigenstates of the neutrinos and those of the charged leptons, or in other words between the mass matrices $L$ and $M_{\nu}$. Thus, there are two obvious ways of obtaining large $U_{\mu 3}$: either $M_{\nu}$ has large off-diagonal elements while $L$ is nearly diagonal, or $L$ has large off-diagonal elements and $M_{\nu}$ is nearly diagonal. Of course this distinction only makes sense in some preferred basis. But in almost every model there is some preferred basis given by the underlying symmetries of that model. This distinction gives the first major division in the classification, between models of what I shall call class I and class II. (It is also possible that the large mixing is due almost equally to large off-diagonal elements in $L$ and $M_{\nu}$, but this possibility seems to be realized in very few published models. I will put them into class II.)
If the large $U_{\mu 3}$ is due to $M_{\nu}$ (class I), then it becomes important whether $M_{\nu}$ arises from a non-see-saw mechanism or the see-saw mechanism. We therefore distinguish these cases as class I(1) and class I(2) respectively. In the see-saw models, $M_{\nu}$ is given by $- N^T M_R^{-1} N$, so a further subdivision is possible: between models in which the large off-diagonal elements are in $M_R$ and those in which they are in $N$. We call these class I(2A) and I(2B) respectively.
If $U_{\mu 3}$ is due to large off-diagonal elements in $L$, while $M_{\nu}$ is nearly diagonal (class II), then the question to ask is why, given that $L$ has large off-diagonal elements, there are not also large off-diagonal elements in the Dirac mass matrices of the other charged fermions, namely $U$ and $D$, causing large CKM mixing of the quarks. In the literature there seem to be two ways of answering this question. One way involves the CKM angles being small due to a cancellation between large angles that are nearly equal in the up and down quark sectors. We call this class II(1). The main examples of this idea are the so-called “flavor democracy models". The other idea is that the matrices $L$ and $D^T$ (related by unified or flavor symmetry) are “lopsided" in such a way that the large off-diagonal elements only affect the mixing of fermions of one handedness: left-handed for the leptons, making $U_{\mu 3}$ large, and right-handed for the quarks, leaving $V_{cb}$ small. We call this approach class II(2).
Schematically, one then has
$$\begin{array}{ll}
I & {\rm Large \; mixing \; from } \; M_{\nu} \\
& (1) \;\; {\rm Non \; see \; saw} \\
& (2) \;\; {\rm See \; saw} \\
& \;\;\;\;\; {\rm A. \; Large \; mixing \; from } \; M_R \\
& \;\;\;\;\; {\rm B. \; Large \; mixing \; from } \; N \\
II & {\rm Large \; mixing \; from } \; L \\
& (1) \;\; {\rm CKM \; small \; by \; cancellation} \\
& (2) \;\; {\rm lopsided } \; L.
\end{array}$$
Now let us examine the different categories in more detail, giving examples from the literature.
[**I(1) Large mixing from $M_{\nu}$, non-see-saw**]{}.
This kind of model gives a natural explanation of the discrepancy between the largeness of $U_{\mu 3}$ and the smallness of $V_{cb}$. $V_{cb}$ comes from Dirac mass matrices, which are all presumably nearly diagonal like $L$, whereas $U_{\mu 3}$ comes from the matrix $U_{\nu}$; and since in non-see-saw models $M_{\nu}$ comes from models the matrix $M_{\nu}$ comes from completely different physics than do the Dirac mass matrices it is not at all surprising if it has a very different form from the others, containing some large off-diagonal elements. While this basic idea is very simple and appealing, these models have the drawback that in non-see-saw models the form of $M_{\nu}$, since it comes from new physics unrelated to the origin of the other mass matrices, is highly unconstrained. Thus, there are few definite predictions, in general, for masses and mixings in such schemes. However, in some schemes constraints can be put on the new physics responsible for $M_{\nu}$.
As we saw, there are a variety of attractive ideas for generating a non-see-saw $M_{\nu}$ at low energy, and there are published models of neutrino mixing corresponding to all these ideas.$^{9-13}$ $M_{\nu}$ comes from triplet Higgs in Ref. 9; from the Zee mechanism in Ref. 10; and from R-parity and lepton-number-violating terms in a SUSY model in Ref. 11. In Ref. 12 a “democratic form" of $M_{\nu}$ is enforced by a family symmetry. Several other models in class I(1) exist in the literature.$^{13}$
[**I(2A) See-saw $M_{\nu}$, large mixing from $M_R$**]{}
In these models, $M_{\nu}$ comes from the see-saw mechanism and therefore has the form $- N^T M_R^{-1} N$. The large off-diagonal elements in $M_{\nu}$ are assumed to come from $M_R$, while the Dirac neutrino matrix $N$ is assumed to be nearly diagonal and hierarchical like the other Dirac matrices $L$, $U$, and $D$. As with the models of class I(1), these models have the virtue of explaining in a natural way the difference between the lepton angle $U_{\mu 3}$ and the quark angle $V_{cb}$. The quark mixings all come from Dirac matrices, while the lepton mixings involve the Majorana matrix $M_R$, which it is quite reasonable to suppose might have a very different character, with large off-diagonal elements.
However, there is a general problem with models of this type, which not all the examples in the literature convincingly overcome. The problem is that if $N$ has a hierarchical and nearly diagonal form, it tends to communicate this property to $M_{\nu}$. For example, suppose we take $N = {\rm diag}( \epsilon', \epsilon, 1) M$, with $1 \gg \epsilon \gg \epsilon'$. And suppose that the $ij^{th}$ element of $M_R^{-1}$ is called $a_{ij}$. Then the matrix $M_{\nu}$ will have the form
$$M_{\nu} \propto \left( \begin{array}{ccc}
\epsilon^{\prime2} a_{11} & \epsilon' \epsilon a_{12} &
\epsilon' a_{13} \\ \epsilon' \epsilon a_{12} & \epsilon^2 a_{22} &
\epsilon a_{23} \\ \epsilon' a_{13} & \epsilon a_{23} &
a_{33} \end{array} \right).$$
If all the non-vanishing elements $a_{ij}$ were of the same order of magnitude, then obviously $M_{\nu}$ is approximately diagonal and hierarchical. The contribution to the leptonic angles coming from $M_{\nu}$ would therefore typically be proportional to the small parameters $\epsilon$ and $\epsilon'$. This suggests that to get a value of $U_{\mu 3}$ that is of order 1, it is necessary to have the small parameter coming from $N$ get cancelled by a correspondingly large parameter from $M_R^{-1}$. The trouble is that to have such a conspiracy between the magnitudes of parameters in $N$ and $M_R$ is unnatural, in general, since these matrices have very different origins. This problem has been pointed out by various authors.$^{14}$ We shall call it the Dirac-Majorana conspiracy problem.
There are several models in the literature that fall into class I(2A).$^{15-17}$ Of these, an especially interesting paper is that of Jezabek and Sumino,$^{15}$ because it shows that a Dirac-Majorana conspiracy can be avoided. Jezabek and Sumino consider the case that the Dirac and Majorana matrices of the neutrinos have the forms
$$N = \left( \begin{array}{ccc} x^2 y & 0 & 0 \\ 0 & x & x \\
0 & O(x^2) & 1 \end{array} \right) m_D, \;\;\;
M_R = \left( \begin{array}{ccc} 0 & 0 & A \\ 0 & 1 & 0 \\
A & 0 & 0 \end{array} \right) m_R,$$
where $x$ is a small parameter. If one computes $M_{\nu}
= - N^T M_R^{-1} N$ one finds that
$$M_{\nu} = - \left( \begin{array}{ccc} 0 & O(x^4 y/A) & x^2 y/A \\
O(x^4 y/A) & x^2 & x^2 \\ x^2 y/A & x^2 & x^2 \end{array} \right)
m_D^2/m_R.$$
Note that this gives a maximal mixing of the second and third families, without having to assume any special relationship between the small parameter in $N$ (namely $x$) and the parameter in $M_R$ (namely $A$). Altarelli and Feruglio$^{16}$ generalize this example, showing that the same effect occurs if $M_R$ is taken to have a triangular symmetric form.
An interesting point about the form of $M_{\nu}$ in Eq. (4) is that it gives bimaximal mixing. This is easily seen by doing a rotation of $\pi/4$ in the 2-3 plane, bringing the matrix to the form
$$M_{\nu}' = \left( \begin{array}{ccc} 0 & z & z' \\ z & 0 & 0 \\
z' & 0 & 2 x^2 \end{array} \right).$$
In the 1-2 block this matrix has a Dirac form, giving nearly maximal mixing of $\nu_e$.
Other published models that fall into class I(2) are given in Ref. 17.
[**I(2B) See-saw $M_{\nu}$, large mixing from $N$**]{}
At least at first glance, this seems to be a less natural approach. the point is that if the large $U_{\mu 3}$ is due to large off-diagonal elements in $N$, it might be expected that the other Dirac mass matrices, $U$, $D$, and $L$, would also have large off-diagonal elements, giving large CKM angles. There are ways around this objection, and a few interesting models that fall into this class have been constructed. However, experience seems to show that this approach is harder to make work than the others, and fewer models of this type exist in the literature.$^{18}$
[**II(1) Large mixing from $L$, CKM small by cancellation**]{}
If the large value of $U_{\mu 3}$ comes from large off-diagonal elements in the mass matrix $L$ of the [*charged*]{} leptons, then it is most natural to assume that the other Dirac mass matrices have large off-diagonal elements also. Why, then, are the CKM angles small? One possibility is that the CKM angles are small because of an almost exact cancellation between large angles needed to diagonalize $U$ and $D$. That, in turn, would imply that $U$ and $D$, even though highly non-diagonal, have nearly identical forms. This is the idea realized in so-called “flavor democracy" models.
In flavor democracy models, a permutation symmetry $S_3 \times S_3$ among the left- and right-handed fermions causes the Dirac mass matrices $L$, $D$, and $U$ to have the form
$$L, D, U \propto \left( \begin{array}{ccc} 1 & 1 & 1 \\
1 & 1 & 1 \\ 1 & 1 & 1 \end{array} \right).$$
Smaller contributions that break the permutation symmetry cause deviations from this form. These flavor-democratic forms are of rank 1, explaining why one family is much heavier than the others. On the other hand, the mass matrix of the neutrinos $M_{\nu}$ is assmed to have, by an $S_3$ symmetry acting on the left-handed neutrinos, the approximate form
$$M_{\nu} \propto \left( \begin{array}{ccc} 1 & 0 & 0 \\ 0 & 1 & 0 \\
0 & 0 & 1 \end{array} \right).$$
If $M_{\nu}$ were [*exactly*]{} proportional to the identity, then the basis of neutrino mass eigenstates would be undefined, and so then would be the MNS angles. However, once the small $S_3$-violating effects are taken into account, a neutrino basis is picked out. It is not surprising that, typically, the neutrino angles that are predicted are of order unity. On the other hand, the fact that $U$ and $D$ are nearly the same in form leads to a cancellation that tends to make the quark mixing angles small.
Exactly what angles are predicted for the neutrinos depends on the form of the small contributions to the mass matrices that break the permutation symmetries. There are many simple forms that might be assumed, and the possibilities are rich. There exists a large and growing literature on these models.$^{19}$
The idea of flavor democracy is an elegant one, especially in that it uses one basic idea to explain the largeness of the leptonic angles, the smallness of the quark angles, and the fact that one family is much heavier than the others. On the other hand, it requires the very specific forms given in Eqs. (6) and (7), which come from very specific symmetries. It is in this sense a narrower approach to the problem of fermion masses than some of the others I have mentioned.
It would be interesting to know whether models of class II(1), in which the CKM angles are small by cancellations of large angles, can be constructed using ideas other than flavor democracy.
[**II(2) Large mixing from “lopsided" $L$**]{}
We now come to what I regard as the most elegant way to explain the largeness of $U_{\mu 3}$: “lopsided" $L$. The basic idea is that the charged-lepton and down-quark mass matrices have the approximate forms
$$L \sim \left( \begin{array}{ccc} 0 & 0 & 0 \\ 0 & 0 & \epsilon \\
0 & \sigma & 1 \end{array} \right) m_D, \;\;
D \sim \left( \begin{array}{ccc} 0 & 0 & 0 \\ 0 & 0 & \sigma \\
0 & \epsilon & 1 \end{array} \right) m_D.$$
The “$\sim$" sign is used because in realistic models these $\sigma$ and $\epsilon$ entries could have additional factors of order unity, such as from Clebsches. The fact that $L$ is related closely in form to the [*transpose*]{} of $D$ is a very natural feature from the point of view of $SU(5)$ or related symmetries, and is a crucial ingredient in this approach. The assumption is that $\epsilon \ll 1$, while $\sigma \sim 1$. In the case of the charged leptons $\epsilon$ controls the mixing of the second and third families of [*right*]{}-handed fermions (which is not observable at low energies), while $\sigma$ controls the mixing of the second and third families of [*left*]{}-handed fermions, which contributes to $U_{\mu 3}$ and makes it large. For the quarks the reverse is the case because of the “$SU(5)$" feature: the small $O(\epsilon)$ mixing is in the left-handed sector, accounting for the smallness of $V_{cb}$, while the large $O(\sigma)$ mixing is in the right-handed sector, where it cannot be observed and does no harm.
In this approach the three crucial elements are these: (a) Large mixing of neutrinos (in particular of $\nu_{\mu}$ and $\nu_{\tau}$) caused by large off-diagonal elements in the [*charged*]{}-lepton mass matrix $L$; (b) this off-diagonal element appearing in a highly asymmetric or lopsided way; and (c) $L$ being similar to the transpose of $D$ by $SU(5)$ or a related symmetry.
To my knowledge the first place that all the elements of this approach appear is in a paper by Babu and Barr$^{20}$ and a sequel by Barr.$^{21}$ In those papers the emphasis was on a particular mechanism (in $SU(5)$ and $SO(10)$) by which the lopsidedness of $L$ and $D$ can arise. So perhaps it was not noticed by some readers that the scheme described in those papers was an instance of a more general mechanism.
The next time that this general idea can be found is in three papers that appeared almost simultaneously: Sato and Yanagida,$^{22}$ Albright, Babu, and Barr,$^{23}$ and Irges, Lavignac, and Ramond.$^{24}$
It is interesting that the same mechanism was arrived at independently by these three groups from completely different points of view. In Sato and Yanagida the model is based on $E_7$, and the structure of the matrices is determined by the Froggatt-Nielson mechanism. In Albright, Babu, and Barr, the model was based on $SO(10)$, and does not use the Froggett-Nielson approach. Rather, the constraints on the form of the mass matrices come from assuming a “minimal" set of Higgs for $SO(10)$ and choosing the smallest and simplest set of Yukawa operators that can give realistic matrices. Though both papers assume a unified symmetry larger than $SU(5)$, in both it is the $SU(5)$ subgroup that plays the critical role in relating $L$ to $D^T$. The model of Irges, Lavignac, and Ramond, like that of Sato and Yanagida, uses the Froggatt-Nielson idea, but is not based on a grand unified group. Rather, the fact that $L$ is related to $D^T$ follows ultimately from the requirement of anomaly cancellation for the various $U(1)$ flavor symmetries of the model. However, it is well known that anomaly cancellation typically enforces charge assignments that can be embedded in unified groups. So that even though the model does not contain an explicit $SU(5)$, it could be said to be “$SU(5)$-like".
In the last two years, the same mechanism has been employed by a large number of authors using a variety of approaches.$^{25}$
[**A PREDICTIVE SO(10) MODEL WITH LOPSIDED L**]{}
The model that I shall now describe briefly was not constructed to explain neutrino phenomenology; rather it emerged from the attempt to find a realistic model of the masses of the charged leptons and quarks in the context of $SO(10)$, In particular, the idea was to take the Higgs sector of $SO(10)$ to be as minimal as possible, and then to find what this implied for the mass matrices of the quarks and leptons. In fact, in the first paper we wrote, we did not pay any attention to the neutrino spectrum. Then we noticed that the model in that paper actually predicted a large mixing of $\nu_{\mu}$ with $\nu_{\tau}$ and published a follow-up paper.$^{23}$ The reason for the large mixing of the mu and tau neutrinos was precisely the fact that the charged lepton mass matrix has a lopsided form.
The reason this lopsided form was built into this model (which I shall refer to as the ABB model henceforth) was that it was necessary to account for certain well-known features of the mass spectrum of the quarks. In particular, the mass matrix entry that is denoted $\sigma$ in Eq. (8) above plays three crucial roles in the ABB model that have nothing to do with neutrino mixing. (1) It is required to get the Georgi-Jarlskog$^{26}$ factor of 3 between $m_{\mu}$ and $m_s$. (2) It explains the value of $V_{cb}$. (3) It explains why $m_c/m_t \ll m_s/m_b$. Remarkably, it turns out not only to perform these three tasks, but also gives mixing of order 1 between $\nu_{\mu}$ and $\nu_{\tau}$. Not often are four birds killed with one stone!
In constructing the model, several considerations guided us. First, we assumed the “minimal" set of Higgs for $SO(10)$. It has been shown$^{27}$ that the smallest set of Higgs that will allow a realistic breaking of $SO(10)$ down to $SU(3) \times SU(2) \times U(1)$, with natural doublet-triplet splitting,$^{28}$ consists of a single adjoint (${\bf 45}$), two pairs of spinors (${\bf 16} + \overline{{\bf 16}}$), a pair of vectors (${\bf 10}$), and some singlets. The adjoint, in order to give the doublet-triplet splitting, must have a VEV proportional to the $SO(10)$ generator $B-L$. This fact is an important constraint. Second, we assumed that the qualitative features of the quark and lepton spectrum should not arise by artificial cancellations or numerical accidents. Third, we required that the Georgi-Jarlskog factor arise in a simple and natural way. Fourth, we assumed that the entries in the mass matrices should come from operators of low-dimension that arise in simple ways from integrating out small representations of fermions.
Having imposed these conditions of economy and naturalness on the model we were led to a structure coming from just six effective Yukawa terms (just five if $m_u$ is allowed to vanish). These gave the following mass matrices:
$$\begin{array}{ll}
U^0 = \left( \begin{array}{ccc} \eta & 0 & 0 \\ 0 & 0 & \frac{1}{3}
\epsilon \\ 0 & - \frac{1}{3} \epsilon & 1 \end{array} \right) m_U, \;\;\;
& D^0 = \left( \begin{array}{ccc} 0 & \delta & \delta' \\ \delta &
0 & \sigma + \frac{1}{3} \epsilon \\ \delta' & - \frac{1}{3} \epsilon
& 1 \end{array} \right) m_D \\
& \\
N^0 = \left( \begin{array}{ccc} \eta & 0 & 0 \\ 0 & 0 & - \epsilon \\
0 & \epsilon & 1 \end{array} \right) m_U, \;\;\; & L^0 = \left(
\begin{array}{ccc} 0 & \delta & \delta' \\ \delta & 0 & - \epsilon \\
\delta' & \sigma + \epsilon & 1 \end{array} \right) m_D.
\end{array}$$
(The first papers$^{23}$ gave only the structures of the second and third families, while this was extended to the first family in a subsequent paper.$^{29}$) Here $\sigma \cong 1.8$, $\epsilon \cong 0.14$, $\delta \cong |\delta'|
\cong 0.008$, $\eta \cong 0.6 \times 10^{-5}$. The patterns that are evident in these matrices are due to the $SO(10)$ group-theoretical characteristics of the various Yukawa terms. Notice several facts about the crucial parameter $\sigma$ that is responsible for the lopsidedness of $L$ and $D$. First, if $\sigma$ were not present, then instead of the Georgi-Jarlskog factor of 3, the ratio $m_{\mu}/m_s$ would be given by 9. (That is, the Clebsch of $\frac{1}{3}$ that appears in $D$ due to the generator $B-L$ gets squared in computing $m_s$.) Since the large entry $\sigma$ overpowers the small entries of order $\epsilon$, the correct Georgi-Jarlskog factor emerges. Second, if $\sigma$ were not present, $U$ and $D$ would be proportional, as far as the two heavier families are concerned, and $V_{cb}$ would vanish. Third, by having $\sigma \sim
1$ one ends up with $V_{cb}$ and $m_s/m_b$ being of the same order ($\epsilon$) as is indeed observed. And since $\sigma$ does not appear in $U$ (for group-theoretical reasons) the ratio $m_c/m_t$ comes out much smaller, of order $\epsilon^2$, also as observed. In fact, with this structure, the mass of charm is predicted correctly to within the level of the uncertainties.
Thus, for several reasons that have nothing to do with neutrinos one is led naturally to the very lopsided form that we found gives an elegant explanation of the mixing seen in atmospheric neutrino data!
From the very small number of Yukawa terms, and from the fact that $SO(10)$ symmetry gives the normalizations of these terms, and not merely order of magnitude estimates for them, it is not surprising that many precise predictions result. In fact there are altogether nine predictions.$^{29}$ Some of these are post-dictions (including the highly non-trivial one for $m_c$). But several predictions will allow the model to be tested in the future, including predictions for $V_{ub}$, and the mixing angles $U_{e2}$ $U_{e3}$.
In the first papers it appeared that the model only gave the small-angle MSW solution to the solar neutrino problem. In fact, if $\eta =0$, or if forms for $M_R$ are chosen that do not involve much mixing of the first-family right-handed neutrino with the others, then a very precise prediction for $U_{e2}$ results that is beautifully consistent with the small-angle MSW solution.$^{29}$ However, in a subsequent paper$^{30}$ we showed that for other simple forms of $M_R$ the model gives bi-maximal mixing. (This happens in a way similar to what we saw above in Eqs. (4) and (5) for the Jezabek-Sumino model.)
For more details of the ABB model and its predictions I refer you the papers I have mentioned.
(The classification given in this talk has been somewhat expanded in a paper by Barr and Dorsner.$^{31}$ That paper also contains a much more complete listing of three-neutrino models that have been published in the last few years. It also gives a general discussion of expectations for the parameter $U_{e3}$.)
[**REFERENCES**]{}
1. J.W.F. Valle, Neutrino physics at the turn of the millenium (hep-ph/9911224); S.M. Bilenky, Neutrino masses, mixings, and oscillations, Lectures at the 1999 European School of High Energy Physics, Casta Papiernicka, Slovakia, Aug. 22-Sept. 4, 1999 (hep-ph/0001311).
2. M.C. Gozalez-Garcia, P.C. de Holanda, C. Peña-Garay, and J.C.W. Valle, Status of the MSW solutions to the solar neutrino problem, hep-ph/9906469
3. V. Barger and K. Whisnant, Seasonal and energy dependence of solar neutrino vacuum oscillations, hep-ph/9903262
4. M.C. Gonzalez-Garcia, talk at International Workshop on Particles in Astrophysics and Cosmology: From Theory to Observation, Valencia, Spain, May 3-8, 1999.
5. M. Gell-Mann, P. Ramond, and Slansky, in [*Supergravity, Proc. Supergravity Workshop at Stony Brook*]{}, ed. P. Van Nieuwenhuizen and D.Z. Freedman (North-Holland, Amsterdam, 1979); T. Yanagida, [*Proc. Workshop on Unified theory and the baryon number of the universe*]{}, ed. O. Sawada and A. Sugamota (KEK, 1979).
6. A. Zee, Phys. Lett. [**B93**]{} (1980) 389; Phys. Lett. [**B161**]{} (1985) 141.
7. C. Froggatt and H.B. Nielson, Nucl. Phys. [**B147**]{} (1979) 277.
8. J.A. Harvey, D.B. Reiss, and P. Ramond, Mass relations and neutrino oscilations in an $SO(10)$ model, Nucl. Phys. [**B199**]{} (1982) 223-268.
9. R.N. Mohapatra and S. Nussinov, Bimaximal neutrino mixing and neutrino mass matrix, Phys. Rev. [**D60**]{} (1999) 013002 (hep-ph/9809415).
10. C. Jarlskog, M. Matsuda, S. Skadhauge, and M. Tanimoto, Zee mass matrix and bimaximal neutrino mixing, Phys. Lett. [**B449**]{} (1999) 240-252 (hep-ph/9812282).
11. M. Drees, S. Pakvasa, X. Tata, T. terVeldhuis, A supersymmetric resolution of solar and atmospheric neutrino puzzles, Phys. Rev. [**D57**]{} (1998) 5335-5339 (hep-ph/9712392).
12. K. Fukuura, T. Miura, E. Takasugi, and M. Yoshimura, Maximal CP violation, large mixings of neutrinos and democratic type neutrino mass matrix, Osaka Univ. preprint, OU-HET-326 (hep-ph/9909415).
13. G.K. Leontaris and J. Rizos, New fermion mass textures from anomalous $U(1)$ symmetries with baryon and lepton number conservation, CERN-TH-99-268 (hep-ph/9909206).
14. M. Jezabek and Y. Sumino, Neutrino mixing and seesaw mechanism, Phys. Lett. [**B440**]{} (1998) 327-331 (hep-ph/9807310); G. Altarelli and F. Feruglio, Neutrino mass textures from oscillations with maximal mixing, Phys. Lett. [**B439**]{} (1998) 112-118 (hep-ph/9807353).
15. M. Jezabek and Y. Sumino, Neutrino mixing and seesaw mechanism, Phys. Lett. [**B440**]{} (1998) 327-331 (hep-ph/9807310).
16. G. Altarelli and F. Feruglio, Phys. Lett. [**B439**]{} (1998) 112-118 (hep-ph/9807353).
17. B. Stech, Are the neutrino masses and mixings closely related to the masses and mixings of quarks?, talk at 23rd Johns Hopkins Workshop on Current Problems in Particle Theory: Neutrinos in the Next Millenium, Baltimore, MD, 10-12 June 1999 (hep-ph/9909268). M. Bando, T. Kugo, and K. Yoshioki, Neutrino mass textures with large mixing, Phys. Rev. Lett. [**80**]{} (1998) 3004-3007 (hep-ph/9710417). M. Abud, F. Buccella, D. Falcone, G. Ricciardi, and F. Tramontano, Neutrino masses and mixings in $SO(10)$, DSF-T-99-36 (hep-ph/9911238).
18. Q. Shafi and Z. Tavartkiladze, Proton decay, neutrino oscillations and other consequences from supersymmetric $SU(6)$ with pseudogoldstone Higgs, BA-99-39 (hep-ph/9905202); D.P. Roy, talk at 6th Topical Seminar on Neutrino and AstroParticle Physics, San Miniato, Italy, 17-21 May 1999 (hep-ph/9908262).
19. M. Fukugita, M. Tanimoto, and T. Yanagida, Atmospheric neutrino oscillation and a phenomenological lepton mass matrix, Phys. Rev. [**D57**]{} (1998) 4429-4432 (hep-ph/9709388); M. Tanimoto, Vacuum neutrino oscillations of solar neutrinos and lepton mass matrix, Phys. Rev. [**D59**]{} (1999) 017304 (hep-ph/9807283); H. Fritzsch and Z.-z. Xing, Large leptonic flavor mixing and the mass spectrum of leptons, Phys. Lett. [**B440**]{} (1998) 313-318 (hep-ph/9808272); S.K. Kang and C.S. Kim, Bimaximal lepton flavor mixing and neutrino oscillation, Phys. Rev. [**D59**]{} (1999) 091302 (hep-ph/9811379).
20. K.S. Babu and S.M. Barr, Phys. Lett. [**B381**]{} (1996) 202 (hep-ph/9511446).
21. S.M. Barr, Phys. Rev. [**D55**]{} (1997) 1659 (hep-ph/9607419).
22. J. Sato and T. Yanagida, Large lepton mixing in a coset space family unification on $E(7)/SU(5)\times U(1)^3$, Phys. Lett. [**B430**]{} (1998) 127-131 (hep-ph/9710516).
23. C.H. Albright, K.S. Babu, and S.M. Barr, Phys. Rev. Lett. [**81**]{} (1998) 1167 (hep-ph/9802314); C.H. Albright and S.M. Barr, Fermion masses in $SO(10)$ with a single adjoint Higgs field, Phys. Rev. [**D58**]{} (1998) 013002 (hep-ph/9712488).
24. N. Irges, S. Lavignac, and P. Ramond, Predictions from an anomalous $U(1)$ model of Yukawa hierarchies, Phys. Rev. [**D58**]{} (1998) 035003 (hep-ph/9802334).
25. Y. Nomura and T. Yanagida, Bimaximal neutrino mixing in $SO(10)$ (GUT), Phys. Rev. [**D59**]{} (1999) 017303 (hep-ph/9807325); Z. Berezhiani and A. Rossi, Grand unified textures for neutrino and quark mixings, JHEP 9903:002 (1999) (hep-ph/9811447); K. Hagiwara and N. Okamura, Quark and lepton flavor mixings in the $SU(5)$ grand unification theory, Nucl. Phys. [**B548**]{} (1999) 60-86 (hep-ph/9811495); G. Altarelli and F. Feruglio, A simple grand unification view of neutrino mixing and fermion mass matrices, Phys. Lett. [**B451**]{} (1999) 388-396 (hep-ph/9812475); K.S. Babu, J. Pati, and F. Wilczek, Fermion masses, neutrino oscillations and proton decay in the light of SuperKamiokande, (hep-ph/9812538); R. Barbieri, L.J. Hall, G.L. Kane, and G.G. Ross, Nearly degenarate neutrinos and broken flavor symmetry, OUTP-9901-P (hep-ph/9901228); K.I. Izawa, K. Kurosawa, Y. Nomura, and T. Yanagida, Grand unification scale generation through anomalous $U(1)$ breaking, Phys. Rev. [**D60**]{} (1999) 115016 (hep-ph/9904303); E. Ma, Permutation symmetry for neutrino and charged lepton mass matrices, Phys. Rev. [**D61**]{} (2000) 033012 (hep-ph/9909249); Q. Shafi and Z. Tavartkiladze, Bimaximal neutrino mixings and proton decay in $SO(10)$ with anomalous flavor $U(1)$, BA-99-63 (hep-ph/9910314); P. Frampton and A. Rasin, Non-abelian discrete symmetries, fermion mass textures and large neutrino mixing, IFP-777-UNC (hep-ph/9910522).
26. H. Georgi and S.L. Glashow, Phys. Lett. [**B86**]{} (1979) 297.
27. S.M. Barr and S. Raby, Minimal $SO(10)$ unification, Phys. Rev. Lett. [**79**]{} (1998) 4748-4751.
28. S. Dimopoulos and F. Wilczek, report No. NSF-ITP-82-07 (1981), in [*The unity of fundamental interactions*]{} Proceedings of the 19th Course of the International School of Subnuclear Physics, Erice, Italy, 1981 ed. A. Zichichi (Plenum Press, New York, 1983); K.S. Babu and S.M. Barr, Phys. Rev. [**D48**]{} (1993) 5354 (hep-ph/9306242); K.S. Babu and S.M. Barr, Phys. Rev. [**D50**]{} (1994) 3529 (hep-ph/9402291).
29. C.H. Albright and S.M. Barr, Phys. Lett. [**B452**]{} (1999) 287 (hep-ph/9901318).
30. C.H. Albright and S.M. Barr, minimal Higgs model, Phys. Lett. [**B461**]{} (1999) 218 (hep-ph/9906297).
31. S.M. Barr and Ilja Dorsner, hep-ph/0003058.
[^1]: Talk given at Orbis Scientiae Conference, Coral Gables, FLA, Dec. 16-19, 1999
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Delivering hands-on practice laboratories for introductory courses on operating systems is a difficult task. One of the main sources of the difficulty is the sheer size and complexity of the operating systems software. Consequently, some of the solutions adopted in the literature to teach operating systems laboratory consider smaller and simpler systems, generally referred to as instructional operating systems. This work continues in the same direction and is threefold. First, it considers a simpler hardware platform. Second, it argues that a minimal operating system is a viable option for delivering laboratories. Third, it presents a laboratory teaching platform, whereby students build a minimal operating system for an embedded hardware platform. The proposed platform is called MiniOS. An important aspect of MiniOS is that it is sufficiently supported with additional technical and pedagogic material. Finally, the effectiveness of the proposed approach to teach operating systems laboratories is illustrated through the experience of using it to deliver laboratory projects in the Operating Systems course at the University of Northern British Columbia. Finally, we discuss experimental research in computing education and considered the qualitative results of this work as part of a larger research endeavour.'
author:
- RAFAEL ROMAN OTERO ALEX ARAVIND
bibliography:
- 'acmsmall-sample-bibfile.bib'
title: 'MiniOS: an instructional platform for teaching operating systems labs'
---
<ccs2012> <concept> <concept\_id>10003456.10003457.10003527.10003531.10003533</concept\_id> <concept\_desc>Social and professional topics Computer science education</concept\_desc> <concept\_significance>500</concept\_significance> </concept> <concept> <concept\_id>10003456.10003457.10003527.10003531.10003537</concept\_id> <concept\_desc>Social and professional topics Computational science and engineering education</concept\_desc> <concept\_significance>300</concept\_significance> </concept> <concept> <concept\_id>10003456.10003457.10003527</concept\_id> <concept\_desc>Social and professional topics Computing education</concept\_desc> <concept\_significance>100</concept\_significance> </concept> <concept> <concept\_id>10010520.10010553.10010562.10010563</concept\_id> <concept\_desc>Computer systems organization Embedded hardware</concept\_desc> <concept\_significance>300</concept\_significance> </concept> <concept> <concept\_id>10011007.10010940.10010941.10010949</concept\_id> <concept\_desc>Software and its engineering Operating systems</concept\_desc> <concept\_significance>300</concept\_significance> </concept> </ccs2012>
Introduction
============
Operating Systems is a central topic in undergraduate computer science curricula. Comprehension of subsequent computer science courses relies on proper understanding of operating systems concepts. Whilst this is similar to many other undergraduate courses, what makes operating systems peculiar is the difficulty of delivering laboratory assignments. Due to its complexity and scope, OS courses are delivered in several styles. On one extreme, several universities across the world deliver purely theoretical OS courses. Contrarily, other universities, particularly top western Universities, offer OS courses with a heavy project component. This is the other extreme, and a more effective way of teaching OS, as it gives students the opportunity for a hands-on experience. (The rest fit in between these two extremes.) In the words of M. Ben-Ari:
> “Programming is the fundamental activity of computing. As such it must be a major component of courses for students of computing. Courses should not be purely descriptive; instead, they must require students to construct implementations.” [@benari]
Nonetheless, delivering labs where students write or modify an operating system in a semester is a challenge. Operating systems are large, intricate, concurrent, low-level pieces of software. Writing one requires dealing with: i) asynchronous interrupts; ii) direct access to memory and registers; iii) the inner details of the target computer architecture; iv) the size of the OS itself; and v) the concepts and ideas behind each of the different OS components. Thus, offering the same kind of practical exposure present in some other computer science undergraduate courses is, at best, impractical.
Several approaches for teaching OS laboratories have been proposed in literature. Given that concurrency and low-level programming (i.e. (i) and (ii) above) are inherent to the hardware platform programming model, efforts in the computer science education community have focused on creating smaller and simpler instructional OS (i.e. they have focused on (iv), (v), and less on (iii)). Continuing in this direction, this work takes the small-size philosophy of instructional OS further, and proposes a minimal system to deliver laboratory assignments. In addition, it attempts to lessen the difficulties that originate from programming a complex machine (i.e. (iii)). Specifically, it does so without opting for either simulated or emulated hardware, nor hiding it behind software abstraction. It instead proposes the use of less complex hardware. Then it combines everything together in a laboratory teaching platform called MiniOS. Lastly, we discuss the effectiveness of the proposed approach and our experience using it over the past few years.
Section 2 briefly reviews the literature related to this work. It categorizes different approaches and presents where our work stands in relation to it. Section 3 traces the origins of the main difficulties in teaching OS labs in terms of software system, hardware platform, and lack of expertise of students in OS development. Then it uses them as the basis to propose a new solution called MiniOS based on minimal software, minimal hardware, and a guide specifying the construction of the system. Section 4 describes MiniOS and its architecture, discusses the embedded target platform, and elaborates on the accompanying guide. Subsequently, a set of laboratory assignments together with recommendations of its delivery are provided. Section 5 discusses the evaluation of the final product, and the experience in using it to deliver laboratory projects. Lastly, in section 6, we conclude with some remarks and future research directions.
Related Work
============
The teaching of OS lab projects can be broadly classified into four approaches: (i) those where the OS is partially or entirely simulated; (ii) those modifying or extending a full-fledged operating system, either desktop, mobile, or embedded; (iii) those where a toy operating system is built from bare metal; and (iv) those modifying or extending an instructional OS (whether they execute on simulated, emulated, or actual hardware).
Simulation based approaches are attractive as they capture high-level functionality, which can be presented in a visual and intuitive manner. Yet simulations are unrealistic, thereby limiting the learning experience. Conversely, modifying or extending a full-fledged operating system, such as GNU/Linux, does provide the experience of working with a *real system*. This is, however, at the cost of a steep learning curve, which results in students only having time to modify a limited number of components in a superficial manner. It is our opinion that these two methods are inadequate, and we will not consider them further. Our views on production operating systems as teaching tools comply with those found in literature [@Survey]. Since our approach relates more with (iii) and (iv), they are described in more detail in the remainder of this section.
Building a Toy OS from the ground up
------------------------------------
Building a toy OS from the ground up involves students designing their own simple OS. Out of convenience, a virtual machine (e.g., bochs) is typically used as development platform; though it is possible, with some assistance, to have students execute their OS in actual hardware. Examples of instructional operating systems following this philosophy are the uMPS/Kaya platform [@kaya], the TempOS platform [@tempos], GeekOS [@GeekOS], VIREOS [@vireos], Black’s OS [@black], and Chadwick’s OS [@cambridge]. The building of a toy OS approach has the following advantages:
- There is no pre-existing OS to assimilate;
- Building the system from the ground up demonstrates how the system fits together, thereby providing a holistic view of it;
- Letting students build an OS of their own gives them a gratifying feeling (this is from our experience, and similar views are reported in [@kaya]).
The disadvantage, on the other hand, is that students need to work directly with hardware that is intricate and has a steep learning curve. Complex hardware together with the difficulties of writing an OS from the ground up, leaves no opportunity to cover more than a few topics in their most rudimentary forms. Consider the case of Chadwick’s OS, where four out of eleven lessons are dedicated solely to controlling the screen. The final OS then not only has little resemblance with a production OS, but is also a tiny toy—in the sense of not being developed enough to have any practical purpose.
Moreover, the lack of device drivers availability is a problem for both students and instructors; as they constitute a bulk of operating systems code. Writing drivers is a difficult technical task, beyond the skill set of anyone who is not an experienced kernel developer. Without having proper drivers support, it is impossible to go ahead with projects, be they lab assignments or final projects. For instance, in [@black], a lab project on user and kernel mode separation was rendered impossible due to the use of BIOS for accessing I/O. Because of such complexity, some systems such as KayaOS and VIREOS have opted for running on simpler simulated hardware. Although it does bring the complexity down, it is at the cost of realism.
Modifying/Extending an Instructional OS
---------------------------------------
In this approach, students are given the task of manipulating an instructional OS; namely, adding functionality or modifying the existing one. Unlike production operating systems, pedagogic ones are more compact. That is, the number of concepts, amount of code, and technical details that must be comprehended are fewer. Examples in this category are: Nachos [@nachos], Pintos [@pintos], PortOS [@portos], BabyOS [@babyos], OS/161 [@OS161], Topsy [@topsy], among others. There are some advantages to this approach:
- Their smaller size makes them more approachable than production operating systems.
- Interaction with hardware is not direct, as interaction occurs with pre-written lower abstraction layers.
- They (some more, some less) resemble operating systems as built in reality.
On the down side, resembling real operating systems entails complexity. Thus, these systems deal with the issue of how much realism must be traded for simplicity. Consider, for example, the case of modifying a FAT32 file system. Students ought to have some understanding of the file format itself, the actual—often non-trivial—code used to implement it and its interaction with other system modules. On the other hand, a simpler ad-hoc file format, which lends itself to easier comprehension, is not a file format used in deployed systems. In other words, there is no single system that fits both simplicity and realism. Instructors must, therefore, select one that adequate to their teaching objectives.
Like build-your-own approaches, most instructional systems cannot be used for any purpose other than instruction; as often they do not run on actual hardware, or they simply are not developed enough. Those that can are complex systems. A survey of instructional OSes can be found in [@Survey].
The Xinu approach
-----------------
An intermediate approach that fits in the two previous categories is Xinu [@comer]. Xinu is an instructional OS and is peculiar in that a guide (in the form of a book) to its design and implementation is available. While its design and inner workings are detailed, its implementation is also given a rationale and demonstrated in code. This makes Xinu more self-contained; meaning, that the necessary knowledge to put the system together is part of the guide. Altogether, the book removes the mystery surrounding the OS.
Instructors may have options for students to either extend/modify the system or build it in its entirety. So, the advantages and disadvantages are a mixture of the two previous approaches. Importantly, the system developed is not a toy, but a complete and functional operating system, and for the same reason, it is complex. In fact, it is intended to be used for advanced courses with a focus on production operating systems. Consider one of the highlight remarks from the back cover of the Xinu Book (Lynksys version):
> “Designed for advanced undergraduate or graduate courses, the book prepares students for the increased demand for operating system expertise in industry.” [@comer]
Similarly, from embedded Xinu’s website:
> “A student built operating system puts the student in the trenches of operating system development. The student will become intimately involved with the inner workings of an operating system.” [@xinuwiki]
Hence, this approach is less suited for any introductory operating systems course. Further, the guide does not touch on any hardware-related details, leaving to students and/or instructors the task of filling the considerable gap between the OS guide and the hardware documentation. One might point out that other operating systems such as Minix [@minix], Kaya, TempOS and Topsy also come with a document describing the system. Yet, none is as self-contained, nor offers the amount of detail as that of Xinu.
With this brief review of the related work, we next present the rationale for our approach.
Rationale for MiniOS
====================
A student-built production-like operating system that can run on real hardware is the ideal under the philosophy of hands-on experiential learning. Similar views are expressed in [@kaya] and [@OS161]. Whether the system of choice is mobile, desktop, embedded, or some other will depend on the specific instructional objectives of courses. In either case, a dichotomy exists between the ideal and fitting the workload into one semester. In this section, we attempt to explain this dichotomy and then provide the basis of our solution.
The issues of building a complex system
---------------------------------------
Operating systems are complex. Mosley et al [@tarpit] point out that complexity has a direct impact on one’s attempts to understand a system. They also identify the three main sources of complexity in software: [*state, flow of control,*]{} and [*code volume*]{}. Given the sizes of modern operating systems, as well as the topic in question (OS instruction), we focus our attention on code volume.
Modern operating system sizes are typically in the order of millions lines of code (LoC), and it is not surprising to find them in the list of the largest softwares [@AndroidLOC]. Consider, for instance, the latest versions of the mobile operating systems Android and Symbian, or the latest versions of the desktop systems GNU/Linux, Mac OS, and Windows; all of them are composed of millions of LoC.
For this reason, the computer science education community has favoured the use of smaller instructional operating systems. That is, simpler systems whose main purpose is to serve as a teaching tool. To put this into perspective, consider the size, in LoC, of some popular instructional systems: Minix and Xinu, with tens of thousands; Kaya OS with over 7,000; Pintos with over 5,000; and Nachos with approximately 2,500. Their smaller size enables labs to be carried out in one semester’s time (some with more difficulty than others).
Instructional systems may or may not have what we consider two important characteristics. In particular, being complete and functional. Complete in the sense of implementing the typical components of an OS, and functional in the sense of supporting execution of real applications on real hardware (e.g., a teller machine system, a basic laptop, or a robotic system). Unfortunately, a complete and functional system is more realistic, and thereby, more complex. For instance, Minix and Xinu are complete and functional systems; hence, unsuited for undergraduate instruction. From those remaining, none of them are functional, and their degree of completeness varies. Among them, nachos is arguably the smallest yet complete, due to their philosophy of minimal implementations. In fact, we use the term *minimal* to mean exactly what [@nachos] describe in their implementation philosophy: “Our approach was to build the simplest implementation we could think of for each sub-system” [@nachos].
Accordingly, we argue that, by means of minimal implementations, we can build a system with further reduced code volume. Moreover, we can use this reduction in size and complexity as an opportunity to:
- cover (i.e. implement) components that are otherwise “out of scope”, and
- build a system capable of serving a purpose using actual hardware.
In other words, we make the case for a minimal—yet complete and functional—instructional operating system. Thus far we have discussed the difficulties of dealing with OS software. Now, we consider another source of complexity: the target hardware platform.
The issues of complex hardware
------------------------------
Present time computers are intricate pieces of hardware. Manuals detailing the functionings (from a programmer perspective) of a modern 32- or 64-bit processor add up to at least a few thousand pages. To that, one must add the documentation detailing the functioning of the rest of the computer hardware, e.g., interrupt controller, BIOS/UEFI, timer, real-time clock, and others. For these reasons, writing non-trivial bare-metal applications (such as operating systems) is a technical, tedious, error-prone, and laborious task. One ought to know the precise inner workings of the computer if one hopes to direct it to do anything. Even though OS courses are customarily preceded by architecture or organization courses, these inner workings are often too advanced, and there are too many details to be covered in their entirety. Additionally, the machine exposes a programming model of asynchronous interrupts. Concurrent code accessing arbitrary memory and registers is one of the most challenging code students will encounter during their studies.
Then, how can a student possibly aspire to build an operating system, even a simple one, in one semester? The answer is simple—they cannot. For this particular source of complexity, solutions have been proposed in literature. One solution is to build the system for a hardware simulator or emulator that is simpler to interact with. This is advocated and demonstrated in Kaya OS, OS/161, VIREOS, PortOS, and Nachos. For example:
> “Simulators are used to eliminate the burden of working on a bare machine, which, given the time frame of a single term, is outside the scope of an undergraduate’s ability.” [@kaya]
A second solution is to abstract away hardware via a software layer. While this is indirectly followed by any instructional OS not meant to be built from bare metal, GeekOS explicitly follows this approach:
> “Working at the hardware level has two main disadvantages. First, hardware devices can be tricky to program correctly. A more fundamental problem is that debugging kernel code running on real hardware is difficult, even for experts. The contribution of our work is to show that both of these difficulties can be overcome without requiring heroic measures from students or instructors. We have implemented a tiny OS kernel, called GeekOS, which provides a sufficient abstraction layer over the hardware to hide the genuinely difficult details.” [@GeekOS]
A third solution is to compromise on the level of sophistication of the system, so as to simplify the technical (hardware) details required to build it. This is put into practice in Black’s OS and BabyOS, where students build a toy OS from bare metal.
We are of the opinion that exposing the students to real hardware is not only essential for a holistic understanding of the system, but also increases their engagement. Similar views are expressed by Pfaff et al [@pintos]. Therefore, we consider only the latter approach of compromising on the level of sophistication. Unfortunately, such compromise results in a system that does execute in real hardware, but is far from being complete and/or functional.
Yet, we argue that by targeting a simpler real hardware platform, we can decrease the technical knowledge required to build an instructional system; then, use that as an opportunity to build a complete and functional system. In other words, we make the case for a minimal—yet complete and functional—instructional operating system for a minimal hardware platform.
Instructional OSes achieve simplicity by trading the capabilities of full-fledged real systems. Next, we elaborate on it.
A minimal instructional OS for a minimal platform
-------------------------------------------------
If realism must be traded for simpler minimal implementations, the question that follows is, where is the ideal trade-off point between one and another? This is a difficult question, and it is (directly or indirectly) explored in each and all of the different instructional operating system proposals. For instance, Holland et al elaborate:
> “For teaching, a certain amount of realism is desirable. Too much realism, however, becomes both too complicated and, sometimes, realistically painful. \[...\] \[R\]eal OSes are immensely large and complicated, and are full of complexities and constructs for coping with real-world issues that have little instructional value.” [@OS161]
Liu et al also elaborate:
> “In the process of using BabyOS, we found that it is really difficult to make tradeoff between realism and simplicity. A certain amount of realism is desirable, otherwise BabyOS feels like an unreal OS. Too much realism, however, becomes too complicated and, student would fail to finish their projects.” [@babyos]
Even though we do not know where the ideal trade-off point resides, it is our intention to explore it by implementing a minimal instructional OS for a minimal hardware platform, which we call MiniOS.
Real being impractical, we focus on the minutiae that can preserve “relevant realism” in trade of “less relevant realism” (as far as undergraduate instruction goes). In particular, MiniOS is complete and functional; it is targeted for a real hardware platform; and it follows the design, layout, and mechanisms of real systems. Meanwhile, fault tolerance, robustness, efficiency, reliability, sophistication, and other attributes in deployed systems are not considered.
To put it bluntly, whilst MiniOS should not be deployed as part of an aircraft computer or an X-ray device, it is perfectly suitable for less important applications, such as a gardening system, or a small mobile robot—and such system, we believe, is well suited for instruction.
Thus far we have used the term *real hardware* generically, now it is time to specify a target platform.
### A minimal embedded hardware platform
We use the term minimal hardware to refer to those computers with the least amount of sophistication still capable of hosting an OS. For the sake of exploration, we have selected what we consider to be one of the smallest among them; more specifically, a 32-bit ARM low-end embedded platform. This choice is partly arbitrary and partly influenced by ARM’s popularity in the mobile and embedded systems industries.
It is worth noting that the term *embedded* does not imply simplicity. While there exists basic 8-bit microcontroller (MCU)-based embedded computers (e.g. a coffee maker’s computer), there too exists sophisticated 64-bit microprocessor (MPU)-based embedded computers (e.g. an industrial robot’s computer).
Despite the fact that some instructional systems, such as Minix, Xinu, and BabyOS are targeted for (or have been ported to) embedded platforms, they differ from our philosophy of minimal hardware. In fact, to our knowledge, there is not an existing instructional OS with similar views on hardware.
It is also important to clarify that MiniOS is not intended to be an embedded production OS. Like desktop systems, embedded production OSes are complex. They tend to be plagued with intricacies that make them adequate for deployment in life-critical applications such as aircraft and military. A representative sample, and in the smaller side of the size espectrum, is FreeRTOS [@freertos]; which, intended for low-end embedded platforms, has over 9,000 LoC [@FreeRTOSLOC].
A low-end embedded OS may seem as an over simplification, and naturally one raises the question of whether such a simple system has any instructional value outside the embedded systems realm.
### A low-end embedded OS as a teaching tool
The purpose of MiniOS is not to serve as a tool for teaching embedded systems, but to serve as a tool for teaching general principles that apply to operating systems. In fact, MiniOS is not well suited for teaching labs in embedded systems, as embedded-specific details are deliberately overlooked. With few exceptions where it is impossible, it is emphasized how they contrast with general purpose computers. Consider, for instance, the case of a MCU-based low-end embedded platform (a Hardvard architecture) with Flash as program memory; it must be brought to the students’ attention that general purpose systems do not, customarily, have non-volatile program memory in their address space. Thus a boot-loader for a MCU will be different than one for, say, a desktop computer.
Fortunately, the similarities are greater than the differences, and this is why we believe a simpler low-end embedded system can be used as a teaching tool. That is, for a course with no intention of preparing students for real-world OS development (whether embedded, desktop, or other).
One important benefit of working with MCU-based embedded platforms is the availability of device drivers. Hardware manufacturers typically release open source bare-metal middle-ware (mostly drivers) to be used on their platforms.
Finally, we argue that, recently, there has been a switch from traditional desktop systems to mobile and embedded systems (e.g. internet of things and wireless sensor networks). An embedded instructional system with wireless capabilities can be a tool for introducing students to the latter. A similar argument is expressed by Atkin and Sirer [@portos].
An equally important aspect of MiniOS is its guide. It covers building the system from nothing, and it is described in the following section.
From the ground-up: a guide to MiniOS design
--------------------------------------------
MiniOS is intended to be built from the ground up, on bare metal. For this, a guide to its design is primary. In a comprehensive and thorough manner the guide must—step by step—detail the construction of the system from nothing. All the technical details dealing with the hardware, the compiler, as well as OS concepts and their specific implementations should be covered; including details such as exceptions, memory mapped IO, linking of relocatable code, calling conventions, memory segmentation, and context switching.
Other instructional systems also advocate for the use of a guide or manual [@GeekOS], [@topsy], [@kaya], [@tempos], [@vireos], [@minix], [@pintos], and [@nachos]; some with more details and code than others. None, however, go to the amount of detail (instruction) that we consider necessary for building an OS from the ground up. (XINU is the exception; the amount of instruction offered as written material in [@comer] is near to what we advocate for.)
Guzdials [@guzdial] argues that the amount of instruction matters when teaching computer science to beginners. In particular, “putting introductory students in the position of discovery information for themselves is a bad idea.” Although this argument is given in the context of introductory programming (100 level courses), the instruction in question (operating systems) is introductory, as both systems programming and programming at such low-level of abstraction are substantially different from what students have encountered in preceding courses.
From experience we have noticed that, at this introductory stage, most students lack the experience, the patience, and the right approach to meticulously construct and debug low-level systems’ code. Moreover, they are faced with programming patterns and tricks specific to the machine’s programming model. While many of these patterns are simple and of common use, it can be difficult to re-invent them if one has never encountered them before. Also, in contrast with higher-level programming, bugs manifest differently (typically the CPU faults and does nothing) in low level. Code is highly dependent on a great number of machine-specific details, all of which must be set correctly, and access to raw memory requires precise knowledge of its organization and how instructions access it. Moreover, it is practically impossible for students to obtain all of the required details for OS construction from the thousands of pages included in the documentation, for they are not at the level of understanding the technicalities therein. The end result is that students are prone to get hopelessly stuck.
Consequently, we consider that a guide demonstrating how to build the system from the ground-up: as well as specifying, in a comprehensive manner, the technical details relevant to OS writing is a necessity for the delivery of OS labs. With this background, we now proceed to describe our proposed system called MiniOS.
MiniOS—Proposed OS instructional platform
=========================================
The proposed OS instructional platform consists of the system, the target hardware, and its construction guide. This chapter describes them and gives a set of suggested laboratory projects, as well as recommendations for their delivery.
The system
----------
First we present the high level architecture of the system, and then describe the different parts that constitute the system.
### Architecture
From an architectural point of view it is unclear the parts that must be included in a presumably minimal, complete, and functional operating system. It cannot be composed of too many parts (layers or modules) as to become too complex, nor it should have too few as to be incomplete or non-functional. Our approach on this is to incorporate components typically found in production systems, and offer the choice of what components make it into the system. Specifically, the system is built as a set of loosely coupled modules categorized in base modules and optional modules, as shown in Figure \[fig:arch\].
![MiniOS’ Architecture.[]{data-label="fig:arch"}](./minios-arch.png)
As their names suggest, *base modules* form the foundation of the system and must be implemented, whereas *optional modules* add specific functionality that may or may not be integrated in the system. This configuration gives lab instructors and students the flexibility to start with a minimal base and add modules to accommodate to their instructional objectives. Complying with the minimality principle, the system has as few lines of code and as few components as possible.
From a design perspective, we classify the modules into two types: [*primary*]{} and [*secondary*]{}. *Primary modules* represent an identifiable OS component: hardware abstraction layer (HAL), fault manager, memory protection, file system, scheduler, IO manager, network stack, system calls interface, IO event dispatcher, minilib, thread synchronizer, and command-line interface (CLI). *Secondary components* offer some abstraction or functionality but do not represent an OS components: context switcher, disk abstraction, network interfaces, app loader, IO, CPU, and interrupts. Every component is mapped to a source file of the same name. There are as many C or assembly files as there are components in the architecture.
From a software engineering perspective, a modular architecture has additional benefits. First, it improves modifiability of the system. It allows students to add or remove modules with little or no modification of others. Second, it improves local reasoning, hence aiding our main objective of making the system easier to comprehend. Such design is typically achieved with support from a programming language. However, since the system is written in C and assembly, we rely merely on discipline. Particularly, we strongly advise students to keep state confined to the scope of a module, and let module interaction occur only via interfaces; practices, which we demonstrate throughout the construction guide. It is worth noting that, although instructional systems are more or less designed in this manner, often modules end up keeping global state used by other modules. More importantly, we want to make it explicit that these software engineering practices are essential for our purpose.
An important aspect of MiniOS architecture is that, unlike production systems, device drivers are in direct contact with hardware. This means, the system is built on top of them, instead of them being part of the system itself. Although some instructional systems follow this design for simplicity purposes, we do it explicitly to support integration of open source third-party firmware that is often only available as bare-metal. With this small design choice, MiniOS benefits from available code from chip vendors and/or embedded systems enthusiasts. With this higher level description, next we will describe the individual components.
### Components description
Let us begin with base modules.
- [*HAL.*]{} This is the lowest layer of the system and it is responsible for providing sensible machine-independent abstractions to upper layers. Particularly, it implements three abstractions: CPU, interrupts, and IO.
- [*System.*]{} System is central to the rest of the modules, and is in control of all the system-related tasks, such as system initialization and kernel panics. Additionally, it offers implementations of various data structures to aid in the development of the kernel.
- [*Application loader.*]{} This module is responsible for the loading of applications from the SD Card.
- [*Console.*]{} The Console is a kernel shell that runs after system initialization. It allows to run applications by name. In addition, it supports a small number of commands, such as *ls*, *cd*, *cat*, and *netstat*; minimal versions of GNU/Linux’s commands with the same name.
- [*System calls interface.*]{} After configuring the CPU to run in user mode, the system calls interface serves as the only gateway to the system. Invocation of system calls is via software interrupts.
- [*Minilib.*]{} This small library module sits in between applications and system calls. Minilib’s purpose is to wrap up system calls and present user applications with a more intelligible interface. It also provides support for buffered IO operations in the presence of the IO manager.
Now we describe optional modules.
- [*Fault manager.*]{} It is a small module whose only task is to raise kernel panics on the occurrence of CPU faults (e.g. div-by-zero fault).
- [*Memory protection module.*]{} This module protects kernel code and data from code running in user mode. It restricts applications from accessing specific parts of memory, generating a segmentation fault if boundaries are violated.
- [*Thread synchronizer.*]{} Albeit part of minilib, thread synchronizer is a module on its own. It contains common implementations of thread synchronization mechanisms: lock, semaphore, monitor, and barrier synchronizations.
- [*Scheduler.*]{} The scheduler is a limited, but functional, priority-based pre-emptive thread scheduler. It supports a fixed number of threads with fixed stack sizes. While termination for a given thread is supported, freeing of its memory is not (mainly to avoid handling complex memory details). Threads can yield, can signal other threads, and can sleep. For portability, platform-dependent code for context switch is part a context-switcher, and not the scheduler itself.
- [*File system.*]{} A functional operating system must have a file system to start with. MiniOS uses FatFS [@fatfs] as file system. FatFS is a small FAT file system for resource-constrained devices.
- [*CLI.*]{} The command-line interface is a user-level shell whereby users can access kernel services.
- [*Network.*]{} As for networking capabilities, the network stack supports a very simple, inefficient, but functional network protocol over IEEE 802.15.4. Namely, it uses a flooding algorithm (Trickle [@trickle]) to form a network of ad-hoc connected devices. To avoid dependencies, the network stack purposefully overpasses the IO manager and handles its own buffers and radio interrupts.
- [*IO manager.*]{} The IO manager controls access of I/O devices. When interrupt-based devices notify the system of available data, it is responsible for a) placing the incoming data in an intermediate buffer accessible to both minilib and the IO event dispatcher; and b) notifying the scheduler of new incoming IO data.
- [*IO event dispatcher.*]{} The IO event dispatcher enhances the system with *IO events*. Whenever the scheduler is notified of new incoming IO data, the event dispatcher runs and executes the corresponding user-level event handler. Unlike other modules that can be implemented on top of base modules, the IO event dispatcher requires the scheduler and the IO manager to be part of the system.
For a more concrete idea, consider the following sample program.
#include "minilib/thread.h"
#include "minilib/display.h"
#include "minilib/network.h"
#include "minilib/sensor.h"
#include "minilib/led.h"
#include "minilib/ioevents.h"
void salute_thread( void* params ){
thread_set_priority( (uint32_t)params );
while( true ){
//print salute to USB
usb_write( "Hello, I'm thread %s \n", thread_get_current() );
thread_sleep( 200 );
}
}
IOEvent io_in_network( NetFrame* frame ){
//echo
net_write( frame );
}
int main(){
//Create salute threads
thread_create( salute_thread, "Mariana", 128, THREAD_PRTY_MIN );
thread_create( salute_thread, "Cafe", 128, THREAD_PRTY_MIN );
uint32_t state = 1;
while( true ){
//print sensor information to OLED display
display_write( "Light level (%%): %d \n", sensor_light_read() );
display_write( "Temperature (C): %d \n", sensor_temp_read() );
//blink LED0
led_write( Led0, state++ % 2 == 0 ? LedOn : LedOff );
thread_sleep( 500 );
}
}
This program is composed of four threads, one of which is main. Two of them print their name approximately five times a second over a CDC USB connection; one waits for an incoming network message and echoes it back to the same sender; and main prints sensor information on the OLED screen and blinks an LED approximately every half second.
An operating system works closely with a specific hardware. The following section discusses the target hardware platform.
The target hardware platform
----------------------------
Among all the different available ARM processor cores on the market, the Cortex-M series are those with the least sophistications that still offer support for operating systems. Among them, we have opted for the Cortex-M4, which was the most sophisticated in the Cortex-M series at the time MiniOS was initially conceived. Some of these OS-supporting features are software interrupts, memory protection, different CPU modes (kernel and user), separate user and kernel stacks. In fact, the only missing feature to fully support a conventional OS, capable of executing applications, is a memory management unit (MMU).
![MiniOS’ target hardware platform.[]{data-label="fig:platform"}](./platform.jpg)
Cortex-M cores are only available in MCUs, and because a MCU by itself is of no use, a MCU prototyping (evaluation) board must be used. Although it is possible to carry out labs with tailor-made hardware, an off-the-shelf board has its advantages. First, there are available device drivers from manufacturers. Second, these boards typically integrate an on-board chip debugger and programmer, thereby eliminating the need of an expensive JTAG emulator that does the same. Third, they can be purchased by anyone interested in taking or delivering the course. Lastly, being official boards, they integrate seamlessly with manufacturers’ development tools.
A variety of MCU prototyping boards exist in the market from different vendors. Based partly on its low cost, and partly in nothing in particular (as they all are quite similar), we have selected the *Atmel SAM4S Xplained Pro Starter Kit*. Its main board runs at 120 Mhz, and together with its three expansion boards integrate enough peripherals for laboratory projects. They include a small OLED screen, buttons, LEDs, a light sensor, a temperature sensor, a microSD card slot (and the microSD card), a USB device port, an on-board 256 MB Flash memory, and exposed pins for various on-chip peripherals. Figure \[fig:platform\] shows the main SAM4S board and its daughter boards, together with the *REB233 board* (acquired separately) for IEEE 802.15.4 connectivity. This is the hardware assumed by the construction guide.
Being an embedded platform, development of software is somewhat distinct. The following section attempts to offer more details in this regard.
### Development Environment
![MiniOS development environment[]{data-label="fig:devenvironment"}](./dev-env.png)
Clearly one cannot (easily) use the system’s target platform to develop the system itself. Instead a separate *host computer* is necessary for development of the system. In particular, using a cross-compiler, first the source code is compiled to an executable in the host. Then, the executable is flashed to the target’s program memory by a flashing tool. Finally, for debugging, an on-board hardware debugger interfaces with software in the host to enable source-level debugging. All of the different host-side software tools, including the GNU toolchain are integrated in Atmel’s IDE: Atmel Studio. Communication between the target platform and host tools is via USB. Figure \[fig:devenvironment\] depicts the described programming environment. Incidentally, Atmel Studio was built with Microsoft Visual Studio Shell. So, the programming environment is the same as that from Microsoft Visual Studio.
The Atmel debugging facilities, when used correctly, allow debugging of firmware running in the MCU as if it was a regular desktop application. It allows pausing (possibly at breakpoints) of the CPU for inspection and modification of memory, registers, IO interfaces, and source-level variables (including not primitive types). It also allows to step through both assembly and C code, as well as dis-assembled code.
The final piece in the development platform is the system’s guide to its construction, which is discussed in the following section.
The MiniOS Book
---------------
The idea of the MiniOS guide (or book) is to:
- Cover in-detail all the technical material that is necessary to build MiniOS.
- Guide students in the process of developing it themselves.
Consequently, the guide is intended to be self-contained, in the sense that a student could rely solely on it to build the system (characteristic not present in other similar OS books). The style of the guide was initially inspired by the tutorial *Write Yourself a Scheme in 48hrs* [@scheme], and later by the more textook-like style of the Xinu Book [@comer]. Our guide is divided in two parts, as shown in Table \[tab:book\].
The intention of the first part is to instruct on computer architecture using the ARM Cortex-M4 and the SAM4S board. The second part is dedicated entirely to the system, and it assumes some working knowledge of what is covered in the first section. Ideally, a student should complete the first section of the book, and then engage in building the system. However, if this is not the case (as we have experienced), working knowledge of a different computer architecture suffices. At worst, students will take extra time to learn certain Cortex-M4 technicalities. Importantly, all these required technicalities are available for consultation in the architecture section, and when used in the systems section, they are referenced.
An important aspect of the guide is the great amount of details offered. This is because it was written with the purpose to not leave students in the situation of discovering neither advanced topics nor topics pertaining to other subjects by themselves. Among others, it covers topics and information related to data structures, drivers, CPU, peripherals, the SAM4S board, the SAM4S MCU, the C language, assembly, the linker, and even programming patterns that are particular of systems or low-level programming. For instance, the guide explains and demonstrates the following: how to use callbacks to push data (coming from interrupts) from a lower layer to an upper layer; how to load a pre-compiled application from permanent storage to RAM for execution; how to write a linker script; how to do context switch; how to change CPU privileges; where in the documentation to find the mapping between physical pins and logical IO bits; and so forth. Some of this information is too technical or advanced to be left for discovery, and some does not pertain to OS instruction per se. Appendices \[sec:appendix-apps\] and \[sec:appendix-syscalls\] show excerpts from the Executing Applications and System Calls Chapters.
Importantly, most pieces of code that are given, are not just given, but derived, meaning, the book explains the steps in obtaining it from documentation or other assumed background knowledge. This gives interested and motivated students the tools to modify those parts, should they want to (e.g. for a final project).
Additionally, a device driver integration section was added to the guide. This sub-guide demonstrates the process of integrating third-party drivers, and shows working sample code. It can be challenging to write working code for an IO peripheral out of poor, and often buggy or incomplete, third-party documentation. Appendix \[sec:appendix-drivers\] shows one sample entry from the device drivers section.
In addition to text material we have prepared demonstrative videos. These are videos made to strengthen the text material, by showing explained concepts, techniques, processes, solutions to labs, or running sample driver code. For example, Figure \[fig:samplevideo\] shows a debugging session right before a system call. In this video, the control register is highlighted to demonstrate that, in fact, the CPU is in both user and unprivileged mode. Upon execution of the software interrupt, it is shown again, but this time specifying kernel privileged mode.
![Video demo: Entering kernel mode[]{data-label="fig:samplevideo"}](./sample-video1.png)
Lastly, we would like to emphasize that all this extra instruction goes in accordance with what is argued by Guzdial et al in [@guzdial] in favour of strong instructional guidance for novice learners. In particular, he mentions that there is strong evidence that the minimal guidance approach we typically use in computer science instruction is inadequate. One can argue that operating systems and computer architecture are typically second and third year courses, and therefore students are not novice programmers. While this is true, students are still considered novice learners from a low-level programming and OS development perspective.
Based on the described system, hardware platform, and guide thus far, the next section presents suggestions on how to accommodate the material in actual laboratory projects.
Laboratory Projects
-------------------
There is a total of twelve labs, with different suggested durations. The first lab is an optional short introduction. The next two labs are also short, and, since they involve base modules, they cannot be skipped nor their order can be altered. The remaining eight are optional, and most of them can be implemented regardless of order. In case a module is considered to be good to have, but not of interest as to dedicate a lab to it, there is the possibility to hand it in to students. For instance, the fault manager can be a useful module to have as it outputs human-readable messages when the CPU faults, and it could be given to students.
### Lab 1 - Basic IO and Booting
In this lab students are introduced to the booting process, the use of third-party firmware as basic input-output, and the programming environment (including debugging facilities). The recommended time for solving this lab is one week and is optional, albeit recommended. Another way of looking at this lab is that it enhances bare-metal applications with bare-metal firmware, as depicted in Figure \[fig:lab1\].
![Architecture Goal for Intro Lab[]{data-label="fig:lab1"}](./intro-arch.png)
The learning outcomes for this lab are to familiarize students with the development environment; to provided some guidelines on how to make efficient use of the debugging facilities; and to show the process of integrating third-party firmware to be used as basic IO.
Lab 2 - Hardware Abstraction Layer
----------------------------------
For this lab students write the HAL, and the system module. Some of the implementations expected from this lab are, for example, an IO device abstraction composed of a read function and a write function, an abstraction for registering callbacks of interrupt-based IO, among others. The recommended time for solving this lab is one week, and it is mandatory. Figure \[fig:lab2\] shows the result of completing the HAL.
![Architecture Goal for HAL Lab[]{data-label="fig:lab2"}](./hal-arch.png)
The objective of this lab is to give some insight and hands-on experience on interaction with bare-metal IO peripherals, and in the process convey students the importance that a) abstraction plays in development of the system, and b) the repercussions that a HAL has in portability.
Lab 3 - System Calls
--------------------
Provided hardware-specific information on how to establish a kernel and user mode separation, students must add code to support software-interrupt-based system calls via minilib. The separation is made even clearer by splitting compilation of OS and app. MiniOS is compiled and flashed to the MCU, while applications are compiled and moved to an SD Card from where they are loaded into RAM and executed. Since loader code is given, students are asked to write a rudimentary version of the console that supports listing of files and execution of applications. The console is a serial terminal, thus can be accessed via any serial terminal (e.g. PuTTY). Special emphasis is placed on conveying the importance of CPU mode separation in security. In fact the chapter begins by demonstrating an app that changes the user name displayed int he console by directly manipulating a kernel data structure.
Optional tasks involve writing of various different system calls (e.g. execv for child process creation), or taks related to buffered output (e.g. implementation of a line-buffered display\_write function). In the process, the inability of user code to directly access data from interrupt-based input is emphasized; although nothing is done about it until later labs. The recommended time for solving this lab is one to two weeks.
![Architecture Goal for System Calls Lab[]{data-label="fig:lab3"}](./syscalls-arch.png)
Upon completion of this lab students are expected to have some insight and working knowledge of:
- The need and implementation of kernel-application separation, as well as the mechanisms used by operating systems to interface them both.
- The role and place of libraries such as the GNU C Library in an operating system.
- The limitations of poll-based IO.
- Buffered IO.
The resulting architecture is in Figure \[fig:lab3\].
### Optional Lab – Fault Manager
The fault manager is another short lab. Here students are given guidance on CPU faults, and are asked to add support for fatal system errors—the mini black screen of death. If memory protection is part of the MiniOS version for this lab, a more complex task involves termination of the offending application and continue of execution. The recommended time for solving this lab is one week.
For this lab, students are expected to gain insight on error detection, and have some experience in the process of handling and reporting them.
### Optional Lab – Memory Protection
For this lab students must implement memory protection to prevent user code from accessing system code and data in memory. If the fault manager has been implemented, an extra task of enabling segmentation faults is available. The recommended time for solving this lab is one week.
The idea of this lab is to supply students with insight and working knowledge on the use of memory protection to prevent bugs and malicious code to mess with the system, as well as let them experience first hand what a segfault is.
### Optional Lab – Scheduler
The scheduler is perhaps the most technically challenging lab. Starting from the system timer, a single-threaded scheduler and a yield function are derived and demonstrated. Available tasks include extending it to support multiple threads, priorities, round-robin scheduling, a sleep function, thread signalling, and different scheduling policies, among other tasks. The recommended time for solving this lab is two weeks, or three if extra tasks are added.
The learning outcomes for this lab are to provide students with experience of the obscure inner workings of a thread scheduler, to let them experience first hand how sharing CPU is made possible by a set of small clever tricks done by the operating system; also, to get some working knowledge on a) implementation of different scheduling policies; b) how different threads queues can be used to support sleeping threads, priorities, and thread signalling, among other tasks.
### Optional Lab – IO events
In this lab students write the IO manager and the IO event dispatcher to add support for user-level run-to-completion IO events. Every time new data is received from interrupt-based IO, the IO manager stores it in intermediate kernel buffers and notifies the scheduler to wake up and run the event dispatcher thread. This is a short lab, and its recommended time is one week.
On completion of this lab, students are expected to have an understanding of a) the implementation of events from threads; b) the benefits of a hybrid thread-event approach to concurrent programming.
### Optional Lab – Thread synchronization
Here students implement thread synchronization mechanisms: locks, semaphores, monitors, and barrier synchronizations. The recommended time for solving this lab is one week or two depending on the number of mechanisms implemented and the amount of help offered.
At the end of this lab students are expected to understand, from an implementation perspective, synchronization mechanisms used in multi-threaded programming.
### Optional Lab – Network Stack
A MAC layer interface is delivered as part of this lab’s material. Students must, then, use the trickle algorithm [@trickle] to enhance the system with network capabilities. A more complicated task includes writing a host application that transmits an executable, having a node receive it and execute it. Since trickle is straightforward to implement, the recommended time for solving this lab is one week, or two with more complicated tasks.
The objective of this lab is to demonstrate, in a rudimentary manner, how computer networks are built out of layer-2 point-to-point communication; also, to show the difficulties of a) providing applications with networking services, and b) dealing with unreliable wireless communications.
### Optional Lab – Console and CLI
In this lab students write either a more advanced console (than what they have in Lab 3) or a CLI, or both. The console is internal to MiniOS, it is launched on system start up, and enables:
- to print information during system initialization;
- user login;
- execution of basic commands; and
- to browse and execute applications stored in the SD Card.
System initialization messages include CPU speed and peripherals found. More advanced features involve basic managing of user accounts, and enabling applications to exit and give control back to the console. The CLI, on the other hand, is an application that runs in user mode and allows similar functionality. More advanced tasks include the implementation of privileges for user accounts; replacement of PuTTY for a custom made terminal, so as to give the terminal a more traditional feeling; or a host panel board that shows sensor information.
This is a short lab and the recommended time for solving this lab is one week. The purpose of this lab is to demonstrate how command-line interpreters, whether they run in kernel or user mode, fit in with the rest of the system.
### Final Project
As final projects, students may form teams and create something of their own. Any idea involving an embedded OS, or extension of the OS itself are eligible choices. Unlike all the remaining labs, this project has no rigid specification. It is open ended and students are encouraged to implement something of their interest. A complete version of the system can be handed in to those teams who need it. In the end, exact specifications differ depending on instructors’ preference. The idea is for students to put all acquired knowledge to practice and hopefully deepen their knowledge in some specific OS aspect of their choice.
This concludes the presentation of the instructional platform.
Evaluation
==========
The idea of MiniOS was not to replace other instructional systems, but to create an alternative system that could adhere itself to the already existing set. More precisely, it was meant to be a small, complete, and functional MCU-based system that could be used for teaching operating systems concepts, while lessening students’ struggle.
Due to our policy of minimal implementations and minimal hardware, we were able to write the entire system in less than 3,500 lines of C code and less than 250 lines of assembly. Following the principle of not placing students in the position of discovering new information, we have developed a self-contained book covering the construction of the system. All this together has enabled MiniOS to be:
- functional and complete
- small
- built-from the ground up, and
- simple enough to reduce students’ struggle in building an instructional system.
For further evaluation, next we report our observations in using MiniOS to deliver operating systems labs at our university. In addition, we present student’s feedback, sample final laboratory projects (showing their quality), and student’s laboratory grades. Finally, we elaborate on the use of MiniOS as a prototyping and research platform.
Observations
------------
Reflecting on the experience to date, MiniOS has served well as an instructional system. Previous versions of the teaching platform (Figure \[fig:minios-years\]) were used as the laboratory component for Operating Systems in Fall 2013, 2014 and 2015, and as the laboratory component for Computer Architecture in Winter 2015 and 2016.
![MiniOS architecture in different years[]{data-label="fig:minios-years"}](./minios-2013.png)
![MiniOS architecture in different years[]{data-label="fig:minios-years"}](./minios-2014.png)
![MiniOS architecture in different years[]{data-label="fig:minios-years"}](./minios-2015.png)
Figures \[fig:minios-years\](a), \[fig:minios-years\](b), and \[fig:minios-years\](c) illustrate the evolution of MiniOS. Overall the delivery of the material went without problems, and in the process we gathered experiences.
### Book experience
The role of the book seemed to have served its purpose. In particular, we think it seems to have allowed students to learn more complex OS topics (that would otherwise be out of scope). In other words, it seems that giving them all the basic details of OS construction allowed them to focus on more advanced features.
The amount of details offered in its latest version allowed for students to complete assignments with little necessity to rely on other sources. In early offerings where the guide covered less material, students consistently indicated being frustrated for the lack of related external sources. Moreover, the amount of details seemed to have enabled students to complete projects. In Fall 2013 two of the five teams using the SAM4S board (two teams used different hardware) were not able to present working projects due problems of technical nature. In 2014, the number went down to one (that team used different hardware). In 2015, it went down to zero, and all teams used the SAM4S board.
The system part of the guide assumes that students have had some experience working with the SAM4S board and Cortex-M architecture. When we delivered Operating Systems in 2015 this was not the case for every student, as some had taken Computer Architecture one year earlier (in 2014) using a different CPU architecture. Interestingly, the lack of previous Cortex-M experience did not seem to matter considerably. Three of the thirteen students attending labs did not have previous Cortex-M experience. Still, based on marks and personal appreciation, they performed similar to the rest of the class. In fact, one of them went to obtain the highest marks in labs. It is quite possible and reasonable that they dedicated more time to get on track with the new hardware platform and programming environment.
### Instruction and tutorial experience
Based on previous experience teaching labs by building an OS for x86, MCU embedded hardware seemed to have allowed us to dedicate less instruction to present students with hardware details. Specially with the use of the guide, the required concepts previously introduced in Computer Architecture were just referenced and not re-introduced.
Tutorials were offered in a classroom once a week. Despite the material being covered in the guide, many times students needed further clarification. While some were able to finish labs with minimal or no consultation at all, others did not. Thus, it is recommended to have dedicated lab or tutorial sessions, where the lab instructor gives an oral presentation of the material. To gain insight into what is difficult and what is not, and what could use extra instruction, it is advised that the lab instructor solves the labs in advance.
### Language experience
Java is the language used to teach most courses at UNBC. This means that, for many students, Operating Systems was their first encounter with the C language. Among all the C-specific concepts, pointers and pointers to pointers demonstrated to be difficult to decipher. In fact, students consistently reported much of the struggle with assignments came from inexperience with C. To compensate for it, tutorials covered the use of pointers, callbacks, structures, organization of code in modules, use of header files, compiler attributes, among other relevant C specifics. Often students were not able to proceed further due to a specific language detail they were confused about or not knowledgeable of. While some students were prompt to ask, other were not. Those that did not ask reported spending a considerable amount of extra hours figuring out by themselves. Thus, students were encouraged to ask for language related doubts. In general, it is recommended that the lab instructor does not hesitate in assisting students with language problems.
### Debugger experience
The presence of debugging facilities, albeit circumstantial, showed to be very important for solving the laboratory assignments. On occasions, the debugger was the difference between students completing an assignment or not. Often assistance was given in the form of debugging sessions. Sometimes because the lab instructor was unsure where the mistake was, and some other times because the debugger allowed for a demonstration of a concept that was otherwise not being understood from an oral explanation. Also, we have found that most of students’ bugs are due not to wrong logic (they usually get it right), but due to a missing technical detail or a wrong assumption of technical nature. Debugging was very useful in finding those mistakes, as it enabled to verify step by step the details and assumptions of what is supposed to happen versus what is actually happening. This contrasts with typical remote debugging, which is a rather limited way of debugging (similar views are expressed by Holland et al [@OS161]).
We also noted that in spite of previous debugging experience, in most cases, students lacked the debugging abilities to take advantage of the facilities available. In this regard, videos showing effective use of the debugger, as well as personal assistance, were provided. Interestingly, students seemed to have improved their bug-finding skills after only a couple short sessions of personal debugging assistance.
### Hardware experience
In the first offering of the course, the MCU platform had a neutral reception. This, however, has changed for the later two offerings of the course. Perhaps the guide played a role in that. For the latest offerings of both more than half of students showed interest and enthusiasm of working with hardware.
Overall, boards behave well. On occasions, albeit not often, boards would simply fail to be recognized by Atmel Studio. Some times resetting the host computer or re-plugging the board would fix the problem, but other times the board would continue to fail and a replacement had to be given. So it is recommended to have extra boards in case they are needed.
Working with external peripherals can sometimes be a problem. There was a few incidents where Xbee modules and one board where burnt due to wrong wiring. Being computer science majors, a good number of students showed problems with wiring of external peripherals. For example, late in the course one student started having problems integrating an Xbee for his final project, and expressed that the OS course was (before the issues started) his favourite course in that semester. These problems repeated in a few occasions, and as a consequence, the latest version of MiniOS has stopped requiring any use of peripherals that are not expansion boards, since they simply plug into an expansion slot. Concretely, Xbee radio modules have been replaced by the REB233 board. At the time adding external support for PS/2 keyboards was also being considered, but had to be dismissed for the same reasons.
For final projects, students choose something of their interest, and often the wiring problem is impossible to avoid since they choose external hardware that requires wiring. Not only wiring, but finding the right piece of hardware have been consistently mentioned (in project reports) to be challenging. Our solution was to provide extra assistance. In some cases the assistance included videos showing how to wire the peripherals, and in other occasions sitting down with them and explaining the wiring in detail.
### Drivers experience
An important aspect of MiniOS is the integration of third-party open source firmware. This enabled seamless integration of a variety of different peripherals. With available drivers, the adding of hardware functionality to the system became a mechanical task. We found this to be good for student engagement, as driver availability is the main limitation in using external hardware in projects. In fact, during project proposals we advised students to check for driver availability before acquiring any peripheral. It is worth noting that by themselves, drivers are of little use as their use is difficult to figure out from documentation. The device driver guide played an important role in simplifying it, and turning it into a mechanical task.
Students’ feedback
------------------
In Fall 2014, students were asked to elaborate in what they thought were strengths and weaknesses of laboratories projects. These are some of the answers, organized by student number S1, S2, and so forth.
- S1 strengths: *The assignments are very hands on and we get to see the things we discuss in class. The example programs show the functionality of the board.*
- S1 weaknesses: *there seems to be little documentation for the ASF. The coding can be hard to follow.*
- S2: *Assignments are a good way to see the complexity and challenge in dealing with the hardware level. They are nice in that you can access the hardware directly, and use the debugger that is provided to actually “see” the registers and the hex or binary values stored here, and how things are interacting.*
*However, that is also its main weakness. being that they are reasonable complex pieces of software it is very challenging to understand how all the components are interacting at times. \[...\] Along with the large amount of digging that needs to be done to understand the documentation, the other challenge is understanding C as I have not used it much to although things are similar it is still not Java. however it is kinda nice \[illegible\] to use something other than Java.*
- S3 strengths: *in my opinion, the programming assignments are a mixed bag. I think its a good way to show incrementally how each part of an OS is designed and implemented in practice. The assignment themselves strengthen the knowledge learned in class.*
- S3 ‘weaknesses: *the downside, however, is the language implemented. Its a minor point, but it is an issue with which I struggle. Until this course, I’ve never used/been/seen exposed to C. It just makes understanding and implementing ideas needlessly complex.*
- S4: *\[...\]It is too easy to get stuck on a simple task specially when the student is using a new language and development environment that is foreign.*
*Being a Java university, the first assignment should a strictly C assignment. Designed to get an understanding of the differences from Java, and features required to use the Atmel libraries. This can be assigned day 1 of the term.*
- S5 strengths: *Got to see and develop an entire OS. Get to use a simple enough platform to feasibly develop all components. The interactive nature of the platform makes progress easy to see and rewarding. Tutorials are well written and provide detailed instructions. Code base is quite small and it is easy to hold entire system in your head.*
- S5 weaknesses: *Lack of documentation and online support for platform. Each assignment is very involved and requires a large time commitment.*
- S6 strengths: *The projects are fun, engaging and rewarding. Interacting with real hardware is great. The resources provided by the TA are complete and helpful*
- S6 strengths: *C is not something I am particularly familiar with. Atmel resources (documentation) are not always helpful. C language guides are not always applicable*
For fall 2015 substantial changes were made to the guide. In particular, it was made more self-contained, in the sense that it made little or no reference to external documentation. Additional missed necessary details dealing with the architecture were covered. Tutorial time was devoted for looking at specific C knowledge required to complete assignments. Certain embedded systems specific details not relevant to operating systems were removed. Students were again asked to elaborate in what they thought were strengths and weaknesses of laboratories projects. These are some of the answers (here we group them).
- Strengths:
- *Physical depiction of what we’re doing (interactive buttons & oled screen)*
- *Overall great assignment layout. I don’t understand why people need extensions!*
- *Everyone loves bonus questions!*
- *Nice to have many examples/viewpoints*
- *I’ve heard from many that they’re having trouble with their board. Not me personally though.*
- *Great examples to draw from.*
- *Much prior use of board/software*
- *enforces understanding of:*
- *Interrupt*
- *HAL*
- *better code*
- *better structure*
- *FUN!*
- *The documentation provided is top notch. It makes the world of a difference having lab documents and driver documents written in PLAIN LANGUAGE which speeds up learning.*
- *Unified system. The SAM4S is easy to work with, we have been using it for a few years, so students aren’t totally new to developing for it.*
- *Software support. It’s undeniable Atmel Studio is useful in learning how to code for embedded systems. Visually stepping through code and viewing memory being modified helps intuition. As well as debugger.*
- *Relevant work, simplistic design. it’s easy to develop on ARM and learn the ropes. Jumping to x86 would be more difficult. ARM is also very popular and won’t be going away soon.*
- *Assignments are relevant to course. It’s easier to break down OS concepts and learn how to code them \[illegible\]. it solidifies the ideas and makes classes easier to understand.*
- Weaknesses:
- *lack of any useful documentation about the SAM4S online*
- *C is a bit tough when no taught any prior C (pointers, memory is odd) (still not so hard)*
- *Atmel Studio 6.0, 6.2 a bit finicky and error prone (better with 7.0 now!)*
- *\*\*For me\* Many others would disagree:*
- *not so hard enough sometimes*
- *would like to build some driver from scratch (camera, touchscree, etc)*
- *Assignment are long. A lot of time is required to complete. Due to bugs it can sometimes take more than a week. Two weeks is usually required.*
- *Atmel Studio is buggy, it leads to a lot of wasted time messing around.*
- *It can sometimes be difficult to tie into classroom lectures. It would almost help to focus lecture on how ARM systems can have an OS build on them to have more insight.*
Sample Final Projects
---------------------
At the end of the semester we held, for both courses, a final project presentation, where students had to give a short presentation and demonstration of their projects. Figure \[fig:student-projects\] shows two sample projects. Figure \[fig:student-projects\](a) shows a clock alarm project that runs MiniOS. Figure \[fig:student-projects\](b) shows a project named *Pinto pipes*, a rudimentary command line interpreter that supports redirection. Other projects include *Ascii at a distance* (a communication API for wireless devices), *SOS (simple operating system)*, *Remote sensor drone* (a remote rover with sensing capabilities), thread signalling for MiniOS, and *System Security* (secure user account management).
![Sample student projects[]{data-label="fig:student-projects"}](./situ-project.png)
![Sample student projects[]{data-label="fig:student-projects"}](./adam-project.png)
Laboratory and Project Grades
-----------------------------
Laboratory grading had two components. A programming component that accounted for around 80% of the whole laboratory assignment grade, and a writing component accounting for 20%. So, to get full marks students required not only working code, but also an understanding of the topic discussed in each lab. Non-working code was highly penalized with two thirds of the complete value for a given task (lab assignments were divided into tasks). Projects’ grades were dependent on the difficulty and quality of the project, the amount of help asked for, and how it related to OS concepts (e.g. extending the system is preferred over using it as a tool).
Laboratory and project grades are available for 2013 and 2015 (in 2014 a different lab instructor taught labs) with a total of 17 and 13 attendants, respectively. Figure \[fig:grades\]a shows distribution of laboratory grades, while Figure \[fig:grades\]b shows distribution of project grades.
![Student Grades[]{data-label="fig:grades"}](./marks_assignments.png)
![Student Grades[]{data-label="fig:grades"}](./marks_projects.png)
The results show an adequate level of proficiency from students in both years. Incidentally, in 2015 marking was purposefully made more strict than in 2013 for both assignments and projects. That decision reflects in lab assignment’s media dropping from 87% in 2013 to 77% in 2015. Interestingly, in 2015 project marks were higher, in spite of the harsher marking. The media went from 77% in 2013 to 85% in 2015. Moreover, 2015 was the only year when all students used the SAM4S board in their projects. Improved project marks may be due to (a) better student engagement in 2015 than in 2013 (we noticed students seemed more dedicated to their projects in 2015); and (b) a more developed guide that translated in better understanding of OS concepts and their implementation, which in turn resulted in projects of better quality, more complex, and more related to OS concepts.
MiniOS as a prototyping research platform
-----------------------------------------
The amount of functionality built in, together with the ease of hardware integration, made MiniOS a good alternative for embedded systems prototyping. In particular, applications have access to OS facilities, as well as straightforward sensor and actuator integration. Prototyping platforms with similar capabilities are Microsoft’s .NET gadgeteer [@NetGadgeteer] and mbed [@mbed]. As example, consider Figure \[fig:ruperts\], which shows two mobile robots part of an experimental multi-robot platform based on MiniOS.
![Multi-robot platform running MiniOS[]{data-label="fig:ruperts"}](./ruperts.png)
Likewise, given that MiniOS’ inner working are well documented and are comparatively simpler that other systems, it could serve well as systems research platform where system designs can be tried in relatively small time frame.
Up to this point, we have discussed our appreciation of the teaching experience. The next section discusses feedback received from students.
Concluding Remarks
==================
Conclusions
-----------
The main contribution of this thesis is MiniOS, an instructional platform for the delivery of operating systems laboratories. MiniOS follows on the steps of instructional systems that attempt to deal with complexity by lowering the code volume. We go further and identify the target hardware platform as an additional source of complexity that can also be account for. The result is a MCU instructional operating system, and to our knowledge, the first one of its type. In addition, the platform includes a step-by-step guide to its construction whose purpose is to offer all the necessary details for the construction of the system.
The platform was used in three different occasions to deliver laboratory assignments for an introductory course in operating systems. Student feedback was overall favourable. We presented this feedback together with other experiences.
MiniOS cannot—and is not intended to—replace more traditional desktop-centered design approaches; instead it serves as an alternative. For example, if a course has as objective to provide students with experience in topics related to the MMU (e.g. virtual memory), or to teach OSes as they are built in industry, the MiniOS approach is a poor fit.
University laboratories are not the only place where MiniOS has found use. Given the amount of built-in functionality and the ease with which hardware can be integrated, it can, and has, been used as a rapid embedded prototyping platform. Likewise, being well-documented, it serves as a systems research test bed. Currently the only existing ports is the SAM4S Xplained Pro board, but there is no reason that would not allow MiniOS to be ported to other MCU platforms. In fact, it was designed to be ported.
Future Work
-----------
For future work, we would like to explore the possibility of porting the instructional platform to other MCU platforms like Arduino Zero, a Cortex-M0 based Arduino. The integration of bare-metal drivers as part of MiniOS might work well with the plethora of available Arduino code online. Also, since there is existing infrastructure for wireless connectivity, it is possible to add internet connectivity and customize MiniOS to work as an IoT OS.
Finally, there is space for further experimental exploration of some of the observations made during the use of the platform in the past three years. For instance, the authors would like to test whether a smaller MCU platform does enable learning of the same concepts while requiring less instruction. Also, the impact that using the debugger has in explaining of technical concepts; among other observations made in this work.
Excerpt from Executing Applications Chapter {#sec:appendix-apps}
===========================================
Excerpt from System Calls Chapter {#sec:appendix-syscalls}
=================================
Excerpt from Device Driver’s Section {#sec:appendix-drivers}
====================================
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Blue compact dwarf galaxies (BCDs) are faint (M$_B\le$-17 mag) compact (R$<$1 kpc), at least qualitatively very blue galaxies due to active star formation, and have low metallicities. Found serendipitously as part of a redshift survey of faint galaxies with the Keck Telescope (DEEP), SA 68-6597 is at a distance of 80 Mpc, and is one of the faintest, -12.4 mag, lowest metallicity, $\sim$0.05 Z$_\odot$, BCDs known. Its H$\beta$ linewidth of $\sigma =$27 [kms$^{-1}$]{} and small size, R$_{eff}\sim$190 pc, suggest that it is an extremely low mass galaxy. We have used the Arecibo telescope to measure the properties of SA 68-6597 in order to better constrain its total mass and its potential for future star formation. SA 68-6597 has a [$M_{HI}$]{}=(1.4$\pm$0.4)$\times$10$^7$[$M_\odot$]{} and an FWHM linewidth of 33$\pm^{60}_{12}$. Combining the linewidth with an estimate of the size of the disk, we derive a [$M_{dyn}$]{}$\gtrsim$3$\times$10$^7$[$M_\odot$]{}. The [$M_{HI}$]{}/[$L_B$]{}=1.0$\pm$0.3[$M_\odot$]{}/[$L_\odot$]{}, [$M_{dyn}$]{}/[$L_B$]{}$\ge$2[$M_\odot$]{}/[$L_\odot$]{} and [$M_{HI}$]{}/[$M_{dyn}$]{}$\lesssim$0.47 values are typical for BCDs. Combining the measured star formation rate of 0.003 [$M_\odot$]{}/yr with the mass, we derive a gas depletion timescale of 5$\pm$2 Gyr. While SA 68-6597 is a fainter, lower-mass, higher metallicity counterpart to other BCDs like I Zw 18 and SBS 0335-052, its properties suggest it will not evolve dramatically in the near future. Given the limits on its gaseous and dynamical masses, SA 68-6597 may be able to evolve into a moderately massive dwarf spheroidal galaxy.'
author:
- 'D.J. Pisano, David C. Koo, Christopher N.A. Willmer, Kai Gerhard Noeske,and A.C. Phillips'
title: ' Observations of SA 68-6597: the faintest Blue Compact Dwarf Galaxy.'
---
Introduction {#intro}
============
Blue Compact Dwarfs (BCDs) are faint [M$_B\le$-17 mag, e.g. @kon02], compact [diameters of the high surface brightness regions of less than 1 kpc, e.g. @thu81], and blue enough to suggest active star formation [e.g. @gor81; @thu81]. They are typically low metallicity systems [@izo99]. Two of the most extreme BCDs, in terms of luminosity, mass, and metallicity, are I Zw 18 and SBS 0335-052. I Zw 18 and SBS 0335-052 have luminosities of -12.8 mag and -14.3 mag, total masses of 10$^{8.5-9.5}$[$M_\odot$]{}, HI masses of 10$^{7.8-9.3}$[$M_\odot$]{} [@vz98a; @pus01], and oxygen abundances of the ionized gas of $12+log(O/H) =$ 7.17 & 7.34 [@izo99]–the lowest known in the universe. Because of these properties, I Zw 18 and SBS 0335-052 are believed to be undergoing early bursts of star formation [@izo04; @lip99]. While much more luminous, BCD-like, galaxies at moderate redshift may evolve into galaxies like NGC 205 [@koo94; @koo95; @guz96], these low luminosity, low mass BCDs may be the progenitors of dwarf spheroidal galaxies like Carina.
SA 68-6597 was discovered serendipitously during the first DEEP[^1] run using the Keck LRIS instrument[@oke95] with a 1200 l/mm grating (Koo et al. 2005, in preparation). This galaxy was selected because of its blue color and visually estimated compact non-stellar appearance and faint apparent magnitude. The LRIS data show that SA 68-6597 has a redshift of $z$=0.0186 implying that it is intrinsically extremely faint. Assuming a Hubble constant of 70 [kms$^{-1}$]{} Mpc$^{-1}$, SA 68-6597 is located at a distance of 80 Mpc and has a B magnitude of -12.4 mag as measured by the DEEP.team. Using the combination of the $[\lambda 5007]$/H$\beta$ and the $[\lambda
6583]$/H$\alpha$ line ratios as measured with HIRES, Koo et al. find that SA 68-6597 has an extremely low excitation temperature, $\sim$14,500 K, and metallicity $12+log(O/H) \sim$7.4 ($\sim$0.05 Z$_\odot$). This places SA 68-6597 well away from the well-defined locus of galaxies [e.g. @lee04] and Local Group dwarf irregulars [@mat98] in this parameter space. The most similar galaxy to SA 68-6597 in this space is the BCD SBS 0335-052. Follow-up observations (Koo et al. 2005, in preparation) with Keck HIRES [@vog94] also show that the emission lines have Gaussian velocity dispersions of $\sim$ 27 km/s. HST WFPC2 images reveal a very small galaxy, R$_{25}$ of 1.0$\pm^{0.1}_{0.05}$= 400$\pm^{40}_{20}$ pc. Combined with the small optical linewidth, this implies that SA 68-6597 is a very low mass galaxy, $\sim$10$^7$[$M_\odot$]{}. Based on the flux of the H$\alpha$ line in the LRIS spectra, the star formation rate of this galaxy, 0.003 [$M_\odot$]{} yr$^{-1}$, is similar to what is expected for a BCD given its inferred low total mass [@hop02]. These properties from Koo et al. (2005, in preparation) and summarized in Table \[props\], strongly suggest that SA 68-6597 is a blue compact dwarf that is fainter and smaller than other BCDs [e.g. @thu81; @sal02]. It is a fainter, lower-mass, but slightly higher metallicity counterpart to the more famous BCDs I Zw 18 and SBS 0335-052. It is the faintest known BCD (Koo et al. 2005, in preparation). Its extreme nature makes this galaxy a particularly interesting probe of low mass galaxy formation [see @pus01 and references therein for a discussion].
Observations of the 21-cm line of neutral hydrogen () can help us better understand the nature of SA 68-6597. The content of SA 68-6597 is an important measure of its potential for future star formation. The current burst of star formation is small in an absolute sense, 0.003 [$M_\odot$]{} yr$^{-1}$, but if the mass is also low then it can still rapidly consume its and subsequently passively evolve, fade and possibly become a galaxy like the Carina dwarf spheroidal in the Local Group. If the mass is much higher, then it is more likely to continue forming stars for a long time and will retain its current appearance. Furthermore, while the optical emission lines have a width of 27 [kms$^{-1}$]{}, this linewidth may not trace the entire gravitational potential of the galaxy. The gas tends to trace the gravitational potential to larger radii than the stars or ionized gas. In addition, ionized gas may be tracing galactic outflows; this is less likely for the neutral gas. For all of these reasons, provides the best measure of the total, dynamical mass of a galaxy. The dynamical mass is an important constraint on the evolution of a galaxy. If SA 68-6597 has a low total mass, then it may eject its neutral gas before it can consume it in star formation [e.g. @mac99]. If it is higher, then it may be too massive to evolve into a dwarf spheroidal galaxy. In this paper, we report on our Arecibo observations of in SA 68-6597. These observations help constrain the potential for future star formation in the galaxy and provide a more robust measure of the total mass of the galaxy constraining the current nature and future evolution of SA 68-6597, and its relation to other BCDs like I Zw 18.
Arecibo Observations & Reductions {#obs}
==================================
We observed SA 68-6597 with the Arecibo[^2] 305 m telescope on 2004 August 1–4 and October 3–4. We observed only at night to minimize solar interference. We used the L-wide receiver for all our observations. This receiver has a system temperature of $\sim$27 K, and a gain of $\sim$10 K Jy$^{-1}$ as measured by the Arecibo staff. Both values are weakly dependent on the zenith angle of the observation. Data were processed through the interim correlator in both linear polarizations over total bandwidths of 25 MHz and 12.5 MHz, corresponding to a velocity range of $\sim$5000 [kms$^{-1}$]{} and $\sim$2500 [kms$^{-1}$]{}. Each bandwidth and polarization had 9-level sampling and 2048 channels, resulting in a velocity resolution per channel of 5.2 [kms$^{-1}$]{} and 2.6 [kms$^{-1}$]{}. Our observations utilized a high pass filter to block interference below 1370 MHz contaminating our band. The beam size of the L-wide receiver according to the Arecibo documentation is 3.1$\arcmin \times$ 3.5$\arcmin$. At the distance of SA 68-6597, 80 Mpc, this corresponds to a linear size of 72 kpc$\times$ 81 kpc. As the effective radius of the stellar emission is only 190 pc, this should be more than sufficient to encompass all of the associated with this galaxy. Yet this beam size is small enough that there are no known galaxies at a similar redshift that can contaminate our measurements; the closest galaxy is $\gtrsim$400 kpc away and there are no known groups within 3 Mpc based on a NED[^3] search. It is further unlikely that there are any -rich, optically invisible galaxies exist that could contaminate our measurements [@doy05].
We used the standard position switching algorithm for our observations spending 5 minutes on SA 68-6597 followed by 5 minutes offset to blank sky by 5 minutes in right ascension such that we tracked the same azimuth and zenith angle as the on-source scan. This was repeated for all six nights for an on-source integration time of 240 minutes. Our data were reduced using standard Arecibo IDL routines written by Phil Perrilat. Each bandwidth of each scan was calibrated separately and the polarizations were averaged together before a first or second order baseline was fit across the portion of the spectrum clean of any interference. All scans were then averaged together to produce the final spectrum. While some scans showed signatures of interference around 1400 MHz, since SA 68-6597 is at a frequency of 1394 MHz this should not affect our ability to detect the galaxy. The bandpass, however, was significantly better for the 12.5 MHz band as a result of the RFI at 1400 MHz, and so we proceeded only with this band. Because the two bands are split after the first amplification, their noise is not independent and, therefore, it would not have helped to combine these bands. The resulting noise was 0.22 mJy per 2.6 [kms$^{-1}$]{}channel. For analysis, we binned this spectrum by 4 channels to a resolution of 10.8 [kms$^{-1}$]{}, improving the noise to 0.10 mJy per channel. The observational details are listed in Table \[props\]. Our flux measurements of a bright galaxy, UGC 199, are within 17% of previously published values [@sch90]. This is a much smaller source of error than that of random noise and can be disregarded for this work.
[@lrl]{} Optical Properties: & &\
Right Ascension (J2000)& 00:17:15.41 & h:m:s\
Declination (J2000) & 15:48:38.36 & $^\circ : \arcmin\ : \arcsec\ $\
Optical Heliocentric Velocity & 5595$\pm$40 & [kms$^{-1}$]{}\
Distance & 80 & Mpc\
Optical Velocity Dispersion & 27$\pm^1_2$ & [kms$^{-1}$]{}\
Absolute B Magnitude & -12.4 & mag\
Luminosity & 1.4 & 10$^7$[$L_\odot$]{}\
R$_{25}$ & 400$\pm^{40}_{20}$ & pc\
Star Formation Rate & 0.003 & [$M_\odot$]{}/yr\
Observational Properties: & &\
Time on Source & 240 & min.\
Channel Width (binned) & 10.8 & [kms$^{-1}$]{}\
RMS noise per channel & 0.10 & mJy\
Properties: & &\
Peak Flux & 0.31 & mJy\
Heliocentric Recession Velocity & 5557$\pm$5 & [kms$^{-1}$]{}\
Velocity Width (20%) & 51$\pm^{93}_{19}$ & [kms$^{-1}$]{}\
Integrated Flux & 9.5$\pm$2.5 & mJy [kms$^{-1}$]{}\
Mass & 1.4$\pm$0.4 & 10$^7$M$_\odot$\
[$M_{HI}$]{}/[$L_B$]{}& 1.0$\pm$0.3 & [$M_\odot$]{}/[$L_\odot$]{}\
Gas Depletion Time & 5$\pm$2 & Gyr\
[$M_{dyn}$]{}& $\ge$3.0 & 10$^7$[$M_\odot$]{}\
[$M_{dyn}$]{}/[$L_B$]{}& $\ge$2 & [$M_\odot$]{}/[$L_\odot$]{}\
[$M_{HI}$]{}/[$M_{dyn}$]{}& $\lesssim$0.47 &\
Results
=======
Figure \[hispec\] shows the binned spectrum of SA 68-6597 near the known optical velocity of the galaxy. The small inset in the lower left corner shows the entire spectrum (excluding the edges of the bandpass). The emission is very weak, but clearly detected. The line has a peak flux of 0.31 mJy located at a velocity of 5552 [kms$^{-1}$]{}. This is only a 3$\sigma$ detection, but it is the brightest feature in the spectrum and is located within $\sim$40 [kms$^{-1}$]{} of the optical velocity of SA 68-6597; which is within the 1$\sigma$ uncertainties of the optical redshift. To check its reality, we split the raw data into various subsets (e.g. by sets of days, polarization, etc.) and searched for the line in these data. The line was either visible in the subsets or its absence was consistent with the noise levels; thus we believe that it is a real emission line from SA 68-6597.
The vertical dotted lines in Figure \[hispec\] indicate the region over which we measured the properties of SA 68-6597. The integrated flux is 0.0095$\pm$0.0025 Jy [kms$^{-1}$]{}, which translates to a [$M_{HI}$]{} of $(1.4\pm0.4)\times$10$^7$[$M_\odot$]{}–a 3.8$\sigma$ detection over five channels. We know of no galaxies within the 3.5$\arcmin$ beam of Arecibo that may be contaminating this measurement, so we believe that this is truly associated with SA 68-6597. Combining these data with the optical properties, we find a [$M_{HI}$]{}-to-[$L_B$]{}ratio of 1.0$\pm$0.3[$M_\odot$]{}/[$L_\odot$]{}. The gas depletion timescale, $\tau$=[$M_{HI}$]{}/SFR, is 5$\pm$2 Gyr without accounting for helium, molecular gas, or recycling.
The linewidth at 50% of the peak flux, W$_{50}$ (FWHM), is measured to be 32$\pm$10 [kms$^{-1}$]{}, centered at 5557$\pm$5 [kms$^{-1}$]{}. Again, this is within the uncertainties of the optical recession velocity. Because of the low signal-to-noise ratio of the detection, this value is highly imprecise and potentially inaccurate. To address this issue, we have used a Monte Carlo simulation of a Gaussian line with a peak signal-to-noise ratio of 3$\sigma$ to relate the measured FWHM to the true FWHM. We find that the the true FWHM = 33$\pm^{60}_{12}$ [kms$^{-1}$]{}. Converting this to a W$_{20}$ yields 51$\pm^{93}_{19}$ [kms$^{-1}$]{}, assuming a Gaussian lineshape.
One of the main goals of our project is to determine the dynamical mass, [$M_{dyn}$]{}, of SA 68-6597 using the line. As discussed in Section \[intro\], the line is generally believed to be a more reliable tracer of the gravitational potential than the H$\beta$ line. Because of the large uncertainties associated with our measurement, and the additional uncertainties from the unknown inclination of SA 68-6597, we are practically restricted to calculating a lower limit to [$M_{dyn}$]{}. We follow the same procedure to derive the dynamical mass as in @pis01 using the following standard formula assuming the is in circular rotation: [$M_{dyn}$]{}$(<R)~\ge~V_{rot}^2 \times R/G$.
In this case we take V$_{rot}$ to be half of the lower limit on W$_{20}$ uncorrected for inclination, for the radius we scale R$_{25}$ using a canonical scaling factor from @bro97 to get R$_{HI}$=680$\pm^{210}_{200}$ pc. If we use these values to calculate a lower limit to [$M_{dyn}$]{}, we find it is greater than 3.0$\times$10$^7$[$M_\odot$]{}. This yields [$M_{HI}$]{}/ M$_{dyn} \le$ 0.47, and [$M_{dyn}$]{}/[$L_B$]{}$\ge$2. See @pis01 for a discussion of the uncertainties involved in this calculation. All of these measured and derived properties of SA 68-6597 are summarized in Table \[props\].
Discussion
==========
Our observations have revealed that SA 68-6597 is not only a low luminosity galaxy, but also has a low [$M_{HI}$]{}, and probably a low [$M_{dyn}$]{}as well. By all three measures, SA 68-6597 reveals itself to be an extreme cousin of other BCDs. BCDs typically have [$M_{HI}$]{}$\sim$10$^{8-9}$[$M_\odot$]{}, with only a few as low as 10$^7$[$M_\odot$]{} or as high as 10$^{10}$[$M_\odot$]{}. They have [$M_{dyn}$]{}$\sim$10$^{8-10}$[$M_\odot$]{}, and [$L_B$]{}$\sim$10$^{8-10}$[$L_\odot$]{} [@cha77; @thu81; @hof89; @sta92; @thu99; @sal02]. SA 68-6597 represents the extreme low mass, low luminosity end of BCDs and is not a distinctly different class of galaxy as generally evidenced by its scale-free properties, such as the -mass-to-light ratio, [$M_{HI}$]{}/[$L_B$]{}, and the gas-mass fraction, [$M_{HI}$]{}/[$M_{dyn}$]{}.
SA 68-6597 has an [$M_{HI}$]{}/[$L_B$]{} ratio that is consistent with that of the large samples of BCDs studied by @sta92 [@vz98b; @vz01; @sal02; @hof03; @thu04] who found ratios ranging from $\sim$0.33-1.46 [$M_\odot$]{}/[$L_\odot$]{}. SA 68-6597’s mass-to-light ratio is even within the range for luminous compact blue galaxies studied by @gar04b, but is about twice the median value. Only the study of @hof89 found a significantly lower [$M_{HI}$]{}/[$L_B$]{} value for 11 Virgo cluster BCDs of 0.04 [$M_\odot$]{}/[$L_\odot$]{}. The [$M_{HI}$]{}/[$M_{dyn}$]{} values for BCDs are also generally consistent with SA 68-6597’s upper limit of 0.47. A variety of studies of BCDS find [$M_{HI}$]{}/[$M_{dyn}$]{} ratios ranging from 0.01-0.78 [@hof89; @vz98b; @hof03; @thu04]. The ratio of dynamical mass-to-light, [$M_{dyn}$]{}/[$L_B$]{}, for other BCDs ranges from 0.18-2.62 [$M_\odot$]{}/[$L_\odot$]{} [@hof89; @sta92; @thu04], which is also consistent with the lower limit of 2 [$M_\odot$]{}/[$L_\odot$]{} for SA 68-6597. It’s SFR and [$M_{HI}$]{} are consistent with expectations for BCDs based on SA 68-6597’s linewidth [@hop02]. Even luminous compact blue galaxies have a median [$M_{dyn}$]{}/[$L_B$]{} only slightly higher (5 [$M_\odot$]{}/[$L_\odot$]{}) than that of SA 68-6597[@gar04b]. All of these ratios suggest that SA 68-6597 is an extremely faint, extremely low-mass version of a typical blue compact dwarf.
In terms of individual BCDs, SA 68-6597 is quite similar in its gaseous properties to I Zw 18 and Haro 4. I Zw 18 is still slightly more massive and more luminous with a [$M_{HI}$]{} = 2.6$\times$10$^7$[$M_\odot$]{}, [$M_{dyn}$]{}=2.6$\times$10$^8$[$M_\odot$]{}, and [$L_B$]{} = 3.5$\times$10$^7$[$L_\odot$]{}[@vz98a]. Nevertheless, with [$M_{HI}$]{}/[$L_B$]{} = 0.7 and [$M_{dyn}$]{}/[$L_B$]{} = 5 I Zw 18 has mass-to-light ratios nearly identical to SA 68-6597. Its gas-mass fraction of 0.1 is also similar to SA 68-6597. Haro 4 is slightly less similar with [$M_{HI}$]{} = 2$\times$10$^7$[$M_\odot$]{}, [$M_{HI}$]{}/[$L_B$]{}= 0.17[$M_\odot$]{}/[$L_\odot$]{}, [$M_{dyn}$]{}=5$\times$10$^8$[$M_\odot$]{}, [$M_{dyn}$]{}/[$L_B$]{} = 4.8[$M_\odot$]{}/[$L_\odot$]{}, and [$M_{HI}$]{}/[$M_{dyn}$]{}=0.03 [@bra04]. While [$M_{HI}$]{}, [$M_{dyn}$]{}, and [$M_{dyn}$]{}/[$L_B$]{} of Haro 4 are close to those of SA 68-6597, the [$M_{HI}$]{}/[$L_B$]{} is lower, and the gas-mass fraction is lower than SA 68-6597, but still consistent with it.
The question is then raised: “what do these properties say about the evolutionary path of SA 68-6597?” Our derived gas depletion timescale, $\tau$, for SA 68-6597 is 5$\pm$2 Gyr. This value provides a rough measure of the time it will take for SA 68-6597 to consume all of its gas at its current rate of star formation [@ken83]. This is similar to many of the measured values for a sample of 21 BCDs from @hop02, but is much greater than a sample of 15 BCDs studied by @sag92. It has a gas depletion timescale equal to the median value for the sample of field and cluster galaxies studied by @ken83. SA 68-6597 has a larger $\tau$ than either I Zw 18 or SBS 0355-052 [@hop02]. It has a shorter $\tau$ than all Local Group dwarf irregular galaxies except IC 10 and NGC 6822 [@mat98]. Our estimate does not account for the contribution of helium, recycling of gas, molecular gas, a decreasing SFR or less than 100% star formation efficiency–all of which would increase $\tau$–or the possible effects of outflow–which would decrease $\tau$. @mac99 suggest that galaxies with gas masses below 10$^6$[$M_\odot$]{} can suffer complete blowout of their gas, while galaxies between 10$^6$ and 10$^7$[$M_\odot$]{} may suffer partial blowout. Because the derived gas mass of SA 68-6597 is $\sim$10$^7$[$M_\odot$]{}, we expect that it could only have a partial outflow [@mac99]. The results of @mac99 are based on a dark matter halo approximately 10-100$\times~$ larger than the gas mass. Overall, this means that, SA 68-6597 should evolve in a similar fashion to many less extreme BCDs and normal spiral and irregular galaxies; it will not rapidly consume its gas and passively evolve in the near future. The large [$M_{dyn}$]{}of SA 68-6597 implies that if and when it consumes all its gas, it may be able to evolve into a moderate mass analog of the Local Group dwarf spheroidals [@mat98].
While the low signal-to-noise ratio of our observations make our velocity widths very uncertain, it is worth noting a potentially interesting property of SA 68-6597. The H$\beta$ velocity dispersion of SA 68-6597 is 27$\pm^1_2$ [kms$^{-1}$]{}, while the dispersion is 14$\pm^{25}_{5}$ [kms$^{-1}$]{}. If the linewidth is actually smaller than that of the ionized gas, then we may be seeing evidence of a galactic outflow in SA 68-6597. Such outflows could result in partial blowout of the ISM in the galaxy of anywhere between $\lesssim$1% to 100% [@mac99] potentially reducing the gas depletion time. Outflows can also be particularly efficient at ejecting metals into the intergalactic medium [@mac99; @fer00] permitting a galaxy like SA 68-6597 to form many generations of stars while maintaining its very low metallicity. Clearly more sensitive and detailed observations are needed to address this issue.
Conclusions {#conc}
===========
We have observed the recently discovered blue compact dwarf galaxy, SA 68-6597, with the Arecibo telescope to determine its gas content and better constrain its total mass using the line. SA 68-6597 is one of the faintest, lowest metallicity BCDs known.
SA 68-6597 has properties which indicate it is a typical blue compact dwarf galaxy in all ways, except for its extremely low luminosity and small and dynamical masses. In this way it represents the faint, low mass tail of the distribution of BCD properties. It is slightly fainter and less massive than the famous BCD I Zw 18 and only slightly more metal rich. SA 68-6597’s gas depletion timescale is similar to the value for other BCDs and normal field galaxies, yet is shorter than most dwarf irregulars in the Local Group. Nevertheless, SA 68-6597 can continue to form stars at its current, prolific rate for almost 5 Gyr and would, therefore, be unlikely to fade significantly in that time. When it does fade, its relatively large dynamical mass suggests it may be able to evolve into a massive dwarf spheroidal galaxy. Its future evolutionary path remains murky.
Because of the combination of the distance and low mass of SA 68-6597, our detection was only at the 3$\sigma$ level, meaning that the measured linewidth and derived dynamical mass are poorly constrained. Nevertheless, the most probable value of the linewidth is less than the measured H$\beta$ linewidth indicating that galactic outflows may be present in SA 68-6597. Because of the potential implications of such a situation on SA 68-6597’s evolution, more sensitive, spatially resolved observations of SA 68-6597 are essential to unravel its current nature and reveal its evolutionary path.
This research has made use of the NASA/IPAC Extragalactic Database (NED) which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration. The authors wish to thank the staff at Arecibo, especially Tapasi Ghosh, for their help with our observations. We also wish to thank the Arecibo observatory for granting us more observing time to improve the quality of our observations. We thank Eric Wilcots for assistance with the observing and helpful discussions. This research was performed while D.J.P. held a National Research Council Research Associateship Award at the Naval Research Laboratory. Basic research in astronomy at the Naval Research Laboratory is funded by the Office of Naval Research. D.J.P. also acknowledges generous support from the ATNF via a Bolton Fellowship and from NSF MPS Distinguished International Research Fellowship grant AST 0104439. D.C.K. acknowledges support from NSF AST-0071198 and HST GO-07339.01-96A.
Bravo-Alfaro, H., Brinks, E., Baker, A. J., Walter, F., & Kunth, D. 2004, , 127, 264
Broeils, A. H. & Rhee, M.-H. 1997, , 324, 877
Chamaraux, P., 1977, , 60, 67
Doyle, M. T., et al., 2005, , 361, 34
Ferrara, A. & Tolstoy, E. 2000, , 313, 291
Garland, C. A., Pisano, D. J., Williams, J. P., Guzm[' a]{}n, R., & Castander, F. J., 2004, , 615, 689
Gebhardt, K., et al. 2000, , 539, L13
Gordon, D., Gottesman, S.T., 1981, , 86, 161
Guzmán, R., Koo, D.C., Faber, S.M., Illingworth, G.D., Takamiya, M., Kron, R., Bershady, M.A., 1996, , 460, L5
Hoffman, G. L., Helou, G., Salpeter, E. E., & Lewis, B. M. 1989, , 339, 812
Hoffman, G. L., Brosch, N., Salpeter, E. E., & Carle, N. J. 2003, , 126, 2774
Hopkins, A. M., Schulte-Ladbeck, R. E., & Drozdovsky, I. O. 2002, , 124, 862
Izotov, Y.I., Chaffee, F.H., Foltz, C.B., Green, R.F., Guseva, N.G. & Thuan, T.X. 1999, , 527, 757
Izotov, Y.I., & Thuan, T.X., 2004, , in press (astro-ph/0408391)
Kennicutt, R. C. 1983, , 272, 54
Kong, X., Cheng, F.Z., 2002, , 389, 845
Koo, D.C., Bershady, M.A., Wirth, G.D., Stanford, S.A., Majewski, S.R., 1994, , 427, L9
Koo, D. C., Guzman, R., Faber, S. M., Illingworth, G. D., Bershady, M. A., Kron, R. G., & Takamiya, M. 1995, , 440, L49
Lee, J. C., Salzer, J. J., & Melbourne, J. 2004, , 616, 752
Lipovetsky, V. A., Chaffee, F. H., Izotov, Y. I., Foltz, C. B., Kniazev, A. Y., & Hopp, U. 1999, , 519, 177
Mac Low, M.-M., Ferrara, A., 1999, , 513, 142
Mateo, M. L. 1998, , 36, 435
Oke, J. B. et al. 1995, , 107, 3750
Pisano, D.J., Kobulnicky, H.A., Guzmán, R., Gallego, J., Bershady, M.A., 2001, , 122, 1194
Pustilnik, S.A., Brinks, E., Thuan, T.X., Lipovetsky, V.A., Izotov, Y.I., 2001, , 121, 1413
Sage, L. J., Salzer, J. J., Loose, H.-H., & Henkel, C. 1992, , 265, 19
Salzer, J. J., Rosenberg, J. L., Weisstein, E. W., Mazzarella, J. M., & Bothun, G. D. 2002, , 124, 191
Schneider, S.E., Thuan, T.X., Magri, C., & Wadiak, J.E., 1990, , 72, 245
Staveley-Smith, L., Davies, R. D., & Kinman, T. D. 1992, , 258, 334
Thuan, T.X., Martin, G.E., 1981, , 247, 823
Thuan, T. X., Lipovetsky, V. A., Martin, J.-M., & Pustilnik, S. A. 1999, , 139, 1
Thuan, T. X., Hibbard, J. E., & L[' e]{}vrier, F. 2004, , 128, 617
, 58, 67
van Zee, L., Westpfahl, D., Haynes, M. P., & Salzer, J. J. 1998, , 115, 1000
van Zee, L., Skillman, E.D., Salzer, J.J., 1998, , 116,1186
van Zee, L., Salzer, J. J., & Skillman, E. D. 2001, , 122, 121
Vogt, S. S. et al. 1994, Proc. SPIE Instrumentation in Astronomy VIII, David L. Crawford; Eric R. Craine; eds., 2198, 362
[^1]: Deep Extragalactic Evolutionary Probe: see URL <http://deep.ucolick.org>
[^2]: The Arecibo Observatory is part of the National Astronomy and Ionosphere Center which is operated by Cornell University under a Cooperative Agreement with the National Science Foundation.
[^3]: The NASA/IPAC Extragalactic Database (NED) is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The Fermat principle is used to define trajectories in nonhomogenous optical media. The Poincaré model of the Lobačevskii geometry is derived. The index of refraction is determined for the light confined in the circular trajectory in the optical medium.'
author:
- |
Miroslav Pardy\
Department of Physical Electronics\
Masaryk University\
Kotlářská 2, 611 37 Brno, Czech Republic\
e-mail:[email protected]
title: Special light trajectories in optical medium
---
=500 =2500 -15mm
[**Key words:**]{} The Fermat principle, ray trajectories, geometrical optics, the Poincaré model of the Lobačevskii geometry, confinement of light.
Introduction
============
Trajectories of elementary particles are the basic ingredients of physics of elementary particles and cosmical rays. No elementary particle can exist without its trajectory. While in particle physics the trajectories of particles are determined by their parameters as mass, charge, spin, velocity and by the influence of the magnetic and electric fields on its motion, the trajectories of light in geometrical optics are determined by the index of refraction of the optical media.
Geometrical optics considers the propagation of waves of light, as the propagation of rays, completely divorced from their wave properties. In other words, geometrical optics corresponds to the limiting case of small wavelength, $\lambda \rightarrow 0$.
The fundamental equations of geometrical optics for the direction of the rays are derived for instance by Landau et al. (1988, 1982) for any quantity $f$ describing the field of the wave (any component of ${\bf E}$ or ${\bf H})$. For a plane monochromatic wave, we have for $f$ the form
$$f=ae^{i({\bf k.r}-\omega t +\alpha)} =ae^{i(-k_{r}x^{r} + \alpha)},\eqno(1)$$ (we omit the Re as the real part of all expressions). We write the expression for the field in the form
$$f = ae^{i\psi}.\eqno(2)$$
In case the wave is not plane, but geometrical optics is applicable, the amplitude $a$ is, generally speaking, a function of the coordinates and time, and the phase $\psi $ which is called the eikonal, does not have a simple form, as in (1). It is essential, however, that $\psi$ be a large quantity. This is clear immediately from the fact that eometrical optics corresponds to the limit $\lambda \rightarrow 0$.
Over small space regions and time intervals the eikonal $\psi$ can be expanded in series. The time derivative of $\psi$ gives the frequency of the wave:
$$\frac{\partial \psi}{\partial t} = -\omega \eqno(3)$$ and the space derivatives give the wave vector
$${\bf grad}\;\psi = {\bf k}\eqno(4)$$ and consequently the direction of the ray through any point in space. For a steady monochromatic wave, the frequency is a constant and the time dependence of the eikonal is given by a term $-\omega t$. We then introduce a new function $\psi_{1}$ (also called the eikonal), such that
$$\psi = -\omega t + (\omega/c) \psi_{1}(x,y,z).\eqno(5)$$.
Then $\psi_{1}$ is a function of the coordinates only, and its gradient is
$${\bf grad}\;\psi_{1} = {\bf n},\eqno(6)$$ where [**n**]{} is a vector such that
$${\bf k} = \omega {\bf n}.\eqno(7)$$
The magnitude of ${\bf n}$ is equal to the refractive index $n$ of the medium. Hence the equation for the eikonal in ray propagation in a medium with refractive index $n(x,y,z)$ ($a$ given function of the coordinates) is
$$|{\bf grad}\;\psi_{1}|^{2} = \left(\frac{\partial\psi_{1}}{\partial x}\right)^{2} +
\left(\frac{\partial\psi_{1}}{\partial y}\right)^{2} +
\left(\frac{\partial\psi_{1}}{\partial z}\right)^{2} = n^{2}. \eqno(8)$$
The equation of ray propagation in a steady state can also be derived from Fermat’s principle, according to which the integral
$$\psi_{1} = \int_{A}^{B}{\bf n}.d{\bf l} = \int_{A}^{B}ndl\eqno(9)$$ along the path of the ray between two given points $A$ and $B$ has a value less than for any other path between $A$ and $B$.
Equating to zero the variation of this integral, we have
$$\delta\psi_{1} = \int_{A}^{B}({\delta n}.d{l} +n\delta dl) = 0.\eqno(10)$$
Let $\delta {\bf r}$ be a displacement of the ray path under the variation. Then $\delta n = \delta{\bf r}.{\bf grad}\; n$, $\delta dl = {\bf 1}.d\delta{\bf r}$, where ${\bf 1}$ is a unit vector tangential to the ray. Substituting in $\delta \psi_{1}$ and integrating by parts in the second term (using the fact that $\delta {\bf r} = 0$ at $A$ and $B$), we have
$$\delta \psi_{1} = \int_{A}^{B}\delta{\bf r} .{\bf grad}\;{n}\;d{l} +
\int_{A}^{B}n{\bf l}.d\delta {\bf r} = \eqno(11)$$
$$\int_{A}^{B}\left({\bf grad}\; n - \frac{d(n{\bf l})}{dl}\right).\delta {\bf r}dl. \eqno(12)$$
Hence
$${\bf grad}\; n.d{l} = \frac{d(n{\bf l})}{dl}.\eqno(13)$$
Expanding the derivative and putting $dn/dl = {\bf 1}. {\bf grad}\; n$, we obtain
$$\frac{d{\bf 1}}{dl} = \frac{1}{n}[{\bf grad}\;n -
{\bf 1}({\bf l} . {\bf grad}\;n)].\eqno(14)$$
This is the equation giving the form of the rays.
We know from differential geometry that the derivative $d{\bf 1}/dl$ along the ray is equal to ${\bf N}/R$, where ${\bf N}$ is the unit vector along the principal normal and R the radius of curvature. Taking the scalar product of both sides of (14) with ${\bf N}$, and using the fact that ${\bf N}$ and ${\bf 1}$ are Perpendicular, we have
$$\frac{1}{R} = {\bf N}.\frac{{\bf grad}\;n}{n}. \eqno(15)$$
The rays are therefore bent in the direction of increasing refractive index.
The velocity of propagation of rays in geometrical optics is in the direction of $\bf 1$ and is given by the derivative
$${\bf u} = \partial\omega/\partial{\bf k}.\eqno(16)$$
This is also called the group velocity, the ratio $\omega/k$ being called the phase velocity. However, the latter is not the velocity of physical propagation of any quantity.
Light trajectory from the Fermat principle
==========================================
The point $A$ and $B$ can be joined by the infinite number of lines passing from $A$ to $B$. There is the shortest distance between $A$ and $B$ forming the segment of the straight line passing through $A$ and $B$ (Hilbert, 1902). The shortest distance between point $A$ and $B$ can be physically realized by the flexible but non-elastic fibre. The segment $AB$ of a straight line can be prolongated in the direction $AB$, or $BA$ in order to generate the straight line. The shortest trajectory $y = y(x)$ between two point $A(x_{1}, y_{1})$ and $B(x_{2}, y_{2})$ in the Euclidean plane x-y is $y = ax + b$, $a, b$ being some constants which can be find as the solution of the Bernoulli izoperimetric problem of the variational calculus with functional
$$F = \int_{x_{1}}^{x_{2}}[1 + y'^{2}]^{1/2}dx ; \quad y' = \frac{dy}{dx}\eqno(17)$$ after insertion of it into the Euler-Lagrange equation
$$\frac{\partial F}{\partial y} - \frac{d}{dx}\left(\frac{\partial F}{\partial y'}\right) = 0\eqno(18)$$ and after its solution (Lavrentjev et al., 1950).
The geodetic line $y = y(x), z = z(x)$ from point $A(x_{1}, y_{1}, z_{1})$ to point $B(x_{2}, y_{2}, z_{2})$ on the surface $\varphi(x, y, z) = 0$ is the solution of the izoperimetric problem with the functional
$$F = \int_{x_{1}}^{x_{2}}\left\{[1 + y'^{2} + z'^{2}]^{1/2} - \lambda
\varphi(x, y, z)\right\}dx ; \quad y' = \frac{dy}{dx}, z' = \frac{dz}{dx},\eqno(19)$$ where $\lambda$ is the Lagrange multiplicator.
The Fermat optical theorem states that the trajectory of light from point $A$ to $B$ in the optical medium is of the shortest optical length. At the same time the trajectory of the optical ray from point $A$ to point $B$ with reflection on the mirror in point $C$ is of the shortest optical length in optical medium.
The trajectory of light passing from $A(x_{1}, y_{1})$ to $B(x_{2},
y_{2})$ is determined by the Fermat principle, which states that the time from $A(x_{1}, y_{1})$ to $B(x_{2}, y_{2})$ is the result of the minimization of the functional
$$T(y,y') = \int_{x_{1}}^{x_{2}}\frac{ds}{v(y)} = \int_{x_{1}}^{x_{2}}\frac{\sqrt{1 + y'^{2}}}{v(y)}dx, \eqno(20)$$ where $v(y)$ is the velocity of light.
The functional $T(y,y')$ is the solution of the Euler-Lagrange equations (18).
If $v = Ay$, the solution of eq. (18) is in the form of circles forming the Poincaré model of the Lobačevskii geometry:
$$(x - C)^{2} + y^{2} = r^{2}.\eqno(21)$$
The last equation can be also derived for velocity $v = Ax$. Then the functional (20) is of the form
$$T(y,y') = \int_{x_{1}}^{x_{2}}\frac{ds}{v(x)} =
\int_{x_{1}}^{x_{2}}\frac{\sqrt{1 + y'^{2}}}{v(x)}dx, \eqno(22)$$
We get from the Euler equation with $y' = dy/dx$
$$T_{y} - \frac{d}{dx}T_{y'} = 0 \eqno(23)$$ we get $T_{y'} = const = C_{1}$, or
$$\frac{y'}{x\sqrt{1 + y'^{2}}} = C_{1}.\eqno(24)$$
We get with the substitution $t = \arctan y' $,
$$x = \frac{y'}{C_{1}\sqrt{1 + y'^{2}}} = \frac{1}{C_{1}}\sin t =
C_{2}\sin t.\eqno(25)$$
It follows from $y' = \tan t$ that $dy = C_{2}\sin t dt$. Or
$$y = - C_{2}\cos t + C_{3}.\eqno(26)$$ Or,
$$x = C_{2}\sin t, \quad y - C_{3} = - C_{2}\cos t. \eqno(27)$$
If we eliminate the variable $t$ in the last equation, we get
$$x ^{2} + (y - C_{3})^{2} = C_{2}^{2} = r^{2},\eqno(28)$$ which is a circle with the center on the y-axis.
Trajectory of light in the stratified medium
============================================
Let us consider the stratified 2D medium in the plane $x-y$, where every layer is parallel with the x-axis. Then, $dx \rightarrow (\Delta x)_{i}, dy \rightarrow H/n$, in (20) where i = 1, 2, 3, ..., and $H$ is the height of the stratified medium. In this case the velocity is $v_{i} = v(y_{0} + iH/n)$. Then, instead of functional $T(y,y')$ we write the following sum:
$$T = \sum_{i=0}^{\infty} \frac{\sqrt{((\Delta x)_{i})^{2} + (\Delta y)^{2}}}{v_{i}} = \sum_{i=0}^{\infty} \frac{\sqrt{(a_{i+1} - a_{i})^{2} + (H/n)^{2}}}{v_{i}}.\eqno(29)$$
The last sum has the stationary value if and only if $\partial T/\partial a_{i} = 0$. Or,
$$\partial T/\partial a_{i} =$$
$$-\frac{(a_{i+1} - a_{i})}{v_{i}\sqrt{(a_{i+1} - a_{i})^{2} + (H/n)^{2}}} +
\frac{(a_{i} - a_{i-1})}{v_{i-1}\sqrt{(a_{i} - a_{i-1})^{2} + (H/n)^{2}}} =$$
$$-\frac{\cos\varphi_{i}}{v_{i}} + \frac{\cos\varphi_{i-1}}{v_{i-1}} = 0.\eqno(30)$$ Or,
$$\frac{\cos\varphi_{i}}{v_{i}} = \frac{\cos\varphi_{i-1}}{v_{i-1}} = const = \frac{1}{k}.\eqno(31)$$ Or, in the continual limit,
$$\frac{\cos\varphi}{v(y)} = \frac{1}{k} \eqno(32)$$
However, for $y = y(x)$ it is $\tan\varphi = y'$ and $\cos\varphi = \frac{1}{\sqrt{1 + y'^{2}}}$
Then we have
$$x - x_{0} = \int\frac{v(y)dy}{\sqrt{k^{2} - v^{2}}}.\eqno(33)$$ We get for $v = Ay$:
$$x = \int_{y_{0}}^{y}\frac{Aydy}{\sqrt{k^{2} - A^{2}y^{2}}} + C\eqno(34)$$ with the solution
$$(x - C)^{2} + y^{2} = \left(\frac{k}{A}\right)^{2} = R^{2},
\eqno(35)$$ which is the circle with the centre on the x-axis.
For $v= A/y$, we get instead of equation (34):
$$x = \int\frac{Ady}{\sqrt{k^{2}y^{2} - A^{2}}} + C.\eqno(36)$$
Let us remark that the physical meaning of the relation (36) with $v = A/y$ is the problem of the catenary in the homogeneous gravitational field.
The Poincaré optical model of the Lobačevskii geometry
======================================================
The Lobačevskii geometry is the integral part of the general geometry called non-euclidean geometry. The name non-Euclidean was used by Gauss to describe a system of geometry which differs from Euclid’s in its properties of parallelism. Such a system was developed independently by Bolyai in Hungary and Lobačevskii in Russia, about 120 years ago. Another system, differing more radically from Euclid’s, was suggested later by Riemann in Germany and Schlafli in Switzerland. The subject was unified in 1871 by Klein, who gave the names parabolic, hyperbolic, and elliptic to the respective systems of Euclid, Bolyai-Lobačevskii, and Riemann-Shlafli (Coxeter, 1998).
The substantial mathematical object in the Lobačevskii geometry is the angle of parallelism defined by Lobačevskii as follows. Given a point $P$ and a line $q$. The Intersection of the perpendicular through $P$ let be $Q$ and $PQ = x$. The intersection of line p passing through $P$, with $q$, let be $R$ and $QR = k$. Then, the angle $RPQ$ for perpendicular distance $x$
$$\Pi(x)=2\tan^{-1}e^{-x/k}.\eqno(37)$$ is known as the Lobačevskii formula for the angle of parallelism (Coxeter, 1998; Lobačevskii, 1914).
The Poincar' e model of the Lobačevskii geometry is the physical model of the optical trajectories in a medium with the velocity of light $v = Ay$.
According to Hilbert (Hilbert, 1903), it is not possible to realize the Lobačevskii geometry globally on surface with the constant negative curvature. The Beltrami realization of the Lobačevskii geometry is only partial. The Lobačevskii geometry is the partial geometry on the pseudosphere with the parametric equations
$$x = a\sin u \cos v, \quad y = a\sin u \sin v, \quad z = a\left(\ln \tan\frac{u}{2} + \cos u\right).\eqno(38)$$ (Kagan, 1947, ibid. 1948; Efimov, 2004; Norden, 1956; Klein, 2004; Manning, 1963).
The pseudosphere is the surface generated by the rotation of the Leibniz tractrix with equation
$$x = a\sin u,\quad y = 0, \quad z = a\left(\ln \tan\frac{u}{2} + \cos u\right).\eqno(39)$$
The pseudosphere is in the half geodetic coordinates given by the squared element
$$ds^{2} = du^{2} + \cosh^{2}\frac{u}{a}dv^{2}.\eqno(40)$$
The pseudosphere is in the izothermical coordinates given by the Poincar' e squared element
$$ds^{2} = \frac{a(dx^{2} + dy^{2})}{y^{2}}.\eqno(41)$$
The Leibniz solution of the tractrix problem is as follows:
$$y = a\frac{\ln (a + \sqrt{a^{2} - x^{2}})}{x} - \sqrt{a^{2} - x^{2}}.\eqno(42)$$
The trajectory of light in the Poincar' e model is a trajectory passing from $A(x_{1}, y_{1})$ to $B(x_{2}, y_{2})$ and determined by the minimal time from $A(x_{1}, y_{1})$ to $B(x_{2}, y_{2})$. It is the result of the minimum of the functional (20, 22). The Poincaré circles in his model are analogue of the for straight lines in the Euclidean geometry. The following theorems are valid in the Poincar' e model of the Lobačevskii geometry:
[**Theorem 1:**]{}. Only one half-circle passes through two points $A, B$ in the Poincar' e plane.
[**Theorem 2:**]{} The curvilinear segment $AB$ in the Poincar' e plane is of the shortest length.
[**Theorem 3:**]{} The parallels are two half-circles with the intersections on the x-axis.
[**Theorem 4:**]{} If point $A \notin q$ then there are $q_{1} \parallel q$, $q_{2} \parallel q$ passing through $A$, with $q_{1} \neq q_{2}$.
[**Theorem 5:**]{} If point $A \notin q$, $q_{1} \parallel q, q_{2} \parallel q$, then $q_{1}, q_{2}$ divide the Poincar' e plane in four different sectors I, II, II, IV.
Let us remark that he optical distance between point $A$ and $B$ is not equivalent to the mechanical distance realized by the nonelastic flexible fibre as the shortest distance between point $A$ and $B$. The Poincar' e model of geometry where the light velocity is $v = Ay$ is the interaction model of light with the optical medium.
It is elementary to see that if we define the Poincar' e problem on a sphere, then we get so called spherical Poincar' e model of the Lobaěvskii geometry.
Let us use the index of refraction $n(r)$ in the Euclidean plane and polar coordinates $r, \varphi$ to derive the Poincar' e model.
The the explicit form of the Fermat principle
$$\delta\int n(r)ds = 0\eqno(43)$$ is (Marklund et al., 2002)
$$\delta\int n(r)\sqrt{1 + r^{2}\left(\frac{d\varphi}{dr}\right)^{2}} = 0\eqno(44)$$
The last equation is equivalent to the Euler-Lagrange variational equation for the functional $F(\varphi, \varphi')$
$$F_{\varphi} - \frac{d}{dr}F_{\varphi'} = 0.\eqno(45)$$ Or,
$$\frac{d}{dr}\left[n(r)\frac{r^{2}d\varphi dr}{\sqrt{1 + r^{2}\left(\frac{d\varphi}{dr}\right)^{2}}}\right] = 0;\eqno(46)$$
It is evident that the elimination of $d\varphi/dr$ is as follows:
$$\frac{d\varphi}{dr} = \pm \frac{C}{\sqrt{r^{4}n^{2}(r) - C^{2}r^{2}}} = 0.\eqno(47)$$
The circular trajectory is defined by equation $dr/d\varphi = 0$ from which follows the index of refraction as
$$n(r) = \frac{const}{r}.\eqno(48)$$
The experiments has been performed in Bose-Einstein condensate with the result that the optical light pulses can travel with extremely small group velocity about 17 meters per second (Hau et all. 1999).
Discussion
==========
We explained the Fermat principle in the geometrical optics and in the variational calculus. We determined the optical trajectories in the optical media. We also derived the Poincar' e model of the Lobačevskii geometry. The Poincar' e model of the Lobačevskii geometry is the optical model based on the index of refraction, which is the consequence of the interaction of light with medium. Also, space, when considered independently of measuring instruments, has neither metric nor projective properties; it has only topological properties. It is amorphous (Poincar' e, 1963). That which cannot be measured cannot be an object of science. We applied the Fermat principle to the determination of the index of refraction of the medium to get the circular optical trajectories. This effect is the confinement of light by optical medium.
There is no doubt that the monochromatic optical beam is composed from photons of energy $E = \hbar\omega$. While the rest mass of photon is zero, the relativistic mass follows from the Einstein relation $E = mc^{2}$. After identifying the relativity energy and quantum energy of photon we have
$$m = \frac{\hbar\omega}{c^{2}}. \eqno(49)$$
The centrifugal force acting on photon moving with velocity $v$ in optical medium along the circle with radius $R$ is for the photon mass as follows:
$$F_{centrifugal} = \frac{\hbar\omega}{c^{2}} \frac{v^{2}}{R}.\eqno(50)$$
The centrifugal force is the origin of the unstability of the photon trajectory in the optical medium. The further origin of the unstability of the trajectory are the thermal fluctuations of the index of refraction. We know, that the Kapitza effect is based on the thermal fluctuations of the index of refraction (Landau, et al., 1982). So, the experimental investigation of the confinement of photon in the optical medium is meaningful at $T \approx 0$. There is no doubt that the investigation of the photon trajectories is the crucial problem of the optical physics and it is interesting for all optical laboratories over the world.
[**References**]{}
Coxeter, H.S. M. [*Non-Euclidean Geometry*]{}, The mathematical association of America, Washington, D.C. 20036 , 6-th ed., Printed in USA, (1998).\
Efimov, N. V. [*The Higher geometry*]{}, 7-th ed. Nauka, Moscow, (2004). (in Russsian).\
Hau, L. V., Harris, E., Dutton, Z. and Behroozi, C. H, Light speed reduction to 17 metres per second in an ultracold atomic gas, Nature [**397**]{}, 594-598 (18 February 1999).\
Hilbert, D. [*The foundation of geometry*]{}, Chicago, London agents, (1902).\
Hilbert, D. (1903). [*[" U]{}ber Flächen von constanter Gauss’schen Krümmung*]{} ([*On the surfaces of the constant Gauss curvature*]{}), Giornale di Matematiche, [**6**]{}, p. 1868.\
Kagan, V. F. [*Foundation of the theory of sourfaces*]{} I, OGIZ, GITL, Moscow, St. Petersburg, (1947); ibid. II, (1948). (in Russian).\
Klein, F., Vorlesungen über höhere Geometrie, Dritte Auflage bearbeitet und herausgegeben von W. Blaschke, (Moscow), (2004). (in Russsian).\
Landau, L. D. and Lifshitz, E. M. [*The classical theory of fields*]{}, 7-th ed., Moscow, Nauka, (1988). (in Russian).\
Landau, L. D. and Lifshitz, E. M. [*Electrodynamics of continuous media*]{}. —2nd ed. —(Course of theoretical physics; V. 8), Moscow, Nauka, (1982).\
Lavrentjev, M. A. and Lyusternik, L. A. [*Lectures on the variational calculus*]{}, GITL, Moscow, (1950). (in Russian).\
Lobačevskii, N. [*The theory of parallels*]{}, La Salle, Illinois, Open court publishing Company, (1914).\
Manning, H. P. [*Introductory Non-Euclidean Geometry*]{}. New York: Dover, (1963).\
Marklund, M., Anderson, D., Cattani, F., Lisak, M. and Lundgren, L. Fermat’s principle and variational analysis of an optical model for light propagation exhibiting a critical radius, ArXiv: physics/0102019v2, 12 Aug 2002.\
Norden, A. P. Editor, [*On the foundations of geometries*]{}, Moscow, (1956). (in Russian).\
Poincar' e, H. [*Mathematics and Science: Last Essays*]{} (Dernieres Pensees) Dover Publications, Inc., New York, (1963).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We develop the formalism for canonical reduction of $(1+1)$–dimensional gravity coupled with a set of point particles by eliminating constraints and imposing coordinate conditions. The formalism itself is quite analogous to the $(3+1)$–dimensional case; however in $(1+1)$ dimensions an auxiliary scalar field is shown to have an important role. The reduced Hamiltonian is expressed as a form of spatial integral of the second derivative of the scalar field. Since in $(1+1)$ dimensions there exists no dynamical degree of freedom of the gravitational field ([*i.e.*]{} the transverse-traceless part of the metric tensor is zero), the reduced Hamiltonian is completely determined in terms of the particles’ canonical variables (coordinates and momenta). The explicit form of the Hamiltonian is calculated both in post-linear and post-Newtonian approximations.'
author:
- |
Tadayuki Ohta\
Department of Physics, Miyagi University of Education\
Aoba-Aramaki, Sendai 980, Japan
- |
Robert Mann\
Department of Physics, University of Waterloo\
Waterloo, Ontario, Canada N2L 3G1
title: |
Canonical reduction of two-dimensional gravity\
for Particle Dynamics
---
Introduction
============
Early versions of $(1+1)$–dimensional gravity [@r1; @r2] have in recent years led to intensive study of a wide variety of such theories, in large part because problems in quantum gravity become much more mathematically tractable in this context [@HS92]. In contrast to $(3+1)$–dimensional gravity, the action integrals for these theories must incorporate some dynamics in the form of an auxiliary (or dilaton) field since the Einstein tensor is topologically trivial in two dimensions. Although most of these theories can be written in a generic form [@BanksMann] whose general properties can then be studied [@LMK; @2djol; @jchan], there is often much to be learned by focussing on specific models within this general class [@RST].
One such theory has been extensively studied in many respects, including gravitational collapse, black holes, cosmological solutions and quantization [@r3; @r4; @r5; @rtquant]. Referred to as $R=T$ theory, the specific form of the coupling of the auxiliary field $\Psi$ to gravity is chosen so that it decouples from the classical field equations in such a way as to ensure that the evolution of the gravitational field is determined only by the matter stress-energy (and reciprocally) [@r3; @r4]. In this manner $R=T$ theory captures in two spacetime dimensions the essence of classical general relativity (as opposed to classical scalar-tensor theories), and has $(1+1)$–dimensional analogs of many of its properties [@r4; @r5]. Indeed, the theory can be understood as the $D\rightarrow 2$ limit of general relativity [@2dross].
One of the fundamental problems of (1+1)-dimensional gravity is its relationship to Newtonian gravity. This is in general a problematic issue [@jchan]. For a system of particles the Hamiltonian in Newtonian gravity in two dimensions is $$H=\sum_{a}\frac{p^{2}_{a}}{2m_{a}}
+\pi G\sum_{a}\sum_{b}m_{a}m_{b}\mid z_{a}-z_{b}\mid$$ where $m_{a}$, $z_{a}$ and $p_{a}$ are the rest mass, the coordinate and the momentum of $a$-th particle, respectively, and $G$ is the gravitational constant. We note that the $R=T$ theory has been shown to have a Newtonian limit [@r3; @r5]. However the dynamical role of the auxiliary field has not yet been fully analysed, a task which we investigate in this paper.
More generally we shall, in the context of $R=T$ theory, formulate a general framework for deriving a Hamiltonian for a system of particles, which coincides in a slow motion, weak field limit with the Hamiltonian (1). In addition to the properties stated above, an advantage of the $R=T$ model is that it needs only one auxiliary field. As with the ADM formalism in $(3+1)$–dimensional theory, we develop a canonical reduction by eliminating constraints and imposing coordinate conditions, which are quite analogous to the ADM conditions. In its final form the reduced Hamiltonian of the system is given as a spatial integral of the second derivative of the auxiliary scalar field $\Psi$. Since in two dimensions there exists no transverse-traceless part of the metric tensor (which are the real dynamical variables of gravitational field), the scalar field $\Psi$ is given as a function of the dynamical variables ($z_{a}$, $p_{a}$) of the particles by solving the constraint equations. Then the Hamiltonian is completely determined in terms of the coordinates and momenta of the particles.
The outline of our paper is as follows. Details of the reduction process are described in section 2, where much attention is devoted to the transformation of the total generator of the whole system and the choice of the coordinate conditions, from which the reduced Hamiltonian is defined. The consistency of the canonical reduction is proved in section 3. It guarantees that the canonical equations of motion given by the reduced Hamiltonian are identical with the original geodesic equations. The explicit form of the Hamiltonian is calculated in section 4 in a post-linear approximation (a series expansion in the coupling constant $G$) and in section 5 in a post-Newtonian approximation. While the former approximation is appropriate for the analysis of fast motion, the latter is adequate for treating the slow motion, weak field case. Section 6 is reserved for concluding remarks. The role of the $\Psi$ field is discussed in the simplest case of a single static source. Finally, an appendix illustrates the relation between the equations of motion in our coordinate conditions and those in other conditions.
Canonical formalism for particle dynamics
=========================================
The action integral for the gravitational field coupled with point particles is [@r4]
$$\begin{aligned}
I&=&\int dx^{2}\left[
\frac{1}{2\kappa}\sqrt{-g}
\left\{\Psi R+\frac{1}{2}g^{\mu\nu}\nabla_{\mu}\Psi\nabla_{\nu}\Psi\right\}
\right.
\nonumber \\
&&\makebox[2em]{}-\left.\sum_{a} m_{a}\int d\tau_{a}
\left\{-g_{\mu\nu}(x)\frac{dz^{\mu}_{a}}{d\tau_{a}}
\frac{dz^{\nu}_{a}}{d\tau_{a}}\right\}^{1/2}\delta^{2}(x-z_{a}(\tau_{a}))
\right] \end{aligned}$$
where $\Psi$ is the auxiliary scalar field. Here $g_{\mu\nu}$, $g$, $R$ and $\tau_{a}$ are the metric tensor of spacetime, det$(g_{\mu\nu})$, the Ricci scalar and the proper time of $a$-th particle, respectively, and $\kappa=8\pi G/c^4$. The symbol $\nabla_{\mu}$ denotes the covariant derivative associated with $g_{\mu\nu}$.
The field equations derived from the variations $\delta\Psi$ and $\delta g_{\mu\nu}$ are $$R-g^{\mu\nu}\nabla_{\mu}\nabla_{\nu}\Psi=R-\frac{1}{\sqrt{-g}}\partial_{\mu}
(\sqrt{-g}g^{\mu\nu}\partial_{\nu}\Psi)=0$$ $$\label{e4}
\frac{1}{2}\nabla_{\mu}\Psi\nabla_{\nu}\Psi
-\frac{1}{4}g_{\mu\nu}\nabla^{\lambda}\Psi\nabla_{\lambda}\Psi
+g_{\mu\nu}\nabla^{\lambda}\nabla_{\lambda}\Psi
-\nabla_{\mu}\nabla_{\nu}\Psi=\kappa T_{\mu\nu}$$ where $$\begin{aligned}
T_{\mu\nu}&=&-\frac{2}{\sqrt{-g}}\frac{\delta\cal L\mit_{M}}{\delta g^{\mu\nu}}
\nonumber \\
&=&\sum_{a} m_{a}\int d\tau_{a}\frac{1}{\sqrt{-g}}
g_{\mu\sigma}g_{\nu\rho}\frac{dz^{\sigma}_{a}}{d\tau_{a}}
\frac{dz^{\rho}_{a}}{d\tau_{a}}\delta^{2}(x-z_{a}(\tau_{a}))\;\; ,\end{aligned}$$ $\cal L\mit_{M}$ being the matter Lagrangian given by the second term in the brackets on the right hand side of (2). The geodesic equation derived from the variation $\delta z^{\mu}_{a}$ is $$\label{geo}
\frac{d}{d\tau_{a}}
\left\{g_{\mu\nu}(z_{a})\frac{dz^{\nu}_{a}}{d\tau_{a}}\right\}
-\frac{1}{2}g_{\nu\lambda,\mu}(z_{a})\frac{dz^{\nu}_{a}}{d\tau_{a}}
\frac{dz^{\lambda}_{a}}{d\tau_{a}}=0 \;\;.$$ The trace of Eq.(4) is $$\nabla^{\mu}\nabla_{\mu}\Psi=\kappa T^{\mu}_{\;\;\mu}$$ which yields $$\label{RT}
R=\kappa T^{\mu}_{\;\;\mu} \;\;.$$
Hence particle dynamics in $R=T$ theory may be described in terms of the equations (6) and (8), both of which are independent of the scalar field. Particulate matter (in terms of the trace of its stress-energy tensor) generates spacetime curvature via (\[RT\]); the effects of curvature act back upon matter via (\[geo\]).
Note that all three components of the metric tensor cannot be determined from (\[RT\]), since it is only one equation. The two extra degrees of freedom are related to the choice of coordinates. If the coordinate conditions are chosen to be independent of $\Psi$, equation (\[RT\]) determines the metric tensor completely. More generally however, we need to know the scalar field $\Psi$, through which the metric tensor is determined – it is this field that guarantees conservation of the stress-energy tensor via (\[e4\]). So far little attention has been paid to the role of the scalar field $\Psi$ .
To derive a Hamiltonian for a system of particles, we shall utilize the canonical formalism [@r6]. Writing $\gamma=g_{11}, N_{0}= (-g^{00})^{-1/2}, N_{1}= g_{10}$, the action (2) transforms to $$\label{e9}
I=\int dx^{2}\left\{\sum_{a}p_{a}\dot{z}_{a}\delta(x-z_{a}(x^{0}))
+\pi\dot{\gamma}+\Pi\dot{\Psi}+N_{0}R^{0}+N_{1}R^{1}\right\}$$ where $\pi$ and $\Pi$ are conjugate momenta to $\gamma$ and $\Psi$ respectively, and $$\begin{aligned}
R^{0}&=&-\kappa\sqrt{\gamma}\gamma\pi^{2}+2\kappa\sqrt{\gamma}\pi\Pi
+\frac{1}{4\kappa\sqrt{\gamma}}(\Psi^{\prime})^{2}
-\frac{1}{\kappa}\left(\frac{\Psi^{\prime}}{\sqrt{\gamma}}\right)^{\prime}
-\sum_{a}\sqrt{\frac{p^{2}_{a}}{\gamma}+m^{2}_{a}}\;
\delta(x-z_{a}(x^{0}))
\nonumber \\
\\
R^{1}&=&\frac{\gamma^{\prime}}{\gamma}\pi-\frac{1}{\gamma}\Pi\Psi^{\prime}
+2\pi^{\prime}
+\sum_{a}\frac{p_{a}}{\gamma}\delta(x-z_{a}(x^{0})) \;\;.\end{aligned}$$ Here and in the following we denote $\partial_{0}$ by a symbol $(\;\dot{}\;)$ and $\partial_{1}$ by a symbol $(\;^{\prime}\;)$.
Taking variations $\delta\gamma, \delta\pi, \delta N_{0},
\delta N_{1}, \delta\Psi, \delta\Pi, \delta z_{a}$ and $\delta p_{a}$ of the action (\[e9\]), we have $$\begin{aligned}
\dot{\pi}&+&N_{0}\left\{\frac{3\kappa}{2}\sqrt{\gamma}\pi^{2}
-\frac{\kappa}{\sqrt{\gamma}}\pi\Pi
+\frac{1}{8\kappa\sqrt{\gamma}\gamma}(\Psi^{\prime})^{2}
-\sum_{a}\frac{p^{2}_{a}}{2\gamma^{2}\sqrt{\frac{p^{2}_{a}}{\gamma}
+m^{2}_{a}}}\;\delta(x-z_{a}(x^{0}))\right\}
\nonumber \\
&+&N_{1}\left\{-\frac{1}{\gamma^{2}}\Pi\Psi^{\prime}
+\frac{\pi^{\prime}}{\gamma}
+\sum_{a}\frac{p_{a}}{\gamma^{2}}\;\delta(x-z_{a}(x^{0}))\right\}
\nonumber \\
&+&N^{\prime}_{0}\frac{1}{2\kappa\sqrt{\gamma}\gamma}\Psi^{\prime}
+N^{\prime}_{1}\frac{\pi}{\gamma}=0 \end{aligned}$$ $$\dot{\gamma}-N_{0}(2\kappa\sqrt{\gamma}\gamma\pi-2\kappa\sqrt{\gamma}\Pi)
+N_{1}\frac{\gamma^{\prime}}{\gamma}-2N^{\prime}_{1}=0$$ $$R^{0}=0$$ $$R^{1}=0$$ $$\dot{\Pi}+\partial_{1}(-\frac{1}{\gamma}N_{1}\Pi
+\frac{1}{2\kappa\sqrt{\gamma}}N_{0}\Psi^{\prime}
+\frac{1}{\kappa\sqrt{\gamma}}N^{\prime}_{0})=0$$ $$\dot{\Psi}+N_{0}(2\kappa\sqrt{\gamma}\pi)-N_{1}(\frac{1}{\gamma}
\Psi^{\prime})=0$$ $$\begin{aligned}
\dot{p}_{a}&+&\frac{\partial N_{0}}{\partial z_{a}}\sqrt{\frac{p^{2}_{a}}
{\gamma}+m^{2}_{a}}-\frac{N_{0}}{2\sqrt{\frac{p^{2}_{a}}{\gamma}+m^{2}_{a}}}
\frac{p^{2}_{a}}{\gamma^{2}}\frac{\partial\gamma}{\partial z_{a}}
\nonumber \\
&-&\frac{\partial N_{1}}{\partial z_{a}}\frac{p_{a}}{\gamma}
+N_{1}\frac{p_{a}}{\gamma^{2}}\frac{\partial\gamma}{\partial z_{a}}=0 \end{aligned}$$ $$\dot{z_{a}}-N_{0}\frac{\frac{p_{a}}{\gamma}}{\sqrt{\frac{p^{2}_{a}}{\gamma}
+m^{2}_{a}}}
+\frac{N_{1}}{\gamma}=0 \;\;.$$ In the equations (18) and (19), all metric components ($N_{0}$, $N_{1}$, $\gamma$) are evaluated at the point $x=z_{a}$ and $$\frac{\partial f}{\partial z_{a}}\equiv
\left.\frac{\partial f(x)}{\partial x}\right|_{x=z_{a}}$$ It can be shown that this set of equations is equivalent to the equations (3), (4) and (6). Here we show explicitly the equivalence of (18) and (19) to the geodesic equation (6). We express the $(\mu=1)$ component of (6) in terms of $N_{0}, N_{1}$ and $\gamma$ and their derivatives. Using the relations $$d\tau_{a}=dt\left\{N^{2}_{0}-\frac{1}{\gamma}(N_{1}+\gamma\dot{z}_{a})^{2}
\right\}^{\frac{1}{2}}$$ we get $$\begin{aligned}
\frac{d}{dt}\left\{\frac{N_{1}+\gamma\dot{z}_{a}}{\left[N^{2}_{0}
-\frac{1}{\gamma}(N_{1}+\gamma\dot{z}_{a})^2\right]^{\frac{1}{2}}}\right\}
+\frac{1}{\left[N^{2}_{0}-\frac{1}{\gamma}(N_{1}+\gamma\dot{z}_{a})^2\right]
^{\frac{1}{2}}}\left\{N_{0}\frac{\partial N_{0}}{\partial z_{a}}
-\left(\frac{N_{1}}{\gamma}+\dot{z}_{a}\right)\frac{\partial N_{1}}
{\partial z_{a}}\right.
\nonumber \\
\makebox[10em]{}\left.+\frac{1}{2}\left[\frac{N^{2}_{1}}{\gamma^{2}}
-(\dot{z}_{a})^{2}\right]\frac{\partial\gamma}{\partial z_{a}}\right\}=0\;\;.\end{aligned}$$ Eliminating $p_{a}$ from (18) and (19) yields an equation identical with (20).
From the expression (9), the total generator (obtained from variations at the end point) is given by $$\label{gen1}
G=\int dx\left\{\sum_{a}p_{a}\delta(x-z_{a})\delta z_{a}
+\pi\delta h-\Psi\delta\Pi\right\} \;\;\;,$$ where the above form has been obtained by adding a total time derivative $-\partial_{0}(\Pi\Psi)$ to the original action (9), and where the constraint equations (14) and (15) have been taken into account. In the above $h=1+\gamma$; the advantage of using $h$ will become clear later [@r7; @r8].
We must now identify the dynamic and gauge character of the variables which appear in the generator (\[gen1\]). By considering the constraint equations (14) and (15), we see that the only linear terms there are $\frac{\Psi^{\prime}}{\sqrt{h-1}}$ and $\pi^{\prime}$ respectively. The equations may therefore be solved for these quantities in terms of the dynamical and gauge ([*i.e.*]{} co-ordinate) degrees of freedom. By writing $\Pi = \frac{1}{\triangle}\Pi^{\prime\prime}$, where $1/\triangle$ is the inverse of the operator $\triangle=\partial^{2}/\partial x^{2}$ with appropriate boundary condition, we find that the generator becomes $$\begin{aligned}
G&=&\int dx\left\{\sum_{a}p_{a}\delta(x-z_{a})\delta z_{a}
-\left[-\frac{1}{\kappa}\left(\frac{\Psi^{\prime}}{\sqrt{h-1}}
\right)^{\prime}\right]
\delta\left[-\frac{\kappa}{\triangle}\left(\sqrt{h-1}\frac{1}{\triangle}
\Pi^{\prime}\right)^{\prime}\right]
\right.
\nonumber \\
&&\makebox[5em]{}\left.-\left[2\pi^{\prime}-\left(\frac{\Psi^{\prime}}{h-1}
\right)^{\prime}\frac{1}{\triangle}\Pi^{\prime}-\frac{1}{h-1}\Pi\Psi^{\prime}
\right]\delta\left(\frac{1}{2\triangle}h^{\prime}\right)\right\} \;\;.
\label{e22}\end{aligned}$$ where in obtaining this form, we have discarded surface terms.
The form (\[e22\]) of the generator is analogous what is obtained in $(3+1)$ dimensions under an orthogonal decomposition of the hypersurface metric and its conjugate momentum [@r7; @r8]. We therefore propose adopting the coordinate conditions $$\begin{aligned}
x&=&\frac{1}{2\triangle}h^{\prime}
\\
t&=&-\frac{\kappa}{\triangle}\left(\sqrt{h-1}\frac{1}{\triangle}\Pi^{\prime}
\right)^{\prime} \;\;\end{aligned}$$ which will then allow us to identify the Hamiltonian and momentum densities as the respective coefficients of $\delta t$ and $\delta x$ in the canonical form of the generator. Taking a spatial derivative of these coordinate conditions yields their differential form, which does not explicitly depend upon the coordinates $(t,x)$. This leads to the choices $$\begin{aligned}
h=2 &\longrightarrow& \gamma=1
\\
\frac{1}{\triangle}\Pi^{\prime}=0 &\longrightarrow& \Pi=0 \;\; \end{aligned}$$ which may then be inserted in equations (12) and (19) to solve for the relevant physical degrees of freedom.
Specifically, we may insert these choices into (14) and (15) to solve for the Hamiltonian and momentum densities. We then find that the generator (\[e22\]) may be expressed in the canonical form $$G=\int dx\left\{\sum_{a}p_{a}\delta(x-z_{a})\delta z_{a}
-\cal T\mit_{0 \mu}\delta x^{\mu}\right\}$$ where $$\begin{aligned}
\cal T\mit_{0 0}&=& \cal H\mit =-\frac{1}{\kappa}\triangle\rm\Psi\mit
\\
\cal T\mit_{0 1}&=&2\pi^{\prime} \;\;.\end{aligned}$$ Thus, $\cal H\mit$ is the Hamiltonian density of the system and $\cal T\mit_{0 1}$ is the momentum density. Note that since the constraints (14) and (15) have already been imposed, ${\cal T}_{0\mu}$ is expressed in terms of the canonical variables $z_{a}$ and $p_{a}$ as ${\cal T}_{0\mu}(x, z_{a}, p_{a})$, by solving these constraints. With the coordinate choices (25) and (26), they lead to $$\triangle\Psi-\frac{1}{4}(\Psi^{\prime})^{2}
+\kappa^{2}\pi^{2}+\kappa\sum_{a}\sqrt{p^{2}_{a}+m^{2}_{a}}
\delta(x-z_{a})=0 \;\;,$$ $$2\pi^{\prime}+\sum_{a}p_{a}\delta(x-z_{a})=0 \;\;.$$
Following a similar procedure as that used to transform the generator $G$, we rewrite the action integral (9) as $$\begin{aligned}
I&=&\int d^{2}x\left\{\sum_{a}p_{a}\dot{z}_{a}\delta(x-z_{a})
-\left[-\frac{1}{\kappa}\left(\frac{\Psi^{\prime}}{\sqrt{h-1}}\right)^{\prime}
\right]\partial_{0}\left[-\frac{\kappa}{\triangle}\left(\sqrt{h-1}
\frac{1}{\triangle}\Pi^{\prime}\right)^{\prime}\right]
\right.
\nonumber \\
&&\makebox[2em]{}\left.
-\left[2\pi^{\prime}-\left(\frac{\Psi^{\prime}}{h-1}\right)^{\prime}
\frac{1}{\triangle}\Pi^{\prime}-\frac{1}{h-1}\Pi\Psi^{\prime}\right]
\partial_{0}\left(\frac{1}{2\triangle}h^{\prime}\right)+N_{\mu}R^{\mu}
\right\}\label{e32}\end{aligned}$$ where we have discarded surface terms. Eliminating the constraints (14) and (15) and imposing the coordinate conditions (23) and (24), the action integral reduces to $$I_{R}=\int dx^{2}\left\{\sum_{a}p_{a}\dot{z}_{a}\delta(x-z_{a})
-\cal H\mit\right\}\;\;.$$
Thus the reduced Hamiltonian for the system of particles is $$H=\int dx \cal H\mit =-\frac{1}{\kappa}\int dx \triangle\Psi$$ where $\Psi$ is a function of $z_{a}$ and $p_{a}$ and is determined by solving the constraint equations (30) and (31). This expression is analogous to the reduced Hamiltonian in $(3+1)$ dimensional general relativity [@r6; @r7; @r8].
We pause to comment on the relationship between the integral forms (23) and (24) of the coordinate conditions and the differential forms (25) and (26). These two forms are equivalent only when one retains the appropriate boundary conditions for the integral operator $1/\triangle$. A proper treatment [@r9; @r10] entails the insertion of a regulator $\exp(-\alpha |x|)$ in the left-hand sides of (23) and (24), yielding $${g} := 2\,{\rm e}^{(\, - { \alpha}\,{|x|}\,)}\,(\, 1 -
{ \alpha}\,{|x|}\,) - 1$$ and $${ \Pi} := \frac{d}{dx}\left({\displaystyle \frac {{-\alpha}\,{t}\,{\rm e}
^{(\, - { \alpha}\,{|x|}\,)}{\mbox{sgn}}(x)}{(\, 2\,{\rm e}^{(\, - { \alpha}\,{|x|}\,
)} - 2\,{|x|}\,{ \alpha}\,{\rm e}^{(\, - { \alpha}\,{|x|}\,)} - 1\,)
^{1/2}\,{ \kappa}}}\right)$$ in place of (25) and (26) respectively. We must then insert these quantities in (12) – (19), taking the limit $\alpha \to 0$ at the end of the calculation (implicitly assuming $\alpha|x| < 1$). This turns out to be equivalent to inserting (25) and (26) in these equations. The action (33) is recovered by a similar limiting procedure.
This situation is analogous with that in (3+1) dimensions dimensions as discussed by ADM in [@r10] (cf. Eqs. (80), (81), (84) and (85) in section 6). We shall discuss these points further in section 6. Note that the coordinate conditions in integral form are needed only at the stage of transforming the generator and defining the Hamiltonian density. After fixing the formalism, we need only the differential forms to solve the constraints and to evaluate the Hamiltonian, a problem which is treated in the following sections.
Consistency of the canonical reduction
======================================
Following a procedure similar to the original ADM argument [@r7; @r9], we shall demonstrate here that the canonical equations of motion derived from the reduced Hamiltonian (34) are identical with equations (18) and (19), which are in turn equivalent to the original geodesic equation (6).
We start from the action integral (9). The variations with respect to $p_{a}$ and $z_{a}$ lead to the equations of motion $$\begin{aligned}
\dot{z}_{a}=-\int d^{2}y N_{\mu}(y)\frac{\delta R^{\mu}(y)}{\delta p_{a}(t)}
\\
\dot{p}_{a}=\int d^{2}y N_{\mu}(y)\frac{\delta R^{\mu}(y)}
{\delta z_{a}(t)}\;\;.\end{aligned}$$ These equations are identical with equations (18) and (19). Our purpose is to prove that these equations lead, when they are combined with the constraints (14) and (15), to the equations in the reduced formalism.
Defining $$\begin{aligned}
P_{0}(x)&\equiv&-\frac{1}{\kappa}\triangle\Psi(x)
\\
P_{1}(x)&\equiv&2\pi^{\prime}(x)\end{aligned}$$ and imposing the coordinate conditions, the action (\[e32\]) may be expressed as $$I=\int d^{2}x\left\{\sum_{a}p_{a}\dot{z}_{a}\delta(x-z_{a})-P_{0}
+N_{\mu}R^{\mu}\right\}\;\;.$$ where we have not yet imposed the constraints $R^{\mu}=0$. By solving the constraint equations, we get $$\begin{aligned}
P_{0}(x)&=&{\cal T}_{0 0}(x,z_{a},p_{a})\equiv \cal H\mit
\\
P_{1}(x)&=&{\cal T}_{0 1}(x,z_{a},p_{a})\;\;\end{aligned}$$ where ${\cal T}_{0 \mu}(x, z_{a}, p_{a})$ was described in the previous section.
Expanding $R^{\mu}$ in a functional Taylor series about the point $P_{\mu}={\cal T}_{0 \mu}$ gives $$R^{\mu}(x)=\int d^{2}y\left[P_{\nu}(y)-{\cal T}_{0 \nu}
(y, z_{a}, p_{a})\right]
\left[\frac{\delta R^{\mu}(x)}{\delta P_{\nu}(y)}\right]_{P=\cal T\mit}
+\cdot\cdot\cdot\;\;.$$ where the coordinate conditions have already been imposed. Substituting this expansion into the right hand side of (39) and taking the variation of the action $I$ with respect to $P^{\nu}$, yields $$-\delta_{\nu 0}+\int d^{2}y N_{\mu}(y)\left\{\left[\frac{\delta R^{\mu}(y)}
{\delta P_{\nu}(x)}\right]_{P=\cal T\mit}+\cdot\cdot\cdot\right\}
=0$$ where the terms represented by $\cdot\cdot\cdot$ contain $[P_{\nu}-{\cal T}_{0 \nu}]$ as a factor.
Now we need the relation which is valid after imposing the constraints. Then, by requiring $R^{\mu}=0$, we have $$-\delta_{\nu 0}+\int d^{2}y N_{\mu}(y)\left[\frac{\delta R^{\mu}(y)}
{\delta P_{\nu}(x)}\right]_{P=\cal T}=0\;\;.$$ Insertion of (42) into (35) leads to $$\begin{aligned}
\dot{z}_{a}&=&-\int d^{2}y N_{\mu}(y)\int d^{2}x\left\{
-\frac{\partial {\cal T}_{0 \nu}}
{\partial p_{a}(t)}\left[\frac{\delta R^{\mu}(y)}{\delta P_{\nu}(x)}\right]
_{P=\cal T}\right.
\nonumber \\
&&\makebox[5em]{}\left.+(P_{\nu}(x)-{\cal T}_{0 \nu}(x,z_{a},p_{a}))
\frac{\partial}{\partial p_{a}(x)}\left[\frac{\delta R^{\mu}(y)}{\delta P_{\nu}
(x)}\right]_{P=\cal T\mit}+\cdot\cdot\cdot\right\}\end{aligned}$$ Imposing the constraint equations $R^{\mu}=0$ yields for (45) $$\dot{z}_{a}
=\int d^{2}x\frac{\partial {\cal T}_{0 \nu}}{\partial p_{a}(t)}
\int d^{2}y N_{\mu}(y)\left[\frac{\delta R^{\mu}(y)}{\delta P_{\nu}(x)}
\right]_{P=\cal T}$$ As a consequence of the relation (44), (46) becomes $$\begin{aligned}
\dot{z}_{a}&=&\int d^{2}x\frac{\partial {\cal T}_{0 0}}
{\partial p_{a}}
\nonumber \\
&=&\frac{\partial H}{\partial p_{a}}\end{aligned}$$
In the same way we can show that the equations of motion (36) is identical with $$\dot{p}_{a}=-\frac{\partial H}{\partial z_{a}}$$ Thus the consistency of the reduced canonical formalism is proved.
Post-linear approximation
=========================
For a direct calculation of the reduced Hamiltonian (34), we have to solve the constraint equations (30) and (31). However, it seems quite a difficult task to get an exact solution except in the case of a single static source. So we need some approximation method.
In this section we apply an iterative method successively to the integrand of the expression (34) with the use of the equations (30) and (31). This provides us with an expansion of (34) in powers of $\kappa$, which we refer to as the post-linear approximation.
First, substituting $\triangle\Psi$ given by the equation (30) into the integrand and performing a partial integration, we have $$\begin{aligned}
H&=&\int dx\left\{-\frac{1}{4\kappa}\left(\Psi^{\prime}\right)^{2}
+\kappa\left(\chi^{\prime}\right)^{2}+\sum_{a}\sqrt{p^{2}_{a}+m^{2}_{a}}
\;\delta(x-z_{a})\right\}
\nonumber \\
&=&\sum_{a}\sqrt{p^{2}_{a}+m^{2}_{a}}+\int dx\left\{
\frac{1}{4\kappa}\Psi\triangle\Psi-\kappa\chi\triangle\chi\right\}+S_{1}\end{aligned}$$ where $\chi$ is defined by $\chi^{\prime}\equiv\pi$ and $$S_{1}=\left[-\frac{1}{4\kappa}\Psi\Psi^{\prime}+\kappa\chi\chi^{\prime}
\right]^{\infty}_{-\infty}\;\;.$$ Noting that $\Psi$ is of order $\kappa$, we substitute (30) for $\triangle\Psi$ and (31) for $\triangle\chi$ into the integrand of the second term on the right hand side of (49). Iterating twice in terms of $\kappa$ yields the expression $$\begin{aligned}
H&=&\sum_{a}\sqrt{p^{2}_{a}+m^{2}_{a}}\left\{1-\frac{1}{4}\Psi(z_{a})
+\frac{1}{32}\Psi(z_{a})\Psi(z_{a})\right\}
\nonumber \\
&&\makebox[2em]{}+\frac{\kappa}{2}\sum_{a}p_{a}\chi(z_{a})
\left\{1-\frac{1}{4}\Psi(z_{a})\right\}+\frac{\kappa^{2}}{8}\sum_{a}
\sqrt{p^{2}_{a}+m^{2}_{a}}\;\chi(z_{a})\chi(z_{a})+S_{2}
\nonumber \\
&&\makebox[2em]{}+\int dx\left\{\frac{\kappa^{3}}{8}\chi^{2}
\left(\chi^{\prime}\right)^{2}
-\frac{\kappa}{32}\chi^{2}\left(\Psi^{\prime}\right)^{2}
+\frac{\kappa}{32}\Psi^{2}\left(\chi^{\prime}\right)^{2}
-\frac{1}{128\kappa}\Psi^{2}\left(\Psi^{\prime}\right)^{2}\right\}
\label{e51}\end{aligned}$$ where $$\label{e52}
S_{2}=\left[-\frac{1}{4\kappa}\Psi\Psi^{\prime}+\kappa\chi\chi^{\prime}
-\frac{\kappa}{8}\left\{\Psi\left(\chi^{2}\right)^{\prime}
-\Psi^{\prime}\chi^{2}\right\}+\frac{1}{32\kappa}\Psi^{2}\Psi^{\prime}
\right]^{\infty}_{-\infty}\;\;.$$ The last term in (\[e51\]) contributes at order $\kappa^{3}$ . This iteration can be continued successively and corresponds to a series expansion in $\kappa$.
In contrast to the $(3+1)$–dimensional situation, we encounter a subtle problem when we try to extract an explicit form of the Hamiltonian under this iteration scheme. In $(1+1)$ dimensions the surface terms such as $S_{1}$ and $S_{2}$ arising in the process of the calculation do not necessarily vanish and make the Hamiltonian indefinite. This is because the dimensionless potential $Gmr/c^{2}$ becomes infinite at spatial infinity, in contrast to $(3+1)$ dimensions where the corresponding quantity $Gm/rc^{2}$ vanishes at spatial infinity thereby assuring that the associated surface terms vanish. It is therefore an important task to find a solution of $\Psi$ and $\chi$ which makes the surface term vanish at a given order in $\kappa$. This is related to the choice of boundary condition.
From the expressions of $S_{1}$ and $S_{2}$, we can infer a simple choice of the boundary condition as follows.
For $f(x)\equiv\Psi^{2}-4\kappa^{2}\chi^{2}$, $$\makebox[2em]{}f(x)=0 \makebox[1em]{}\mbox{and}\makebox[1em]{}
f^{\prime}(x)=0 \makebox[2em]{}\mbox{in a region}\makebox[1em]{}
\mid x\mid>>\mid z_{a}\mid \makebox[1em]{} \mbox{for all}
\makebox[0.5em]{} a.$$
It is easily checked that under this boundary condition, the surface terms $S_{1}$ and $S_{2}$ exactly vanish, since they are proportional to $f^\prime$ and $(4-\Psi)f^\prime + f\Psi^\prime$ respectively.
Now let us try to obtain the Hamiltonian up to $\kappa^{2}$ (the 2nd-post-linear approximation). First we expand $\Psi$ and $\chi$ in a power series in $\kappa$ $$\begin{aligned}
\Psi&=&\kappa\Psi^{(1)}+\kappa^{2}\Psi^{(2)}+\cdot\cdot\cdot
\\
\chi&=&\chi^{(0)}+\kappa\chi^{(1)}+\cdot\cdot\cdot\;\;. \end{aligned}$$ Substituting this expansion into the equations (30) and (31), we get $$\begin{aligned}
\triangle\Psi^{(1)}&=&-\sum_{a}\sqrt{p^{2}_{a}+m^{2}_{a}}\;\delta(x-z_{a})
\\
\triangle\Psi^{(2)}&=&-\left(\chi^{(0)\prime}\right)^{2}
+\frac{1}{4}\left(\Psi^{(1)\prime}\right)^{2}
\\
\triangle\chi^{(0)}&=&-\frac{1}{2}\sum_{a}p_{a}\delta(x-z_{a})
\\
\triangle\chi^{(1)}&=&0\;\;.\end{aligned}$$
The solutions which satisfy the boundary condition (53) are $$\begin{aligned}
\Psi^{(1)}&=&-\frac{1}{2}\sum_{a}\left\{\sqrt{p^{2}_{a}+m^{2}_{a}}\;r_{a}
+\epsilon\;p_{a}(x-z_{a})\right\}
\\
\chi^{(0)}&=&-\frac{1}{4}\sum_{a}\left\{p_{a}r_{a}
+\epsilon\;\sqrt{p^{2}_{a}+m^{2}_{a}}\;(x-z_{a})\right\}\end{aligned}$$
$$\begin{aligned}
\Psi^{(2)}&=&-\frac{1}{2}\left(\frac{1}{4}\right)^{2}\left\{\left[
\sum_{a}p_{a}r_{a}+\epsilon\;\sum_{a}\sqrt{p^{2}_{a}+m^{2}_{a}}\;
(x-z_{a})\right]^{2}
-\left[\sum_{a}\sqrt{p^{2}_{a}+m^{2}_{a}}\;r_{a}
+\epsilon\;\sum_{a}p_{a}(x-z_{a})\right]^{2}\right\}
\nonumber \\
&&-\left(\frac{1}{4}\right)^{2}\left\{\sum_{a}\sqrt{p^{2}_{a}+m^{2}_{a}}
\;r_{a}
\left[\sum_{b}\sqrt{p^{2}_{b}+m^{2}_{b}}\;r_{ab}
+\epsilon\;\sum_{b}p_{b}(z_{a}-z_{b})\right]
\right.
\nonumber \\
&&\makebox[4em]{}\left.-\sum_{a}p_{a}r_{a}\left[\sum_{b}p_{b}r_{ab}
+\epsilon\;\sum_{b}\sqrt{p^{2}_{b}+m^{2}_{b}}\;(z_{a}-z_{b})\right]\right\}
\\
\chi^{(1)}&=&-\frac{\epsilon}{2}\left(\frac{1}{4}\right)^{2}\left\{
\sum_{a}\sqrt{p^{2}_{a}+m^{2}_{a}}\;(x-z_{a})
\left[\sum_{b}\sqrt{p^{2}_{b}+m^{2}_{b}}\;r_{ab}
+\epsilon\;\sum_{b}p_{b}(z_{a}-z_{b})\right]\right.
\nonumber \\
&&\makebox[4em]{}\left.-\sum_{a}p_{a}(x-z_{a})\left[\sum_{b}p_{b}r_{ab}
+\epsilon\;\sum_{b}\sqrt{p^{2}_{b}+m^{2}_{b}}\;(z_{a}-z_{b})\right]
\right\}\;\;.\end{aligned}$$
where $r_a \equiv |x-z_a|$.
In these solutions we introduced a constant of integration $\epsilon$, satifying $\epsilon^{2}=1$. We have two types of solutions corresponding to $\epsilon=1$ and $\epsilon=-1$, which are related to each other under time reversal. This $\epsilon$ factor guarantees the invariance of the whole theory under the time reversal.
The boundary condition is checked as follows. Up to order $\kappa^{3}$ we have $$\begin{aligned}
\lefteqn{\Psi^{2}-4\kappa^{2}\chi^{2}}
\nonumber \\
&=&\frac{\kappa^{2}}{4}\sum_{a}\sum_{b}
\left\{\left[\sqrt{p^{2}_{a}+m^{2}_{a}}\sqrt{p^{2}_{b}+m^{2}_{b}}-p_{a}p_{b}
\right]\left(1-\frac{\kappa}{8}\sum_{c}\left[\sqrt{p^{2}_{c}+m^{2}_{c}}\;r_{c}
+\epsilon\;p_{c}(x-z_{c})\right]\right)\right.
\nonumber \\
&&\makebox[4em]{}\left.+\frac{\kappa}{4}\sqrt{p^{2}_{a}+m^{2}_{a}}
\sum_{c}\left[\sqrt{p^{2}_{b}+m^{2}_{b}}\;A_{bc}-p_{b}B_{bc}\right]\right\}
[r_{a}r_{b}-(x-z_{a})(x-z_{b})]
\nonumber \\
&&+\epsilon\;\frac{\kappa^{2}}{2}\sum_{a}\sum_{b}
\left\{\sqrt{p^{2}_{a}+m^{2}_{a}}\;p_{b}
\left(1-\frac{\kappa}{8}\sum_{c}\left[\sqrt{p^{2}_{c}+m^{2}_{c}}\;r_{c}
+\epsilon\;p_{c}(x-z_{c})\right]\right)\right.
\nonumber \\
&&\makebox[4em]{}\left.-\frac{\kappa}{8}p_{a}\sum_{c}
\left[\sqrt{p^{2}_{b}+m^{2}_{b}}\;A_{bc}-p_{b}B_{bc}\right]\right\}
[r_{a}(x-z_{b})-(x-z_{a})r_{b}]
\nonumber\end{aligned}$$ where $$\begin{aligned}
A_{bc}&=&\sqrt{p^{2}_{c}+m^{2}_{c}}\;r_{bc}+\epsilon\;p_{c}(z_{b}-z_{c})
\nonumber \\
B_{bc}&=&p_{c}r_{bc}+\epsilon\;\sqrt{p^{2}_{c}+m^{2}_{c}}(z_{b}-z_{c})\;\;.
\nonumber \end{aligned}$$ Since both $[r_{a}r_{b}-(x-z_{a})(x-z_{b})]$ and $[r_{a}(x-z_{b})-(x-z_{a})r_{b}]$ vanish in a region $\mid x\mid >> \mid z_{a}\mid, \mid z_{b}\mid$, the boundary condition is satisfied.
The 2nd-post-linear Hamiltonian is therefore unambiguously determined to be $$\begin{aligned}
H&=&\sum_{a}\sqrt{p^{2}_{a}+m^{2}_{a}}\left\{1-\frac{1}{4}\Psi(z_{a})
+\frac{1}{32}\Psi(z_{a})\Psi(z_{a})\right\}
\nonumber \\
&&\makebox[2em]{}+\frac{\kappa}{2}\sum_{a}p_{a}\chi(z_{a})
\left\{1-\frac{1}{4}\Psi(z_{a})\right\}
+\frac{\kappa^{2}}{8}\sum_{a}\sqrt{p^{2}_{a}+m^{2}_{a}}\;\chi(z_{a})
\chi(z_{a})
\nonumber \\
&=&\sum_{a}\sqrt{p^{2}_{a}+m^{2}_{a}}
+\frac{\kappa}{8}\sum_{a}\sum_{b}\left(\sqrt{p^{2}_{a}+m^{2}_{a}}
\sqrt{p^{2}_{b}+m^{2}_{b}}-p_{a}p_{b}\right)r_{ab}
\nonumber \\
&&+\frac{\epsilon\kappa}{8}\sum_{a}\sum_{b}
\left(\sqrt{p^{2}_{a}+m^{2}_{a}}\;p_{b}
-p_{a}\sqrt{p^{2}_{b}+m^{2}_{b}}\right)(z_{a}-z_{b})
\nonumber \\
&&+\frac{1}{4}\left(\frac{\kappa}{4}\right)^{2}
\left\{\sum_{a}\sqrt{p^{2}_{a}+m^{2}_{a}}
\left[\sum_{b}p_{b}r_{ab}
+\epsilon\;\sum_{b}\sqrt{p^{2}_{b}+m^{2}_{b}}(z_{a}-z_{b})\right]^{2}\right.
\nonumber \\
&&\left.-\sum_{a}p_{a}\left[\sum_{b}p_{b}r_{ab}
+\epsilon\;\sum_{b}\sqrt{p^{2}_{b}+m^{2}_{b}}(z_{a}-z_{b})\right]
\left[\sum_{c}\sqrt{p^{2}_{c}+m^{2}_{c}}\;r_{ac}
+\epsilon\;\sum_{c}p_{c}(z_{a}-z_{c})\right]\right.
\nonumber \\
&&\left.+\sum_{a}\sum_{b}\left[\sqrt{p^{2}_{a}+m^{2}_{a}}\sqrt{p^{2}_{b}
+m^{2}_{b}}\;r_{ab}
-\epsilon\;p_{a}\sqrt{p^{2}_{b}+m^{2}_{b}}(z_{a}-z_{b})\right]\right.
\nonumber \\
&&\makebox[8em]{}\left.\times\left[\sum_{c}\sqrt{p^{2}_{c}+m^{2}_{c}}\;r_{bc}
+\epsilon\;\sum_{c}p_{c}(z_{b}-z_{c})
\right]\right.
\nonumber \\
&&\left.-\sum_{a}\sum_{b}\left[\sqrt{p^{2}_{a}+m^{2}_{a}}\;p_{b}r_{ab}
-\epsilon\;p_{a}p_{b}(z_{a}-z_{b})\right]\left[\sum_{c}p_{c}r_{bc}
+\epsilon\;\sum_{c}\sqrt{p^{2}_{c}+m^{2}_{c}}(z_{b}-z_{c})\right]\right\}\;\;.
\nonumber \\\end{aligned}$$
As shown in the above development, the $\kappa$-expansion is the successive approximation in the background of Minkowskian space-time and the terms in each order of $\kappa$ preserve relativistic forms. This approximation is therfore appropriate for describing relativistic fast-motion of the particles and can be carried out to any desired order of $\kappa$.
Post-Newtonian Hamiltonian and the redefinition of canonical variables
======================================================================
To compare $R=T$ theory with Newtonian gravity, we need an approximation method applicable to a slow motion in a weak field. It is provided by the so-called $c^{-1}$ expansion. All terms appropriate to the post-Newtonian approximation (up to the order of $c^{-4}$) are included in the post-linear Hamiltonian (64). Noting that both $p_a^2/m_a^2$ and $\sqrt{\kappa}$ are of the order of $c^{-2}$, we find from (64) $$\begin{aligned}
H&=&\sum_{a}m_{a}+\sum_{a}\frac{p^{2}_{a}}{2m_{a}}+\frac{\kappa}{8}\sum_{a}
\sum_{b}m_{a}m_{b}r_{ab}
+\frac{\epsilon\kappa}{8}\sum_{a}\sum_{b}(m_{a}p_{b}-m_{b}p_{a})(z_{a}-z_{b})
\nonumber \\
&&-\sum_{a}\frac{p^{4}_{a}}{8m^{3}_{a}}+\frac{\kappa}{8}\sum_{a}\sum_{b}
m_{a}\frac{p^{2}_{b}}{m_{b}}r_{ab}-\frac{\kappa}{8}\sum_{a}\sum_{b}p_{a}p_{b}
r_{ab}
\nonumber \\
&&+\frac{1}{4}\left(\frac{\kappa}{4}\right)^{2}\sum_{a}\sum_{b}\sum_{c}
m_{a}m_{b}m_{c}r_{ab}r_{ac}
+\frac{1}{4}\left(\frac{\kappa}{4}\right)^{2}\sum_{a}\sum_{b}\sum_{c}
m_{a}m_{b}m_{c}(z_{a}-z_{b})(z_{a}-z_{c})\;\;.
\nonumber \\\end{aligned}$$ The corresponding solutions of $\Psi$ and $\chi$ are $$\begin{aligned}
\Psi&=&-\frac{\kappa}{2}\sum_{a}m_{a}r_{a}
-\frac{\epsilon\kappa}{2}\sum_{a}p_{a}(x-z_{a})
-\frac{\kappa}{4}\sum_{a}\frac{p^{2}_{a}}{m_{a}}r_{a}
+\frac{1}{2}\left(\frac{\kappa}{4}\right)^{2}\left(\sum_{a}m_{a}r_{a}\right)
^{2}
\nonumber \\
&&-\frac{1}{2}\left(\frac{\kappa}{4}\right)^{2}\left(\sum_{a}m_{a}(x-z_{a})
\right)^{2}-\left(\frac{\kappa}{4}\right)^{2}\sum_{a}\sum_{b}m_{a}m_{b}
r_{a}r_{ab}
\\
\chi&=&-\frac{\epsilon}{4}\sum_{a}m_{a}(x-z_{a})
-\frac{1}{4}\sum_{a}p_{a}r_{a}\;\;.\end{aligned}$$
The canonical equations of motion yield $$\begin{aligned}
\dot{z}_{a}&=&\frac{\partial H}{\partial p_{a}}
\nonumber \\
&=&\frac{p_{a}}{m_{a}}-\frac{\epsilon\kappa}{4}\sum_{b}m_{b}(z_{a}-z_{b})
-\frac{p^{3}_{a}}{2m^{3}_{a}}+\frac{\kappa}{4}\sum_{b}m_{b}\frac{p_{a}}{m_{a}}
r_{ab}-\frac{\kappa}{4}\sum_{b}p_{b}r_{ab}
\\
\dot{p}_{a}&=&-\frac{\partial H}{\partial z_{a}}
\nonumber \\
&=&-\frac{\kappa}{4}\sum_{b}m_{a}m_{b}\frac{\partial r_{ab}}{\partial z_{a}}
-\frac{\epsilon\kappa}{4}\sum_{b}(m_{a}p_{b}-m_{b}p_{a})-\frac{\kappa}{8}
\sum_{b}\left(m_{a}\frac{p^{2}_{b}}{m_{b}}+m_{b}\frac{p^{2}_{a}}{m_{a}}\right)
\frac{\partial r_{ab}}{\partial z_{a}}
\nonumber \\
&&+\frac{\kappa}{4}\sum_{b}p_{a}p_{b}\frac{\partial r_{ab}}{\partial z_{a}}
-\frac{1}{2}\left(\frac{\kappa}{4}\right)^{2}\sum_{b}\sum_{c}m_{a}m_{b}m_{c}
\left(\frac{\partial r_{ab}}{\partial z_{a}}r_{bc}+\frac{\partial r_{ab}}
{\partial z_{a}}r_{ac}\right)
\nonumber \\
&&-\frac{1}{2}\left(\frac{\kappa}{4}\right)^{2}\sum_{b}\sum_{c}m_{a}m_{b}m_{c}
(z_{a}-z_{b})\;\;.\end{aligned}$$ By eliminating $p_{a}$ in the equations (68) and (69), we get the equations of motion to second order $$\begin{aligned}
\lefteqn{m_{a}\ddot{z}_{a}}
\nonumber \\
&=&-\frac{\kappa}{4}\sum_{b}m_{a}m_{b}\frac{\partial r_{ab}}
{\partial z_{a}}+\frac{\kappa}{8}\sum_{b}m_{a}m_{b}\left\{4(\dot{z}_{a})^{2}
-2\dot{z}_{a}\dot{z}_{b}+(\dot{z}_{b})^{2}\right\}\frac{\partial r_{ab}}
{\partial z_{a}}
\nonumber \\
&&-\left(\frac{\kappa}{4}\right)^{2}\sum_{b}\sum_{c}m_{a}m_{b}m_{c}
\left(\frac{\partial r_{ac}}{\partial z_{a}}-\frac{\partial r_{bc}}
{\partial z_{b}}\right)r_{ab}-\frac{1}{2}\left(\frac{\kappa}{4}\right)^{2}
\sum_{b}\sum_{c}m_{a}m_{b}m_{c}\left(\frac{\partial r_{ab}}{\partial z_{a}}
r_{bc}+\frac{\partial r_{ab}}{\partial z_{a}}r_{ac}\right)
\nonumber \\
&&+\frac{1}{2}\left(\frac{\kappa}{4}\right)^{2}\sum_{b}\sum_{c}m_{a}m_{b}m_{c}
(z_{a}-z_{b})\end{aligned}$$
We pause to show that these canonical equations of motion are equivalent to the geodesic equation (6). To evaluate the geodesic equation, we have to know all components of the metric tensor, of which $g_{11}=\gamma$ is fixed and $N_{0}$ and $N_{1}$ components are determined by combining the time derivatives of the coordinate conditions (25) and (26) with the equations (13) and (16). These equations are $$\begin{aligned}
\kappa\chi^{\prime}N_{0}+N^{\prime}_{1}=0
\\
\partial_{1}\left(\frac{1}{2}N_{0}\Psi^{\prime}+N^{\prime}_{0}\right)=0\;\;.\end{aligned}$$ The solutions in the post-Newtonian approximation are $$\begin{aligned}
N_{0}&=&1+\frac{\kappa}{4}\sum_{a}m_{a}r_{a}
+\frac{\epsilon\kappa}{4}\sum_{a}p_{a}(x-z_{a})
+\frac{\kappa}{8}\sum_{a}\frac{p^{2}_{a}}{m_{a}}r_{a}
+\frac{1}{4}\left(\frac{\kappa}{4}\right)^{2}
\left(\sum_{a}m_{a}r_{a}\right)^{2}
\nonumber \\
&&+\frac{1}{4}\left(\frac{\kappa}{4}\right)^{2}\left(\sum_{a}m_{a}
(x-z_{a})\right)^{2}
+\frac{1}{2}\left(\frac{\kappa}{4}\right)^{2}\sum_{a}\sum_{b}m_{a}m_{b}
r_{a}r_{ab}
\\
N_{1}&=&\frac{\epsilon\kappa}{4}\sum_{a}m_{a}(x-z_{a})+\frac{\kappa}{4}
\sum_{a}p_{a}r_{a}\;\;.\end{aligned}$$ The boundary conditions in the previous section are incorporated in these solutions through (12) and (17), which serve as consistency equations.
Since we have proved in section 2 the equivalence of the geodesic equation (6) to a set of equations (18) and (19), we have only to evaluate the latter equations. Under the coordinate conditions (25) and (26), the equations (18) and (19) become $$\begin{aligned}
\dot{z}_{a}&=&\frac{p_{a}}{\sqrt{p^{2}_{a}+m^{2}_{a}}}
N_{0}(z_{a})-N_{1}(z_{a})
\\
\dot{p}_{a}&=&-\sqrt{p^{2}_{a}+m^{2}_{a}}\frac{\partial N_{0}}{\partial z_{a}}
+p_{a}\frac{\partial N_{1}}{\partial z_{a}}\end{aligned}$$ which lead to the equations (68) and (69), after the insertion of the solutions (73) and (74) and expanding in powers of $c^{-1}$.
The Hamiltonian (65) contains a term proportional to $c^{-1}$, namely, $$\frac{\epsilon\kappa}{8}\sum_{a}\sum_{b}(m_{a}p_{b}-m_{b}p_{a})(z_{a}-z_{b})
\;\;.$$ The appearance of such a term may seem unnatural, because in $(3+1)$ dimensions terms in odd powers of $c^{-1}$ , for example $c^{-5}$ terms, are related with gravitational radiation; yet there exists no graviton degree of freedom in (1+1)-dimensions. However, by a suitable redefinition of canonical variables $$\begin{aligned}
z_{a}&\longrightarrow& \tilde{z}_{a}=z_{a}
\nonumber \\
\\
p_{a}&\longrightarrow& \tilde{p}_{a}=p_{a}-\frac{\epsilon\kappa}{4}\sum_{b}
m_{a}m_{b}(z_{a}-z_{b})\;\;
\nonumber \\\end{aligned}$$ this term can be eliminated. Under this redefinition the ”Poisson brackets” among the canonical variables are kept unchanged. The Hamiltonian expressed in terms of the redefined canonical variables is then $$\begin{aligned}
H&=&\sum_{a}m_{a}+\sum_{a}\frac{\tilde{p}^{2}_{a}}{2m_{a}}
+\frac{\kappa}{8}\sum_{a}\sum_{b}m_{a}m_{b}\tilde{r}_{ab}
-\sum_{a}\frac{\tilde{p}^{4}_{a}}{8m^{3}_{a}}
+\frac{\kappa}{8}\sum_{a}\sum_{b}m_{a}\frac{\tilde{p}^{2}_{b}}{m_{b}}
\tilde{r}_{ab}
\nonumber \\
&&-\frac{\kappa}{8}\sum_{a}\sum_{b}\tilde{p}_{a}\tilde{p}_{b}\tilde{r}_{ab}
+\frac{1}{4}\left(\frac{\kappa}{4}\right)^{2}\sum_{a}\sum_{b}\sum_{c}
m_{a}m_{b}m_{c}\left\{\tilde{r}_{ab}\tilde{r}_{ac}
-(\tilde{z}_{a}-\tilde{z}_{b})(\tilde{z}_{a}-\tilde{z}_{c})\right\}\;\;.\end{aligned}$$ It is straightforward to show that the equations of motion derived from the Hamiltonian (79) are identical with those derived from the original variables.
Discussion
==========
We have performed the canonical reduction of a set of point particles coupled to a gravitational field in $(1+1)$ dimensions in the context of the $R=T$ theory, and have shown how to carry out series expansions of such equations in powers of $\kappa$ (the post-linear approximation) and powers of $c^{-1}$ (the post-Newtonian approximation). We have obtained explicit expressions for the Hamiltonian to order $\kappa^2$ and $c^{-4}$ respectively, along with an expansion of the geodesic equation to order $c^{-4}$.
We have considered in this reduction the dynamical role of the auxiliary scalar field $\Psi$. Despite the decoupling of $\Psi$ from the classical gravity/matter field equations, the Hamiltonian of the system is defined in terms of $\Psi$. One of the essential points in this analysis is the choice of the coordinate conditions, which we chose in the forms of (23) and (24) through the transformation of the generator. We note that our coordinate conditions are quite analogous to the ADM coordinate conditions in (3+1)-dimensional gravity [@r6]. These coordinate conditions (frequently used for particle dynamics [@r7; @r8]) are $$\begin{aligned}
x^{i}&=&g_{i}-\frac{1}{4\triangle}g^{T}_{\;,i}
\\
t&=&-\frac{1}{2\triangle}\pi^{ii}\end{aligned}$$ where $$\begin{aligned}
g_{i}&=&\frac{1}{\triangle}\left(g_{ij,j}-\frac{1}{2\triangle}
g_{jk,jki}\right)
\\
g^{T}&=&g_{ii}-\frac{1}{\triangle}g_{ij,ij}\end{aligned}$$ and $\pi^{ij}$ is a conjugate field to $g_{ij}$, $\triangle$ being the Laplacian operator in three dimensions.
The differential forms of the conditions (80) and (81) are $$\begin{aligned}
\triangle g_{ij,j}-\frac{1}{4}g_{jk,ijk}&-&\frac{1}{4}\triangle g_{jj,i}=0
\\
\pi^{ii}&=&0\;\;.\end{aligned}$$ The first condition (84) relates space-derivatives of spatial components $g_{ij}$ of the metric tensor. In $(1+1)$ dimensions there is only one spatial component $\gamma=g_{11}$; from (23) the condition analogous to (84) is $\partial_{1}\gamma=0$. Therefore we chose the condition (25). Since $\pi^{ij}$ in the second condition (85) is the conjugate momenta to $g_{ij}$, the corresponding condition in (1+1)-dimension might at first appear to be $\pi=0$. However, the form of the generator (21) indicates that when the variation $\delta h$ is related to one of the coordinate conditions, the variation of the conjugate $\pi$ can no longer be exploited for the other condition. The candidates would be $\delta\Pi$ or $\delta\Psi$. Judging from the fact that $\pi^{ij}$ is essentially a time-derivative of $g_{ij}$ and $\Pi$ is also a time-derivative of $\gamma$, we have no choice to take other than $\delta\Pi$.
We turn now to a discussion of the consistency of the full set of equations (12) - (19). As discussed in previous sections, to determine the Hamiltonian for a system of particles, we have only to solve the constraint equations (14) and (15) for $\Psi$ and $\pi$. All informations on the dynamics of particles are included in the constraint equations. Equations (13) and (16) are used to determine $N_{0}$ and $N_{1}$, and equations (18) and (19) lead to the equations of motion. The remaining two equations (12) and (17) are not necessary in describing the dynamics of the particles. However they are necessary in providing a full description of the gravitational field.
To check the consistency of the whole formalism, consider substituting the post- linear solutions of $\Psi, \pi, N_{0}$ and $N_{1}$ into (12) and (17). After some calculation, it is straightforward to see that (12) is consistently satisfied but (17) is not. The reason is as follows. In the general solution to the constraint equation (14) (or (30)) $\Psi$ may contain in general an $x$-independent function $f(t)$. Since all other equations except (17) contain only spatial derivatives of $\Psi$, $f(t)$ does not contribute to either the Hamiltonian or to the equations of motion. The function $f(t)$ is necessary only for the consistency of (17). The boundary condition and the consideration on the surface term in the section 3 can be applied with $\Psi$ replaced by $(\Psi - f(t))$, the latter function having no explicit dependence on $t$. After lengthy and complicated calculation of the equation (17), we can determine the explicit form of $f(t)$ to the order $\kappa^{2}$.
To illustrate this situation, let us consider a single static source at the origin. In this case the fundamental equations are $$\begin{aligned}
-\frac{1}{\kappa}\triangle\Psi&=&\kappa\pi^{2}-\frac{1}{4\kappa}
\left(\Psi^{\prime}\right)^{2}+M\delta(x)
\\
\pi^{\prime}&=&0\end{aligned}$$ $$\begin{aligned}
\kappa\pi N_{0}+N^{\prime}_{1}&=&0
\\
\partial_{1}\left(\frac{1}{2}N_{0}\Psi^{\prime}+N^{\prime}_{0}\right)&=&0\end{aligned}$$ $$\begin{aligned}
&&\dot{\pi}+N_{0}\left[\frac{3\kappa}{2}\pi^{2}+\frac{1}{8\kappa}
\left(\Psi^{\prime}\right)^{2}\right]+N_{1}\pi^{\prime}
+\frac{1}{2\kappa}N^{\prime}_{0}\Psi^{\prime}+N^{\prime}_{1}\pi=0
\\
&&\makebox[5em]{}\dot{\Psi}+2\kappa N_{0}\pi-N_{1}\Psi^{\prime}=0\;\;.\end{aligned}$$ Solutions to the constraint equations (86) and (87), which satisfy the same boundary condition in the section 3, are $$\begin{aligned}
\Psi&=&-\frac{\kappa M}{2}\mid x\mid + f(t)
\\
\pi&=&\chi^{\prime}=-\frac{\epsilon}{4}M \makebox[2em]{}
( \chi=-\frac{\epsilon M}{4}x )\;\;.\end{aligned}$$ Here we included $f(t)$ in $\Psi$.
The solutions of $N_{0}$ and $N_{1}$ are $$\begin{aligned}
N_{0}&=&e^{\frac{\kappa M}{4}\mid x\mid}
\\
N_{1}&=&\epsilon\;\frac{x}{\mid x\mid}
\left(e^{\frac{\kappa M}{4}\mid x\mid}-1\right)
\;\;.\end{aligned}$$ Equation (90) is satisfied, whereas (91) leads to $$\dot{f}(t)=\frac{\epsilon\kappa M}{2}\;\;.$$ Then the solution of $\Psi$ becomes $$\Psi=-\frac{\kappa M}{2}\mid x\mid+\frac{\epsilon\kappa M}{2}t\;\;.$$ This shows that even for a static source the auxiliary field is not static.
The Hamiltonian is $$\begin{aligned}
H&=&-\frac{1}{\kappa}\int dx\triangle\Psi
\nonumber \\
&=&M\;\;.\end{aligned}$$ Since $$\begin{aligned}
g_{00}&=&-N^{2}_{0}+N^{2}_{1}
\nonumber \\
&=&1-2e^{\frac{\kappa M}{4}\mid x\mid}
\nonumber \\
&<&-1
\nonumber \end{aligned}$$ this solution has no event horizon.
The canonical formalism we have derived can be utilized in a number of ways. The most obvious of these is to obtain an explicit solution for the $N$-body problem in $(1+1)$ dimensions (the preceding discussion illustrates the solution when $N=1$). In the large–$N$ limit with fixed proper distance between the most widely separated particles one might expect to recover the fluid collapse problem studied in ref. [@r5]. More generally, one could study the $N$-body problem where gravity is coupled to additional matter fields. These all remain interesting subjects for future investigation.
**Appendix\
Relation between the equations of motion in the\
coordinate conditions $\gamma=1, \Pi=0$ and those in the\
conditions $g_{00}=-g^{-1}_{11}, g_{01}=0$**
One of the most popular coordinate conditions in $(1+1)$ dimensions is the one for which the metric tensor has the form $$g_{\mu\nu}=\left(
\begin{array}{cc}
-\alpha & 0 \\
0 & \alpha^{-1}
\end{array}
\right)\;\;.$$ The difference between this condition and the condition we adopted in this paper is that the former is independent of the auxiliary scalar field $\Psi$ but the latter is not. This $\Psi$-dependence of the coordinate condition played an important role in the canonical reduction and the general expression of the Hamiltonian.
In this appendix we shall compare the equations of motion in these two coordinate conditions. Expanding $\alpha$ and the energy-momentum tensor (5) in a power series of $c^{-1}$ we obtain $$\alpha=1+2\phi-g^{(4)}_{00}+\cdot\cdot\cdot$$ $$\begin{aligned}
T_{00}&=&\sum_{a}\frac{m_{a}}{\sqrt{1-\alpha^{-2}(\dot{z}_{a})^{2}}}
\alpha^{3/2}\delta(x-z_{a})
\nonumber \\
&=&\sum_{a}m_{a}\left\{1+3\phi+\frac{1}{2}(\dot{z}_{a})^{2}
+\cdot\cdot\cdot\right\}
\delta(x-z_{a})
\\
T_{01}&=&-\sum_{a}\frac{m_{a}\alpha^{-1/2}\dot{z}_{a}}{\sqrt{1-\alpha^{-2}
(\dot{z}_{a})^{2}}}\delta(x-z_{a})
\nonumber \\
&=&\sum_{a}m_{a}\left\{-\dot{z}_{a}+\cdot\cdot\cdot\right\}\delta(x-z_{a})
\\
T_{11}&=&\sum_{a}\frac{m_{a}\alpha^{-5/2}(\dot{z}_{a})^{2}}
{\sqrt{1-\alpha^{-2}(\dot{z}_{a})^{2}}}\delta(x-z_{a})
\nonumber \\
&=&\sum_{a}m_{a}(\dot{z}_{a})^{2}\left\{1+\cdot\cdot\cdot\right\}
\delta(x-z_{a})\end{aligned}$$ where the number in the upper parenthesis denotes the order of $c^{-1}$ and $\phi\equiv-\frac{1}{2}g^{(2)}_{00}$.
The post-Newtonian equation of motion of the reference [@r5] reads $$m_{1}\ddot{z}_{1}=-m_{1}\phi^{\prime}-m_{1}\psi^{\prime}
-4m_{1}\phi\phi^{\prime}+3m_{1}\dot{\phi}\dot{z}_{1}
+3m_{1}\phi^{\prime}(\dot{z}_{1})^{2}\;\;.$$ Here $\phi$ and $\psi$ satisfy the equations $$\begin{aligned}
\phi^{\prime\prime}&=&\frac{\kappa}{2}T^{(0)}_{00}
\nonumber \\
&=&\frac{\kappa}{2}\sum_{a}m_{a}\delta(x-z_{a})
\\
\psi^{\prime\prime}&=&-\ddot{\phi}-2(\phi^{\prime})^{2}+\frac{\kappa}{2}
\left(T^{00(2)}-T^{11(2)}\right)
\nonumber \\
&=&-\ddot{\phi}-2(\phi^{\prime})^{2}-\frac{\kappa}{2}\sum_{a}m_{a}\left\{
\frac{\kappa}{4}\sum_{b}m_{b}r_{ab}+\frac{1}{2}(\dot{z}_{a})^{2}\right\}
\delta(x-z_{a})\;\;.\end{aligned}$$ The solutions are $$\begin{aligned}
\phi&=&\frac{\kappa}{4}\sum_{a}m_{a}r_{a}
\\
\psi&=&\frac{\kappa}{8}\sum_{a}m_{a}\ddot{z}_{a}(x-z_{a})r_{a}
-\frac{\kappa}{4}\sum_{a}m_{a}(\dot{z}_{a})^{2}r_{a}
\nonumber \\
&&-\left(\frac{\kappa}{4}\right)^{2}\sum_{a}\sum_{b}m_{a}m_{b}r_{a}r_{b}
+2\left(\frac{\kappa}{4}\right)^{2}\sum_{a}\sum_{b}m_{a}m_{b}r_{ab}r_{a}
\nonumber \\
&&-\frac{\kappa^{2}}{16}\sum_{a}\sum_{b}m_{a}m_{b}r_{ab}r_{a}
-\frac{\kappa}{8}\sum_{a}m_{a}(\dot{z}_{a})^{2}r_{a}\;\;.\end{aligned}$$ The equation of motion (104) for a system of two particles becomes $$\begin{aligned}
m_{1}\ddot{z}_{1}&
=&-\frac{\kappa}{4}m_{1}m_{2}\frac{\partial r_{12}}{\partial z_{1}}
-\frac{\kappa^{2}}{8}m_{1}m_{2}(m_{1}+m_{2})(z_{1}-z_{2})
\nonumber \\
&&+\frac{\kappa}{8}m_{1}m_{2}\left\{6(\dot{z}_{1})^{2}-6\dot{z}_{1}\dot{z}_{2}
+3(\dot{z}_{2})^{2}\right\}\frac{\partial r_{12}}{\partial z_{1}}\;\;.\end{aligned}$$
To investigate the relation between this equation of motion and the equation obtained in our canonical formalism, we consider coordinate transformations $x^{\mu}\longrightarrow \underline{x}^{\mu}$, which connect two forms of the line element : $$\begin{aligned}
ds^{2}&=&-\alpha dt^{2}+\frac{1}{\alpha}dx^{2}
\\
&=&-(N^{2}_{0}-N^{2}_{1})d\underline{t}^{2}
+2N_{1}d\underline{t}d\underline{x}+d\underline{x}^{2}\;\;,\end{aligned}$$ The explicit form of the transformations is given by $$\begin{aligned}
t&=&\underline{t}-\frac{\epsilon\kappa}{8}\sum_{a}m_{a}(\underline{x}
-\underline{z}_{a})^{2}+\cdot\cdot\cdot
\\
x&=&\underline{x}+\frac{\kappa}{8}\sum_{a}m_{a}(\underline{x}
-\underline{z_{a}})
\mid\underline{x}-\underline{z_{a}}\mid+\cdot\cdot\cdot\;\;.\end{aligned}$$
For a system of two particles the above transformation (113) leads to $$\begin{aligned}
z_{1}&=&\underline{z}_{1}+\frac{\kappa}{8}m_{2}(\underline{z}_{1}
-\underline{z}_{2})
\mid\underline{z}_{1}-\underline{z}_{2}\mid+\cdot\cdot\cdot
\nonumber \\
z_{2}&=&\underline{z}_{2}-\frac{\kappa}{8}m_{1}(\underline{z}_{1}
-\underline{z}_{2})
\mid\underline{z}_{1}-\underline{z}_{2}\mid+\cdot\cdot\cdot
\nonumber \\
z_{1}-z_{2}&=&(\underline{z}_{1}-\underline{z}_{2})\left\{1+\frac{\kappa}{8}
(m_{1}+m_{2})\mid\underline{z}_{1}-\underline{z}_{2}\mid
+\cdot\cdot\cdot\right\}
\nonumber \\
r_{12}&=&\underline{r}_{12}\left\{1+\frac{\kappa}{8}(m_{1}+m_{2})
\underline{r}_{12}+\cdot\cdot\cdot\right\}\;\;.
\nonumber \end{aligned}$$ Since the second term of the right hand side of the transformation (112) contributes to $\underline{t}$ as a correction of the order $c^{-3}$, in the post-Newtonian approximation we can set $t=\underline{t}$. Then we have $$\begin{aligned}
\dot{z}_{1}&=&\dot{\underline{z}}_{1}+\frac{\kappa}{8}m_{2}\left\{
(\dot{\underline{z}}_{1}-\dot{\underline{z}}_{2})\mid\underline{z}_{1}
-\underline{z}_{2}\mid
+(\underline{z}_{1}-\underline{z}_{2})(\dot{\underline{z}}_{1}
-\dot{\underline{z}}_{2})
\frac{\partial\underline{r}_{12}}{\partial\underline{z}_{1}}\right\}
+\cdot\cdot\cdot
\nonumber \\
&=&\dot{\underline{z}}_{1}+\frac{\kappa}{4}m_{2}(\dot{\underline{z}}_{1}
-\dot{\underline{z}}_{2})\mid\underline{z}_{1}-\underline{z}_{2}\mid
+\cdot\cdot\cdot
\nonumber \\
\ddot{z}_{1}&=&\ddot{\underline{z}}_{1}+\frac{\kappa}{4}m_{2}\left\{
(\ddot{\underline{z}}_{1}-\ddot{\underline{z}}_{2})\mid\underline{z}_{1}
-\underline{z}_{2}\mid
+(\dot{\underline{z}}_{1}-\dot{\underline{z}}_{2})^{2}
\frac{\partial\underline{r}_{12}}
{\partial\underline{z}_{1}}\right\}+\cdot\cdot\cdot
\nonumber \end{aligned}$$ $$\begin{aligned}
\frac{\partial r_{12}}{\partial z_{1}}&=&\frac{\partial}{\partial z_{1}}
\left\{\underline{r}_{12}\left[1+\frac{\kappa}{8}(m_{1}+m_{2})
\underline{r}_{12}+\cdot\cdot\cdot\right]\right\}
\nonumber \\
&=&\left\{1-\frac{\kappa}{4}(m_{1}+m_{2})\underline{r}_{12}
+\cdot\cdot\cdot\right\}
\left\{\frac{\partial\underline{r}_{12}}{\partial\underline{z}_{1}}
+\frac{\kappa}{4}(m_{1}+m_{2})\frac{\partial\underline{r}_{12}}
{\partial\underline{z}_{1}}\underline{r}_{12}+\cdot\cdot\cdot\right\}
\nonumber \\
&=&\frac{\partial\underline{r}_{12}}{\partial\underline{z}_{1}}
+\cal{O}\mit(\kappa^{2})\;\;.
\nonumber\end{aligned}$$
By the use of these relations the equation of motion (109) is transformed to $$m_{1}\ddot{\underline{z}}_{1}=-\frac{\kappa}{4}m_{1}m_{2}
\frac{\partial\underline{r}_{12}}{\partial\underline{z}_{1}}+\frac{\kappa}{8}
\left\{4(\dot{\underline{z}}_{1})^{2}-2\dot{\underline{z}}_{1}
\dot{\underline{z}}_{2}
+(\dot{\underline{z}}_{2})^{2}\right\}\frac{\partial\underline{r}_{12}}
{\partial\underline{z}_{1}}-\frac{\kappa^{2}}{16}m_{1}m_{2}(m_{1}+m_{2})
(\underline{z}_{1}-\underline{z}_{2})
\\$$ This equation is identical with the equation of motion (70) we obtained in the canonical formalism.
Acknowledgements {#acknowledgements .unnumbered}
================
This work was supported in part by the Natural Sciences and Engineering Research Council of Canada.
[1]{}
Teitelboim C 1984 [*Quantum Theory of Gravity*]{} ed S Christensen (Bristol: Hilger) p 327
Jackiw R 1984 [*Quantum Theory of Gravity*]{} ed S Christensen (Bristol: Hilger) p 403; 1985 [*Nucl. Phys*]{} B **252343**
Marnelius R 1983 [*Nucl. Phys.*]{} **B 21114**
Torre C G 1989 [*Phys. Rev.*]{} **D 402588**
Mann R B, Shiekh A and Tarasov L 1990 [*Nucl. Phys.*]{} **B 341134**
Harvey J and Strominger A 1992 “Quantum Aspects of Black Holes,” [hep-th/9209055]{}
Banks T and O’ Loughlin M 1991 [*Nucl. Phys.*]{} [**B362** ]{} 649
Mann RB 1993 [*Phys. Rev.*]{} [**D47**]{} 4438
Louis-Martinez D and Kunstatter G 1995 [*Phys. Rev.*]{} [**D52**]{} 3494
Creighton JDE and Mann RB 1995 “Thermodynamics of Dilatonic Black Holes in N-dimensions” [gr-qc/9511012]{}
Chan SFJ and Mann RB 1995 [*Class. Quant. Grav.* ]{} [**12**]{} 351
Russo J, Susskind L and Thorlacius L 1993 [*Phys. Rev.*]{} [**D47**]{} 533
Mann R B 1991 [*Found. Phys. Lett.*]{} **4425**
Mann R B 1992 [*Gen. Rel. Grav.*]{} **24433**
Mann R B, Morsink S M, Sikkema A E and Steele T G 1991 [*Phys. Rev.*]{} **D 433948**
Morsink S and Mann RB 1991 [*Class. Quantum Grav.*]{} **8 2257**
Mann RB and Steele TG 1992 [*Class. Quantum Grav.*]{} **9 475**
Sikkema A E and Mann R B 1991 [*Class. Quantum Grav.*]{} **8219**
Christensen JD and Mann RB 1992 [*Class. Quantum Grav.*]{} **9 1769**
Chan KCK and Mann RB 1993 [*Class. Quantum Grav.*]{} **10 913**
Mann RB 1994 [*Nucl. Phys.*]{} [**B418**]{} 231
Mann RB and Ross SF 1993 [*Class. Quantum Grav.*]{} **10 1405**
Arnowitt R, Deser S and Misner C W 1962 [*Gravitation: An Introduction to Current Research, Chap.7*]{} (New York: Wiley) This is a review work of a series of their papers, which are listed in the references of this paper.
Kimura T 1961 [*Prog. Theor. Phys.*]{} **26157**
Ohta T and Kimura T 1989 [*Classical and Quantum Gravity, Chap.6 (in Japanese)*]{} (Tokyo: MacGrawhill)
Ohta T, Okamura H, Kimura T and Hiida K 1974 [*Prog. Theor. Phys*]{} **511598**
Arnowitt R, Deser S and Misner C W 1960 [*J. Math. Phys.*]{} **1434**
Arnowitt R, Deser S and Misner C W 1960 [*Phys. Rev.*]{} **1171595**
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider the problem of deep neural net compression by quantization: given a large, reference net, we want to quantize its real-valued weights using a codebook with $K$ entries so that the training loss of the quantized net is minimal. The codebook can be optimally learned jointly with the net, or fixed, as for binarization or ternarization approaches. Previous work has quantized the weights of the reference net, or incorporated rounding operations in the backpropagation algorithm, but this has no guarantee of converging to a loss-optimal, quantized net. We describe a new approach based on the recently proposed framework of *model compression as constrained optimization* [@Carreir17a]. This results in a simple iterative “learning-compression” algorithm, which alternates a step that learns a net of continuous weights with a step that quantizes (or binarizes/ternarizes) the weights, and is guaranteed to converge to local optimum of the loss for quantized nets. We develop algorithms for an adaptive codebook or a (partially) fixed codebook. The latter includes binarization, ternarization, powers-of-two and other important particular cases. We show experimentally that we can achieve much higher compression rates than previous quantization work (even using just 1 bit per weight) with negligible loss degradation.'
author:
- |
Miguel Á. Carreira-Perpiñán Yerlan Idelbayev\
Electrical Engineering and Computer Science, University of California, Merced\
[<http://eecs.ucmerced.edu>]{}
date: 'July 13, 2017'
title: |
Model compression as constrained optimization,\
with application to neural nets.\
Part II: quantization.
---
Introduction {#s:intro}
============
The widespread application of deep neural nets in recent years has seen an explosive growth in the size of the training sets, the number of parameters of the nets, and the amount of computing power needed to train them. At present, very deep neural nets with upwards of many million weights are common in applications in computer vision and speech. Many of these applications are particularly useful in small devices, such as mobile phones, cameras or other sensors, which have limited computation, memory and communication bandwidth, and short battery life. It then becomes desirable to compress a neural net so that its memory storage is smaller and/or its runtime is faster and consumes less energy.
Neural net compression was a problem of interest already in the early days of neural nets, driven for example by the desire to implement neural nets in VLSI circuits. However, the current wave of deep learning work has resulted in a flurry of papers by many academic and particularly industrial labs proposing various ways to compress deep nets, some new and some not so new (see related work). Various standard forms of compression have been used in one way or another, such as low-rank decomposition, quantization, binarization, pruning and others. In this paper we focus on quantization, where the ordinarily unconstrained, real-valued weights of the neural net are forced to take values within a codebook with a finite number of entries. This codebook can be adaptive, so that its entries are learned together with the quantized weights, or (partially) fixed, which includes specific approaches such as binarization, ternarization or powers-of-two approaches.
Among compression approaches, quantization is of great interest because even crudely quantizing the weights of a trained net (for example, reducing the precision from double to single) produces considerable compression with little degradation of the loss of the task at hand (say, classification). However, this ignores the fact that the quantization is not independent of the loss, and indeed achieving a really low number of bits per weight (even just 1 bit, i.e., binary weights) would incur a large loss and make the quantized net unsuitable for practical use. Previous work has applied a quantization algorithm to a previously trained, reference net, or incorporated ad-hoc modifications to the basic backpropagation algorithm during training of the net. However, none of these approaches are guaranteed to produce upon convergence (if convergence occurs at all) a net that has quantized weights and has optimal loss among all possible quantized nets.
In this paper, our primary objectives are: 1) to provide a mathematically principled statement of the quantization problem that involves the loss of the resulting net, and 2) to provide an algorithm that can solve that problem up to local optima in an efficient and convenient way. Our starting point is a recently proposed formulation of the general problem of model compression as a constrained optimization problem [@Carreir17a]. We develop this for the case where the constraints represent the optimal weights as coming from a codebook. This results in a “learning-compression” (LC) algorithm that alternates SGD optimization of the loss over real-valued weights but with a quadratic regularization term, and quantization of the current real-valued weights. The quantization step takes a form that follows necessarily from the problem definition without ad-hoc decisions: $k$-means for adaptive codebooks, and an optimal assignment for fixed codebooks such as binarization, ternarization or powers-of-two (with possibly an optimal global scale). We then show experimentally that we can compress deep nets considerably more than previous quantization algorithms—often, all the way to the maximum possible compression, a single bit per weight, without significant error degradation.
Related work on quantization of neural nets {#s:related}
===========================================
Much work exists on compressing neural nets, using quantization, low-rank decomposition, pruning and other techniques, see @Carreir17a and references therein. Here we focus exclusively on work based on quantization. Quantization of neural net weights was recognized as an important problem early in the neural net literature, often with the goal of efficient hardware implementation, and has received much attention recently. The main approaches are of two types. The first one consists of using low-precision, fixed-point or other weight representations through some form of rounding, even single-bit (binary) values. This can be seen as quantization using a fixed codebook (i.e., with predetermined values). The second approach learns the codebook itself as a form of soft or hard adaptive quantization. There is also work on using low-precision arithmetic directly during training (see [@Gupta_15a] and references therein) but we focus here on work whose goal is to quantize a neural net of real-valued, non-quantized weights.
Quantization with a fixed codebook {#s:related:quant-fixed}
----------------------------------
Work in the 1980s and 1990s explored binarization, ternarization and general powers-of-two quantization [@Fiesler_90a; @Marches_93a; @TangKwan93a]. These same quantization forms have been revisited in recent years [@HwangSung14a; @Courbar_15a; @Rasteg_16a; @Hubara_16b; @Li_16b; @Zhou_16b; @Zhu_17a], with impressive results on large neural nets trained on GPUs, but not much innovation algorithmically. The basic idea in all these papers is essentially the same: to modify backpropagation so that it encourages binarization, ternarization or some other form of quantization of the neural net weights. The modification involves evaluating the gradient of the loss $L(\w)$ at the quantized weights (using a specific quantization or “rounding” operator that maps a continuous weight to a quantized one) but applying the update (gradient or SGD step) to the continuous (non-quantized) weights. Specific details vary, such as the quantization operator or the type of codebook. The latter has recently seen a plethora of minor variations: $\{-1,0,+1\}$ [@HwangSung14a], $\{-1,+1\}$ [@Courbar_15a], $\{-a,+a\}$ [@Rasteg_16a; @Zhou_16b], $\{-a,0,+a\}$ [@Li_16b] or $\{-a,0,+b\}$ [@Zhu_17a].
One important problem with these approaches is that their modification of backpropagation is ad-hoc, without guarantees of converging to a net with quantized weights and low loss, or of converging at all. Consider binarization to $\{-1,+1\}$ for simplicity. The gradient is computed at a binarized weight vector $\w \in \{-1,+1\}^P$, of which there are a finite number ($2^P$, corresponding to the hypercube corners), and none of these will in general have gradient zero. Hence training will never stop, and the iterates will oscillate indefinitely. Practically, this is stopped after a certain number of iterations, at which time the weight distribution is far from binarized (see fig. 2 in [@Courbar_15a]), so a drastic binarization must still be done. Given these problems, it is surprising that these techniques do seem to be somewhat effective empirically in quantizing the weights and still achieve little loss degradation, as reported in the papers above. Exactly how effective they are, on what type of nets and why this is so is an open research question.
In our LC algorithm, the optimization essentially happens in the continuous weight space by minimizing a well-defined objective (the penalized function in the L step), but this is regularly corrected by a quantization operator (C step), so that the algorithm gradually converges to a truly quantized weight vector while achieving a low loss (up to local optima). The form of both L and C steps, in particular of the quantization operator (our compression function $\bPi(\w)$), follows in a principled, optimal way from the constrained form of the problem . That is, given a desired form of quantization (e.g. binarization), the form of the C step is determined, and the overall algorithm is guaranteed to converge to a valid (binary) solution.
Also, we emphasize that there is little practical reason to use certain fixed codebooks, such as $\{-1,+1\}$ or $\{-a,+a\}$, instead of an adaptive codebook such as $\{c_1,c_2\}$ with $c_1,c_2 \in \bbR$. The latter is obviously less restrictive, so it will incur a lower loss. And its hardware implementation is about as efficient: to compute a scalar product of an activation vector with a quantized weight vector, all we require is to sum activation values for each centroid and to do two floating-point multiplications (with $c_1$ and $c_2$). Indeed, our experiments in section \[s:expts:LeNet\] show that using an adaptive codebook with $K=2$ clearly beats using $\{-1,+1\}$.
Quantization with an adaptive codebook {#s:related:quant-adaptive}
--------------------------------------
Quantization with an adaptive codebook is, obviously, more powerful than with a fixed codebook, even though it has to store the codebook itself. Quantization using an adaptive codebook has also been explored in the neural nets literature, using approaches based on soft quantization [@NowlanHinton92a; @Ullric_17a] or hard quantization [@Fiesler_90a; @Marches_93a; @TangKwan93a; @Gong_15a; @Han_15a], and we discuss this briefly.
Given a set of real-valued elements (scalars or vectors), in adaptive quantization we represent (“quantize”) each element by exactly one entry in a codebook. The codebook and the assignment of values to codebook entries should minimize a certain distortion measure, such as the squared error. Learning the codebook and assignment is done by an algorithm, possibly approximate (such as $k$-means for the squared error). Quantization is related to clustering and often one can use the same algorithm for both (e.g. $k$-means), but the goal is different: quantization seeks to minimize the distortion rather than to model the data as clusters. For example, a set of values uniformly distributed in $[-1,1]$ shows no clusters but may be subject to quantization for compression purposes. In our case of neural net compression, we have an additional peculiarity that complicates the optimization: the quantization and the weight values themselves should be jointly learned to minimize the loss of the net on the task.
Two types of clustering exist, hard and soft clustering. In hard clustering, each data point is assigned to exactly one cluster (e.g. $k$-means clustering). In soft clustering, we have a probability distribution over points and clusters (e.g. Gaussian mixture clustering). Likewise, two basic approaches exist for neural net quantization, based on hard and soft quantization. We review each next.
In hard quantization, each weight is assigned to exactly one codebook value. This is the usual meaning of quantization. This is a difficult problem because, even if the loss is differentiable over the weights, the assignment makes the problem inherently combinatorial. Previous work [@Gong_15a; @Han_15a] has run a quantization step ($k$-means) as a postprocessing step on a reference net (which was trained to minimize the loss). This is suboptimal in that it does not learn the weights, codebook and assignment jointly. We call this “direct compression” and discuss it in more detail in section \[s:DC\]. Our LC algorithm does learn the weights, codebook and assignment jointly, and converges to a local optimum of problem .
In soft quantization, the assignment of values to codebook entries is based on a probability distribution. This was originally proposed by @NowlanHinton92a as a way to share weights softly in a neural net with the goal of improving generalization, and has been recently revisited with the goal of compression [@Ullric_17a]. The idea is to penalize the loss with the negative log-likelihood of a Gaussian mixture (GM) model on the scalar weights of the net. This has the advantage of being differentiable and of coadapting the weights and the GM parameters (proportions, means, variances). However, it does not uniquely assign each weight to one mean, in fact the resulting distribution of weights is far from quantized; it simply encourages the creation of Gaussian clusters of weights, and one has to assign weights to means as a postprocessing step, which is suboptimal. The basic problem is that a GM is a good model (better than $k$-means) for noisy or uncertain data, but that is not what we have here. Quantizing the weights for compression implies a constraint that certain weights must take exactly the same value, without noise or uncertainty, and optimize the loss. We seek an optimal assignment that is truly hard, not soft. Indeed, a GM prior is to quantization what a quadratic prior (i.e., weight decay) is to sparsity: a quadratic prior encourages all weights to be small but does not encourage some weights to be exactly zero, just as a GM prior encourages weights to form Gaussian clusters but not to become groups of identical weights.
Neural net quantization as constrained optimization and the “learning-compression” (LC) algorithm {#s:LC}
=================================================================================================
As noted in the introduction, compressing a neural net optimally means finding the compressed net that has (locally) lowest loss. Our first goal is to formulate this mathematically in a way that is amenable to nonconvex optimization techniques. Following @Carreir17a, we define the following *model compression as constrained optimization* problem: $$\label{e:compression-problem}
\textcolor{blue}{\min_{\w,\bTheta}{ L(\w) } \quad \text{s.t.} \quad \w = \bDelta(\bTheta)}$$ where $\w \in \bbR^P$ are the real-valued weights of the neural net, $L(\w)$ is the loss to be minimized (e.g. cross-entropy for a classification task on some training set), and the constraint $\w = \bDelta(\bTheta)$ indicates that the weights must be the result of decompressing a low-dimensional parameter vector . This corresponds to quantization and will be described in section \[s:quant\]. Problem is equivalent to the unconstrained problem “$\min_{\bTheta}{ L(\bDelta(\bTheta)) }$”, but this is nondifferentiable with quantization (where is a discrete mapping), and introducing the auxiliary variable will lead to a convenient algorithm.
Our second goal is to solve this problem via an efficient algorithm. Although this might be done in different ways, a particularly simple one was proposed by @Carreir17a that achieves separability between the data-dependent part of the problem (the loss) and the data-independent part (the weight quantization). First, we apply a penalty method to solve . We consider here the augmented Lagrangian method [@NocedalWright06a], where $\blambda \in \bbR^P$ are the Lagrange multiplier estimates[^1]: $$\begin{aligned}
\label{e:augLag}
\calL_A(\w,\bTheta,\blambda;\mu) &= L(\w) - \blambda^T (\w - \bDelta(\bTheta)) + \frac{\mu}{2} \norm{\w - \bDelta(\bTheta)}^2 \\
\label{e:augLag2}
&= L(\w) + \frac{\mu}{2} \norm[\Big]{\w - \bDelta(\bTheta) - \frac{1}{\mu} \blambda}^2 - \frac{1}{2\mu} \norm{\blambda}^2.\end{aligned}$$ The augmented Lagrangian method works as follows. For fixed $\mu \ge 0$, we optimize $\calL_A(\w,\bTheta,\blambda;\mu)$ over $(\w,\bTheta)$ accurately enough. Then, we update the Lagrange multiplier estimates as $\blambda \leftarrow \blambda - \mu (\w - \bDelta(\bTheta))$. Finally, we increase $\mu$. We repeat this process and, in the limit as $\mu \rightarrow \infty$, the iterates $(\w,\bTheta)$ tend to a local KKT point (typically, a local minimizer) of the constrained problem . A simpler but less effective penalty method, the quadratic penalty method, results from setting $\blambda = \0$ throughout; we do not describe it explicitly, see @Carreir17a.
Finally, in order to optimize $\calL_A(\w,\bTheta,\blambda;\mu)$ over $(\w,\bTheta)$, we use alternating optimization. This gives rise to the following two steps:
- **L step: learning** $$\label{e:Lstep}
\textcolor{blue}{\min_{\w}{ L(\w) + \frac{\mu}{2} \norm[\Big]{\w - \bDelta(\bTheta) - \frac{1}{\mu} \blambda}^2 }.}$$ This involves optimizing a regularized version of the loss, which pulls the optimizer towards the currently quantized weights. For neural nets, it can be solved with stochastic gradient descent (SGD).
- **C step: compression (here, quantization)** $$\label{e:Cstep}
\textcolor{blue}{\min_{\bTheta}{ \norm[\Big]{\w - \frac{1}{\mu} \blambda - \bDelta(\bTheta)}^2 } \quad \Longleftrightarrow \quad \bTheta = \bPi \Big(\w - \frac{1}{\mu} \blambda \Big)}.$$ We describe this in section \[s:quant\]. Solving this problem is equivalent to quantizing optimally the current real-valued weights $\w - \frac{1}{\mu} \blambda$, and can be seen as finding their orthogonal projection $\bPi \big(\w - \frac{1}{\mu} \blambda \big)$ on the feasible set of quantized nets.
This algorithm was called the “learning-compression” (LC) algorithm by @Carreir17a.
We note that, throughout the optimization, there are two weight vectors that evolve simultaneously and coincide in the limit as $\mu \rightarrow \infty$: (or, more precisely, $\w - \frac{1}{\mu} \blambda$) contains real-valued, non-quantized weights (and this is what the L step optimizes over); and $\bDelta(\bTheta)$ contains quantized weights (and this is what the C step optimizes over). In the C step, $\bDelta(\bTheta)$ is the projection of the current on the feasible set of quantized vectors. In the L step, optimizes the loss while being pulled towards the current $\bDelta(\bTheta)$.
The formulation and the LC algorithm have two crucial advantages. The first one is that we get a convenient separation between learning and quantization which allows one to solve each step by reusing existing code. The data-dependent part of the optimization is confined within the L step. This part is the more computationally costly, requiring access to the training set and the neural net, and usually implemented in a GPU using SGD. The data-independent part of the optimization, i.e., the compression of the weights (here, quantization), is confined within the C step. This needs access only to the vector of current, real-valued weights (not to the training set or the actual neural net).
The second advantage is that the form of the C step is determined by the choice of quantization form (defined by $\bDelta(\bTheta)$), and the algorithm designer need not worry about modifying backpropagation or SGD in any way for convergence to a valid solution to occur. For example, if a new form of quantization were discovered and we wished to use it, all we have to do is put it in the decompression mapping form $\bDelta(\bTheta)$ and solve the compression mapping problem (which depends only on the quantization technique, and for which a known algorithm may exist). This is unlike much work in neural net quantization, where various, somewhat arbitrary quantization or rounding operations are incorporated in the usual backpropagation training (see section \[s:related\]), which makes it unclear what problem the overall algorithm is optimizing, if it does optimize anything at all.
In section \[s:quant\], we solve the compression mapping problem for the adaptive and fixed codebook cases. For now, it suffices to know that it will involve running $k$-means with an adaptive codebook and a form of rounding with a fixed codebook.
Geometry of the neural net quantization problem {#s:geom}
-----------------------------------------------
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
\[\]\[\][ ]{} \[l\]\[Bl\] \[l\]\[Bl\] \[l\]\[l\][-space]{} \[\]\[\] ![Schematic representation of the idea of model compression by constrained optimization, in general (top 3 plots) and in particular for quantization (bottom 2 plots). The top of this figure is adapted from @Carreir17a.\ \[t\]\[\][$\overline{\w}_1$]{} \[t\]\[\][$\overline{\w}_2$]{} \[l\]\[Bl\][$\w^*$]{} \[l\]\[Bl\][$\bDelta(\bTheta^{\text{DC}}_1)$]{} \[l\]\[Bl\][$\bDelta(\bTheta^{\text{DC}}_2)$]{} \[l\]\[l\][-space]{} \[\]\[\][$\calF_{\w}$]{} ![Schematic representation of the idea of model compression by constrained optimization, in general (top 3 plots) and in particular for quantization (bottom 2 plots). The top of this figure is adapted from @Carreir17a.\ \[t\]\[\][$\overline{\w}$]{} \[l\]\[Bl\][$\w^*_1$]{} \[Bl\]\[l\][$\w^*_2$]{} \[\]\[l\][$\w^*_3$]{} \[l\]\[Bl\][$\w^*_4$]{} \[l\]\[l\]\[0.55\][$\bDelta(\bTheta^{\text{DC}}_1)$]{} \[l\]\[l\]\[0.55\][$\bDelta(\bTheta^{\text{DC}}_2)$]{} \[bl\]\[l\]\[0.55\][$\bDelta(\bTheta^{\text{DC}}_3)$]{} \[lt\]\[l\]\[0.55\][$\bDelta(\bTheta^{\text{DC}}_4)$]{} \[l\]\[Bl\][$\calF^1_{\w}$]{} \[l\]\[Bl\][$\calF^2_{\w}$]{} \[l\]\[Bl\][$\calF^3_{\w}$]{} \[l\]\[Bl\][$\calF^4_{\w}$]{} \[l\]\[l\][-space]{} ![Schematic representation of the idea of model compression by constrained optimization, in general (top 3 plots) and in particular for quantization (bottom 2 plots). The top of this figure is adapted from @Carreir17a.\
*Plots 1–3 (top row)*: illustration of the uncompressed model space (-space $= \bbR^P$), the contour lines of the loss $L(\w)$ (green lines), and the set of compressed models (the feasible set $\calF_{\w} = \{\w \in \bbR^P\mathpunct{:}\ \w = \bDelta(\bTheta) \text{ for } \bTheta \in \bbR^Q\}$, grayed areas), for a generic compression technique . The -space is not shown. $\overline{\w}$ optimizes $L(\w)$ but is infeasible (no can decompress into it). The direct compression $\w^{\text{DC}} = \bDelta(\bTheta^{\text{DC}})$ is feasible but not optimal compressed (not optimal in the feasible set). $\w^* = \bDelta(\bTheta^*)$ is optimal compressed. Plot 2 shows two local optima $\overline{\w}_1$ and $\overline{\w}_2$ of the loss $L(\w)$, and their respective DC points (the contour lines are omitted to avoid clutter). Plot 3 shows several feasible sets, corresponding to different compression levels ($\calF^1_{\w}$ is most compression).\ *Plots 1–3 (top row)*: illustration of the uncompressed model space (-space $= \bbR^P$), the contour lines of the loss $L(\w)$ (green lines), and the set of compressed models (the feasible set $\calF_{\w} = \{\w \in \bbR^P\mathpunct{:}\ \w = \bDelta(\bTheta) \text{ for } \bTheta \in \bbR^Q\}$, grayed areas), for a generic compression technique . The -space is not shown. $\overline{\w}$ optimizes $L(\w)$ but is infeasible (no can decompress into it). The direct compression $\w^{\text{DC}} = \bDelta(\bTheta^{\text{DC}})$ is feasible but not optimal compressed (not optimal in the feasible set). $\w^* = \bDelta(\bTheta^*)$ is optimal compressed. Plot 2 shows two local optima $\overline{\w}_1$ and $\overline{\w}_2$ of the loss $L(\w)$, and their respective DC points (the contour lines are omitted to avoid clutter). Plot 3 shows several feasible sets, corresponding to different compression levels ($\calF^1_{\w}$ is most compression).\ *Plots 1–3 (top row)*: illustration of the uncompressed model space (-space $= \bbR^P$), the contour lines of the loss $L(\w)$ (green lines), and the set of compressed models (the feasible set $\calF_{\w} = \{\w \in \bbR^P\mathpunct{:}\ \w = \bDelta(\bTheta) \text{ for } \bTheta \in \bbR^Q\}$, grayed areas), for a generic compression technique . The -space is not shown. $\overline{\w}$ optimizes $L(\w)$ but is infeasible (no can decompress into it). The direct compression $\w^{\text{DC}} = \bDelta(\bTheta^{\text{DC}})$ is feasible but not optimal compressed (not optimal in the feasible set). $\w^* = \bDelta(\bTheta^*)$ is optimal compressed. Plot 2 shows two local optima $\overline{\w}_1$ and $\overline{\w}_2$ of the loss $L(\w)$, and their respective DC points (the contour lines are omitted to avoid clutter). Plot 3 shows several feasible sets, corresponding to different compression levels ($\calF^1_{\w}$ is most compression).\
*Plots 4–5 (bottom row)*: illustration when corresponds to quantization, in the particular case of a codebook of size $K=1$ and a 2-weight net, so $\w = (w_1,w_2) \in \bbR^2$, $\bTheta = c \in \bbR$ and $\bDelta(\bTheta) = \binom{c}{c} \in \bbR^2$. Plot 4 is the joint space $(\w,c)$ and plot 5 is its projection in -space (as in plot 1). In plot 4, the black line is the feasible set $\calF = \{(\w,\bTheta) \in \bbR^P \times \bbR^Q\mathpunct{:}\ \w = \bDelta(\bTheta)\}$, corresponding to the constraints $w_1 = w_2 = c$. In plot 5, the black line is the feasible set $\calF_{\w} = \{\w \in \bbR^P\mathpunct{:}\ \w = \bDelta(\bTheta) \text{ for } \bTheta \in \bbR^Q\}$, corresponding to the constraint $w_1 = w_2$. The red line is the quadratic-penalty method path $(\w(\mu),c(\mu))$, which for this simple case is a straight line segment from the point $(\overline{\w},c^{\text{DC}})$ to the solution $(\w^*,c^*)$. We mark three points: blue represents the reference net $\overline{\w}$ at the DC codebook $\bTheta = c^{\text{DC}}$ (the beginning of the path); red is the solution $(\w^*,c^*)$ (the end of the path); and white $\circ$ is the direct compression point $(\bDelta(\bTheta^{\text{DC}}),\bTheta^{\text{DC}}) = \smash{\big( \binom{c^{\text{DC}}}{c^{\text{DC}}},c^{\text{DC}} \big)}$.[]{data-label="f:LC-illustration"}](LC-illustration1.eps "fig:"){height="0.35\linewidth"} *Plots 4–5 (bottom row)*: illustration when corresponds to quantization, in the particular case of a codebook of size $K=1$ and a 2-weight net, so $\w = (w_1,w_2) \in \bbR^2$, $\bTheta = c \in \bbR$ and $\bDelta(\bTheta) = \binom{c}{c} \in \bbR^2$. Plot 4 is the joint space $(\w,c)$ and plot 5 is its projection in -space (as in plot 1). In plot 4, the black line is the feasible set $\calF = \{(\w,\bTheta) \in \bbR^P \times \bbR^Q\mathpunct{:}\ \w = \bDelta(\bTheta)\}$, corresponding to the constraints $w_1 = w_2 = c$. In plot 5, the black line is the feasible set $\calF_{\w} = \{\w \in \bbR^P\mathpunct{:}\ \w = \bDelta(\bTheta) \text{ for } \bTheta \in \bbR^Q\}$, corresponding to the constraint $w_1 = w_2$. The red line is the quadratic-penalty method path $(\w(\mu),c(\mu))$, which for this simple case is a straight line segment from the point $(\overline{\w},c^{\text{DC}})$ to the solution $(\w^*,c^*)$. We mark three points: blue represents the reference net $\overline{\w}$ at the DC codebook $\bTheta = c^{\text{DC}}$ (the beginning of the path); red is the solution $(\w^*,c^*)$ (the end of the path); and white $\circ$ is the direct compression point $(\bDelta(\bTheta^{\text{DC}}),\bTheta^{\text{DC}}) = \smash{\big( \binom{c^{\text{DC}}}{c^{\text{DC}}},c^{\text{DC}} \big)}$.[]{data-label="f:LC-illustration"}](LC-illustration3.eps "fig:"){height="0.35\linewidth"} *Plots 4–5 (bottom row)*: illustration when corresponds to quantization, in the particular case of a codebook of size $K=1$ and a 2-weight net, so $\w = (w_1,w_2) \in \bbR^2$, $\bTheta = c \in \bbR$ and $\bDelta(\bTheta) = \binom{c}{c} \in \bbR^2$. Plot 4 is the joint space $(\w,c)$ and plot 5 is its projection in -space (as in plot 1). In plot 4, the black line is the feasible set $\calF = \{(\w,\bTheta) \in \bbR^P \times \bbR^Q\mathpunct{:}\ \w = \bDelta(\bTheta)\}$, corresponding to the constraints $w_1 = w_2 = c$. In plot 5, the black line is the feasible set $\calF_{\w} = \{\w \in \bbR^P\mathpunct{:}\ \w = \bDelta(\bTheta) \text{ for } \bTheta \in \bbR^Q\}$, corresponding to the constraint $w_1 = w_2$. The red line is the quadratic-penalty method path $(\w(\mu),c(\mu))$, which for this simple case is a straight line segment from the point $(\overline{\w},c^{\text{DC}})$ to the solution $(\w^*,c^*)$. We mark three points: blue represents the reference net $\overline{\w}$ at the DC codebook $\bTheta = c^{\text{DC}}$ (the beginning of the path); red is the solution $(\w^*,c^*)$ (the end of the path); and white $\circ$ is the direct compression point $(\bDelta(\bTheta^{\text{DC}}),\bTheta^{\text{DC}}) = \smash{\big( \binom{c^{\text{DC}}}{c^{\text{DC}}},c^{\text{DC}} \big)}$.[]{data-label="f:LC-illustration"}](LC-illustration2.eps "fig:"){height="0.35\linewidth"}
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
\
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
\[\]\[\][$w_1$]{} \[\]\[\][$w_2$\[0pt\]\[0pt\]]{} \[\]\[r\]\[1\]\[-90\][$c$]{} ![Schematic representation of the idea of model compression by constrained optimization, in general (top 3 plots) and in particular for quantization (bottom 2 plots). The top of this figure is adapted from @Carreir17a.\ \[\]\[B\][$w_1$]{} \[\]\[\]\[1\]\[-90\][$w_2$]{} \[br\]\[b\] \[lt\]\[lb\][ $\w^*$\[0pt\]\[0pt\]]{} \[l\]\[l\]\[0.9\][ $\bDelta(\bTheta^{\text{DC}})$$=$$\binom{c^{\text{DC}}}{c^{\text{DC}}}$]{} ![Schematic representation of the idea of model compression by constrained optimization, in general (top 3 plots) and in particular for quantization (bottom 2 plots). The top of this figure is adapted from @Carreir17a.\
*Plots 1–3 (top row)*: illustration of the uncompressed model space (-space $= \bbR^P$), the contour lines of the loss $L(\w)$ (green lines), and the set of compressed models (the feasible set $\calF_{\w} = \{\w \in \bbR^P\mathpunct{:}\ \w = \bDelta(\bTheta) \text{ for } \bTheta \in \bbR^Q\}$, grayed areas), for a generic compression technique . The -space is not shown. $\overline{\w}$ optimizes $L(\w)$ but is infeasible (no can decompress into it). The direct compression $\w^{\text{DC}} = \bDelta(\bTheta^{\text{DC}})$ is feasible but not optimal compressed (not optimal in the feasible set). $\w^* = \bDelta(\bTheta^*)$ is optimal compressed. Plot 2 shows two local optima $\overline{\w}_1$ and $\overline{\w}_2$ of the loss $L(\w)$, and their respective DC points (the contour lines are omitted to avoid clutter). Plot 3 shows several feasible sets, corresponding to different compression levels ($\calF^1_{\w}$ is most compression).\ *Plots 1–3 (top row)*: illustration of the uncompressed model space (-space $= \bbR^P$), the contour lines of the loss $L(\w)$ (green lines), and the set of compressed models (the feasible set $\calF_{\w} = \{\w \in \bbR^P\mathpunct{:}\ \w = \bDelta(\bTheta) \text{ for } \bTheta \in \bbR^Q\}$, grayed areas), for a generic compression technique . The -space is not shown. $\overline{\w}$ optimizes $L(\w)$ but is infeasible (no can decompress into it). The direct compression $\w^{\text{DC}} = \bDelta(\bTheta^{\text{DC}})$ is feasible but not optimal compressed (not optimal in the feasible set). $\w^* = \bDelta(\bTheta^*)$ is optimal compressed. Plot 2 shows two local optima $\overline{\w}_1$ and $\overline{\w}_2$ of the loss $L(\w)$, and their respective DC points (the contour lines are omitted to avoid clutter). Plot 3 shows several feasible sets, corresponding to different compression levels ($\calF^1_{\w}$ is most compression).\
*Plots 4–5 (bottom row)*: illustration when corresponds to quantization, in the particular case of a codebook of size $K=1$ and a 2-weight net, so $\w = (w_1,w_2) \in \bbR^2$, $\bTheta = c \in \bbR$ and $\bDelta(\bTheta) = \binom{c}{c} \in \bbR^2$. Plot 4 is the joint space $(\w,c)$ and plot 5 is its projection in -space (as in plot 1). In plot 4, the black line is the feasible set $\calF = \{(\w,\bTheta) \in \bbR^P \times \bbR^Q\mathpunct{:}\ \w = \bDelta(\bTheta)\}$, corresponding to the constraints $w_1 = w_2 = c$. In plot 5, the black line is the feasible set $\calF_{\w} = \{\w \in \bbR^P\mathpunct{:}\ \w = \bDelta(\bTheta) \text{ for } \bTheta \in \bbR^Q\}$, corresponding to the constraint $w_1 = w_2$. The red line is the quadratic-penalty method path $(\w(\mu),c(\mu))$, which for this simple case is a straight line segment from the point $(\overline{\w},c^{\text{DC}})$ to the solution $(\w^*,c^*)$. We mark three points: blue represents the reference net $\overline{\w}$ at the DC codebook $\bTheta = c^{\text{DC}}$ (the beginning of the path); red is the solution $(\w^*,c^*)$ (the end of the path); and white $\circ$ is the direct compression point $(\bDelta(\bTheta^{\text{DC}}),\bTheta^{\text{DC}}) = \smash{\big( \binom{c^{\text{DC}}}{c^{\text{DC}}},c^{\text{DC}} \big)}$.[]{data-label="f:LC-illustration"}](toy_quant3D.eps "fig:"){height="0.33\linewidth"} *Plots 4–5 (bottom row)*: illustration when corresponds to quantization, in the particular case of a codebook of size $K=1$ and a 2-weight net, so $\w = (w_1,w_2) \in \bbR^2$, $\bTheta = c \in \bbR$ and $\bDelta(\bTheta) = \binom{c}{c} \in \bbR^2$. Plot 4 is the joint space $(\w,c)$ and plot 5 is its projection in -space (as in plot 1). In plot 4, the black line is the feasible set $\calF = \{(\w,\bTheta) \in \bbR^P \times \bbR^Q\mathpunct{:}\ \w = \bDelta(\bTheta)\}$, corresponding to the constraints $w_1 = w_2 = c$. In plot 5, the black line is the feasible set $\calF_{\w} = \{\w \in \bbR^P\mathpunct{:}\ \w = \bDelta(\bTheta) \text{ for } \bTheta \in \bbR^Q\}$, corresponding to the constraint $w_1 = w_2$. The red line is the quadratic-penalty method path $(\w(\mu),c(\mu))$, which for this simple case is a straight line segment from the point $(\overline{\w},c^{\text{DC}})$ to the solution $(\w^*,c^*)$. We mark three points: blue represents the reference net $\overline{\w}$ at the DC codebook $\bTheta = c^{\text{DC}}$ (the beginning of the path); red is the solution $(\w^*,c^*)$ (the end of the path); and white $\circ$ is the direct compression point $(\bDelta(\bTheta^{\text{DC}}),\bTheta^{\text{DC}}) = \smash{\big( \binom{c^{\text{DC}}}{c^{\text{DC}}},c^{\text{DC}} \big)}$.[]{data-label="f:LC-illustration"}](toy_quant2D.eps "fig:"){height="0.33\linewidth"}
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Problem can be written as $\min_{\w,\bTheta}{ L(\w) }$ s.t. $\w,\bTheta \in \calF$, where the objective function is the loss $L(\w)$ on the real-valued weights and the feasible set on and the low-dimensional parameters is: $$\label{e:feasible-set}
\calF = \{(\w,\bTheta) \in \bbR^P \times \bbR^Q\mathpunct{:}\ \w = \bDelta(\bTheta)\}.$$ We also define the feasible set in -space: $$\label{e:feasible-set-w}
\calF_{\w} = \{\w \in \bbR^P\mathpunct{:}\ \w = \bDelta(\bTheta) \text{ for } \bTheta \in \bbR^Q\}$$ which contains all high-dimensional models that can be obtained by decompressing some low-dimensional model . Fig. \[f:LC-illustration\] (plots 1–3) illustrates the geometry of the problem in general.
Solving the C step requires minimizing (where we write instead of $\w - \frac{1}{\mu} \blambda$ for simplicity of notation): $$\label{e:compression-mapping}
\textcolor{blue}{\bPi(\w) = \argmin_{\bTheta}{\norm{\w - \bDelta(\bTheta)}^2}.}$$ We call $\bDelta\mathpunct{:}\ \bTheta \in \bbR^Q \rightarrow \w \in \bbR^P$ the *decompression mapping* and $\bPi\mathpunct{:}\ \w \in \bbR^P \rightarrow \bTheta \in \bbR^Q$ the *compression mapping*. In quantization, this has the following meaning:
- $\bTheta = \{\calC,\Z\}$ consists of the codebook (if the codebook is adaptive) and the assignments of weight-to-codebook-entries. The assignments can be encoded as 1-of-$K$ vectors $\Z^T = (\z_1,\dots,\z_P)$ or directly as $P$ indices in $\{1,\dots,K\}$ for a codebook with $K$ entries.
- The decompression mapping $\w_{\calC} = \bDelta(\calC,\Z)$ uses the codebook and assignments as a lookup table to generate a real-valued but quantized weight vector $\w_{\calC}$. This vector is used in the L step as a regularizer.
- The compression mapping $\{\calC,\Z\} = \bPi(\w)$ learns optimally a codebook and assignments given a real-valued, non-quantized weight vector (using $k$-means or a form of rounding, see section \[s:quant\]). All the C step does is solve for the compression mapping.
As shown by @Carreir17a, the compression mapping $\bPi(\w)$ finds the orthogonal projection of on the feasible set $\calF_{\w}$, which we call $\w_{\calC}$.
For quantization, the geometry of the constrained optimization formulation is as follows. The feasible set can be written as the union of a combinatorial number of linear subspaces $\calS_j$ (containing the origin), where $\calS_j$ is of the form $\{ \w_i = \c_k,\ \forall i=1,\dots,P,\ k \in \{1,\dots,K\} \}$. Each such subspace defines a particular assignment of the $P$ weights to the $K$ centroids $\calC = \{\c_1,\dots,\c_K\}$. There are $K^P$ assignments. If we knew the optimal assignment, the feasible set would be a single linear subspace, and the weights could be eliminated (using $\w_i = \c_k$) to yield an unconstrained objective $L(\calC)$ of $K$ tunable vectors (shared weights in neural net parlance), which would be simple to optimize. What makes the problem hard is that we do not know the optimal assignment. Depending on the dimensions $P$ and $K$, these subspaces may look like lines, planes, etc., always passing through the origin in $(\w,\calC)$ space. Geometrically, the union of these $K^P$ subspaces is a feasible set with both a continuous structure (within each subspace) and a discrete one (the number of subspaces is finite but very large).
Fig. \[f:LC-illustration\] (plots 4–5) shows the actual geometry for the case of a net with $P = 2$ weights and a codebook with $K=1$ centroid. This can be exactly visualized in 3D $(w_1,w_2,c)$ because the assignment variables $z_{11} = z_{21} = 1$ are redundant and can be eliminated: $\min_{w_1,w_2,c}{ L(w_1,w_2) }$ s.t. $w_1 = c$, $w_2 = c$. The compression mapping is easily seen to be $\bPi(\w) = \frac{w_1+w_2}{2} = c$, and $\bDelta(\bPi(\w)) = \frac{w_1+w_2}{2} \binom{1}{1}$ is indeed the orthogonal projection of onto the diagonal line $w_1 = w_2$ in -space (the feasible set). This particular case is, however, misleading in that the constraints involve a single linear subspace rather than the union of a combinatorial number of subspaces. It can be solved simply and exactly by setting $w_1 = w_2 = c$ and eliminating variables into $L(w_1,w_2) = L(c,c)$.
Convergence of the LC algorithm {#s:conv}
-------------------------------
Convergence of the LC algorithm to a local KKT point (theorem 5.1 in [@Carreir17a]) is guaranteed for smooth problems (continuously differentiable loss $L(\w)$ and decompression mapping $\bDelta(\bTheta)$) if $\mu \rightarrow \infty$ and optimization of the penalty function is done accurately enough for each $\mu$. However, in quantization the decompression mapping $\bDelta(\bTheta)$ is discrete, given by a lookup table, so the theorem does not apply.
In fact, neural net quantization is an NP-complete problem even in simple cases. For example, consider least-squares linear regression with weights in $\{-1,+1\}$. This corresponds to binarization of a single-layer, linear neural net. The loss $L(\w)$ is quadratic, so the optimization problem is a binary quadratic problem over the weights, which is NP-complete [@GareyJohnson79a]. However, the LC algorithm will still converge to a “local optimum” in the same sense that the $k$-means algorithm is said to converge to a local optimum: the L step cannot improve given the C step, and vice versa. While this will generally not be the global optimum of problem , it will be a good solution in that the loss will be low (because the L step continuously minimizes it in part), and the LC algorithm is guaranteed to converge to a weight vector that satisfies the quantization constraints (e.g. weights in $\{-1,+1\}$ for binarization). Our experiments confirm the effectiveness of the LC algorithm for quantization, consistently outperforming other approaches over a range of codebook types and sizes.
Practicalities of the LC algorithm
----------------------------------
We give pseudocode for three representative cases of the resulting LC algorithms: adaptive codebook (fig. \[f:pseudocode-adaptive\]), fixed codebook (fig. \[f:pseudocode-fixed\]) and binarization with global scale (fig. \[f:pseudocode-bin-scale\]).
As usual with path-following algorithms, ideally one would follow the path of iterates $(\w(\mu),\bTheta(\mu))$ closely until $\mu \rightarrow \infty$, by increasing the penalty parameter $\mu$ slowly. In practice, in order to reduce computing time, we increase $\mu$ more aggressively by following a multiplicative schedule $\mu_k = \mu_0 a^k$ for $k = 0,1,2\dots$ where $\mu_0>0$ and $a>1$. However, it is important to use a small enough $\mu_0$ that allows the algorithm to explore the solution space before committing to specific assignments for the weights.
The L step with a large training set typically uses SGD. As recommended by @Carreir17a, we use a clipped schedule $\{\eta'_{t}\}^{\infty}_{t=0}$ for the learning rates of the form $\eta'_{t} = \smash{\min{ \big( \eta_{t},\frac{1}{\mu} \big) }},\ t = 0,1,2\dots$, where $t$ is the epoch index and $\smash{\{\eta_{t}\}^{\infty}_{t=0}}$ is a schedule for the reference net (i.e., for $\mu = 0$). This ensures convergence and avoids erratic updates as $\mu$ becomes large.
We initialize $\blambda = \0$ and $(\w,\bTheta) = (\overline{\w},\bTheta^{\text{DC}})$, i.e., to the reference net and direct compression, which is the exact solution for $\mu \rightarrow 0^+$, as we show in the next section. We stop the LC algorithm when $\norm{\w - \bDelta(\calC,\Z)}$ is smaller than a set tolerance, i.e., when the real-valued and quantized weights are nearly equal. We take as solution $\w_{\calC} = \bDelta(\calC,\Z)$, i.e., the quantized weights using the codebook and assignments .
The runtime of the C step is negligible compared to that of the L step. With a fixed codebook, the C step is a simple assignment per weight. With an adaptive codebook, the C step runs $k$-means, each iteration of which is linear on the number of weights $P$. The number of iterations that $k$-means runs is a few tens in the first $k$-means (initialized by $k$-means++, on the reference weights) and just about one in subsequent C steps (because $k$-means is warm-started), as seen in our experiments. So the runtime is dominated by the L steps, i.e., by optimizing the loss.
Direct compression and iterated direct compression {#s:DC}
--------------------------------------------------
The quadratic-penalty and augmented-Lagrangian methods define a path of iterates $(\w(\mu),\bTheta(\mu))$ for $\mu \ge 0$ that converges to a local solution as $\mu \rightarrow \infty$. The beginning of this path is of special importance, and was called *direct compression (DC)* by @Carreir17a. Taking the limit $\mu \rightarrow 0^+$ and assuming an initial $\blambda=\0$, we find that $\w(0^+) = \argmin_{\w}{ L(\w) } \equiv \overline{\w}$ and $\bTheta(0^+) = \bPi(\overline{\w}) = \argmin_{\bTheta}{ \smash{\norm{\overline{\w} - \bDelta(\bTheta)}^2} } \equiv \smash{\bTheta^{\text{DC}}}$. Hence, this corresponds to training a reference, non-quantized net $\overline{\w}$ and then quantizing it regardless of the loss (or equivalently projecting $\overline{\w}$ on the feasible set). As illustrated in fig. \[f:LC-illustration\], this is suboptimal (i.e., it does not produce the compressed net with lowest loss), more so the farther the reference is from the feasible set. This will happen when the feasible set is small, i.e., when the codebook size $K$ is small (so the compression level is high). Indeed, our experiments show that for large $K$ (around 32 bits/weight) then DC is practically identical to the result of the LC algorithm, but as $K$ decreases (e.g. 1 to 4 bits/weight) then the loss of DC becomes larger and larger than that of the LC algorithm.
A variation of direct compression consists of “iterating” it, as follows. We first optimize $L(\w)$ to obtain $\overline{\w}$ and then quantize it with $k$-means into $\bTheta^{\text{DC}}$. Next, we optimize $L(\w)$ again but initializing from $\bDelta(\bTheta^{\text{DC}})$, and then we compress it; etc. This was called “iterated direct compression (iDC)” by @Carreir17a. iDC should not improve at all over DC if the loss optimization was exact and there was a single optimum: it simply would cycle forever between the reference weights $\overline{\w}$ and the DC weights $\bDelta(\bTheta^{\text{DC}})$. However, in practice iDC may improve somewhat over DC, for two reasons. 1) With local optima of $L(\w)$, we might converge to a different optimum after the quantization step (see fig. \[f:LC-illustration\] plot 2). However, at some point this will end up cycling between some reference net (some local optimum of $L(\w)$) and its quantized net. 2) In practice, SGD-based optimization of the loss with large neural nets is approximate; we stop SGD way before it has converged. This implies the iterates never fully reach $\overline{\w}$, and keep oscillating forever somewhere in between $\overline{\w}$ and $\bDelta(\bTheta^{\text{DC}})$.
DC and iDC have in fact been proposed recently for quantization, although without the context that our constrained optimization framework provides. @Gong_15a applied $k$-means to quantize the weights of a reference net, i.e., DC. The “trained quantization” of @Han_15a tries to improve over this by iterating the process, i.e., iDC. In our experiments, we verify that neither DC not iDC converge to a local optimum of problem , while our LC algorithm does.
Solving the C step: compression by quantization {#s:quant}
===============================================
The C step consists of solving the optimization problem of eq. : $\bPi(\w) = \argmin_{\bTheta}{\norm{\w - \bDelta(\bTheta)}^2}$, where $\w \in \bbR^P$ is a vector of real-valued weights. This is a quadratic distortion (or least-squares error) problem, and this was caused by selecting a quadratic penalty in the augmented Lagrangian . It is possible to use other penalties (e.g. using the $\ell_1$ norm), but the quadratic penalty gives rise to simpler optimization problems, and we focus on it in this paper. We now describe how to write quantization as a mapping in parameter space and how to solve the optimization problem .
Quantization consists of approximating real-valued vectors in a training set by vectors in a codebook. Since in our case the vectors are weights of a neural net, we will write the training set as $\{\w_1,\dots,\w_P\}$. Although in practice with neural nets we quantize scalar weight values directly (not weight vectors), we develop the formulation using vector quantization for generality. Hence, if we use a codebook $\calC = \{\c_1,\dots,\c_K\}$ with $K \ge 1$ entries, the number of bits used to store each weight vector $\w_i$ is $\ceil{\log_2{K}}$.
We consider two types of quantization: using an adaptive codebook, where we learn the optimal codebook for the training set; and using a fixed codebook, which is then not learned (although we will consider learning a global scale).
Adaptive codebook {#s:quant-adaptive}
-----------------
The decompression mapping is a table lookup $\w_i = \c_{\kappa(i)}$ for each weight vector $i=1,\dots,P$ in the codebook $\calC = \{\c_1,\dots,\c_K\}$, where $\kappa\mathpunct{:}$ $\{1,\dots,P\}$ $\rightarrow$ $\{1,\dots,K\}$ is a discrete mapping that assigns each weight vector to one codebook vector. The compression mapping results from finding the best (in the least-squares sense) codebook and mapping $\kappa$ for the “dataset” $\w_1,\dots,\w_P$, i.e., from solving the optimization problem $$\label{e:quant-mapping}
\min_{\calC,\kappa}{ \sum^P_{i=1}{ \norm{\w_i - \c_{\kappa(i)}}^2 } } \qquad \equiv \textcolor{blue}{\qquad \min_{\calC,\Z}{ \sum^{P,K}_{i,k=1}{ z_{ik} \norm{\w_i - \c_k}^2 } } \quad \text{s.t.} \quad
\begin{cases}
\Z \in \{0,1\}^{P\times K} \\
\sum^K_{k=1}{ z_{ik} } = 1,\ i = 1,\dots,P
\end{cases}}$$ which we have rewritten equivalently using binary assignment variables $\Z^T = (\z_1,\dots,\z_P)$. This follows by writing $\c_{\kappa(i)} = \smash{\sum^K_{k=1}{ z_{ik} \c_k }}$ where $z_{ik} = 1$ if $k = \kappa(i)$ and $0$ otherwise, and verifying by substituting the $z_{ik}$ values that the following holds: $$\norm{\w_i - \c_{\kappa(i)}}^2 = \norm[\Big]{\smash{\w_i - \sum^K_{k=1}{ z_{ik} \c_k }}}^2 = \sum^K_{k=1}{ z_{ik} \norm{\w_i - \c_k}^2 }.$$ So in this case the low-dimensional parameters are , the decompression mapping can be written elementwise as for $i = 1,\dots,P$, and the compression mapping results from running the $k$-means algorithm. The low-dimensional parameters are of two types: the assignments $\z_1,\dots,\z_P$ are “private” (each weight $\w_i$ has its own $\z_i$), and the codebook is “shared” by all weights. In the pseudocode of fig. \[f:pseudocode-adaptive\], we write the optimally quantized weights as $\w_{\calC} = \bDelta(\calC,\Z)$.
Problem is the well-known quadratic distortion problem [@GershoGray92a]. It is NP-complete and it is typically solved approximately by $k$-means using a good initialization, such as that of $k$-means++ [@ArthurVassil07a]. As is well known, $k$-means is an alternating optimization algorithm that iterates the following two steps: in the assignment step we update the assignments $\z_1,\dots,\z_P$ independently given the centroids (codebook); in the centroid step we update the centroids $\c_1,\dots,\c_K$ independently by setting them to the mean of their assigned points. Each iteration reduces the distortion or leaves it unchanged. The algorithm converges in a finite number of iterations to a local optimum where cannot improve given and vice versa.
In practice with neural nets we quantize scalar weight values directly, i.e., each $w_i$ is a real value. Computationally, $k$-means is considerably faster with scalar values than with vectors. If the vectors have dimension $D$, with $P$ data points and $K$ centroids, each iteration of $k$-means takes $\calO(PKD)$ runtime because of the assignment step (the centroid step is $\calO(PD)$, by scanning through the $P$ points and accumulating each mean incrementally). But in dimension $D=1$, each iteration can be done exactly in $\calO(P \log{K})$, by using a binary search over the sorted centroids in the assignment step, which then takes $\calO(K \log{K})$ for sorting and $\calO(P \log{K})$ for assigning, total $\calO(P \log{K})$.
### Why $k$-means?
The fact that we use $k$-means in the C step is not an arbitrary choice of a quantization algorithm (among many possible such algorithms we could use instead). It is a necessary consequence of two assumptions: 1) The fact that we want to assign weights to elements of a codebook, which dictates the form of the decompression mapping $\w = \bDelta(\calC,\Z)$. This is not really an assumption because any form of quantization works like this. 2) That the penalty used in the augmented Lagrangian is quadratic, so that the C step is a quadratic distortion problem.
We could choose a different penalty instead of the quadratic penalty $\smash{\norm{\w - \bDelta(\calC,\Z)}^2_2}$, as long as it is zero if the constraint $\w = \bDelta(\calC,\Z)$ is satisfied and positive otherwise (for example, the $\ell_1$ penalty). In the grand scheme of things, the choice of penalty is not important, because the role of the penalty is to enforce the constraints gradually, so that in the limit $\mu \rightarrow \infty$ the constraints are satisfied and the weights are quantized: $\w = \bDelta(\calC,\Z)$. Any penalty satisfying the positivity condition above will achieve this. The choice of penalty does have two effects: it may change the local optimum we converge to (although it is hard to have control on this); and, more importantly, it has a role in the optimization algorithm used in the L and C steps: the quadratic penalty is easier to optimize. As an example, imagine we used the $\ell_1$ penalty $\norm{\w - \bDelta(\calC,\Z)}_1$. This means that the L step would have the form: $$\min_{\w}{ L(\w) + \frac{\mu}{2} \norm[\Big]{\w - \bDelta(\bTheta) - \frac{1}{\mu} \blambda}_1 },$$ that is, an $\ell_1$-regularized loss. This is a nonsmooth problem. One can develop algorithms to optimize it, but it is harder than with the quadratic regularizer. The C step would have the form (again we write instead of $\w - \smash{\frac{1}{\mu}} \blambda$ for simplicity of notation): $$\min_{\bTheta}{ \norm{\w - \bDelta(\bTheta)}_1 } \qquad \Longleftrightarrow \qquad \min_{\calC,\Z}{ \sum^{P,K}_{i,k=1}{ z_{ik} \norm{\w_i - \c_k}_1 } } \quad \text{s.t.} \quad
\begin{cases}
\Z \in \{0,1\}^{P\times K} \\
\sum^K_{k=1}{ z_{ik} } = 1,\ i = 1,\dots,P.
\end{cases}$$ With scalar weights $w_1,\dots,w_P$, this can be solved by alternating optimization as in $k$-means: the assignment step is identical, but the centroid step uses the median instead of the mean of the points assigned to each centroid ($k$-medians algorithm). There are a number of other distortion measures developed in the quantization literature [@GershoGray92a section 10.3] that might be used as penalty and are perhaps convenient with some losses or applications. With a fixed codebook, as we will see in the next section, the form of the C step is the same regardless of the penalty.
On the topic of the choice of penalty, a possible concern one could raise is that of outliers in the data. When used for clustering, $k$-means is known to be sensitive to outliers and nonconvexities of the data distribution. Consider the following situations, for simplicity using just $K=1$ centroid in 1D. First, if the dataset has an outlier, it will pull the centroid towards it, away from the rest of the data (note this is not a local optima issue; this is the global optimum). For compression purposes, it may seem a waste of that centroid not to put it where most of the data is. With the $\ell_1$ penalty, the centroid would be insensitive to the outlier. Second, if the dataset consists of two separate groups, the centroid will end up in the middle of both, where there is no data, for both $k$-means and the $\ell_1$ penalty. Again, this may seem a waste of the centroid. Other clustering algorithms have been proposed to ensure the centroids lie where there is distribution mass, such as the $k$-modes algorithm [@CarreirWang13a; @WangCarreir14c]. However, these concerns are misguided, because neural net compression is not a data modeling problem: one has to consider the overall LC algorithm, not the C step in isolation. While in the C step the centroids approach the data (the weights), in the L step the weights approach the centroids, and in the limit $\mu \rightarrow \infty$ both coincide, the distortion is zero and there are no outliers. It is of course possible that the LC algorithm converge to a bad local optimum of the neural net quantization, which is an NP-complete problem, but this can happen for various reasons. In section \[s:expts:regression\] of the experiments we run the LC algorithm in a model whose weights contain clear outliers and demonstrate that the solution found makes sense.
Fixed codebook {#s:quant-fixed}
--------------
Now, we consider quantization using a fixed codebook[^2], i.e., the codebook entries $\calC = \{\c_1,\dots,\c_K\}$ are fixed and we do not learn them, we learn only the weight assignments $\Z^T = (\z_1,\dots,\z_P)$. In this way we can derive algorithms for compression of the weights based on approaches such as binarization or ternarization, which have been also explored in the literature of neural net compression, implemented as modifications to backpropagation (see section \[s:related:quant-fixed\]).
The compression mapping $\bPi(\w)$ of eq. now results from solving the optimization problem $$\label{e:quant-fixed-mapping}
\textcolor{blue}{\min_{\Z}{ \sum^{P,K}_{i,k=1}{ z_{ik} \norm{\w_i - \c_k}^2 } } \quad \text{s.t.} \quad
\begin{cases}
\Z \in \{0,1\}^{P\times K} \\
\sum^K_{k=1}{ z_{ik} } = 1,\ i = 1,\dots,P.
\end{cases}}$$ This is not NP-complete anymore, unlike in the optimization over codebook and assignments jointly in . It has a closed-form solution for each $\z_i$ separately where we assign $\w_i$ to , with ties broken arbitrarily, for $i = 1,\dots,P$. That is, each weight $\w_i$ is compressed as its closest codebook entry $\c_{\kappa(i)}$ (in Euclidean distance). Therefore, we can write the compression mapping $\bTheta = \bPi(\w)$ explicitly as separately for each weight $\w_i$, $i = 1,\dots,P$.
So in this case the low-dimensional parameters are (or ), the decompression mapping can be written elementwise as for $i = 1,\dots,P$ (as with the adaptive codebook), and the compression mapping can also be written elementwise as (or ) for $i = 1,\dots,P$. The low-dimensional parameters are all private (the assignments $\z_1,\dots,\z_P$ or $\kappa(1),\dots,\kappa(P)$). The codebook is shared by all weights, but it is not learned. In the pseudocode of fig. \[f:pseudocode-fixed\], we use the notation $\w_{\calC} = \bDelta(\calC,\kappa) = (\c_{\kappa(1)},\dots,\c_{\kappa(P)})$ to write the optimally quantized weights.
This simplifies further in the scalar case, i.e., when the weights $\w_i$ to be quantized are scalars. Here, we can write the codebook $\calC = \{c_1,c_2,\dots,c_K\}$ as an array of scalars sorted increasingly, $-\infty < c_1 < c_2 < \dots < c_K < \infty$. The elementwise compression mapping $\Pi(w_i) = \kappa(i) = \argmin_{k = 1,\dots,K}{ \abs{w_i - c_k} }$ can be written generically for $t \in \bbR$ as: $$\label{e:quant-fixed-mapping-scalar}
\textcolor{blue}{
\Pi(t) =
\begin{cases}
1, & \text{if } t < \frac{1}{2}(c_1+c_2) \\
2, & \text{if } \frac{1}{2}(c_1+c_2) \le t < \frac{1}{2}(c_2+c_3) \\
\dots \\
K, & \text{if } \frac{1}{2}(c_{K-1}+c_K) \le t
\end{cases}}$$ since the codebook defines Voronoi cells that are the intervals between midpoints of adjacent centroids. This can be written more compactly as $\Pi(t) = \kappa$ where $\kappa \in \{1,\dots,K\}$ satisfies $\frac{1}{2} (c_{\kappa-1} + c_{\kappa}) \le t < \frac{1}{2} (c_{\kappa} + c_{\kappa+1})$ and we define $c_0 = -\infty$ and $c_{K+1} = \infty$. Computationally, this can be done in $\calO(\log{K})$ using a binary search, although in practice $K$ is small enough that a linear search in $\calO(K)$ makes little difference. To use the compression mapping $\Pi(t)$ in the C step of the LC algorithm given in section \[s:LC\], $t$ equals either a scalar weight $w_i$ for the quadratic-penalty method, or a shifted scalar weight $w_i - \smash{\frac{1}{\mu}} \lambda_i$ for the augmented Lagrangian method. The L step of the LC algorithm always takes the form given in eq. .
Again, this quantization algorithm in the C step is not an arbitrary choice, it follows necessarily from the way any codebook-based quantization works. Furthermore, and unlike the adaptive codebook case, with scalar weights the solution is independent of the choice of penalty, because the order of the real numbers is unique (so using a quadratic or an $\ell_1$ penalty will result in the same step).
#### Application to binarization, ternarization and powers-of-two
Some particular cases of the codebook are of special interest because their implementation is very efficient: binary $\{-1,+1\}$, ternary $\{-1,0,+1\}$ and general powers-of-two $\{0,\pm 1,\pm 2^{-1},\dots,\pm 2^{-C}\}$. These are all well known in digital filter design, where one seeks to avoid floating-point multiplications by using fixed-point binary arithmetic and powers-of-two or sums of powers-of-two multipliers (which result in shift or shift-and-add operations instead). This accelerates the computation and requires less hardware.
We give the solution of the C step for these cases in fig. \[f:fixed-codebook\] (see proofs in the appendix). Instead of giving the compression mapping $\Pi(t)$, we give directly a *quantization operator* $q\mathpunct{:}\ \bbR \rightarrow \calC$ that maps a real-valued weight to its optimal codebook entry. Hence, $q$ corresponds to compressing then decompressing the weights, elementwise: $q(t) = \Delta(\calC,\Pi(t))$, where $t \in \bbR$ is a scalar weight. In the expressions for $q(t)$, we define the floor function for $t \in \bbR$ as $\floor{t} = i$ if $i \le t < i+1$ and $i$ is integer, and the sign function as follows: $$\label{e:sgn}
\sgn{t} =
\begin{cases}
-1, & \text{if } t < 0 \\
+1, & \text{if } t \ge 0.
\end{cases}$$ Note that the generic $k$-means algorithm (which occurs in the C step of our LC algorithm) solves problem , and hence its particular cases, exactly in one iteration: the centroid step does nothing (since the centroids are not learnable) and the assignment step is identical to the expressions for $\Pi(t)$ in eq. or for $q(t)$ in fig. \[f:fixed-codebook\]. However, the expressions in fig. \[f:fixed-codebook\] are more efficient, especially for the powers-of-two case, which runs in $\calO(1)$ (while the generic $k$-means assignment step would run in $\calO(\log{C})$).
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
\[t\]\[t\][$t$]{} \[\]\[\]\[1\]\[-90\][$q(t)$]{} \[r\]\[r\][ ]{} \[tl\]\[bl\] ![Scalar quantization operator $q\mathpunct{:}\ \bbR \rightarrow \calC$ using a fixed codebook $\calC = \{c_1,\dots,c_K\} \subset \bbR$ for some particular cases of interest (the general case without scale is given by eq. ). The input to the quantization is a set of $P$ real-valued weights; for the ternarization with scale these weights must be sorted in decreasing magnitude: $\abs{w_1} \ge \abs{w_2} \ge \dots \ge \abs{w_P}$. In $q(t)$, $t$ represents any weight $w_i$, and $q(w_i)$ gives the optimally quantized $w_i$. This solves the C step in the LC algorithm for quantization. See proofs in the appendix.[]{data-label="f:fixed-codebook"}](fixed-codebook.eps "fig:"){width="0.60\linewidth"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
\
[@lcl@]{} Name & codebook &\
Binarization & $\{-1,{+1}\}$ & $\alpha(t) = 1$\
Binarization with scale$^{\dagger}$ & $\{-a,{+a}\}$ & $\alpha(t) = a$\
\
Ternarization & $\{-1,0,{+1}\}$ & $\alpha(t) =
\begin{cases}
0, & \abs{t} < \frac{1}{2} \\
1, & \abs{t} \ge \frac{1}{2}.
\end{cases}$\
Ternarization with scale$^{\dagger}$ & $\{-a,0,{+a}\}$ & $\alpha(t) =
\begin{cases}
0, & \abs{t} < \frac{1}{2} a \\
a, & \abs{t} \ge \frac{1}{2} a.
\end{cases}$\
\
Powers of two & & $\alpha(t) =
\begin{cases}
0, & f > C+1 \\
1, & f \le 0 \\
2^{-C}, & f \in (C,C+1] \\
2^{-\floor{f + \log_2{\frac{3}{2}}}}, & \text{otherwise}
\end{cases}$\
& & where $f = -\log_2{\abs{t}}$.\
\
### Fixed codebook with adaptive scale {#s:quant-fixed:scale}
Fixed codebook values such as $\{-1,+1\}$ or $\{-1,0,+1\}$ may produce a large loss because the good weight values may be quite bigger or quite smaller than $\pm 1$. One improvement is to rescale the weights, or equivalently rescale the codebook elements, by a scale parameter $a \in \bbR$, which is itself learned. The low-dimensional parameters now are , where $a$ is a shared parameter and the $\z_1,\dots,\z_P$ are private. The decompression mapping can be written elementwise as for $i = 1,\dots,P$. The compression mapping results from solving the optimization problem $$\label{e:quant-fixed-mapping-scale}
\textcolor{blue}{\min_{\Z,a}{ \sum^{P,K}_{i,k=1}{ z_{ik} \norm{\w_i - a \, \c_k}^2 } } \quad \text{s.t.} \quad
\begin{cases}
\Z \in \{0,1\}^{P\times K} \\
\sum^K_{k=1}{ z_{ik} } = 1,\ i = 1,\dots,P.
\end{cases}}$$ In general, this can be solved by alternating optimization over and $a$:
- Assignment step: assign $\w_i$ to for $i = 1,\dots,P$.
- Scale step: .
Like $k$-means, this will stop in a finite number of iterations, and may converge to a local optimum. With scalar weights, each iteration is $\calO(P \log{K})$ by using binary search in the assignment step and incremental accumulation in the scale step.
#### Application to binarization and ternarization with scale
For some special cases we can solve problem exactly, without the need for an iterative algorithm. We give the solution for binarization and ternarization with scale in fig. \[f:fixed-codebook\] (see proofs in the appendix). Again, we give directly the scalar quantization operator $q\mathpunct{:}\ \bbR \rightarrow \calC$. The form of the solution is a rescaled version of the case without scale, where the optimal scale $a > 0$ is the average magnitude of a certain set of weights. Note that, given the scale $a$, the weights can be quantized elementwise by applying $q$, but solving for the scale involves all weights $w_1,\dots,w_P$.
Some of our quantization operators are equal to some rounding procedures used in previous work on neural net quantization: binarization (without scale) by taking the sign of the weight is well known, and our formula for binarization with scale is the same as in @Rasteg_16a. Ternarization with scale was considered by [@Li_16b], but the solution they give is only approximate; the correct, optimal solution is given in our theorem \[th:ter-scale\]. As we have mentioned before, those approaches incorporate rounding in the backpropagation algorithm in a heuristic way and the resulting algorithm does not solve problem . In the framework of the LC algorithm, the solution of the C step (the quantization operator) follows necessarily; there is no need for heuristics.
It is possible to consider more variations of the above, such as a codebook $\calC = \{-a,+b\}$ or $\{-a,0,+b\}$ with learnable scales $a,b > 0$, but there is little point to it. We should simply use a learnable codebook $\calC = \{c_1,c_2\}$ or $\{c_1,0,c_2\}$ without restrictions on $c_1$ or $c_2$ and run $k$-means in the C step.
Computing the optimal scale $a$ with $P$ weights has a runtime $\calO(P)$ in the case of binarization with scale and $\calO(P \log{P})$ in the case of ternarization with scale. In ternarization, the sums can be done cumulatively in $\calO(P)$, so the total runtime is dominated by the sort, which is $\calO(P \log{P})$. It may be possible to avoid the sort using a heap and reduce the total runtime to $\calO(P)$.
Experiments {#s:expts}
===========
We evaluate our learning-compression (LC) algorithm for quantizing neural nets of different sizes with different compression levels (codebook sizes $K$), in several tasks and datasets: linear regression on MNIST and classification on MNIST and CIFAR10. We compare LC with direct compression (DC) and iterated direct compression (iDC), which correspond to the previous works of @Gong_15a and @Han_15a, respectively. By using $K=2$ codebook values, we also compare with BinaryConnect [@Courbar_15a], which aims at learning binary weights. In summary, our experiments 1) confirm our theoretical arguments about the behavior of (i)DC, and 2) show that LC achieves comparable loss values (in training and test) to those algorithms with low compression levels, but drastically outperforms all them at high compression levels (which are the more desirable in practice). We reach the maximum possible compression (1 bit/weight) without significant error degradation in all networks we describe (except in the linear regression case).
We used the Theano [@Theano16a] and Lasagne [@Dielem_15a] libraries. Throughout we use the augmented Lagrangian, because we found it not only faster but far more robust than the quadratic penalty, in particular in setting the SGD hyperparameters. We initialize all algorithms from a reasonably (but not necessarily perfectly) well-trained reference net. The initial iteration ($\mu=0$) for LC gives the DC solution. The C step (also for iDC) consists of $k$-means ran till convergence, initialized from the previous iteration’s centroids (warm-start). For the first compression, we use the $k$-means++ initialization [@ArthurVassil07a]. This first compression may take several tens of $k$-means iterations, but subsequent ones need very few, often just one (figs. \[f:regression\] and \[f:kmeans-its\]).
We report the loss and classification error in training and test. We only quantize the multiplicative weights in the neural net, not the biases. This is because the biases span a larger range than the multiplicative weights, hence requiring higher precision, and anyway there are very few biases in a neural net compared to the number of multiplicative weights.
We calculate *compression ratios* as $$\label{e:compression-ratio}
\rho(K) = \text{\#bits(reference) / \#bits(quantized)}$$ where:
- \#bits(reference) $= (P_1+P_0) b$;
- \#bits(quantized) $= P_1 \ceil{\log_2{K}} + (P_0+K) b$, where $K b$ is the size of the codebook;
- $P_1$ and $P_0$ are the number of multiplicative weights and biases, respectively;
- $K$ is the codebook size;
- and we use 32-bit floats to represent real values (so $b=32$). Note that it is important to quote the base value of $b$ or otherwise the compression ratio is arbitrary and can be inflated. For example, if we set $b = 64$ (double precision) all the compression ratios in our experiments would double.
Since for our nets $P_0 \ll P_1$, we have $\rho(K) \approx b/\log_2{K}$.
Interplay between loss, model complexity and compression level
--------------------------------------------------------------
Firstly, we conduct a simple experiment to understand the interplay between loss, model complexity and compression level, here given by classification error, number of hidden units and codebook size, respectively. One important reason why compression is practically useful is that it may be better to train a large, accurate model and compress it than to train a smaller model and not compress it in the first place (there has been some empirical evidence supporting this, e.g. [@Denil_13a]). Also, many papers show that surprisingly large compression levels are possible with some neural nets (in several of our experiments with quantization, we can quantize all the way to one bit per weight with nearly no loss degradation). Should we expect very large compression levels without loss degradation in general?
The answer to these questions depends on the relation between loss, model complexity and compression. Here, we explore this experimentally in a simple setting: a classification neural net with inputs of dimension $D$, outputs of dimension $d$ (number of classes) and $H$ hidden, tanh units, fully connected, trained to minimize the average cross-entropy. We use our LC algorithm to quantize the net using a codebook of size $K$. The size $C(K,H)$ in bits of the resulting nets is as follows (assuming floating-point values of $b = 32$ bits). For the reference (non-quantized net, “$K = \infty$”), $C(\infty,H) = (D+d) H b$ (multiplicative weights) plus $(H+d)b$ (biases), total $C(\infty,H) \approx (D+d) H b$. For a quantized net, this is the sum of $(D+d) H \log_2{K}$ (for the quantized weights), $(H+d) b$ (for the non-quantized biases) and $K b$ (for the codebook), total $C(K,H) \approx (D+d) H \log_2{K}$.
We explore the space of optimal nets over $H$ and $K$ in order to determine what the best operational point $(K^*,H^*)$ is in order to achieve a target loss $L_{\text{max}}$ with the smallest net, that is, we want to solve the following optimization problem: $$\min_{K,H}{ C(K,H) } \qquad \text{s.t.} \qquad L(K,H) \le L_{\text{max}}.$$
We use the entire MNIST training set of 60000 handwritten digit images, hence $D = 784$ and $d = 10$. We train a reference net of $H$ units for $H \in \{2,\dots,40\}$ and compress it using a codebook of size $K$ for $\log_2{K} \in \{1,\dots,8\}$. The training procedure is exactly as for the LeNet300 neural net discussed later. Fig. \[f:L-K-H\] plots the loss $L(K,H)$ and size $C(K,H)$ for each resulting net using $(K,H)$, and the best operational point $(K^*,H^*)$ for target loss values $L_{\text{max}} \in \{0.005,0.01,0.05,0.3\}$.
\[\]\[\][number of hidden units $H$]{} \[\]\[\][$\log_2{K}$]{} \[l\]\[l\][[$L(K,H) \leq 0.3$]{}]{} \[l\]\[l\][[$L(K,H) \leq 0.05$]{}]{} \[l\]\[l\][[$L(K,H) \leq 0.01$]{}]{} \[l\]\[l\][[$L(K,H) \leq 0.005$]{}]{} \[t\]\[t\]
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
optimal loss $L(K,H)$ $L(K,H)$ and contours of $C(K,H)$
\[2ex\] ![Relation between loss, model complexity and quantization level in a single-layer neural net for MNIST classification. We train nets of different complexity (number of hidden units $H \in \{2,\dots,40\}$) and compress them with our LC algorithm using different codebook sizes ($\log_2{K} \in \{1,\dots,8\}$, with “$\infty$” meaning no compression, i.e., the reference net). *Left plot*: the resulting loss $L(K,H)$. We show in color four level sets $L_{\text{max}} \in \{0.005,0.01,0.05,0.3\}$ of $L$, i.e., the points $(K,H)$ satisfying $L(K,H) \le L_{\text{max}}$. *Middle plot*: the net size $C(K,H)$, with the same level sets over $L$. The color markers $\times$ identify the best operational point $(K^*,H^*)$ within each level set, i.e., the point $(K,H)$ having smallest size $C(K,H)$ such that $L(K,H) \le L_{\text{max}}$. *Right plot*: like the left plot but with the contours of $C(K,H)$ superimposed.[]{data-label="f:L-K-H"}](err_n_space_loss_psfragable.eps "fig:"){height="0.41\linewidth"} ![Relation between loss, model complexity and quantization level in a single-layer neural net for MNIST classification. We train nets of different complexity (number of hidden units $H \in \{2,\dots,40\}$) and compress them with our LC algorithm using different codebook sizes ($\log_2{K} \in \{1,\dots,8\}$, with “$\infty$” meaning no compression, i.e., the reference net). *Left plot*: the resulting loss $L(K,H)$. We show in color four level sets $L_{\text{max}} \in \{0.005,0.01,0.05,0.3\}$ of $L$, i.e., the points $(K,H)$ satisfying $L(K,H) \le L_{\text{max}}$. *Middle plot*: the net size $C(K,H)$, with the same level sets over $L$. The color markers $\times$ identify the best operational point $(K^*,H^*)$ within each level set, i.e., the point $(K,H)$ having smallest size $C(K,H)$ such that $L(K,H) \le L_{\text{max}}$. *Right plot*: like the left plot but with the contours of $C(K,H)$ superimposed.[]{data-label="f:L-K-H"}](err_n_space_mem_psfragable.eps "fig:"){height="0.41\linewidth"} ![Relation between loss, model complexity and quantization level in a single-layer neural net for MNIST classification. We train nets of different complexity (number of hidden units $H \in \{2,\dots,40\}$) and compress them with our LC algorithm using different codebook sizes ($\log_2{K} \in \{1,\dots,8\}$, with “$\infty$” meaning no compression, i.e., the reference net). *Left plot*: the resulting loss $L(K,H)$. We show in color four level sets $L_{\text{max}} \in \{0.005,0.01,0.05,0.3\}$ of $L$, i.e., the points $(K,H)$ satisfying $L(K,H) \le L_{\text{max}}$. *Middle plot*: the net size $C(K,H)$, with the same level sets over $L$. The color markers $\times$ identify the best operational point $(K^*,H^*)$ within each level set, i.e., the point $(K,H)$ having smallest size $C(K,H)$ such that $L(K,H) \le L_{\text{max}}$. *Right plot*: like the left plot but with the contours of $C(K,H)$ superimposed.[]{data-label="f:L-K-H"}](err_n_space_loss_psfragable_contour_final.eps "fig:"){height="0.41\linewidth"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Within a given level set, the points with large $H$ and $\log_2{K} = 1$ (top left in the plot) correspond to the regime “train a large reference net and compress it maximally”; the points with small $H$ and $\log_2{K} = \infty$ (bottom right in the plot) correspond to the regime “train a small reference net and do not compress it”; and intermediate points $(K,H)$ correspond to intermediate model sizes and moderate compression levels. As the plot shows, if the target loss is large (i.e., we do not require high classification accuracy) then maximal compression is optimal; but as the target loss increases (i.e., we require more accurate models), then the optimal point $(K^*,H^*)$ moves towards intermediate compression levels. If we require no loss degradation whatsoever, this might be only achievable without compression. Therefore, in general it is not clear what the optimal regime will be, and solving this model selection problem in practice will involve some trial and error of model sizes and compression levels. However, it will be often be the case that significant compression is achievable if we can tolerate a minor loss degradation. Compression also simplifies model selection: the neural net designer can simply overestimate the size of the net required to achieve a target loss, and let the compression find a smaller net with a similar loss. So it seems clear that *a good approximate strategy is to take a large enough model and compress it as much as possible*.
Quantizing linear regression, with a non-gaussian weight distribution {#s:expts:regression}
---------------------------------------------------------------------
\[l\]\[l\][[reference]{}]{} \[l\]\[l\][iDC]{} \[l\]\[l\][LC]{} \[t\]\[\][weight density]{} \[l\]\[l\][0]{} \[l\]\[l\][1]{} \[l\]\[l\][30]{} \[t\]\[\][\# $k$-means its.]{}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
\[b\]\[b\] ![Regression problem using a codebook of size $K = 4$ (row 1) and $K = 2$ (row 2), and training loss for each method (table). *Column 1*: loss over iterations. *Column 2*: weight distribution of the reference model (iteration 0, before quantization), direct compression DC (iteration 1), and the LC algorithm (iteration 30), using a kernel density estimate with manually selected bandwidth. The inset enlarges vertically the distributions to show the small cluster structure. The locations of the codebook centroids are shown below the distributions as markers: $+$ are the centroids fitted to the reference model and $\times$ the centroids at the end of the LC algorithm. *Column 3*: number of $k$-means iterations in each C step of the LC algorithm.[]{data-label="f:regression"}](toy_nested_loss.eps "fig:"){height="0.27\linewidth"} ![Regression problem using a codebook of size $K = 4$ (row 1) and $K = 2$ (row 2), and training loss for each method (table). *Column 1*: loss over iterations. *Column 2*: weight distribution of the reference model (iteration 0, before quantization), direct compression DC (iteration 1), and the LC algorithm (iteration 30), using a kernel density estimate with manually selected bandwidth. The inset enlarges vertically the distributions to show the small cluster structure. The locations of the codebook centroids are shown below the distributions as markers: $+$ are the centroids fitted to the reference model and $\times$ the centroids at the end of the LC algorithm. *Column 3*: number of $k$-means iterations in each C step of the LC algorithm.[]{data-label="f:regression"}](toy_histogram.eps "fig:"){height="0.27\linewidth"} ![Regression problem using a codebook of size $K = 4$ (row 1) and $K = 2$ (row 2), and training loss for each method (table). *Column 1*: loss over iterations. *Column 2*: weight distribution of the reference model (iteration 0, before quantization), direct compression DC (iteration 1), and the LC algorithm (iteration 30), using a kernel density estimate with manually selected bandwidth. The inset enlarges vertically the distributions to show the small cluster structure. The locations of the codebook centroids are shown below the distributions as markers: $+$ are the centroids fitted to the reference model and $\times$ the centroids at the end of the LC algorithm. *Column 3*: number of $k$-means iterations in each C step of the LC algorithm.[]{data-label="f:regression"}](toy_kmeans_n.eps "fig:"){height="0.27\linewidth"}
\[b\]\[b\] \[\]\[\][LC iterations]{} ![Regression problem using a codebook of size $K = 4$ (row 1) and $K = 2$ (row 2), and training loss for each method (table). *Column 1*: loss over iterations. *Column 2*: weight distribution of the reference model (iteration 0, before quantization), direct compression DC (iteration 1), and the LC algorithm (iteration 30), using a kernel density estimate with manually selected bandwidth. The inset enlarges vertically the distributions to show the small cluster structure. The locations of the codebook centroids are shown below the distributions as markers: $+$ are the centroids fitted to the reference model and $\times$ the centroids at the end of the LC algorithm. *Column 3*: number of $k$-means iterations in each C step of the LC algorithm.[]{data-label="f:regression"}](toy_nested_loss_k2.eps "fig:"){height="0.27\linewidth"} \[\]\[B\][weights $w_i$]{} ![Regression problem using a codebook of size $K = 4$ (row 1) and $K = 2$ (row 2), and training loss for each method (table). *Column 1*: loss over iterations. *Column 2*: weight distribution of the reference model (iteration 0, before quantization), direct compression DC (iteration 1), and the LC algorithm (iteration 30), using a kernel density estimate with manually selected bandwidth. The inset enlarges vertically the distributions to show the small cluster structure. The locations of the codebook centroids are shown below the distributions as markers: $+$ are the centroids fitted to the reference model and $\times$ the centroids at the end of the LC algorithm. *Column 3*: number of $k$-means iterations in each C step of the LC algorithm.[]{data-label="f:regression"}](toy_histogram_k2.eps "fig:"){height="0.27\linewidth"} \[\]\[\][LC iterations]{} ![Regression problem using a codebook of size $K = 4$ (row 1) and $K = 2$ (row 2), and training loss for each method (table). *Column 1*: loss over iterations. *Column 2*: weight distribution of the reference model (iteration 0, before quantization), direct compression DC (iteration 1), and the LC algorithm (iteration 30), using a kernel density estimate with manually selected bandwidth. The inset enlarges vertically the distributions to show the small cluster structure. The locations of the codebook centroids are shown below the distributions as markers: $+$ are the centroids fitted to the reference model and $\times$ the centroids at the end of the LC algorithm. *Column 3*: number of $k$-means iterations in each C step of the LC algorithm.[]{data-label="f:regression"}](toy_kmeans_n_k2.eps "fig:"){height="0.27\linewidth"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
\
[@c@c@]{}
------------------------------------------------------------------------------------
$\displaystyle L(\W,\b) = \frac{1}{N} \sum^N_{n=1}{ \norm{\y_n - \W\x_n - \b}^2 }$
\[3ex\]
------------------------------------------------------------------------------------
&
--------------------- -------------------- -------- -------- -------- -------
codebook compression
size $K$ ratio $\rho$
$4$ (2 bits/weight) $\approx\times 16$ 21.531 21.531 10.666 7.779
$2$ (1 bit/weight) $\approx\times 32$ 23.721 23.721 15.026 7.779
--------------------- -------------------- -------- -------- -------- -------
This experiment has two goals: 1) to verify in a controlled setting without local optima and with exact L and C steps that DC and iDC are identical to each other and significantly worse than LC. 2) To test our LC algorithm with a weight distribution that is far from Gaussian (unlike the weight distributions that typically arise with deep nets, which look zero-mean Gaussian). The problem is a simulated “super-resolution” task, where we want to recover a high-resolution image from a low-resolution one, by training a linear regression on pairs $(\x_n,\y_n) =$ (low-res, high-res), i.e., the loss is $L(\W,\b) = \smash{\frac{1}{N} \sum^N_{n=1}{ \norm{\y_n - \W\x_n - \b}^2 }}$, with weights and biases . We construct each low-resolution image by bicubic interpolation (using Matlab) of the high-resolution image . Ignoring border effects and slight nonlinearities, this means that each pixel (component) of is approximately a linear combination with constant coefficients of its corresponding pixels in . Hence, we can write the mapping from high to low resolution approximately as a linear mapping $\y = \A\x + \aa$, where $\aa = \0$ and the $i$th row of contains a few nonzero weights (the coefficients of the linear combination for pixel $x_i$). The ground-truth recovery matrix that optimizes the loss is then $\W = \A^+$, and it has a similar structure: roughly, each row contains only a few nonzeros, whose values are about the same across rows. We also add Gaussian noise when generating each low-resolution image $\x_n$, which spreads the optimal weights $w_{ij}$ around the ideal values above and also spreads the biases around zero. In summary, this means that the reference model weights have a clustered distribution, with a large cluster around zero, and a few small clusters at positive values.
To construct the dataset, we randomly selected $N=1\,000$ MNIST images $\y_n$ of $28 \times 28$, resized them as above to $14\times 14$ and added Gaussian noise to generate the $\x_n$, so that is of $784 \times 196$ ($P_1 =$ 153664 weights) and of $784 \times 1$ ($P_0 =$ 784). We compress using a codebook of size $K=2$ (1 bit per weight value, $\rho \approx \times$32) or $K=4$ (2 bits per weight value, $\rho \approx \times$16). The reference model and the L step have a single closed-form solution given by a linear system. For the LC algorithm, we increase $\mu_k = a^k \mu_0$ with $\mu_0 = 10$ and $a = 1.1$ for 30 iterations.
Fig. \[f:regression\] shows the results for codebook sizes $K=2$ and $4$. Firstly, as expected, DC and iDC do not change past the very first iteration, while LC achieves a much lower loss. The reference weight distribution (blue curve) shows a large cluster at zero and small clusters around 0.25 and 0.75 (see the inset). The small clusters correspond to the (inverse) bicubic interpolation coefficients, and it is crucial to preserve them in order to achieve a low loss. The LC algorithm indeed does a good job at this: with $K=2$ it places one centroid at zero and the other around $0.4$ (striking a balance between the small clusters); with $K=4$ it places one centroid at zero, two near the small clusters, and a fourth around $-0.1$. Note that the location of these centroids does not correspond to the reference model (the reference model quantization are the red $+$ markers), because the LC algorithm optimizes the centroids and the weights to achieve the lowest loss.
Quantizing LeNet neural nets for classification on MNIST {#s:expts:LeNet}
--------------------------------------------------------
We randomly split the MNIST training dataset (60k grayscale images of $28 \times 28$, 10 digit classes) into training (90%) and test (10%) sets. We normalize the pixel grayscales to \[0,1\] and then subtract the mean. We compress all layers of the network but each layer has its own codebook of size $K$. The loss is the average cross-entropy. To train a good reference net, we use Nesterov’s accelerated gradient method [@Nester83a] with momentum 0.9 for 100k minibatches, with a carefully fine-tuned learning rate $0.02 \cdot 0.99^j$ running 2k iterations for every $j$ (each a minibatch of 512 points). The $j$th L step parameters are given below for each net. For LC we also update $\mu$ and $\blambda$ at the end of the C step. Our LC algorithm uses a multiplicative schedule for the penalty parameter $\mu_j = \mu_0 a^j$ with $\mu_0 = 9.76 \cdot 10^{-5}$ and $a = 1.1$, for $0 \le j \le 30$. The batch size is 512 points for all methods.
We use the following neural nets, whose structure is given in table \[t:LeNet\]:
LeNet300
: This is a 3-layer densely connected feedforward net [@LeCun_98a] with tanh activations, having 266610 learnable parameters ($P_1 =$ 266200 weights and $P_0 =$ 410 biases). The $j$th L step (for LC and for iDC) runs 2k SGD iterations with momentum 0.95 and learning rate $0.1 \cdot 0.99^j$. We also trained a BinaryConnect net using the code of @Courbar_15a with deterministic rounding (without batch normalization), with $\alpha = 0.001$ and $\beta = 0.98$ after every 2k minibatch iterations, initialized from the reference net and trained for 120k minibatch iterations.
LeNet5
: This is a variation of the original LeNet5 convolutional net described in @LeCun_98a. It is included in Caffe[^3] [@Jia_14a] and was used by @Han_15a. It has ReLU activations [@NairHinton10a], dropout [@Srivas_14a] with $p = 0.5$ on the densely connected layers, and softmax outputs, total 431080 trainable parameters ($P_1 =$ 430500 weights and $P_0 =$ 580 biases). The $j$th L step (for LC and for iDC) runs 4k SGD iterations with momentum 0.95 and learning rate $\alpha \cdot 0.99^j$, where $\alpha=0.02$ for codebook sizes $K = 2,4,8$ and $\alpha=0.01$ for $K = 16,32,64$. This is because $\alpha = 0.02$ lead to divergence on iDC, even though LC was still able to converge.
[@ll@]{}\
Layer & Connectivity\
Input & $28 \times 28$ image\
1 &\
2 &\
&\
\
[@ll@]{}\
Layer & Connectivity\
Input & $28 \times 28$ image\
1 &\
2 &\
3 &\
4 &\
5 &\
&\
\
\[l\]\[l\][ref]{} \[l\]\[l\][iDC]{} \[l\]\[l\][LC]{} \[l\]\[l\][$\times$30]{} \[l\]\[l\][$\times$15]{} \[l\]\[l\][$\times$10]{} \[l\]\[l\][$\times$7.9]{} \[l\]\[l\][$\times$6.3]{} \[l\]\[l\][$\times$5.2]{}
---------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
LeNet300 LeNet5
\[\]\[\][time (s)]{} ![Learning curves over runtime and over SGD iterations (each marker $\bullet$ indicates one LC iteration, containing 2000 or 4000 SGD iterations), with different compression ratios (codebook sizes $K$), on the LeNet neural nets. Reference: dashed black horizontal line, LC: thick lines with markers, iDC: thin lines.[]{data-label="f:learning-curves"}](lenet300_nested_loss_vs_time.eps "fig:"){width="0.48\linewidth"} \[\]\[\][time (s)]{} ![Learning curves over runtime and over SGD iterations (each marker $\bullet$ indicates one LC iteration, containing 2000 or 4000 SGD iterations), with different compression ratios (codebook sizes $K$), on the LeNet neural nets. Reference: dashed black horizontal line, LC: thick lines with markers, iDC: thin lines.[]{data-label="f:learning-curves"}](lenet5_caffe_nested_loss_vs_time.eps "fig:"){width="0.48\linewidth"}
\[2ex\] ![Learning curves over runtime and over SGD iterations (each marker $\bullet$ indicates one LC iteration, containing 2000 or 4000 SGD iterations), with different compression ratios (codebook sizes $K$), on the LeNet neural nets. Reference: dashed black horizontal line, LC: thick lines with markers, iDC: thin lines.[]{data-label="f:learning-curves"}](lenet300_nested_loss.eps "fig:"){width="0.48\linewidth"} ![Learning curves over runtime and over SGD iterations (each marker $\bullet$ indicates one LC iteration, containing 2000 or 4000 SGD iterations), with different compression ratios (codebook sizes $K$), on the LeNet neural nets. Reference: dashed black horizontal line, LC: thick lines with markers, iDC: thin lines.[]{data-label="f:learning-curves"}](lenet5_caffe_nested_loss.eps "fig:"){width="0.48\linewidth"}
\[-3ex\] \[\]\[B\][SGD iterations $\times$2k]{} ![Learning curves over runtime and over SGD iterations (each marker $\bullet$ indicates one LC iteration, containing 2000 or 4000 SGD iterations), with different compression ratios (codebook sizes $K$), on the LeNet neural nets. Reference: dashed black horizontal line, LC: thick lines with markers, iDC: thin lines.[]{data-label="f:learning-curves"}](lenet300_val_errs.eps "fig:"){width="0.48\linewidth"} \[\]\[B\][SGD iterations $\times$4k]{} ![Learning curves over runtime and over SGD iterations (each marker $\bullet$ indicates one LC iteration, containing 2000 or 4000 SGD iterations), with different compression ratios (codebook sizes $K$), on the LeNet neural nets. Reference: dashed black horizontal line, LC: thick lines with markers, iDC: thin lines.[]{data-label="f:learning-curves"}](lenet5_caffe_val_errs.eps "fig:"){width="0.48\linewidth"}
---------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
---------------- -------------- ------- ----------- -------------------- ------------------- ----------- -------------------- ------------------- ----------- -------------------- -------------------
$\log{L}$ $E_{\text{train}}$ $E_{\text{test}}$ $\log{L}$ $E_{\text{train}}$ $E_{\text{test}}$ $\log{L}$ $E_{\text{train}}$ $E_{\text{test}}$
-3.87 0 2.28
$\times5.3$ 64 -4.33 0 2.25 -3.85 0 2.28 -4.41 0 2.25
$\times6.3$ 32 -4.29 0 2.25 -3.79 0 2.24 -4.37 0 2.24
\[0pt\]\[0pt\] $\times7.9$ 16 -4.20 0 2.25 -3.50 0 2.26 -4.18 0 2.20
$\times10.5$ 8 -3.99 0 2.29 -2.48 0.07 2.46 -3.31 0.004 2.34
$\times15.6$ 4 -3.62 0 2.44 -1.18 2.21 4.58 -1.77 0.543 3.23
$\times30.5$ 2 -3.10 0.009 2.42 -0.13 23.02 23.68 -0.61 5.993 7.98
-4.58 0 0.54
$\times5.3$ 64 -5.38 0 0.47 -4.54 0 0.52 -5.31 0 0.52
$\times6.3$ 32 -5.38 0 0.48 -4.47 0 0.49 -5.22 0 0.49
\[0pt\]\[0pt\] $\times7.9$ 16 -5.26 0 0.54 -4.24 0 0.49 -4.87 0 0.49
$\times10.5$ 8 -5.19 0 0.45 -3.42 0 0.58 -4.56 0 0.54
$\times15.7$ 4 -4.58 0 0.53 -1.94 0.29 0.94 -2.45 0.05 0.66
$\times30.7$ 2 -3.26 0.006 0.57 -0.00 15.77 15.62 -1.09 1.92 2.56
---------------- -------------- ------- ----------- -------------------- ------------------- ----------- -------------------- ------------------- ----------- -------------------- -------------------
\
\[l\]\[l\][DC]{} \[l\]\[l\][iDC]{} \[l\]\[l\][LC]{} \[l\]\[l\][LeNet300]{} \[l\]\[l\][LeNet5]{} \[B\]\[\][training loss $L$]{} \[B\]\[\][test error $E_{\text{test}}$ (%)]{} \[\]\[\][$\xleftarrow{\hspace{5ex}}$ codebook size $K$ $\xrightarrow{\hspace{5ex}}$]{}
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
$\xleftarrow{\hspace{5ex}}$ compression ratio $\rho$ $\xrightarrow{\hspace{5ex}}$ $\xleftarrow{\hspace{5ex}}$ compression ratio $\rho$ $\xrightarrow{\hspace{5ex}}$
![Compression results for the LeNet neural nets using different algorithms, for different codebook sizes $K$ and corresponding compression ratio $\rho$, in two forms: tabular (top) and graph (bottom). We report the training loss $\log_{10}{L}$ and training and test classification error $E_{\text{train}}$ and $E_{\text{test}}$ (%). The curves show the tradeoff between error vs compression ratio. Each LeNet neural net is shown in a different color, and each algorithm is shown in a different line type (LC: thick solid, iDC: thin dashed, DC: thin dotted).[]{data-label="f:tradeoff"}](tradeoff_train_loss.eps "fig:"){width="0.475\linewidth"} ![Compression results for the LeNet neural nets using different algorithms, for different codebook sizes $K$ and corresponding compression ratio $\rho$, in two forms: tabular (top) and graph (bottom). We report the training loss $\log_{10}{L}$ and training and test classification error $E_{\text{train}}$ and $E_{\text{test}}$ (%). The curves show the tradeoff between error vs compression ratio. Each LeNet neural net is shown in a different color, and each algorithm is shown in a different line type (LC: thick solid, iDC: thin dashed, DC: thin dotted).[]{data-label="f:tradeoff"}](tradeoff_val_err.eps "fig:"){width="0.475\linewidth"}
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
\[l\]\[l\][Layer 1]{} \[l\]\[l\][Layer 2]{} \[l\]\[l\][Layer 3]{} \[\]\[\][SGD iterations $\times$2k]{}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
learning-compression (LC) algorithm iterated direct compression (iDC)
\[B\]\[\][number of $k$-means iterations]{} ![Number of iterations ran within $k$-means over training for LeNet300 with a $K=4$ codebook.[]{data-label="f:kmeans-its"}](lenet300_lc_k_means_over_it.eps "fig:"){width="0.475\linewidth"} \[B\]\[\][number of $k$-means iterations]{} ![Number of iterations ran within $k$-means over training for LeNet300 with a $K=4$ codebook.[]{data-label="f:kmeans-its"}](lenet300_idc_k_means_over_it.eps "fig:"){width="0.475\linewidth"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
The following figures and tables show the results. Fig. \[f:learning-curves\] shows the learning curves and fig. \[f:tradeoff\] the error vs compression tradeoff. The runtime for iDC and LC is essentially the same, as is their loss and error at low compression levels. But for high compression LC is distinctly superior. When using $K=2$ (1 bit per weight), LC also outperforms BinaryConnect, as shown in table \[t:binary\]. Note the two weight values found by LC considerably differ from $\pm 1$ and depend on the layer, namely $\{0.089,-0.091\}$ in layer 1, $\{0.157,-0.155\}$ in layer 2 and $\{0.726,-0.787\}$ in layer 3. Indeed, forcing weights to be $\pm 1$ is more limiting, and does not have a practical advantage over using two arbitrary values in terms of storage or runtime when applying the compressed net to an input.
[@c@c@]{}
method
--------------- ------- -------- -------
reference -3.87 0% 2.28%
LC algorithm -3.10 0.009% 2.42%
BinaryConnect -2.33 0.14% 3.76%
: Binarization results in LeNet300 using the LC algorithm and BinaryConnect. LC uses a codebook of size $K=2$ (1 bit per weight) and the resulting codebook values are shown on the right (the BinaryConnect values are always $\pm 1$). The compression ratio is $\rho \approx \times 30.5$ for all methods.[]{data-label="t:binary"}
&
layer codebook values for LC
------- -----------------------------
1 $\calC = \{0.089, -0.091\}$
2 $\calC = \{0.157, -0.155\}$
3 $\calC = \{0.726, -0.787\}$
: Binarization results in LeNet300 using the LC algorithm and BinaryConnect. LC uses a codebook of size $K=2$ (1 bit per weight) and the resulting codebook values are shown on the right (the BinaryConnect values are always $\pm 1$). The compression ratio is $\rho \approx \times 30.5$ for all methods.[]{data-label="t:binary"}
Note how the LC training loss need not decrease monotonically (fig. \[f:learning-curves\]). This is to be expected: the augmented Lagrangian minimizes eq. for each $\mu$, not the actual loss, but it does approach a local optimum in the limit when $\mu \rightarrow \infty$. Also note how some compressed nets actually beat the reference. This is because the latter was close but not equal to a local optimum, due to the long training times required by SGD-type algorithms. Since the compressed nets keep training, they gain some accuracy over the reference.
Figs. \[f:weights-centroids-LC\]–\[f:weights-centroids-iDC\] show the evolution of weights and codebook centroids for $K=4$ for iDC and LC, for LeNet300. While LC converges to a feasible local optimum (note the delta-like weight distribution centered at the centroids), iDC does not. As we argued earlier, this is likely because iDC oscillates in a region half way between the reference net and its direct compression. These oscillations are very noticeable in the weight trajectories (right plots) for iDC in layers 1–2: the weights do not change their centroid assignment but oscillate around the centroid’s value. In contrast, in LC some weights do change their centroid assignment (this happens because the L step moves the weights jointly in $P$-dimensional space), and robustly converge to a centroid. The weight distribution of BinaryConnect (not shown) is also far from the values $\pm 1$ to which the weights should tend.
Fig. \[f:centroids-K\] shows the final locations of the centroids for iDC and LC, for codebook sizes $K = 2$ to $64$, for LeNet300. Although the general distribution of the centroids is similar for both methods, there are subtle differences (particularly for small $K$) which, as seen in fig. \[f:tradeoff\], translate into a significantly lower error for LC. For large enough $K$, the centroid distribution resembles that of the reference net. This is to be expected, because, with enough centroids, the optimally quantized network will be very similar to the reference and all 3 algorithms (DC, iDC and LC) will give the same result, namely quantizing the reference weights directly (regardless of the loss). Hence the centroid distribution will be the optimal quantization of the reference’s weight distribution. Since the latter is roughly Gaussian for the LeNet300 net, the centroids will reflect this, as seen in fig. \[f:centroids-K\] for the larger $K$. However, for compression purposes we are interested in the small-$K$ region, and here the final weight distribution can significantly differ from the reference. For all values of $K$, the distribution of the centroids for LC is usually symmetric around zero (but sometimes its mean is significantly different from zero, as in layer 3), and its spread increases from layer 1 to layer 2 to layer 3 in range but not in standard deviation. Although some of these observations may carry over to other types of neural nets, we emphasize that the reference weight distribution and hence the centroid distribution strongly depend on the problem (dataset, model and even particular local optimum found). Indeed, the clustered distribution for the regression problem of fig. \[f:regression\] was very different from Gaussian. Therefore, it seems risky to anticipate what the optimal codebook distribution may be. Finding a really accurate net must be done for each particular problem with a careful optimization using a good algorithm, and the same is true for finding a really well compressed net.
\[l\]\[l\][0]{} \[l\]\[l\][1]{} \[l\]\[l\][30]{}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
\[\]\[\][density, layer 1]{} ![Evolution of the centroid distribution over iterations for the LC algorithm for each layer of LeNet300 ($K=4$). *Left*: weight distribution for LC iterations 0, 1 and 30, using a kernel density estimate with manually selected bandwidth. The locations of the codebook centroids are shown below the distributions as markers: $+$ are the centroids fitted to the reference net and $\times$ the centroids at the end of the LC algorithm. *Right*: codebook centroids $c_k$ (blue) and 40 randomly chosen weights $w_i$ (red).[]{data-label="f:weights-centroids-LC"}](lenet300_lc_q4_histogram_W1.eps "fig:"){width="0.49\linewidth"} ![Evolution of the centroid distribution over iterations for the LC algorithm for each layer of LeNet300 ($K=4$). *Left*: weight distribution for LC iterations 0, 1 and 30, using a kernel density estimate with manually selected bandwidth. The locations of the codebook centroids are shown below the distributions as markers: $+$ are the centroids fitted to the reference net and $\times$ the centroids at the end of the LC algorithm. *Right*: codebook centroids $c_k$ (blue) and 40 randomly chosen weights $w_i$ (red).[]{data-label="f:weights-centroids-LC"}](lenet300_lc_weights_and_codebooks_layer1.eps "fig:"){width="0.49\linewidth"}
\[-2ex\] \[\]\[\][density, layer 2]{} ![Evolution of the centroid distribution over iterations for the LC algorithm for each layer of LeNet300 ($K=4$). *Left*: weight distribution for LC iterations 0, 1 and 30, using a kernel density estimate with manually selected bandwidth. The locations of the codebook centroids are shown below the distributions as markers: $+$ are the centroids fitted to the reference net and $\times$ the centroids at the end of the LC algorithm. *Right*: codebook centroids $c_k$ (blue) and 40 randomly chosen weights $w_i$ (red).[]{data-label="f:weights-centroids-LC"}](lenet300_lc_q4_histogram_W2.eps "fig:"){width="0.49\linewidth"} ![Evolution of the centroid distribution over iterations for the LC algorithm for each layer of LeNet300 ($K=4$). *Left*: weight distribution for LC iterations 0, 1 and 30, using a kernel density estimate with manually selected bandwidth. The locations of the codebook centroids are shown below the distributions as markers: $+$ are the centroids fitted to the reference net and $\times$ the centroids at the end of the LC algorithm. *Right*: codebook centroids $c_k$ (blue) and 40 randomly chosen weights $w_i$ (red).[]{data-label="f:weights-centroids-LC"}](lenet300_lc_weights_and_codebooks_layer2.eps "fig:"){width="0.49\linewidth"}
\[-2ex\] \[\]\[\][density, layer 3]{} \[\]\[\][weights $w_i$]{} ![Evolution of the centroid distribution over iterations for the LC algorithm for each layer of LeNet300 ($K=4$). *Left*: weight distribution for LC iterations 0, 1 and 30, using a kernel density estimate with manually selected bandwidth. The locations of the codebook centroids are shown below the distributions as markers: $+$ are the centroids fitted to the reference net and $\times$ the centroids at the end of the LC algorithm. *Right*: codebook centroids $c_k$ (blue) and 40 randomly chosen weights $w_i$ (red).[]{data-label="f:weights-centroids-LC"}](lenet300_lc_q4_histogram_W3.eps "fig:"){width="0.49\linewidth"} \[\]\[\][SGD iterations $\times$2k]{} ![Evolution of the centroid distribution over iterations for the LC algorithm for each layer of LeNet300 ($K=4$). *Left*: weight distribution for LC iterations 0, 1 and 30, using a kernel density estimate with manually selected bandwidth. The locations of the codebook centroids are shown below the distributions as markers: $+$ are the centroids fitted to the reference net and $\times$ the centroids at the end of the LC algorithm. *Right*: codebook centroids $c_k$ (blue) and 40 randomly chosen weights $w_i$ (red).[]{data-label="f:weights-centroids-LC"}](lenet300_lc_weights_and_codebooks_layer3.eps "fig:"){width="0.49\linewidth"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
\[l\]\[l\][0]{} \[l\]\[l\][1]{} \[l\]\[l\][30]{}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------
\[\]\[\][density, layer 1]{} ![Like fig. \[f:weights-centroids-LC\] but for iDC.](lenet300_idc_q4_histogram_W1.eps "fig:"){width="0.49\linewidth"} ![Like fig. \[f:weights-centroids-LC\] but for iDC.](lenet300_idc_weights_and_codebooks_layer1.eps "fig:"){width="0.49\linewidth"}
\[-2ex\] \[\]\[\][density, layer 2]{} ![Like fig. \[f:weights-centroids-LC\] but for iDC.](lenet300_idc_q4_histogram_W2.eps "fig:"){width="0.49\linewidth"} ![Like fig. \[f:weights-centroids-LC\] but for iDC.](lenet300_idc_weights_and_codebooks_layer2.eps "fig:"){width="0.49\linewidth"}
\[-2ex\] \[\]\[\][density, layer 3]{} \[\]\[B\][weights $w_i$]{} ![Like fig. \[f:weights-centroids-LC\] but for iDC.](lenet300_idc_q4_histogram_W3.eps "fig:"){width="0.49\linewidth"} \[\]\[B\][SGD iterations $\times$2k]{} ![Like fig. \[f:weights-centroids-LC\] but for iDC.](lenet300_idc_weights_and_codebooks_layer3.eps "fig:"){width="0.49\linewidth"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------
\[f:weights-centroids-iDC\]
\[c\]\[c\][codebook size $K$]{} \[l\]\[l\][iDC]{} \[l\]\[l\][LC]{} \[c\]\[c\][$\infty$]{} \[l\]\[l\][LC-mean]{} \[l\]\[l\][LC-stdev]{} \[l\]\[l\][iDC-mean]{} \[l\]\[l\][iDC-stdev]{} \[l\]\[l\] \[\]\[\][centroids $c_k$]{} \[\]\[\]
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Distribution of the centroids learnt by the LC and iDC algorithms, for layers 1–3 of LeNet300, for codebook sizes $K = 2$ to $64$. *Top row*: actual centroid locations $c_k$, $k = 1,\dots,K$. The distribution of the weights of the reference net is shown at the top as a kernel density estimate. *Bottom row*: mean and standard deviation of the centroid set $\calC = \{c_1,\dots,c_K\}$ (“$\infty$” corresponds to no quantization, i.e., the mean and standard deviation of the reference net).[]{data-label="f:centroids-K"}](lenet300_means_k_layer1.eps "fig:"){width="0.33\linewidth"} ![Distribution of the centroids learnt by the LC and iDC algorithms, for layers 1–3 of LeNet300, for codebook sizes $K = 2$ to $64$. *Top row*: actual centroid locations $c_k$, $k = 1,\dots,K$. The distribution of the weights of the reference net is shown at the top as a kernel density estimate. *Bottom row*: mean and standard deviation of the centroid set $\calC = \{c_1,\dots,c_K\}$ (“$\infty$” corresponds to no quantization, i.e., the mean and standard deviation of the reference net).[]{data-label="f:centroids-K"}](lenet300_means_k_layer2.eps "fig:"){width="0.33\linewidth"} ![Distribution of the centroids learnt by the LC and iDC algorithms, for layers 1–3 of LeNet300, for codebook sizes $K = 2$ to $64$. *Top row*: actual centroid locations $c_k$, $k = 1,\dots,K$. The distribution of the weights of the reference net is shown at the top as a kernel density estimate. *Bottom row*: mean and standard deviation of the centroid set $\calC = \{c_1,\dots,c_K\}$ (“$\infty$” corresponds to no quantization, i.e., the mean and standard deviation of the reference net).[]{data-label="f:centroids-K"}](lenet300_means_k_layer3.eps "fig:"){width="0.33\linewidth"}
\[2ex\] ![Distribution of the centroids learnt by the LC and iDC algorithms, for layers 1–3 of LeNet300, for codebook sizes $K = 2$ to $64$. *Top row*: actual centroid locations $c_k$, $k = 1,\dots,K$. The distribution of the weights of the reference net is shown at the top as a kernel density estimate. *Bottom row*: mean and standard deviation of the centroid set $\calC = \{c_1,\dots,c_K\}$ (“$\infty$” corresponds to no quantization, i.e., the mean and standard deviation of the reference net).[]{data-label="f:centroids-K"}](mean_std_codebook_layer1.eps "fig:"){width="0.33\linewidth"} ![Distribution of the centroids learnt by the LC and iDC algorithms, for layers 1–3 of LeNet300, for codebook sizes $K = 2$ to $64$. *Top row*: actual centroid locations $c_k$, $k = 1,\dots,K$. The distribution of the weights of the reference net is shown at the top as a kernel density estimate. *Bottom row*: mean and standard deviation of the centroid set $\calC = \{c_1,\dots,c_K\}$ (“$\infty$” corresponds to no quantization, i.e., the mean and standard deviation of the reference net).[]{data-label="f:centroids-K"}](mean_std_codebook_layer2.eps "fig:"){width="0.33\linewidth"} ![Distribution of the centroids learnt by the LC and iDC algorithms, for layers 1–3 of LeNet300, for codebook sizes $K = 2$ to $64$. *Top row*: actual centroid locations $c_k$, $k = 1,\dots,K$. The distribution of the weights of the reference net is shown at the top as a kernel density estimate. *Bottom row*: mean and standard deviation of the centroid set $\calC = \{c_1,\dots,c_K\}$ (“$\infty$” corresponds to no quantization, i.e., the mean and standard deviation of the reference net).[]{data-label="f:centroids-K"}](mean_std_codebook_layer3.eps "fig:"){width="0.32\linewidth"}
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Figs. \[f:L1-neurons\]–\[f:L2\_L3-neurons\] show the reference and final weights for LeNet300 compressed by the LC algorithm using a codebook of size $K = 2$ (a separate codebook for each of the 3 layers), which gives binary weights. For many of the weights, the sign of the quantized weight in LC equals the sign of the corresponding reference net weight. However, other weights change side during the optimization, specifically 5.04%, 3.22% and 1% of the weights for layers 1, 2 and 3, respectively.
![Weight vector of the 300 neurons of layer 1 for LeNet300 for the reference net (left image of each horizontal pair, or odd-numbered columns) and for the net compressed with the LC algorithm using $K=2$ (right image of each pair, or even-numbered columns). We show each weight vector as a $28 \times 28$ image. All images have been globally normalized to the interval $[-3.5\sigma_1,3.5\sigma_1]$, where $\sigma_1$ is the standard deviation of the layer–1 reference weights (weights outside this interval are mapped to the respective end of the interval).[]{data-label="f:L1-neurons"}](out_first_last.eps){width="0.83\linewidth"}
\[c\]\[c\][$K$]{} \[l\]\[l\][Layer 2]{} \[l\]\[l\][Layer 3]{} \[l\]\[l\][iDC]{} \[l\]\[l\][LC]{} \[\]\[\][SGD iterations $\times$2k]{}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
reference LC algorithm reference
![Weights of layers 2 and 3 of LeNet300 for the reference net and for the net compressed with the LC algorithm using $K=2$. We show layer 2 as a matrix of $300 \times 100$ and layer 3 as a matrix of $100 \times 10$. The normalization is as in fig. \[f:L1-neurons\], i.e., to the interval $[-3.5\sigma_i,3.5\sigma_i]$, where $\sigma_i$ is the standard deviation of the $i$th layer reference weights, for $i = 2$ or $3$.[]{data-label="f:L2_L3-neurons"}](w2_iteration_01.eps "fig:"){width="1.10\linewidth"} ![Weights of layers 2 and 3 of LeNet300 for the reference net and for the net compressed with the LC algorithm using $K=2$. We show layer 2 as a matrix of $300 \times 100$ and layer 3 as a matrix of $100 \times 10$. The normalization is as in fig. \[f:L1-neurons\], i.e., to the interval $[-3.5\sigma_i,3.5\sigma_i]$, where $\sigma_i$ is the standard deviation of the $i$th layer reference weights, for $i = 2$ or $3$.[]{data-label="f:L2_L3-neurons"}](w2_iteration_32.eps "fig:"){width="1.10\linewidth"} ![Weights of layers 2 and 3 of LeNet300 for the reference net and for the net compressed with the LC algorithm using $K=2$. We show layer 2 as a matrix of $300 \times 100$ and layer 3 as a matrix of $100 \times 10$. The normalization is as in fig. \[f:L1-neurons\], i.e., to the interval $[-3.5\sigma_i,3.5\sigma_i]$, where $\sigma_i$ is the standard deviation of the $i$th layer reference weights, for $i = 2$ or $3$.[]{data-label="f:L2_L3-neurons"}](w3_iteration_01.eps "fig:"){width="1.10\linewidth"} ![Weights of layers 2 and 3 of LeNet300 for the reference net and for the net compressed with the LC algorithm using $K=2$. We show layer 2 as a matrix of $300 \times 100$ and layer 3 as a matrix of $100 \times 10$. The normalization is as in fig. \[f:L1-neurons\], i.e., to the interval $[-3.5\sigma_i,3.5\sigma_i]$, where $\sigma_i$ is the standard deviation of the $i$th layer reference weights, for $i = 2$ or $3$.[]{data-label="f:L2_L3-neurons"}](w3_iteration_32.eps "fig:"){width="1.10\linewidth"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Quantizing a large deep net for classification on CIFAR10
---------------------------------------------------------
We randomly split the CIFAR10 training dataset (60k color images of $32 \times 32 \times 3$, 10 object classes) into training (90%) and test (10%) sets. We normalize the pixel colors to \[0,1\] and then subtract the mean. We train a 12-layer convolutional neural network inspired by the VGG net [@SimonyZisser15a] and described by @Courbar_15a, with the structure shown in table \[t:CIFAR\]. All convolutions are symmetrically padded with zero (padding size 1). The network has 14 million parameters ($P_1 =$ 14022016 weights and $P_0 =$ 3850 biases). Because of time considerations (each experiment takes 18 hours) we only report performance of LC with respect to the reference net. The reference net achieves 13.15% error on the test set and a training loss of $1.1359 \cdot 10^{-7}$. Compressing with a $K=2$ codebook (compression ratio $\rho \approx \times$31.73), LC achieves a *lower* test error of 13.03% and a training loss of $1.6242 \cdot 10^{-5}$.
[@ll@]{} Layer & Connectivity\
Input & $3\times 32 \times 32$ image\
1 &\
2 &\
3 &\
4 &\
5 &\
6 &\
7 &\
8 &\
9 &\
10 &\
11 &\
&\
\
Conclusion {#s:concl}
==========
Neural net quantization involves minimizing the loss over weights taking discrete values, which makes the objective function nondifferentiable. We have reformulated this as optimizing the loss subject to quantization constraints, which is a mixed discrete-continuous problem, and given an iterative “learning-compression” (LC) algorithm to solve it. This alternates two steps: a learning step that optimizes the usual loss with a quadratic regularization term, which can be solved by SGD; and a compression step, independent of the loss and training set, which quantizes the current real-valued weights. The compression step takes the form of $k$-means if the codebook is adaptive, or of an optimal assignment and rescaling if the codebook is (partially) fixed, as for binarization. The algorithm is guaranteed to converge to a local optimum of the quantization problem, which is NP-complete. Experimentally, this LC algorithm beats previous approaches based on quantizing the reference net or on incorporating rounding into backpropagation. It often reaches the maximum possible compression (1 bit/weight) without significant loss degradation.
### Acknowledgements {#acknowledgements .unnumbered}
Work supported by NSF award IIS–1423515, by a UC Merced Faculty Research Grant and by a Titan X Pascal GPU donated by the NVIDIA Corporation.
Theorems and proofs
===================
We prove several results concerning quantization with a fixed codebook (section \[s:quant-fixed\]) and give an alternative formulation for the case of binarization.
Optimal quantization using a fixed codebook with or without scale
-----------------------------------------------------------------
We prove the optimal quantization results for binarization, ternarization and powers-of-two of section \[s:quant-fixed\]. The formulas for binarization and ternarization without scale follow from eq. . The formulas for the powers-of-two and binarization and ternarization with scale are given in the theorems below. Define the sign function $\sgn{}$ as in eq. , and the floor function for $t \in \bbR$ as $\floor{t} = i$ if $i \le t < i+1$ and $i$ is integer.
\[th:pow2\] Let $w \in \bbR$ and $C \ge 0$ integer. The solution $\theta^*$ of the problem $$\label{e:pow2}
\min_{\theta}{ E(\theta) = (w - \theta)^2 } \qquad \text{s.t.} \qquad \theta \in \{0,\pm 1,\pm 2^{-1},\dots,\pm 2^{-C}\}$$ is $\theta^* = \alpha \, \sgn{w}$ where $$\label{e:pow2-sol}
\alpha =
\begin{cases}
0, & f > C+1 \\
1, & f \le 0 \\
2^{-C}, & f \in (C,C+1] \\
2^{-\floor{f + \log_2{\frac{3}{2}}}}, & \text{otherwise}
\end{cases}$$ and $f = -\log_2{\abs{w}}$.
The sign of $\theta^*$ is obviously equal to the sign of $w$, so consider $w > 0$ and call $f = -\log_2{w}$. The solution can be written as a partition of $\bbR^+$ in four intervals $[0,2^{-C-1})$, $[2^{-C-1},2^{-C})$, $[2^{-C},1)$ and $[1,\infty)$ for $w$, or equivalently $(C+1,\infty)$, $(C,C+1]$, $(0,C]$ and $(-\infty,0]$ for $f$. These intervals are optimally assigned to centroids $0$, $2^{-C}$, $2^{-\floor{f + \log_2{\frac{3}{2}}}}$ and $1$, respectively. The solutions for the first, second and fourth intervals are obvious. The solution for the third interval $w \in [2^{-C},1) \Leftrightarrow f \in (0,C]$ is as follows. The interval of $\bbR^+$ that is assigned to centroid $2^{-i}$ for $i \in \{0,1,\dots,C\}$ is $w \in (3 \cdot 2^{-i-2},3 \cdot 2^{-i-1}]$, given by the midpoints between centroids (and breaking ties as shown), or equivalently $f + \log_2{\frac{3}{2}} \in [i,i+1)$, and so $i = \floor{f + \log_2{\frac{3}{2}}}$.
\[th:bin-scale\] Let $w_1,\dots,w_P \in \bbR$. The solution $(a^*,\btheta^*)$ of the problem $$\label{e:bin-scale}
\min_{a,\btheta}{ E(a,\btheta) = \sum^P_{i=1}{ (w_i - a \, \theta_i)^2 } } \qquad \text{s.t.} \qquad a \in \bbR,\quad \theta_1,\dots,\theta_P \in \{-1,+1\}$$ is $$\label{e:bin-scale-sol}
a^* = \frac{1}{P} \sum^P_{i=1}{ \abs{w_i} } \qquad \theta^*_i = \sgn{w_i} =
\begin{cases}
-1, & \text{if } w_i < 0 \\
+1, & \text{if } w_i \ge 0
\end{cases}
\quad \text{for } i = 1,\dots,P.$$
For any $a \in \bbR$, the solution $\btheta^*(a)$ for results from minimizing $E(a,\btheta)$ over (breaking the ties as shown in ). Substituting this into the objective function: $$\begin{gathered}
E(a,\btheta^*(a)) = \sum^P_{i=1}{ (w_i - a \, \sgn{w_i})^2 } = \sum^P_{i=1}{ \left( w^2_i + a^2 (\sgn{w_i})^2 - 2 a \, w_i \sgn{w_i} \right) } \\
= \bigg( \sum^P_{i=1}{ w^2_i } \bigg) + P a^2 - 2 a \, \sum^P_{i=1}{\abs{w_i}}.
\end{gathered}$$ The result for $a^*$ follows from differentiating wrt $a$ and equating to zero.
\[th:ter-scale\] Let $w_1,\dots,w_P \in \bbR$ and assume w.l.o.g. that $\abs{w_1} \ge \abs{w_2} \ge \dots \ge \abs{w_P}$. The solution $(a^*,\btheta^*)$ of the problem $$\label{e:ter-scale}
\min_{a,\btheta}{ E(a,\btheta) = \sum^P_{i=1}{ (w_i - a \, \theta_i)^2 } } \qquad \text{s.t.} \qquad a \in \bbR,\quad \theta_1,\dots,\theta_P \in \{-1,0,+1\}$$ is $$\label{e:ter-scale-sol}
j^* = \argmax_{1 \le j \le P}{ \frac{1}{\sqrt{j}} \sum^j_{i=1}{ \abs{w_i} } } \qquad a^* = \frac{1}{j^*} \sum^{j^*}_{i=1}{ \abs{w_i} } \qquad \theta^*_i =
\begin{cases}
0, & \text{if } \abs{w_i} < a^*/2 \\
\sgn{w_i}, & \text{if } \abs{w_i} \ge a^*/2
\end{cases}
\quad \text{for } i = 1,\dots,P.$$
For any $a \in \bbR$, the solution $\btheta^*(a)$ for results from minimizing $E(a,\btheta)$ over (breaking the ties as shown in ). Substituting this into the objective function: $$E(a,\btheta^*(a)) = \sum_{i \in \overline{\calS}}{ w^2_i } + \sum_{i \in \calS}{ (w_i - a \, \sgn{w_i})^2 } = \bigg( \sum^P_{i=1}{ w^2_i } \bigg) + \abs{\calS} a^2 - 2 a \, \sum_{i \in \calS}{ \abs{w_i} }$$ where $\calS = \{i \in \{1,\dots,P\}\mathpunct{:}\ \abs{w_i} \ge a\}$, $\overline{\calS} = \{1,\dots,P\} \setminus \calS$ and $\abs{\calS}$ is the cardinality of . Differentiating wrt $a$ keeping fixed and equating to zero yields $a = \frac{1}{\abs{\calS}} \sum_{i \in \calS}{ \abs{w_i} }$. It only remains to find the set $\calS^*$ that is consistent with the previous two conditions on and $a$. Since the $w_i$ are sorted in decreasing magnitude, there are $P$ possible sets that can be and they are of the form $\calS_i = \{w_1,\dots,w_i\} = \{w_j\mathpunct{:}\ \abs{w_j} \ge w_i\}$ for $i = 1,\dots,P$, and $\calS^* = \calS_{j^*}$ is such that the objective function is maximal. Hence, calling $a_j = \frac{1}{\abs{\calS_j}} \sum_{i \in \calS_j}{ \abs{w_i} }$ and noting that $$E(a_j,\btheta^*(a_j)) = \bigg( \sum^P_{i=1}{ w^2_i } \bigg) - \abs{\calS_j} a^2_j = \bigg( \sum^P_{i=1}{ w^2_i } \bigg) - j \, a^2_j$$ we have $$j^* = \argmin_{1 \le j \le P}{ E(a_j,\btheta^*(a_j)) } = \argmax_{1 \le j \le P}{ j \, a^2_j } = \argmax_{1 \le j \le P}{ \sqrt{j} \, a_j } = \argmax_{1 \le j \le P}{ \frac{1}{\sqrt{j}} \sum^j_{i=1}{ \abs{w_i} } }.$$ Finally, let us prove that the set $\calS_{j^*}$ is consistent with $a_{j^*}$ and $\btheta(a_{j^*})$, i.e., that $\calS_{j^*} = \{i \in \{1,\dots,P\}\mathpunct{:}\ \abs{w_i} \ge \frac{1}{2} a_{j^*}\}$. Since the $w_i$ are sorted in decreasing magnitude, it suffices to prove that $\abs{w_{j^*}} > \frac{1}{2} a_{j^*} > \abs{w_{j^*+1}}$. Since $j^* = \argmax_{1 \le j \le P}{ \sqrt{j} \, a_j }$, we have (in the rest of the proof we write $j$ instead of $j^*$ to avoid clutter): $$\sqrt{j} \, a_j \ge \sqrt{j+1} \, a_{j+1} = \sqrt{j+1} \, \bigg( \frac{1}{j+1} \sum^{j+1}_{i=1}{ \abs{w_i} } \bigg) = \frac{1}{\sqrt{j+1}} \, ( \abs{w_{j+1}} + j \, a_j ) \Leftrightarrow \abs{w_{j+1}} \le a_j \left( \sqrt{j(j+1)} - j \right).$$ Now $\sqrt{j(j+1)} - j < \frac{1}{2}$ (since $j(j+1) < (j + \frac{1}{2})^2 = j^2 + \frac{1}{4} + j$ $\forall j$), hence $\abs{w_{j+1}} < \frac{1}{2} a_j$. Likewise: $$\sqrt{j} \, a_j \ge \sqrt{j-1} \, a_{j-1} = \frac{1}{\sqrt{j+1}} \, ( j \, a_j - \abs{w_j} ) \Leftrightarrow \abs{w_j} \ge a_j \left( j - \sqrt{j(j-1)} \right).$$ Now $j - \sqrt{j(j-1)} > \frac{1}{2}$ (since $j(j-1) < (j - \frac{1}{2})^2 = j^2 + \frac{1}{4} - j$ $\forall j$), hence $\abs{w_j} > \frac{1}{2} a_j$.
An equivalent formulation for binarization
------------------------------------------
We show that, in the binarization case (with or without scale), our constrained optimization formulation “$\min_{\w,\bTheta}{ L(\w) } \text{ s.t.\ } \w = \bDelta(\bTheta)$” of eq. can be written equivalently without using assignment variables , as follows: $$\begin{aligned}
\text{Binarization: } & \min_{\w,\b}{ L(\w) } \quad \text{s.t.} \quad \w = \b,\ \b \in \{-1,+1\}^P \\
\text{Binarization with scale: } & \min_{\w,\b,a}{ L(\w) } \quad \text{s.t.} \quad \w = a \, \b,\ \b \in \{-1,+1\}^P,\ a > 0.\end{aligned}$$ We can write an augmented-Lagrangian function as (for the quadratic-penalty function, set $\blambda = \0$): $$\begin{aligned}
\calL_A(\w,\b,\blambda;\mu) &= L(\w) - \blambda^T (\w - \b) + \frac{\mu}{2} \norm{\w - \b}^2 \quad \text{s.t.} \quad \b \in \{-1,+1\}^P \\
\calL_A(\w,\b,a,\blambda;\mu) &= L(\w) - \blambda^T (\w - a \, \b) + \frac{\mu}{2} \norm{\w - a \, \b}^2 \quad \text{s.t.} \quad \b \in \{-1,+1\}^P,\ a > 0.\end{aligned}$$ And applying alternating optimization gives the steps:
- L step: $\min_{\w}{ L(\w) + \frac{\mu}{2} \norm{\w - \b - \smash{\frac{1}{\mu}} \blambda}^2 }$ or $\min_{\w}{ L(\w) + \frac{\mu}{2} \norm{\w - a \, \b - \smash{\frac{1}{\mu}} \blambda}^2 }$, respectively.
- C step: $\b = \sgn{\w}$ or $\b = a \, \sgn{\w}$ and $a = \frac{1}{P} \sum^P_{i=1}{\abs{w_i}}$, respectively, an elementwise binarization and global rescaling.
This is identical to our LC algorithm.
[31]{} \[1\][\#1]{} \[1\][`#1`]{} urlstyle \[1\][doi: \#1]{}
D. Arthur and S. Vassilvitskii. `k-means++`: The advantages of careful seeding. In *Proc. of the 18th ACM-SIAM Symposium on Discrete Algorithms (SODA 2007)*, pages 1027–1035, New Orleans, LA, Jan. 7–9 2007.
M. [Á]{}. Carreira-Perpi[ñ]{}[á]{}n. Model compression as constrained optimization, with application to neural nets. [Part]{} [I]{}: General framework. arXiv:1707.01209 \[cs.LG\], July 5 2017.
M. [Á]{}. Carreira-Perpi[ñ]{}[á]{}n and W. Wang. The [$K$]{}-modes algorithm for clustering. arXiv:1304.6478 \[cs.LG\], Apr. 23 2013.
M. Courbariaux, Y. Bengio, and J.-P. David. : Training deep neural networks with binary weights during propagations. In C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett, editors, *Advances in Neural Information Processing Systems (NIPS)*, volume 28, pages 3105–3113. MIT Press, Cambridge, MA, 2015.
M. Denil, B. Shakibi, L. Dinh, M. Ranzato, and N. de Freitas. Predicting parameters in deep learning. In C. J. C. Burges, L. Bottou, M. Welling, Z. Ghahramani, and K. Q. Weinberger, editors, *Advances in Neural Information Processing Systems (NIPS)*, volume 26, pages 2148–2156. MIT Press, Cambridge, MA, 2013.
S. Dieleman, J. Schl[ü]{}ter, C. Raffel, E. Olson, S. K. S[œ]{}nderby, D. Nouri, D. Maturana, M. Thoma, E. Battenberg, J. Kelly, J. D. Fauw, M. Heilman, D. M. de Almeida, B. [McFee]{}, H. Weideman, G. Tak[á]{}cs, P. de Rivaz, J. Crall, G. Sanders, K. Rasul, C. Liu, G. French, and J. Degrave. Lasagne: First release, Aug. 2015.
E. Fiesler, A. Choudry, and H. J. Caulfield. Weight discretization paradigm for optical neural networks. In *Proc. SPIE 1281: Optical Interconnections and Networks*, pages 164–173, The Hague, Netherlands, Aug. 1 1990.
M. R. Garey and D. S. Johnson. *Computers and Intractability: A Guide to the Theory of [NP-Completeness]{}*. W.H. Freeman, 1979.
A. Gersho and R. M. Gray. *Vector Quantization and Signal Compression*. Kluwer Academic Publishers Group, 1992.
Y. Gong, L. Liu, M. Yang, and L. Bourdev. Compressing deep convolutional networks using vector quantization. In *Proc. of the 3rd Int. Conf. Learning Representations (ICLR 2015)*, San Diego, CA, May 7–9 2015.
S. Gupta, A. Agrawal, K. Gopalakrishnan, and P. Narayanan. Deep learning with limited numerical precision. In F. Bach and D. Blei, editors, *Proc. of the 32nd Int. Conf. Machine Learning (ICML 2015)*, pages 1737–1746, Lille, France, July 6–11 2015.
S. Han, J. Pool, J. Tran, and W. Dally. Learning both weights and connections for efficient neural network. In C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett, editors, *Advances in Neural Information Processing Systems (NIPS)*, volume 28, pages 1135–1143. MIT Press, Cambridge, MA, 2015.
I. Hubara, M. Courbariaux, D. Soudry, R. El-Yaniv, and Y. Bengio. Quantized neural networks: Training neural networks with low precision weights and activations. arXiv:1609.07061, Sept. 22 2016.
K. Hwang and W. Sung. Fixed-point feedforward deep neural network design using weights $+1$, $0$, and $-1$. In *2014 IEEE Workshop on Signal Processing Systems (SiPS)*, pages 1–6, Belfast, UK, Oct. 20–22 2014.
Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama, and T. Darrell. Caffe: Convolutional architecture for fast feature embedding. arXiv:1408.5093 \[cs.CV\], June 20 2014.
Y. [LeCun]{}, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. *Proc. IEEE*, 860 (11):0 2278–2324, Nov. 1998.
F. Li, B. Zhang, and B. Liu. Ternary weight networks. arXiv:1605.04711, Nov. 19 2016.
M. Marchesi, G. Orlandi, F. Piazza, and A. Uncini. Fast neural networks without multipliers. *IEEE Trans. Neural Networks*, 40 (1):0 53–62, Jan. 1993.
V. Nair and G. E. Hinton. Rectified linear units improve restricted [Boltzmann]{} machines. In J. F[ü]{}rnkranz and T. Joachims, editors, *Proc. of the 27th Int. Conf. Machine Learning (ICML 2010)*, Haifa, Israel, June 21–25 2010.
Y. Nesterov. A method of solving a convex programming problem with convergence rate [$\calO(1/k^2)$]{}. *Soviet Math. Dokl.*, 270 (2):0 372–376, 1983.
J. Nocedal and S. J. Wright. *Numerical Optimization*. Springer Series in Operations Research and Financial Engineering. Springer-Verlag, New York, second edition, 2006.
S. J. Nowlan and G. E. Hinton. Simplifying neural networks by soft weight-sharing. *Neural Computation*, 40 (4):0 473–493, July 1992.
M. Rastegari, V. Ordonez, J. Redmon, and A. Farhadi. -net: [ImageNet]{} classification using binary convolutional neural networks. In B. Leibe, J. Matas, N. Sebe, and M. Welling, editors, *Proc. 14th European Conf. Computer Vision (ECCV’16)*, pages 525–542, Amsterdam, The Netherlands, Oct. 11–14 2016.
K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. In *Proc. of the 3rd Int. Conf. Learning Representations (ICLR 2015)*, San Diego, CA, May 7–9 2015.
N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov. Dropout: A simple way to prevent neural networks from overfitting. *J. Machine Learning Research*, 15:0 1929–1958, June 2014.
C. Z. Tang and H. K. Kwan. Multilayer feedforward neural networks with single powers-of-two weights. *IEEE Trans. Signal Processing*, 410 (8):0 2724–2727, Aug. 1993.
. Theano: A [Python]{} framework for fast computation of mathematical expressions. arXiv:1605.02688, May 9 2016.
K. Ullrich, E. Meeds, and M. Welling. Soft weight-sharing for neural network compression. In *Proc. of the 5th Int. Conf. Learning Representations (ICLR 2017)*, Toulon, France, Apr. 24–26 2017.
W. Wang and M. [Á]{}. Carreira-Perpi[ñ]{}[á]{}n. The [Laplacian]{} [$K$]{}-modes algorithm for clustering. arXiv:1406.3895 \[cs.LG\], June 15 2014.
S. Zhou, Z. Ni, X. Zhou, H. Wen, Y. Wu, and Y. Zou. Dorefa-net: Training low bitwidth convolutional neural networks with low bitwidth gradients. arXiv:1606.06160, July 17 2016.
C. Zhu, S. Han, H. Mao, and W. J. Dally. Trained ternary quantization. In *Proc. of the 5th Int. Conf. Learning Representations (ICLR 2017)*, Toulon, France, Apr. 24–26 2017.
[^1]: All norms are $\norm{\cdot}_2$ throughout the paper unless indicated otherwise.
[^2]: We can also achieve *pruning* together with quantization by having one centroid be fixed to zero. We study this in more detail in a future paper.
[^3]: <https://github.com/BVLC/caffe/blob/master/examples/mnist/lenet.prototxt>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Quenching of gluonic jets and heavy quark production in Au+Au collisions at RHIC can be understood within the pQCD based 3+1 dimensional parton transport model BAMPS including pQCD bremsstrahlung $2 \leftrightarrow 3$ processes. Furthermore, the development of conical structures induced by gluonic jets is investigated in a static box for the regimes of small and large dissipation.'
author:
- |
I. Bouras, A. El, O. Fochler, F. Reining J. Uphoff C. Wesp\
Z. Xu C. Greiner
title: Collective Flow and Energy Loss with parton transport
---
\[1999/12/01 v1.4c Il Nuovo Cimento\]
Introduction
============
The values of the elliptic flow parameter $v_2$ measured by the experiments at the Relativistic Heavy Ion collider (RHIC) [@Adler:2003kt; @Adams:2003am; @Back:2004mh] suggest that in the evolving QCD fireball a fast local equilibration of quarks and gluon occurs at a very short time scale $\le 1$ fm/c. This locally thermalized state of matter, the quark gluon plasma (QGP), behaves as a nearly perfect fluid, confirmed by viscous hydrodynamics [@Luzum:2008cw; @Song:2008hj] and microscopic transport theory [@Xu:2007jv; @Xu:2008av]. The viscosity to entropy ratio coefficient $\eta/s$ has to be rather small, possibly close to the conjectured lower bound $\eta/s = 1/4 \pi$ from a correspondence between conformal field theory and string theory in an Anti-de-Sitter space [@Kovtun:2004de]. To achieve a rather small $\eta/s$ value in a partonic gas, binary perturbative QCD (pQCD) processes require unphysical large cross sections and thus inelastic radiative interactions [@Xu:2004mz] become important.
The phenomenon of jet-quenching has been another important discovery at RHIC [@Adams:2003kv]. Hadrons with high transverse momenta are suppressed in $Au + Au$ collisions with respect to a scaled $p + p$ reference [@Adler:2002xw; @Adcox:2001jp]. This quenching of jets is commonly attributed to energy loss on the partonic level as the hard partons produced in initial interactions are bound to traverse the QGP created in the early stages of heavy-ion collisions (HIC). In addition, very exciting jet-associated particle correlations have been observed [@Wang:2004kfa], which might be the result of a conical emission of propagating shock waves in form of Mach Cones induced by highly-energetic partons traversing the expanding medium [@Stoecker:2004qu]. Furthermore, heavy quarks are interesting for the investigation of the early stage of the QGP. Due to their large mass, initially produced heavy quarks, can cover – depending on their production point – a long distance through the QGP. Interactions on this way and subsequent modifications on heavy quark distributions can reveal valuable information about the properties of the medium.
A large class of phenomena in heavy-ion collisions can be investigated within the framework of the kinetic transport theory. Among others, the kinetic transport model BAMPS (Boltzmann Approach to Multiparton Scatterings) [@Xu:2004mz] was developed to describe the early QGP of a HIC. Using BAMPS early thermalization of gluons within $\tau < 1$ fm/c was demonstrated in Au+Au collisions at $\sqrt{s_{NN}} = 200$ GeV employing Glauber initial conditions and the coupling constant $a_s = 0.3$. In addition to the elastic pQCD $gg \leftrightarrow gg$ processes, pQCD-inspired bremsstrahlung $gg \leftrightarrow ggg$ was included. This was shown to be essential for the achievment of local thermal equilibrium at that short time scale. The fast thermalization happens also in a similar way using a Color Glass Condensate as inital conditions [@El:2007vg].
BAMPS has been applied to simulate elliptic flow and jet quenching at RHIC energies [@Fochler:2008ts] for the first time using a consistent and fully pQCD–based microscopic transport model to approach both key observables on the partonic level within a common setup. The left panel of Fig. \[fig:v2\_summary\_RAA\_central\] shows that the medium simulated in the parton cascade BAMPS exhibits a sizable degree of elliptic flow in agreement with experimental findings at RHIC as established in [@Xu:2007jv; @Xu:2008av].
The extraction of the shear viscosity over entropy density ratio $\eta/s$ has confirmed the essential importance of inelastic processes. Within the present description bremsstrahlung and back reaction processes lower the shear viscosity to entropy density ratio significantly by a factor of $7$, compared to the ratio when only elastic collisions are considered [@Xu:2007ns; @El:2008yy]. For $a_s = 0.3$ one finds $\eta/s = 0.13$, where for $a_S = 0.6$ the values matches the lower bound of $\eta/s = 1/4\pi$ from the AdS/CFT conjecture.
In this paper we demonstrate the description of different phenomena in relativistic HIC using the relativistic pQCD-based on-shell parton transport model BAMPS. Due to the large momentum scales involved the energy loss of partonic jets can be treated in terms of perturbative QCD (pQCD) and most theoretical schemes attribute the main contribution to partonic energy loss to radiative processes [@Wicks:2005gt]. In addition, the possible propagation of Mach Cones in the QGP induced by such highly-energetic partons can be studied. Considering the earlier work investigating the effects of dissipation on relativistic shock waves [@Bouras:2009nn; @Bouras:2009vs; @Bouras:2010hm] we demonstrate the transition of Mach Cones from ideal to the viscous one. It is a major challenge to combine jet physics on the one hand and bulk evolution on the other hand within a common framework. In the end, the heavy quark production at RHIC and the Large Hadron Collider (LHC) during the evolution of the QGP is studied.
Jet Quenching in Au+Au collisions at 200 AGeV
=============================================
For simulations of Jet Quenching in heavy ion collisions the initial gluon distributions are sampled according to a mini–jet model with a lower momentum cut-off $p_{0} = 1.4\,\mathrm{GeV}$ and a $K$–factor of $2$. The test particle method [@Xu:2004mz] is employed to ensure sufficient statistics. Quarks are discarded after sampling the initial parton distribution since currently a purely gluonic medium is considered. To model the freeze out of the simulated fireball, free streaming is applied to regions where the local energy density has dropped below a critical energy density $\varepsilon_{c}$ ($\varepsilon_{c} = 1.0\, \mathrm{GeV}/\mathrm{fm}^3$ unless otherwise noted).
The right panel of Fig. \[fig:v2\_summary\_RAA\_central\] shows the gluonic $R_{AA}$ simulated in BAMPS for central, $b=0\,\mathrm{fm}$, collisions. It is roughly constant at $R_{AA}^{\mathrm{gluons}} \approx 0.053$ and in reasonable agreement with analytic results for the gluonic contribution to the nuclear modification factor $R_{AA}$ [@Wicks:2005gt], though the suppression of gluon jets in BAMPS appears to be slightly stronger. We expect improved agreement in future studies when employing a carefully averaged $\langle b \rangle$ that will be better suited for comparison to experimental data than the strict $b=0\,\mathrm{fm}$ case.
![Left panel: Elliptic flow $v_2$ as a function of the number of participants for Au+Au at 200 AGeV for different combinations of the strong coupling $\alpha_s$ and the critical energy density $\varepsilon_c$. See [@Xu:2008av] for more information. Right panel: Gluonic $R_{AA}$ at midrapidity ($y \,\epsilon\, [-0.5,0.5]$) as extracted from simulations for central Au+Au collisions at 200 AGeV. For comparison the result from Wicks et al. [@Wicks:2005gt] for the gluonic contribution to $R_{AA}$ and experimental results from PHENIX [@Adare:2008qa] for $\pi^{0}$ and STAR [@Adams:2003kv] for charged hadrons are shown.[]{data-label="fig:v2_summary_RAA_central"}](v2vsnpart){width="\linewidth"}
![Left panel: Elliptic flow $v_2$ as a function of the number of participants for Au+Au at 200 AGeV for different combinations of the strong coupling $\alpha_s$ and the critical energy density $\varepsilon_c$. See [@Xu:2008av] for more information. Right panel: Gluonic $R_{AA}$ at midrapidity ($y \,\epsilon\, [-0.5,0.5]$) as extracted from simulations for central Au+Au collisions at 200 AGeV. For comparison the result from Wicks et al. [@Wicks:2005gt] for the gluonic contribution to $R_{AA}$ and experimental results from PHENIX [@Adare:2008qa] for $\pi^{0}$ and STAR [@Adams:2003kv] for charged hadrons are shown.[]{data-label="fig:v2_summary_RAA_central"}](RAA_compare_wicks_exp){width="\linewidth"}
We have computed the gluonic $R_{AA}$ for non–central Au + Au collisions at the RHIC energy of $\sqrt{s} = 200 \mathrm{AGeV}$ with a fixed impact parameter $b=7\,\mathrm{fm}$ (Fig. \[fig:v2\_RAA\_b7\]), which roughly corresponds to $(20-30)\%$ experimental centrality. A comparison in terms of the magnitude of the jet suppression for $b=7\,\mathrm{fm}$ is difficult since there are no published analytic results available to compare to. Taking the ratio of the $b=7\,\mathrm{fm}$ to the $b=0\,\mathrm{fm}$ results as a rough guess indicates that the decrease in quenching is more pronounced in BAMPS compared to experimental data. The ratio of the nuclear modification factor between central $(0 - 10) \%$ and more peripheral $(20-30)\%$ collisions is $\left. R_{AA}\right|_{0 \% - 10 \%} / \left. R_{AA}\right|_{20 \% - 30 \%} \approx 0.6$ for the experimental data, while for the BAMPS results $\left. R_{AA}\right|_{b=0\,\mathrm{fm}} / \left. R_{AA}\right|_{b=7\,\mathrm{fm}} \approx 0.4$. However, the issue of detailed quantitative comparison needs to be re-investigated once light quarks and a fragmentation scheme are included into the simulations.
![Left panel: Elliptic flow $v_{2}$ for gluons in simulated Au+Au collisions at 200 AGeV with $b=7\,\mathrm{fm}$. $\varepsilon_c = 0.6\, \mathrm{GeV}/\mathrm{fm}^3$. Right panel: Gluonic $R_{AA}$ as extracted from BAMPS simulations for $b=0\,\mathrm{fm}$ and $b=7\,\mathrm{fm}$, $\varepsilon_c = 1.0\, \mathrm{GeV}/\mathrm{fm}^3$. For comparison experimental results from PHENIX [@Adare:2008qa] for $\pi^{0}$ are shown for central $(0 - 10) \%$ and off–central $(20-30)\%$ collisions.[]{data-label="fig:v2_RAA_b7"}](v2_b7_new){width="\linewidth"}
![Left panel: Elliptic flow $v_{2}$ for gluons in simulated Au+Au collisions at 200 AGeV with $b=7\,\mathrm{fm}$. $\varepsilon_c = 0.6\, \mathrm{GeV}/\mathrm{fm}^3$. Right panel: Gluonic $R_{AA}$ as extracted from BAMPS simulations for $b=0\,\mathrm{fm}$ and $b=7\,\mathrm{fm}$, $\varepsilon_c = 1.0\, \mathrm{GeV}/\mathrm{fm}^3$. For comparison experimental results from PHENIX [@Adare:2008qa] for $\pi^{0}$ are shown for central $(0 - 10) \%$ and off–central $(20-30)\%$ collisions.[]{data-label="fig:v2_RAA_b7"}](RAA_compare_centrality_BAMPS_exp){width="\linewidth"}
To complement the investigations of $R_{AA}$ at a non–zero impact parameter $b=7\,\mathrm{GeV}$, we have computed the elliptic flow parameter $v_{2}$ for gluons at the same impact parameter and extended the range in transverse momentum up to roughly $p_{T} \approx 10\,\mathrm{GeV}$, see left panel of Fig. \[fig:v2\_RAA\_b7\]. For this a critical energy density $\varepsilon_{c} = 0.6\, \mathrm{GeV}/\mathrm{fm}^3$ has been used, in order to be comparable to previous calculations of the elliptic flow within BAMPS. The $v_{2}$ of high–$p_{T}$ gluons is at first rising with $p_{T}$, but from $p_{T} \approx 4$ to $5\,\mathrm{GeV}$ on, it begins to slightly decreases again. This behavior is in good qualitative agreement with recent RHIC data [@Abelev:2008ed] that for charged hadrons shows $v_{2}$ to be rising up to $v_{2} \approx 0.15$ at $p_{T} \approx 3\,\mathrm{GeV}$ followed by a slight decrease.
Transition from ideal to dissipative Mach Cones
===============================================
There is an important issue in recent studies of relativistic heavy-ion collisions (HIC) whether the small but finite viscosity allows the development of relativistic shocks in form of Mach Cones. Within the framework of BAMPS studies were finished to answer the question whether shocks can develop with finite viscosity and how this will alter such a picture [@Bouras:2009nn]. Within the relativistic Riemann problem it was shown that one dimensional shocks smears out if viscosity is large [@Bouras:2009vs; @Bouras:2010hm]. However, the expected viscosity in HIC seems to be small enough to allow a significant contribution of shocks in form of Mach Cones into the picture of HIC. In the following we report a very recent study.
Mach Cones, which are special phenomena of shock waves, have their origin in ideal hydrodynamics. A very weak perturbation in a perfect fluid induces sound waves which propagate with the speed of sound $c_s = \sqrt{dp/de}$, where $p$ is the pressure and $e$ is the energy density. In the case where the perturbation with velocity $v_{\rm jet}$ propagates faster than the generated sound waves, the sound waves lie on a cone. Considering a gas of massless particles, where $e = 3p$ and $c_s = 1/\sqrt{3}$, then the emission angle of the Mach Cone is given by $\alpha_w = \arccos ( c_s / v_{\rm jet} ) = 54,73^\circ$.
A stronger perturbation induces the propagation of shock waves exceeding the speed of sound, therefore the emission angle changes and can be approximated by\
$\alpha \approx \arccos ( v_{\rm shock} / v_{\rm jet} )$. Here $$\label{eq:v_shock}
v_{\rm{shock}} = \left [ \frac{(p_1 - p_0)(e_0 + p_1)}
{(e_1 - e_0)(e_1 + p_0)} \right]^{\frac{1}{2}}$$ is the velocity of the shock front, $p_{ \rm 0}$ ($e_{ \rm 0}$) the pressure (energy density) in the shock front region and $p_{ \rm 1}$ ($e_{ \rm 1}$) in the stationary medium itself. Eq. has the following limits: For $p_0 >> p_1$ we obtain $v_{\rm{shock}} \approx 1$, whereas for a small perturbation, $p_0 \approx p_1$, we get $v_{\rm{shock}} \approx c_s$.
We employ the microscopic transport model BAMPS to investigate Mach Cones with different strength of dissipations in the medium using a jet moving in positive $z$-direction, initialized at $t = 0$ fm/c at the position $z = -0.8$ fm. The jet is treated as a massless particle with zero spatial volume and zero transverse momentum, that is, $p_z = E_{\rm jet} = 200$ GeV and $v_{\rm jet} = 1$. The energy and momentum deposition to the medium is realized via collisions with medium particles. In this scenario we neglect the deflection of the jet and it can not be stopped by the medium; its energy and momentum is set to its initial value after every collision.
All simulations are realized within a static and uniform medium of massless Boltzmann particles and $T = 400$ MeV. For this study we consider only binary scattering processes with an isotropic cross section among the bulk particles. To save computational runtime we reduce our problem to two dimensions. Here we choose the $xz$-plane and apply a periodic boundary condition in $y$-direction.
![(Color online) Scenario of a massless jet with $p_z = E_{\rm jet} = 200$ GeV which can not be stopped by the medium - the shape of a Mach Cone shown for different viscosities of the medium, $\eta/s = 1/64\pi$ (left), $\eta/s = 1/4\pi$ (middle), $\eta/s = 1/\pi$ (right). We show the energy density plotted together with the velocity profile. Additionally, in the left panel the linear ideal Mach Cone for a very weak perturbation is shown by a solid line; its emission angle is $\alpha_w = 54,73^\circ$.[]{data-label="fig:machCone"}](multiEDensityVelo_Arrow_3_pro){width="\columnwidth"}
In Fig.\[fig:machCone\] we demonstrate the transition from ideal Mach Cone to a highly viscous one by adjusting the shear viscosity over entropy density ratio in the medium from $\eta/s = 1/64 \pi \approx 0.005$ to $1/\pi \approx 0.32$. The energy deposition of the jet is approximately $dE/dx = 11 - 14$ GeV/fm. We show a snapshot at $t = 2.5$ fm/c.
Using an unphysical small viscosity of $\eta/s = 1/64\pi$ we observe a strong collective behavior in form of a Mach Cone, as shown in the left panel of Fig.\[fig:machCone\]. Due to the fact that the energy deposition is strong, the shock propagates faster than the speed of sound through the medium. For comparison, the ideal Mach Cone caused by a very weak perturbation is given by a solid line with its emission angle $\alpha_w = 54,73^\circ$. Furthermore, a strong diffusion wake in direction of the jet, characterized by decreased energy density, and a head shock in the front are clearly visible.
![(Color online) Scenarion of a deflectable jet with finite energy $p_z = E_{\rm jet} = 20$ GeV - Induced Mach Cone structure for different viscosities of the medium, $\eta/s = 0.01$ (left), $\eta/s = 0.1$ (middle), $\eta/s = 0.5$ (right). We show the energy density plotted together with the velocity profile.[]{data-label="fig:machCone_jetBundle"}](multiEDensityVelo_Arrow_3_pro_jetBundle){width="\columnwidth"}
If we increase the viscosity of the medium to larger values, shown in the middle and left panel of Fig.\[fig:machCone\], the typical Mach Cone structure smears out and vanishes completely. Due to stronger dissipation, the collective behavior gets weaker because of less particle interactions in the medium with a larger $\eta/s$. The results agree qualitatively with earlier studies [@Bouras:2009nn; @Bouras:2010hm], where a smearing-out of the shock profile is observed with higher viscosity.
In addition to the scenario of an unstoppable jet we demonstrate in Fig.\[fig:machCone\_jetBundle\] the scenario of a massless jet with finite energy which can also be deflected. Its initial energy is set to $p_z = E_{\rm jet} = 20$ GeV, where the starting point is $z = -0.3$ fm. We show the results for different viscosities, $\eta/s \approx 0.01$ to $0.5$ at $t = 5.0$ fm/c. In analogy to the results above we observe a clear Mach Cone structure for small viscosities and a smearing out with larger values of $\eta/s$. Only in the ideal case a strongly curved structure in which the building up of a strong vortex is visible. The physical meaning of these phenomena and also jets with the full pQCD cascade have to be explored in future studies.
Heavy quarks in BAMPS
=====================
Initial heavy quark production during hard parton interactions in nucleon-nucleon scatterings and secondary production during the evolution of the QGP are studied. We use the event generator PYTHIA [@Sjostrand:2006za] to determine the initial heavy quark distributions, which agree with the experimental data from PHENIX [@Adare:2006hc_phenix_dsigmady]. Nevertheless, these distributions have large uncertainties due to their sensitivity on the parton distribution functions in nucleons, the heavy quark masses as well as the renormalization and factorization scales (see [@Uphoff:2010sh] for a detailed analysis). For the initial distribution of the gluonic medium we use three different approaches: the mini-jet model, a color glass condensate inspired model and also PYTHIA in combination with the Glauber model.
In the following we give a brief overview of our results on heavy quark production in the QGP. More details concerning this section can be found in [@Uphoff:2010sh].
![Left panel: elliptic flow $v_2$ of heavy quarks with pseudo-rapidity $|\eta|<0.35$ at the end of the QGP phase for Au+Au collisions at RHIC with an impact parameter of $b=8.2 \, {\rm fm}$. For one curve the cross section of $gQ \rightarrow gQ$ is multiplied with a $K$ factor. For comparison, data of heavy flavor electrons [@Adare:2010de] is also shown. In contrast to the left panel, where the theoretical curves for $v_2$ on the quark level are plotted, the right panel depicts the flow of heavy quarks and electrons, which stem from the decay of $D$ and $B$ mesons produced by fragmentation.[]{data-label="fig:v2"}](v2_asRun.eps){width="100.00000%"}
![Left panel: elliptic flow $v_2$ of heavy quarks with pseudo-rapidity $|\eta|<0.35$ at the end of the QGP phase for Au+Au collisions at RHIC with an impact parameter of $b=8.2 \, {\rm fm}$. For one curve the cross section of $gQ \rightarrow gQ$ is multiplied with a $K$ factor. For comparison, data of heavy flavor electrons [@Adare:2010de] is also shown. In contrast to the left panel, where the theoretical curves for $v_2$ on the quark level are plotted, the right panel depicts the flow of heavy quarks and electrons, which stem from the decay of $D$ and $B$ mesons produced by fragmentation.[]{data-label="fig:v2"}](v2_asRun_elect.eps){width="100.00000%"}
Secondary heavy quark production in the QGP is studied within a full BAMPS simulation of Au+Au collisions at RHIC. According to our calculations the charm quark production in the medium lies between 0.3 and 3.4 charm pairs, depending on the model of the initial gluon distribution, the charm mass and whether a $K=2$ factor for higher order corrections of the cross section is employed. However, compared to the initial yield these values are of the order of a few percent for the most probable scenarios. Therefore, one can conclude that charm production at RHIC in the QGP is nearly negligible.
At LHC, however, the picture looks a bit different: Here the charm production in the QGP is a sizeable fraction of the initial yield and is even of the same order for some scenarios (with mini-jet initial conditions for gluons with a high energy density). In numbers, between 11 and 55 charm pairs are produced in the QGP.
Bottom production in the QGP, however is very small both at RHIC and LHC and can be safely neglected. As a consequence, all bottom quarks at these colliders are produced in initial hard parton scatterings.
The elliptic flow and the nuclear modification factor $$\begin{aligned}
\label{elliptic_flow}
v_2=\left\langle \frac{p_x^2 -p_y^2}{p_T^2}\right\rangle \ , \qquad \qquad
R_{AA}=\frac{{\rm d}^{2}N_{AA}/{\rm d}p_{T}{\rm d}y}{N_{\rm bin} \, {\rm d}^{2}N_{pp}/{\rm d}p_{T}{\rm d}y}\end{aligned}$$ ($p_x$ and $p_y$ are the momenta in $x$ and $y$ direction in respect to the reaction plane) of heavy quarks at mid-rapidity are observables which are experimentally measurable and reflect the coupling of heavy quarks to the medium. A large elliptic flow comparable to that of light partons indicates a strong coupling to the medium. On the other hand a small $R_{AA}$ is a sign for a large energy loss of heavy quarks. Experimental results reveal that both quantities are on the same order as the respective values for light particles [@Adare:2010de; @Abelev:2006db; @Adare:2006nq].
As we have recently shown [@Uphoff:2010fz] elastic scatterings of heavy quarks with the gluonic medium using a constant coupling $\alpha_s = 0.3$ and the Debye mass for screening the $t$ channel cannot reproduce the experimentally measured elliptic flow. In order to explain the data one would need a $40-50$ times larger cross section than the leading order one. Of course, this $K$ factor is too large to represent the contribution of higher order corrections. However, as we demonstrated in Ref. [@Uphoff:2010sy] and is shown in the following, the discrepancy with the data can be lowered – even on the leading order level – by a factor of 10 by taking the running of the coupling into account and by improving the incorporation of Debye screening. The remaining factor of 4 difference could then indeed stem from neglecting higher order effects, which, however, must be checked in a future project.
The following calculations are done analogously to [@Uphoff:2010sy; @Gossiaux:2008jv; @Peshier:2008bg]. An effective running coupling is obtained from measurements of $e^+e^-$ annihilation and non-strange hadronic decays of $\tau$ leptons [@Gossiaux:2008jv; @Dokshitzer:1995qm].
Since the $t$ channel of the $g Q \rightarrow g Q$ cross section is divergent it is screened with a mass proportional to the Debye mass, which is calculated by the common definition $m_{D}^2 = 4 \pi \, (1+N_f/6) \,
\alpha_s(t) \, T^2$ with the running coupling. The proportionality factor $\kappa$ of screening mass and Debye mass is mostly set to 1 in the literature without a sophisticated reason. However, one can fix this factor to $\kappa = 0.2$ by comparing the ${\rm d}E/{\rm d}x$ of the Born cross section with $\kappa$ to the energy loss within the hard thermal loop approach to $\kappa
\approx 0.2$ [@Gossiaux:2008jv; @Peshier:2008bg].
Fig. \[fig:raa\] depicts the $R_{AA}$ of heavy quarks, which shows for $K=4$ the same magnitude of suppression as the data.
![As in Fig. \[fig:v2\], but the nuclear modification factor $R_{AA}$ of heavy flavor is shown instead of $v_2$.[]{data-label="fig:raa"}](raa_asRun.eps){width="100.00000%"}
![As in Fig. \[fig:v2\], but the nuclear modification factor $R_{AA}$ of heavy flavor is shown instead of $v_2$.[]{data-label="fig:raa"}](raa_asRun_elect.eps){width="100.00000%"}
These improvements lead to an enhanced cross section which also increases the elliptic flow. The left panel of Fig. \[fig:v2\] shows $v_2$ as a function of the transverse momentum $p_T$ for the leading order cross section without any improvements, with the running coupling, with the corrected Debye screening and with both modifications.
The elliptic flow of the latter reproduces the order of magnitude of the data, if the cross section is multiplied with $K=4$, which is much smaller than the previous employed $K=40-50$ and lies in a region which could account for higher order corrections. However, one has to check if these corrections have indeed a similar effect as a constant $K$ factor of 4. Therefore, the calculation of the next-to-leading order cross section is planned for the near future and will complement $2 \leftrightarrow 3$ interactions for gluons, which are already implemented in BAMPS [@Xu:2004mz]. The shapes of the theoretical curve and of the data points are, however, slightly different. This is an effect of hadronization and decay to electrons, which is not shown in the left panel of Fig. \[fig:v2\]. If one takes those two effects into account, the agreement of the data and the theoretically curve is much better, in particular for high $p_T$ (see right panel of Fig. \[fig:v2\]). We performed the fragmentation of charm (bottom) quarks to $D$ ($B$) mesons via Peterson fragmentation [@Peterson:1982ak] and used PYTHIA for the decay to electrons. At low $p_T$ the agreement between the theoretical curve and data becomes worse, since Peterson fragmentation is not the correct model here and coalescence may play a role.
The authors are grateful to the Center for Scientific Computing (CSC) at Frankfurt University for the computing resources. I. B., J. U. and C. W. are grateful to Helmholtz Graduate School for Hadron and Ion Research .
This work was supported by the Helmholtz International Center for FAIR within the framework of the LOEWE program launched by the State of Hesse.
[10]{} url\#1[`#1`]{}urlprefix
, . , . , . , . , . , . , . , . , . , . , . , . , . , . , . , . , . , . , . , . , . , . , . , . , . , . , . , . , . , . , . , . , . , . , . , .
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Scenarios for the emergence or bootstrap of a lexicon involve the repeated interaction between at least two agents who must reach a consensus on how to name $N$ objects using $H$ words. Here we consider minimal models of two types of learning algorithms: cross-situational learning, in which the individuals determine the meaning of a word by looking for something in common across all observed uses of that word, and supervised operant conditioning learning, in which there is strong feedback between individuals about the intended meaning of the words. Despite the stark differences between these learning schemes, we show that they yield the same communication accuracy in the realistic limits of large $N$ and $H$, which coincides with the result of the classical occupancy problem of randomly assigning $N$ objects to $H$ words.'
author:
- 'José F. Fontanari'
- Angelo Cangelosi
title: 'Cross-situational and supervised learning in the emergence of communication'
---
Introduction
============
How a coherent lexicon can emerge in a group of interacting agents is a major open issue in the language evolution and acquisition research area (Hurford, 1989; Nowak & Krakauer, 1999; Steels, 2002; Kirby, 2002; Smith, Kirby, & Brighton, 2003). In addition, the dynamics in the self-organization of shared lexicons is one of the issues to which computational and mathematical modeling can contribute the most, as the emergence of a lexicon from scratch implies some type of self-organization and, possibly, threshold phenomenon. This cannot be completely understood without a thorough exploration of the parameter space of the models (Baronchelli, Felici, Loreto, Caglioli, & Steels, 2006).
There are two main research avenues to investigate the emergence or bootstrapping of a lexicon. The first approach, inspired by the seminal work of Pinker and Bloom (1990) who argued that natural selection is the main design principle to explain the emergence and complex structure of language, resorts to evolutionary algorithms to evolve the shared lexicon. The key element here is that an improvement on the communication ability of an individual results, in average, in an increase of the number of offspring it produces (Hurford, 1989; Nowak & Krakauer, 1999; Cangelosi, 2001; Fontanari & Perlovsky, 2007, 2008). The second research avenue, which we will follow in this paper, argues for a culturally based view of language evolution and so it assumes that the lexicons are acquired and modified solely through learning during the individual’s lifetime (Steels, 2002; Smith, Kirby, & Brighton, 2003).
Of course, if there is a fact about language which is uncontroversial, it is that the lexicon must be learned from the active or passive interaction between children and language-proficient adults. The issue of whether this ability to learn the lexicon is due to some domain-general learning mechanism, or is an innate ability, unique to humans, is still on the table (Bates & Elman, 1996). In the problem we address here, there is simply no language-proficient individuals, so it is not so far-fetched to put forward a biological rather than a cultural explanation for the emergence of a self-organized lexicon. Nevertheless, in this contribution we will use many insights produced by research on language acquisition by children (see, e.g., Gleitman, 1990; Bloom, 2000) to study different learning strategies.
From a developmental perspective, there are basically two competing schemes for lexicon acquisition by children (Rosenthal & Zimmerman, 1978). The first scheme, termed cross-situational or observational learning, is based on the intuitive idea that one way that a learner can determine the meaning of a word is to find something in common across all observed uses of that word (Pinker, 1984; Gleitman, 1990; Siskind, 1996). Hence learning takes place through the statistical sampling of the contexts in which a word appears. Since the learner receives no feedback about its inferences, we refer to this scheme as unsupervised learning. The second scheme, known generally as operant conditioning, involves the active participation of the agents in the learning process, with exchange of non-linguistic cues to provide feedback on the hearer inferences. This supervised learning scheme has been applied to the design of a system for communication by autonomous robots – the so-called language game in the Talking Heads experiments (Steels, 2003). Despite the technological appeal, the empirical evidence is that most part of the lexicon is acquired by children as a product of unsupervised learning (Pinker, 1984; Gleitman, 1990; Bloom, 2000).
Interestingly, from the perspective of evolving or bootstrapping a lexicon, the unsupervised scheme is very attractive too, since it eliminates altogether the issue of honest signaling (Dawkins & Krebs, 1978), as no signaling is involved in the learning process, which requires only observation and some elements of intuitive psychology (e.g. Theory of Mind).
Many different computational implementations and variants of these two schemes for bootstrapping a lexicon have been proposed in the literature. For example, Smith (2003a, 2003b), Smith, Smith, Blythe, & Vogt (2006), and De Beule, De Vylder, & Belpaeme (2006) have addressed the unsupervised learning scheme, whereas Steels & Kaplan (1999), Ke, Minett, Au, Wang (2002), Smith, Kirby, & Brighton, (2003), and Lenaerts, Jansen, Tuyls, & De Vylder (2005), the supervised scheme. However, except for the extensive statistical analysis of a variant of the supervised learning algorithm which reduces the problem to that of naming a single object (Baronchelli, Felici, Loreto, Caglioli, & Steels, 2006), the study of the effects of changing the parameters of those models have been usually limited to the display of the time evolution of some measure of the communication accuracy of the population. Although at first sight the supervised learning scheme may seem to be clearly superior to the unsupervised one (albeit less realistic in the context of language acquisition by children), we are not aware of any thorough comparison between the performances of these two learning scenarios. In fact, in this contribution we show that in a realistic limit of very large lexicon sizes the supervised and unsupervised learning performances are essentially identical.
In this paper we study minimal models of the supervised and unsupervised learning schemes which preserve the main ingredients of these two classical language acquisition paradigms. For the sake of simplicity, here we interpret the lexicon as a mapping between objects and words (or sounds) rather than as a mapping between meanings (conceptual structures) and sounds. A more complete scenario would involve first the creation of meanings, i.e., the bootstrapping of an object-meaning mapping (Steels, 1996; Fontanari, 2006) and then the emergence of a meaning-sound mapping (see, e.g., Smith, 2003a, 2003b; Fontanari & Perlovsky, 2006).
Model
=====
Following a common assumption in lexicon bootstrapping models, such as the popular iterated learning model (Smith, Kirby, & Brighton, 2003; Brighton, Smith, & Kirby, 2005 ), we consider here only two agents who play in turns the roles of speaker and hearer. The agents live in a fixed environment composed of $N$ objects and have $H$ words available to name these objects. As we are interested in the limit where $N$ and $H$ are very large with the ratio $\alpha \equiv H/N$ finite we do not need to account for the possibility of creation of new words as in some variants of the supervised learning scheme (Baronchelli, Felici, Loreto, Caglioli, & Steels, 2006).
We assume that each agent is characterized by a $N \times H$ verbalization matrix $P$ the entries of which $p_{nh} \in \left [ 0,1 \right ]$, with $p_{nh} \in \left [ 0,1 \right ]$ for all values of $n=1,\ldots,N$, being interpreted as the probability that object $n$ is associated with word $h$. This assumption rules out the existence of objects without names, but it allows for words which are never used to name objects. To describe the communicative behavior of the agents through the verbalization matrix (i.e., the associations between objects and words for use both in production and interpretation) we need to specify how the speaker chooses a word for any given object as well as how the hearer infers the object the speaker intended to name by that word.
To name an object, say object $n$, the speaker simply chooses the word $h^*$ which is associated to the largest entry of row $n$ of the matrix $P$, i.e., $h^* = \max_{h} \left \{ p_{nh}, h = 1, \ldots, H \right \}$. In addition, to guess which object the speaker named by word $h$ the hearer selects the object that corresponds to the largest of the $N$ entries $p_{nh}$, $n=1, \ldots, N$. In other words, the hearer chooses the object that it itself would be most likely to associate with word $h$ (Smith, 2003a, 2003b). This amounts to assuming that the agents are endowed with a ‘Theory of Mind’ (ToM), i.e., that the hearer is somehow able to understand that the speaker thinks similar to itself and hence would behave likewise when facing the same situation (Donald, 1991). We note that the original inference scheme, termed “obverter” (Oliphant & Batali, 1997), assumed that the hearer has access to the verbalization matrix of the speaker (through mind reading, as the critics were ready to point out). Here we follow the more reasonable scheme, dubbed “introspective obverter” (Smith, 2003a), which requires endowing the agents with a Theory of Mind rather than with telepathic abilities.
Effective communication takes place when the two agents reach a consensus on which word must be assigned to each object. To achieve this, we must provide a prescription to modify their initially random verbalization matrices. Here we will consider two learning procedures that differ basically on whether the agents receive feedback (supervised learning) or not (unsupervised learning) about the success of a communication episode. But before doing this we need to set up the language game scenario where the agents interact.
From the list of $N$ objects, the agent who plays the speaker role chooses randomly $C$ objects without replacement. This set of $C$ objects forms the context. Then the speaker chooses randomly one object in the context and produces the word associated to that object, according to the procedure sketched before. The hearer has access to that word as well as to the $C$ objects that comprise the context. Its task is to guess which object in the context is named by that word. This is then an ambiguous language acquisition scenario in which there are multiple object candidates for any word. Once the verbalization matrices are updated the two agents interchange the roles of speaker and hearer and a new context is generated following the same procedure.
To control the convergence properties of the learning algorithms described next we assume that the entries $p_{nh}$ are discrete variables that can take on the values $0,1/M,2/M,\ldots,1-1/M,1$. In our simulations we choose $M=10^4$. The reciprocal of $M$ can be interpreted as the algorithm learning rate. In addition, as there are two agents who alternate in the roles of speaker and hearer, henceforth we will add the superscripts I or J to the verbalization matrix in order to identify the agent it corresponds to. At the beginning of the language game each agent has a different, randomly generated verbalization matrix. More pointedly, to generate the row $n$ of $P^I$ we distribute with equal probability $M$ balls among $H$ slots and set the value of entry $p_{nh}^I$ as the ratio between the number of balls in slot $h$ and the total number of balls $M$. An analogous procedure is used to set the initial value of $P^J$.
Unsupervised learning
---------------------
In this scheme, the list of objects in the context $n_1, \ldots, n_c$ and the accompanying word $h^*$ is the only information fed to the learning algorithm. Hence, in the unsupervised scheme, only the hearer’s verbalization matrix is updated. Of course, since the agents change roles at each learning episode, the verbalization matrices of both agents are updated during the learning stage. For concreteness, let us assume that agent $I$ is the speaker and so agent $J$ is the hearer in a particular learning episode. As pointed out before, the idea here is to model the cross-situational learning scenario (Siskind, 1996) in which the agents infer the meaning of a given word by monitoring its occurrence in a variety of contexts. Accordingly, the learning procedure increases the entries $p_{n_1 h^*}^J, \ldots, p_{n_c h^*}^J$ by the amount $1/M$. In addition, for each object in the context, say $n_1$, a word, say $h$, is chosen randomly and the entry $p_{n_1 h}^J$ is decreased by the same amount $1/M$, thus keeping the correct normalization of the rows of the verbalization matrix. (The possibility that $h=h^*$ is not ruled out.) This procedure which is inspired by Moran’s model of population genetics (Ewens, 2004) guarantees a minimum disturbance in the verbalization matrix and can be interpreted as the lateral inhibition of the competing word-object associations. We note that during the learning stage the agent playing the hearer role does not need to guess which object in the context is named by word $h^*$.
An extra rule is needed to keep the entries $p_{nh}^J$ within the unit interval $\left [ 0,1 \right ]$: we assume that once an entry reaches the values $p_{nh}^J = 1$ or $p_{nh}^J = 0$ it becomes fixed, so the extremes of the unit interval act as absorbing barriers for the stochastic dynamics of the learning algorithm.
Supervised learning
-------------------
The setting is identical to that described before except that now the hearer must guess which object in the context the speaker named by $h^*$ and then communicate its choice to the speaker (using some nonlinguistic means, such as pointing to the chosen object). In turn, the speaker must provide another nonlinguistic hint to indicate which object in the context it named by word $h^*$. Let us assume that the speaker associates word $h^*$ to object $n_1$. If the hearer’s guess happens to be the correct one, then both entries $p_{n_1 h^*}^I$ and $p_{n_1 h^*}^J$ are incremented by the amount $1/M$. Furthermore, two words, say $h_s$ and $h_h$, are chosen randomly and the entries $p_{n_1 h_s}^I$ and $p_{n_1 h_h}^J$ are decreased by $1/M$ so the normalization of row $n_1$ is preserved in both verbalization matrices. Suppose now the hearer’s guess is wrong, say, object $n_2$ instead of $n_1$. Then both entries $p_{n_1 h^*}^I$ and $p_{n_2 h^*}^J$ are decreased by the amount $1/M$ and, as before, two words $h_s$ and $h_h$ are chosen randomly and the entries $p_{n_1 h_s}^I$ and $p_{n_2 h_h}^J$ are increased by $1/M$. As in the unsupervised case, the extremes $p_{nh}^{I,J} = 1$ and $p_{nh}^{I,J} = 0$ are absorbing barriers.
The weak point of this learning scheme is the need for nonlinguistic hints to communicate the success or failure of the communication episode. This implies that, prior to learning, the agents are already capable to communicate (and understand) sophisticated meanings such as success and failure and behave (by updating their verbalization matrices) accordingly. In fact, feedback about the outcome of the communication episode may be seen as a form of telepathic meaning transfer.
Results
=======
Simulation experiments of the two learning algorithms described above show, not surprisingly, that after a transient the two agents become identical, in the sense that they are described by the same verbalization matrix. In addition, in the case of unsupervised learning the stochastic dynamics always leads to binary verbalization matrices, i.e., matrices whose entries $p_{nh}$ can take on the values 1 or 0 only. Of course, once the dynamics produces a binary matrix it becomes frozen. This same outcome characterizes the supervised case as well, except in the cases that the lexicon size $H$ is on the same order of the context size $C$. However, as we focus on the regime where $C$ is finite and $N$ and $H$ are large we can guarantee that the stochastic dynamics leads to binary verbalization matrices regardless of the learning procedure.
Once the dynamics becomes frozen (and so the learning stage is over) we measure the average communication error $\epsilon$ as follows. The speaker chooses object $n$ from the list of $N$ objects and emits the corresponding word (there is a unique word assigned to any given object, i.e., there is a single entry 1 in any row of the verbalization matrix). The hearer must then infer which object is named by that word. Since the same word can name many objects (i.e., there may be many entries 1 in a given column), the probability $\phi_n$ that the hearer’s guess is correct is simply the reciprocal of the number of objects named by that word. This probability is the communication accuracy regarding object $n$. The procedure is repeated for the $N$ objects, so the average communication error is defined as $\epsilon = 1 - \phi$ where $\phi = \sum_n \phi_n/N$ is the average communication accuracy of the algorithm.
As already pointed out, the normalization condition on the rows of the verbalization matrix $P$ allows for the possibility that a certain number of words are not used by the lexicon acquisition algorithms. Let $H_u \leq H$ stand for the actual number of words used by those algorithms. Then we can easily convince ourselves that $H_u = \sum_n^N \phi_n $ simply by noting that $\sum_n' \phi_n = 1$ when the sum is restricted to objects that are associated to the same word. Finally, we note that in the definitions of these communication measures the context plays no role at all; indeed the context is relevant only during the learning stage.
It is important to estimate the optimal (minimum) communication error $\epsilon_m$ in our learning scenario since, in addition to being a lower bound to the communication error produced by the learning algorithms, it allows us to rate their absolute performances. For $H \leq N$ the optimal communication error is obtained by making a one-to-one assignment between $H-1$ words and $H-1$ objects, and then assigning the single remaining word to the remaining $N-H+1$ objects. This procedure yields $\epsilon_m = 1 - H/N = 1 - \alpha$. For $ H > N$ we can obtain $\epsilon_m =0$ simply by discarding $H-N$ words and making a one-to-one word-object assignment with the other $N$ words. In fact, using our finding that $\phi = H_u/N$ we see that, as expected, the optimal performance is obtained by setting $H_u = H$ if $H \leq N$ and $H_u = N$ if $H > N$.
Figure \[fig:1\] shows the comparison between the optimal performance and the actual performances of the two learning algorithms as function of the ratio $\alpha$. In this, as well as in the other figures of this paper, each symbol stands for the average over $10^4$ independent samples or language games. The performance of the supervised algorithm deteriorates as the number of objects $N$ increases, in contrast to that of the unsupervised algorithm which actually shows a slight improvement in this case. For $N \to \infty$, both algorithms produce the same communication error (see Fig. \[fig:2\]), which is shown by the solid line in Fig. \[fig:1\]. We note that a preliminary comparative analysis of these algorithms for $N=8$ led to an incorrect claim about the general superiority of the supervised learning scheme (Fontanari & Perlovsky, 2006). For small values of $\alpha$ the performances of the two learning algorithms are practically indistinguishable from the optimal performance, but as we will argue below the algorithms actually never achieve that performance, except for $\alpha=0$.
It is instructive to calculate the communication error in the case that the $N$ objects are assigned randomly to the $H$ words. This is a classical occupancy problem discussed at length in the celebrated book by Feller (1968). In this occupancy problem, the probability $P_m$ that the number of words $m$ not used in the assignment of the $N$ objects to the $H$ words (i.e., $m = H - H_u$) is $$\label{m}
P_m = \left ( \begin{array}{c} H \\ m \end{array} \right )
\sum_{\nu =0}^{H-m} \left ( \begin{array}{c} H-m \\ \nu \end{array} \right ) \left ( -1 \right )^{\nu}
\left ( 1 - \frac{m + \nu}{H} \right )^N ,$$ which in the limits $N \to \infty $ and $H \to \infty$ reduces to the Poisson distribution $$\label{poisson}
p \left (m; \lambda \right ) = \mbox{e}^{-\lambda} \frac{\lambda^m}{m!}$$ where $\lambda = H \exp \left ( - N/H \right )$ remains bounded (Feller, 1968). Hence the average communication accuracy resulting from the random assignment of objects to words is simply $ \left ( H - \left \langle m \right \rangle \right )/N$, which yields the communication error $$\label{Er}
\epsilon_r = 1 - \alpha + \alpha \mbox{e}^{-1/\alpha} .$$ Surprisingly, this equation describes perfectly the communication error of the two learning algorithms in the limit $N \to \infty$ (solid line in Fig. \[fig:1\]). We note that the (small) discrepancy observed in Fig. \[fig:2\] for the extrapolated data of the unsupervised algorithm and the analytical prediction can be reduced to zero by decreasing the learning rate $1/M$. Equation (\[Er\]) explains also why the performances of the algorithms are practically indistinguishable from the optimal performance for small $\alpha$, since the difference between them vanishes as $\exp \left ( -1/\alpha \right )$. In addition, Eq. (\[Er\]) shows that in the limit of large $\alpha$, the communication error vanishes as $1/\alpha$.
A word is in order about the effect of the context size $C$ on the performance of the two learning algorithms, since Figs. \[fig:1\] and \[fig:2\] exhibit the results for $C=2$ only. Simulations for larger values of $C$ show that this parameter is completely irrelevant for the performance of the supervised algorithm. Of course, this is expected since regardless of the context size, at most two rows (object labels) of the verbalization matrices are updated. But the situation is far from obvious for the unsupervised algorithm since $C$ determines the number of rows to be updated in each round of the game. However, the results summarized in Fig. \[fig:3\] for $C=4$ indicate that, despite strong finite-size effects particularly for small $\alpha$, the communication error ultimately tends to $\epsilon_r$ in the limit of large $N$.
Conclusion
==========
In this paper we have unveiled two remarkable results. First, the supervised and unsupervised schemes for bootstrapping a lexicon yield the same communication accuracy in the limit of very large lexicon sizes. For finite lexicon sizes the supervised scheme always outperforms the unsupervised one, but its performance degrades as the lexicon size increases, whereas the performance of the unsupervised learning algorithm improves slightly with increasing lexicon size (see Fig. \[fig:1\]). Second, those performances tend to the communication accuracy obtained by a random occupancy problem in which the $N$ objects are assigned randomly to the $H$ words. These findings reveal a surprising inefficiency of traditional lexicon bootstrapping scenarios when evaluated in the realistic regime of very large lexicon sizes. It would be most interesting to devise sensible scenarios that reproduce the optimal communication performance or, at least, that exhibit an communication error that decays faster than the random occupancy result, $1/\alpha=N/H$, in the case the number of available words is much greater than the number of objects ($H \gg N$).
The scenarios studied here are easily adapted to model the problem of lexicon acquisition (rather than bootstrapping): we have just to assume that one of the agents, named the master in this case, knows the correct lexicon and so its verbalization matrix is kept fixed during the entire learning procedure; the verbalization matrix of the other agent – the pupil – is allowed to change following the update algorithms described before (see, e.g., Fontanari, Tikhanoff, Cangelosi, Ilin, & Perlovsky, 2009). Most interestingly, in this context, statistical world learning has been observed in controlled experiments involving infants (Smith & Yu, 2008) and adults (Yu & Smith, 2007). Similar experiments, but now aiming at bootstrapping a lexicon, could be easily carried out by replacing our virtual agents by two adults, who would then resort to some conscious or unconscious mechanism to track the co-occurrence of words and objects. Of course, the very emergence of pidgin - a means of communication between two or more groups which lack a common language (Thomason & Kaufman, 1988) - can be seen as a realization of such an experiment and serves as additional justification for the study of lexicon bootstrapping.
Acknowledgments {#acknowledgments .unnumbered}
===============
The research at São Carlos was supported in part by CNPq, FAPESP and SOARD grant FA9550-10-1-0006. J.F.F. thanks the hospitality of the Adaptive Behaviour & Cognition Research Group, University of Plymouth, where this research was initiated. The visit was supported by euCognition.org travel grant NA-097-6. Cangelosi also acknowledges the contribution of the ITALK project from the European Commission (FP7 ICT Cognitive Systems and Robotics).
References {#references .unnumbered}
==========
Baronchelli, A., Felici, M., Loreto, V., Caglioli, E., & Steels, L. (2006). Sharp transition towards shared vocabularies in multi-agent systems. [*Journal of Statistical Mechanics*]{}, P06014.
Bates, E., & Elman, J. (1996). Learning rediscovered. [*Science*]{}, 274, 1849-1850.
Bloom, P. (2000). [*How children learn the meaning of words*]{}. Cambridge, MA: MIT Press.
Brighton, H., Smith, K., & Kirby, S. (2005). Language as an evolutionary system. [*Physics of Life Reviews*]{}, 2, 177-226.
De Beule, J., De Vylder, B., & Belpaeme, T. (2006). A cross-situational learning algorithm for damping homonymy in the guessing game. In L.M. Rocha, M. Bedau, D. Floreano, R. Goldstone, A. Vespignani, & L. Yaeger (Eds.), [*Proceedings of the Xth Conference on Artificial Life*]{} (pp. 466-472). Cambridge, MA: MIT Press.
Cangelosi, A. (2001). Evolution of Communication and Language using Signals, Symbols and Words. [*IEEE Transactions on Evolutionary Computation*]{}, 5, 93-101.
Dawkins, R., & Krebs, J.R. (1978). Animal signals: information or manipulation? In: J.R. Krebs, & N. B. Davies (Eds.), [*Behavioural ecology: an evolutionary approach*]{} (pp. 282-309). Oxford, UK: Blackwel Scientific Publications.
Donald, M. (1991). [*Origins of the Modern Mind*]{}. Cambridge, MA: Harvard University Press.
Ewens, W.J. (2004). [*Mathematical Population Genetics*]{}. New York: Springer-Verlag.
Feller, W. (1968). [*An Introduction to Probability Theory and Its Applications*]{}. Vol. I, 3rd Edition. New York: Wiley.
Fontanari, J.F. (2006). Statistical analysis of discrimination games. [*European Physical Journal B*]{}, 54, 127-130.
Fontanari, J.F., & Perlovsky, L.I. (2006). Meaning creation and communication in a community of agents. In [*Proceedings of the 2006 International Joint Conference on Neural Networks*]{} (pp. 2892-2897). Piscataway, NJ: IEEE Press.
Fontanari, J.F., & Perlovsky, L.I. (2007). Evolving compositionality in evolutionary language games. [*IEEE Transactions on Evolutionary Computation*]{}, 11, 758-769.
Fontanari, J.F., & Perlovsky, L.I. (2008). A game theoretical approach to the evolution of structured communication codes. [*Theory in Biosciences*]{}, 127, 205-214.
Fontanari, J.F., Tikhanoff , V., Cangelosi, A., Ilin, R., & Perlovsky, L.I. (2009). Cross- situational learning of object-word mapping using Neural Modeling Fields. [*Neural Networks*]{}, 22, 579-585.
Gleitman, L. (1990). The structural sources of verb meanings. [*Language Acquisition*]{}, 1, 1-55.
Hurford, J.R. (1989). Biological evolution of the Saussurean sign as a component of the language acquisition device. [*Lingua*]{}, 77, 187-222.
Ke, J, Minett, J.W., Au, C.-P., & Wang, W.S.-Y. (2002). Self-organization and Selection in the Emergence of Vocabulary. [*Complexity*]{}, 7, 41-54.
Kirby, S. (2002). Natural language from artificial life. [*Artificial Life*]{}, 8, 185-215.
Lenaerts, T., Jansen, B., Tuyls, K., & De Vylder, B. (2005). The evolutionary language game: An orthogonal approach. [*Journal of Theoretical Biology*]{}, 235, 566-582.
Nowak, M.A., & Krakauer, D.C. (1999).The evolution of language. [*Proceedings of the National Academy of Sciences USA*]{}, 96, 8028-8033.
Oliphant, M., & Batali, J. (1997). Learning and the emergence of coordinated communication, [*Center for Research on Language Newsletter*]{}, 11.
Pinker, S. (1984). [*Language learnability and language development*]{}. Cambridge, MA: Harvard University Press.
Pinker, S., & Bloom, P. (1990). Natural languages an natural selection, [*Behavioral and Brain Sciences*]{}, 13, 707-784.
Rosenthal, T., & Zimmerman, B. (1978). [*Social Learning and Cognition*]{}. New York: Academic Press.
Siskind, J.M. (1996). A computational study of cross-situational techniques for learning word-to-meaning mappings. [*Cognition*]{}, 61, 39-91.
Smith, A.D.M. (2003a). Semantic generalization and the inference of meaning. [*Lecture Notes in Artificial Intelligence*]{}, 2801, 499-506.
Smith, A.D.M. (2003b). Intelligent meaning creation in a clumpy world helps communication. [*Artificial Life*]{}, 9, 557-574.
Smith, K., Kirby, S., & Brighton, H. (2003). Iterated Learning: a framework for the emergence of language. [*Artificial Life*]{}, 9, 371-386.
Smith, K., Smith, A.D.M, Blythe, R.A., & Vogt, P. (2006). Cross-Situational Learning: A Mathematical Approach. [*Lecture Notes in Computer Science*]{}, 4211, 31-44.
Smith, L.B., & Yu, C. (2008). Infants rapidly learn word-referent mappings via cross- situational statistics. [*Cognition*]{}, 106, 1558-1568.
Steels, L. (1996). Perceptually grounded meaning creation. In M. Tokoro (Ed.), [*Proceedings of the Second International Conference on Multi-Agent Systems*]{} (pp. 338-344). Menlo Park, CA: AAAI Press.
Steels, L., & Kaplan, F. (1999). Situated grounded word semantics. In [*Proceedings of the Sixteenth International Joint Conferences on Artificial Intelligence*]{} (pp. 862-867). San Francisco, CA: Morgan Kauffman.
Steels, L. (2002). Grounding symbols through evolutionary language games. In A. Cangelosi, & D. Parisi (Eds.), [*Simulating the Evolution of Language*]{} (pp. 211-226). London: Springer-Verlag.
Steels, L. (2003). Evolving Grounded Communication for Robots. [*Trends in Cognitive Sciences*]{}, 7, 308-312.
Thomason, S.G., & Kaufman, T. (1988). [*Language contact, creolization, and genetic linguistics*]{}. Berkeley: University of California Press.
Yu, C., & Smith, L.B (2007). Rapid word learning under uncertainty via cross-situational statistics. [*Psychological Science*]{}, 18, 414-420.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'R.P. Mignani'
- 'S. Zharikov'
- 'P. A. Caraveo'
date: 'Received ...; accepted ...'
title: 'The Optical Spectrum of the Vela Pulsar[^1] '
---
Introduction
============
The study of the optical emission properties of Isolated Neutron Stars (INSs) is hampered by their intrinsic faintness. For most of them the knowledge of the optical spectrum is still based on the comparison between multi-band photometry measurements. We note here that this comparison is affected by several uncertainties since photometry measurements are often compiled from the literature and, thus, are taken with different instruments and filters, calibrated with different photometric systems (167 used in total; see, e.g. Mono & Munari 2000[^2]), indipendently corrected for the atmospheric and interstellar extinction and, last but not least, converted to spectral fluxes using slightly different techniques. Only for five of the fourteen INSs with identified optical counterparts (see Mignani et al. 2004; Mignani 2005 for updated reviews) medium-resolution spectra have been obtained so far. The first one was the Crab pulsar (Oke 1969), the youngest ($\sim 1000$ years) and the brightest INS ($V=16.6$), for which repeated spectroscopy observations have been performed (Nasuti et al. 1996; Sollerman et al. 2000). The second one was PSR B0540–69 (Hill et al. 1997, Serafimovich et al. 2004), the youngest ($\sim 2000$ years) and brightest ($V=22.4$) INS after the Crab pulsar. Only after the advent of the generation of the 10 m class telescopes, it was possible to obtain spectra of the fainter ($V\sim
25-25.6$) and older ($\ge 10^5$ years) INSs such as Geminga (Martin et al. 1998) and PSR B0656+14 (Zharikov et al. 2007), as well as of the radio-quiet INS RX J1856–3754 (van Kerkwijk & Kulkarni 2001).
The best next target for optical spectroscopy is the Vela Pulsar (PSR B0833–45), the third optically brightest ($V=23.6$) INS. Among the first radio pulsars discovered in the sixties, its optical identification was proposed by Lasker (1976) and confirmed by the detection of optical pulsations at the radio period (Wallace et al. 1977). The high pulsed fraction of the optical lightcurve (see also Gouiffes 1998), naturally pointed towards a non thermal, magnetospheric origin of the optical radiation. This was confirmed by multi-band photometry performed by Nasuti et al. (1997) with the [[*NTT*]{}]{} and by Mignani & Caraveo (2001) with the [[*HST*]{}]{}. The fluxes between $\sim3000$ Å and $\sim8000$ Å follow a flattish, power-law distribution ($F_{\nu} \propto \nu^{-\alpha}; \alpha = -0.2
\pm 0.2$). The non-thermal nature of the optical radiation of the Vela pulsar has been also confirmed by polarisation observations performed with the [[*VLT*]{}]{} (Wagner & Seifert 2000; Mignani et al. 2007), pointing to a relatively high polarisation level. The Vela pulsar has been recently observed in the IR with the [[*VLT*]{}]{} (Shibanov et al. 2003), and in the near-UV (Romani et al. 2005; Kargaltsev & Pavlov 2007) with the [[*HST*]{}]{}. While its multi-band photometry now extends from the IR to the near-UV, no optical spectrum has been published yet. Here, we report on the first spectroscopy observations of the Vela pulsar performed with the [[*VLT*]{}]{}. Observations, data analysis and results are described in §2, while the interpretations are discussed in §3.
Observations and Data Analysis
==============================
Observations
------------
Spectroscopy observations of the Vela pulsar were performed in Service Mode in four different nights between December 2000 and January 2001 with the [[*VLT*]{}]{} at the ESO’s Paranal Observatory (Chile). We have used [[*FORS2*]{}]{} (FOcal Reducer Spectrograph[^3]), a multi-mode camera for imaging and long-slit/multi-object spectroscopy which is identical to [[*FORS1*]{}]{} (Appenzeller et al. 1998) but with the CCD optmized to achieve a higher sensitivity in the Red part of the spectrum. At the time of the observations [[*FORS2*]{}]{} was still mounting the original detector, a Tektronix 2048$\times$2046 pixels CCD with a plate scale of $0\farcs20$ with the Standard Resolution (SR) collimator. The instrument was operated in its single-port, high-gain, read out mode which is the default one for long slit spectroscopy (LSS). To cover the wavelength interval 4000-11000Å the observations were performed with two different grisms: the $300V$ ($\Delta \lambda=4500-8600$ Å) and the $300I$ one ($\Delta \lambda=6000-11100$ Å), equipped with the order separation filter $OG590$. Both grisms have a dispersion of $\sim$2.7Å/pixel and a resolving power $\lambda / \Delta \lambda =
440$. A slit width of $2\farcs5$ was used to collect as much flux as possible from the pulsar (see Fig. 1).
![VLT [[*FORS2*]{}]{} raw $B$-band acquisition image of the Vela pulsar. North to the top, East to the left. The $2\farcs5$ LSS slit is overplotted, with the orientation NS as used in the first night (see Table 1). The position of the Vela pulsar (barely detected in the 60 s exposure acquisition image) is marked by the arrow. The difference between the four quadrants corresponds to the different CCD readout ports. []{data-label="vela_acq"}](7774fig1.ps){width="8.0cm"}
Single LSS science exposures of 2800 s each were obtained for both grisms and repeated for cosmic rays filtering. A total of eight 2800 s exposures were taken for a total integration time of 11200 s for each grism. The complete log of the science observations is reported in Table 1. The first exposure sequence in the $300I$ grism was aborted and is not considered in the following analysis. The seeing conditions were always sub-arcsec ($0\farcs8$ average) and the airmass below 1.2. The slit was oriented NS on the first night and always EW on the following nights. For each night, daytime calibrations (biases, darks, flatfields, arc lamp spectra) were taken to correct for instrumental effects and to perform the wavelength calibration. Multi object spectroscopy (MOS) observations of the spectrophotometric standard stars LTT 3218, Feige 56 and Feige 67 (Hamuy et al. 1994) were acquired at the beginning of each night (with the only exception of the third one) for flux calibration. Because of their brightness, standard stars were observed in MOS mode (slitlet size $22\farcs0$) to avoid flux losses which might occur when using the narrower LSS slit ($2\farcs5$). Thanks to the sub-arcsec seeing conditions during the observations it was possible to resolve the pulsar spectrum in each individual science exposure. As an example, we show in Fig. 2 one of the two $300V$ spectra taken on the first night, where the pulsar is clearly detected. Unfortunately, the pulsar is detected with a much lower significance in the $300I$ spectra due to the higher sky background and to the drop of the CCD sensitivity towards longer wavelengths
yyyy-mm-dd Grism Time (s) N Seeing Airmass
------------ ------- ---------- --- ------------- ---------
2000-12-03 300V 2800 2 $0\farcs79$ 1.14
2001-01-23 300V 2800 2 $0\farcs60$ 1.07
2001-02-25 300I 1400 1 $0\farcs89$ 1.09
2001-01-30 300I 2800 4 $0\farcs73$ 1.11
: Summary of the spectroscopy observations of the Vela pulsar taken with [[*FORS2*]{}]{}. The columns give the observing dates, the used grism, the exposure times, the number of exposures per pointing N, and the average seeing and airmass values during each exposure sequence.
\[tabdatasummary\]
Data Analysis
-------------
Spectral data reduction and calibration were performed using standard suites of tools for treating CCD data available in the MIDAS and IRAF sofware packages. Master calibration frames (master biases and darks, flux-normalized flat fields) were provided by the [[*FORS*]{}]{} data reduction pipeline.[^4] Science and standard stars spectra, as well as arc lamp spectra, were then bias and dark-subtracted, and flatfielded. Wavelength calibration was computed from the reduced arc lamp spectra by fitting a second order polynomial, yielding an rms of 0.4 Å/pixel. The wavelength calibration was then applied row-by-row to both the standard stars and science spectra. For each grism, the flux normalization was computed from the extracted one-dimensional standard star spectrum using the available flux reference table and applying the atmospheric extinction correction with the extinction curves measured for the Paranal Observatory (Patat 2004). For each grism, single science spectra were finally co-added and cosmic rays removed. The pulsar one-dimensional spectrum was then extracted from each of the co-added science spectra using a 4 pixel wide ($0\farcs8$) extraction window centered on the pulsar position where the S/N is higher. The sky background was computed and subtracted from two 6 pixel wide ($1\farcs2$) regions immediately adjacent to the pulsar spectrum. For each extracted one-dimensional spectrum of the pulsar we finally applied a flux correction factor, computed from the extracted spectra of brighter field stars, to account for the width difference between the extraction window and the object’s PSF. The spectra were corrected for the interstellar extinction using as a reference $E(B-V)=0.05$ obtained from the spectral fit to the [[*XMM*]{}]{}X-ray spectrum (e.g. Manzali et al. 2007). The two one-dimensional spectra of the pulsar were finally merged to yield a single spectrum across the 4000-11000Å wavelength range.
Results
-------
The final Vela pulsar’s spectrum is shown in Fig. 3. Due to the different detection significance of the pulsar in the two grisms, the S/N of the resulting (unbinned) spectrum is rather unhomogeneous. In particular, we estimated a S/N $\sim$ 5 (per pixel) in the 4000-5500Å wavelength range and a S/N as low as $\sim 1$ near/above 8000Å. At longer wavelengths the spectrum is entirely dominated by noise. For this reason in the following analyses we have considered only the part of the spectrum between 4000 Å and 8000 Å. Clearly, the spectrum consists of a pure flat continuum, with no evidence for emission or absorption feautures. For comparison, we have overplotted the spectral fluxes recomputed from the optical multi-band photometry (Nasuti et al. 1997; Mignani & Caraveo 2001), obtained between $\sim$ 4000 Å and $\sim$ 8000 Å. To be consistent, we have re-corrected these fluxes for the same value of the interstellar extinction used in this work ($A_V$=0.16), while both Nasuti et al. (1997) and Mignani & Caraveo (2001) have used $A_V$=0.4. For the interstellar extinction correction we have used the coefficients of Fitzpatrick (1999). As seen, within the cross-calibration uncertainties, the multi-band photometry and the spectroscopy flux measurements are substantially consistent with each other. A putative dip at 6500 Å was hinted in the multi-band photometry data of Mignani & Caraveo (2001). However, its existence was not confirmed by the [[*VLT*]{}]{} photometry of Shibanov et al. (2003). No evidence for such a dip is found in our spectral data either. We thus conclude that it was just the result of a cross-calibration problem in the multi-band photometry of Mignani & Caraveo (2001).
We have fitted the [[*FORS2*]{}]{} spectrum between 4000 and 8000 Å with a power-law ($F_{\nu} \propto \nu^{-\alpha}$) and we derived a spectral index $\alpha =-0.04 \pm 0.04$. This value superseeds the one obtained from the multi-band photometry of Mignani & Caraveo (2001) on the base of only five spectral flux measurements. We note that our spectral index is somewhat flatter than the one obtained by Shibanov et al. (2003) by fitting the IR-to-optical spectral fluxes ($\alpha = 0.12 \pm 0.05$), while it compares better with the spectral index $\alpha = 0.01 \pm 0.02$, obtained by Kargaltsev & Pavlov (2007) by extending the fit to the near-UV. Interestingly, the optical power-law is below the extrapolation of the power-law component ($\alpha = 1.2 \pm 0.3$) used to fit the [[*XMM*]{}]{}X-ray spectrum (Manzali et al. 2007; see also Kargaltsev & Pavlov 2007), which clearly suggests that the magnetospheric emission of the Vela pulsar features a spectral turnover at longer wavelengths.
Discussion
==========
---------- ------------- ---------- ---------------- ----------------- ------------------ ----------------- ------------------ ----------------- ----------
Name Log($\tau$) Log($B$) Log($\dot E$) $\alpha_{O,sp}$ $\lambda\lambda$ $\alpha_{O,ph}$ $\lambda\lambda$ $\alpha_X$ Ref.
yrs (G) (erg s$^{-1}$) (Å) (Å)
Crab 3.1 12.58 38.65 $-0.11\pm0.04$ 3300-9250 $1.079\pm0.004$ 1,2
B1509–58 3.2 13.19 37.25 $0.5$ 6000-18000 $0.4\pm0.5$ 3,4
B0540–69 3.2 12.70 38.17 $1.88\pm0.01$ 3605-6060 $1.07\pm0.2$ 3300-8000 $0.88\pm0.11$ 5,4
Vela 4.1 12.53 36.84 $-0.04\pm0.04$ 4000-8000 $0.01\pm0.02$ 1500-16000 $1.2\pm0.3$ 6,7,8
B0656+14 5.0 12.67 34.58 $0.2\pm0.2$ 4600-7000 0.41 4300-18000 $1.1\pm0.3$ 9,10,11
Geminga 5.5 12.21 34.51 $0.8\pm0.5$ 3700-8000 0.46 4300-16000 $0.7\pm0.1$ 12,10,11
B1929+10 6.5 11.71 33.59 $0.5\pm0.5$ 1700-3400 $1.7\pm0.1$ 13,14
B0950+08 7.2 11.39 32.75 $0.65\pm0.4$ 3600-8000 $0.9\pm0.1$ 15,16
---------- ------------- ---------- ---------------- ----------------- ------------------ ----------------- ------------------ ----------------- ----------
\[tabdatasummary\]
\(1) Sollerman et al. (2000), (2) Kirsch et al. (2006), (3) Kaplan & Moon (2006),(4) Gotthelf (2003),(5) Serafimovich et al. (2004), (6) this work, (7) Kargaltsev & Pavlov (2007), (8) Manzali et al. (2007), (9) Zharikov et al.(2007), (10) Koptsevich et al. (2001), (11) De Luca et al. (2005),(12) Martin et al. (1998), (13) Mignani et al. (2002), (14) Becker et al. (2006), (15) Zharikov et al. (2004), (16) Becker et al. (2004)
We have compared the spectrum of the Vela pulsar with those of the other rotation-powered INSs for which either medium-resolution spectroscopy or multi-band photometry is available (Fig. 4). Apparently, the complexity of the spectral flux distribution grows with the INS age. For the young objects the optical spectral energy distribution is dominated by a flat power-law continuum which brings the signature of synchrotron radiation produced by relativistic charged particles in the neutron star’s magnetosphere. For the older ones, a Rayleigh-Jeans component, ascribed to thermal radiation from the cooling neutron star’s surface, is also present. In all cases, no statistically significant evidence of emission or absorption features is found. For all the objects in Fig.4, Tab. 2 reports the measured spectral index $\alpha_O$ of the power-law component, either derived from spectroscopy or from broad-band photometry. In those cases where both values are available the agreement is rather good, with the only exception of PSR B0540–69. However, as discussed in Serafimovich et al. (2004), its spectrum is unrecoverably polluted by the background of the surrounding, bright compact ($\sim 4''$) synchrotron nebula. The comparison of the optical spectral indeces shows that it is difficult to find clear spectral templates for different INS groups[^5]. This is shown in Fig. 5 (top left panel), where we have plotted the optical spectral index $\alpha_O$ as a function of the INS spin-down age. When available, we have assumed as a reference the spectral index obtained from spectroscopy, with the obvious exception of PSR B0540–69 (see above).
Although the spectral index value seems generally to correlate with the spin-down age, with $\alpha_O \approx (0.12 \pm 0.04) Log(\tau)$, the errors are such that most of the points are also consistent with a costant distribution around the average value $<\alpha_O> =0.44 \pm
0.4$. Thus, there is no firm evidence for an evolution of the pulsars’ optical spectral index over four age decades. This is similar to what is found in the X-rays, where the available data also suggest that the value of the spectral index $\alpha_X$ does not depend on the spin-down age (see, e.g. Becker & Trümper 1997). A possible anti correlation between the optical spectral index and the rotational energy loss $\dot E$, with $\alpha_O \approx (-0.07 \pm
0.02) Log(\dot E)$, can also be recognized in the data (Fig. 5, top right panel). However, as before, the large errors do not allow to draw any firm conclusion. Finally, no trend can be recognized between the optical spectral index and the magnetic field $B$ (Fig.5, bottom left panel).
{width="8.0cm"} {width="8.0cm"} {width="8.0cm"} {width="8.0cm"}
As it has been pointed out (e.g. Mignani et al. 2004; Serafimovich et al. 2004), the optical and X-ray magnetospheric emission of rotation-powered pulsars are almost never described by the same spectral parameters. This is clearly shown in Fig. 5 (bottom right panel), where we have plotted the optical spectral index $\alpha_O$ vs. the X-ray one $\alpha_X$. As seen, the optical power-law index is often less steep than the X-ray one, suggesting that the spectra of rotation-power pulsars undergo a general turnover at lower energies. In particular, for the Crab the slope of the power-law swaps from positive to negative when passing from the X-ray to the optical/IR domain, underlying an even more marked spectral turnover. Furthermore, no general positive or negative correlation is found between $\alpha_O$ and $\alpha_X$, which indicates that the X-ray and optical magnetospheric emission, although likely produced by the same physical processes, are not directly correlated. This finding is particularly interesting in comparison with the strong correlation between the optical and X-ray luminosities of rotation-powered pulsars noticed by Zharikov et al. (2004) and by Zavlin and Pavlov (2004).
Summary
=======
We have presented the first optical spectroscopy observations of the Vela pulsar. The pulsar’s spectrum is characterized by a flat power-law with spectral index $\alpha =-0.04 \pm 0.04$, consistent with the values derived from broad-band photometry (Mignani & Caraveo 2001; Shibanov et al. 2003; Kargaltsev & Pavlov 2007). We have compared the newly derived optical spectral index of Vela with those of all rotation-powered INSs for which a power-law component has been identified in the optical/IR spectrum. While a trend can be recognized in the data, the large errors on the spectral index values for most of the objects prevent any claim for an evolution of the magnetospheric emission properties over four age decades. We also found no evidence for a correlation between the optical and X-ray spectral indeces, which indicates that the X-ray and optical magnetospheric emissions are not directly correlated. However, we showed that in the majority of cases, the optical spectral index is flatter than the X-ray one, suggesting a spectral turnover in the INSs’ spectra at low energies.
RPM is supported by a PPARC Rolling Grant. SZ acknowledges the support of the DGAPA/PAPIIT project IN101506 and of CONACYT 48493. We thank Werner Becker for the useful discussions and Oleg Kargaltsev for sending us the STIS/FUV pulsars fluxes. We thank our referee, Yury Shibanov, for his comments which helped to improve the quality of the manuscript.
Appenzeller, I., Fricke, K., Furtig, W., Gassler, W., Haffner, R., et al. 1998, The Messenger, 94, 1
Becker, W., Trümper, J., 1997, A&A, 326, 682
Becker, W., Weisskopf, M. C., Tennant, A. F., Jessner, A., Dyks, J., et al., 2004, ApJ, 615, 908
Becker, W., Kramer, M., Jessner, A., Taam, R. E.. Jia, J.J., et al., 2006, ApJ, 645, 1421
De Luca, A., Caraveo, P. A., Mereghetti, S., Negroni, M., Bignami, G. F., 2005, ApJ, 623, 1051
Fitzpatrick, E.L., 1999, PASP, 111, 63
Gouiffes, C. 1998, Proc. of “Neutron Stars and Pulsars: Thirty years after the discovery”, eds. N. Shibazaki et al., Universal Academic Press, Frontiers Science Series n. 24, p. 363
Gotthelf, E., 2003, ApJ, 591, 361
Hamuy, M., Suntzeff, N. B., Heathcote, S. R., Walker, A. R., Gigoux, P., et al., 1994, PASP, 106, 566
Hill, R.J., Dolan, J. F., Bless, R. C., Boyd, P. T., Percival, J. W., et al., 1997, ApJ, 486, L99
Kaplan, D.L & Moon, D. 2006, ApJ 644, 1056
Kargaltsev, O. Y., Pavlov, G. G., Zavlin, V. E., Romani, R. W., 2005, ApJ, 625, 307
Kargaltsev, O., Pavlov, G, G., 2007, ApSS, 308, 287
Kirsch, M. G. F., Scoönherr, G., Kendziorra, E., Freyberg, M. J., Martin, M., et al., 2006, A&A, 453, 173
Koptsevich, A. B., Pavlov, G. G., Zharikov, S. V., Sokolov, V. V., Shibanov, Yu. A. et al. 2001, A&A 370, 1004
Lasker, B., 1976, ApJ 203, 193
Manzali, A., De Luca, A., Caraveo, P.A., 2007, ApJ accepted, (arXiv:0706.3194)
Martin, C., Halpern, J. P., Schiminovich, D., et al., 1998, ApJ, 494, L211
Mignani, R. P., Caraveo, P. A., 2001, A&A, 376, 213
Mignani, R.P., 1998, Proc. of “Neutron Stars and Pulsars: Thirty Years after the Discovery”, eds. N. Shibazaki et al., Universal Academy Press, Frontiers Science Series, n. 24, p.335 Mignani, R. P., De Luca, A., Caraveo,P.A., Becker, W., 2003, ApJ, 580, L147
Mignani, R. P., De Luca, A., Kargaltsev, O., Pavlov, G. G., Zaggia, S. et al. 2003, ApJ, 594, 419
Mignani, R.P., De Luca, A., Caraveo, P.A. 2004, Proc. of “Young Neutron Stars and Their Environments”, IAU Symp. 218, eds. F. Camilo and B. Gaensler, ASP Conf. Proc., p. 391
Mignani, R.P., 2005, Proc. of “The Electromagnetic Spectrum of Neutron Stars”, eds. A. Baykal et al., Springer, 210, 133
Mignani, R.P., Bagnulo, S., Dyks, J., Lo Curto, G., Słowikowska, A., 2007, A&A, 467, 1157
Moro, D. & Munari, U., 2000, A&A Suppl., 147, 361
Nasuti, F. P., Mignani, R., Caraveo, P. A., Bignami, G. F., 1996, A&A, 314, 849
Nasuti, F.P., Mignani, R., Caraveo, P.A., Bignami, G.F., 1997, A&A, 323,839
Oke, J.B., 1969, ApJ, 156, L49
Patat, F., 2004, The Messenger, 118, 11
Romani, R. W., Kargaltsev, O., Pavlov, G, G., 2005, ApJ, 627, 383
Serafimovich, N.I., Shibanov, Y. A., Lundqvist, P., Sollerman, J., 2004, A&A, 425, 1041
Shibanov, Y. A., Koptsevich, A. B., Sollerman, J., Lundqvist, P., 2003, A&A, 406, 645
Shibanov, Yu. A., Sollerman, J., Lundqvist, P., Gull, T., Lindler, D., 2005, A&A, 440, 693
Sollerman, J., Lundqvist, P., Lindler, D., Chevalier, R. A., Fransson, C., et al., 2000, ApJ, 537, 861
Wagner, S.J., & Seifert, W. 2000, Proc. of “Pulsar Astronomy - 2000 and Beyond”, eds. M. Kramer et al., ASP Conf. Series, 202, 315
Wallace, P.T., Peterson, B. A., Murdin, P. G., Danziger, I. J., Manchester, R. N., et al., 1977, Nature, 266, 692
van Kerkwijk, M. H., Kulkarni, S. R., 2001, A&A, 378, 986
Zavlin, V. E., Pavlov, G. G., 2004, ApJ, 616, 452
Zharikov, S. V., Shibanov, Yu. A., Mennickent, R. E., Komarova, V. N., Koptsevich, A. B., et al., 2004, A&A, 417, 1017
Zharikov, S., Mennickent, R. E., Shibanov, Y., Komarova, V., 2007, ApSS, 308, 545
[^1]: Based on observations collected at the European Southern Observatory, Paranal, Chile under programme ID 66.D-0261(A)
[^2]: http://ulisse.pd.astro.it/Astro/ADPS/
[^3]: www.eso.org/instruments/fors
[^4]: http://www.eso.org/observing/dfo/quality/FORS1
[^5]: We warn here that the values of $\alpha_O$ have been computed over slightly different wavelength ranges, which makes a direct comparison more uncertain
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'It is thought that many characteristics of the gaseous features within the central parsec of our Galaxy, are associated with the accretion of ambient plasma by a central concentration of mass. Using a 3D hydrodynamical code, we have been simulating this process in order to realistically model the gaseous flows in the center of our Galaxy. In the most recent simulation, we have taken into account the multi-point-like distribution of stellar wind sources, as well as the magnetic heating and radiative cooling of these stellar winds. As expected, we find that the structure of the flow is significantly different from that due to a uniform medium. We also investigate the possibility that Sgr A\* is due to a distributed mass concentration instead of the canonical point mass of a black hole. We discuss the physical state of the accreting gas and how our results suggest that Sgr A\* is unlikely to be associated with a “dark cluster”.'
author:
- 'R.F. Coker'
- Fulvio Melia
title: 3D Simulation of the Gas Dynamics in the Central Parsec of the Galaxy
---
@scaling[.95]{} \#1[@scaling[\#1]{}]{} \#1[=@scaling]{}
Introduction
============
The hydrodynamics of the interstellar medium (ISM) within the central parsec of the Galaxy has long been thought to be dominated by the gravitational potential due to a central mass concentration (for a review, see Mezger, Duschl, & Zylka, 1996). For example, the Western Arc, one of the dominant kinematic features of the region, appears to be in circular rotation about Sgr A\*, a unique compact radio source whose lack of proper motion suggests it lies at the dynamical heart of the Galaxy (Roberts & Goss, 1993; Backer, 1994). Stellar motions (Genzel, et. al., 1997), gas kinematics (Herbst, et. al., 1993), and velocity dispersion measurements (Eckart & Genzel, 1998) together suggest the presence of $\sim 2.6\times 10^6 M_\odot$ of dark mass located within $\sim .01$ parsecs of Sgr A\*. Note that the Galactic center (GC) is at a distance of $\sim 8$ kpc so that $1$ arcsecond $\sim 0.04$ parsecs.
However, showing that the GC must contain a centralized dark mass concentration does not necessarily imply that it is in the form of a single compact object nor does it imply that Sgr A\* must be associated with it. Stellar kinematic arguments (Genzel et. al., 1996) rule out a distribution of neutron stars or white dwarfs. One possibility is that the dark mass distribution consists of $\sim 10 M_\odot$ black holes; cluster evolution calculations (Lee, 1995) have shown that this is at least feasible although stability arguments (Maoz, 1998) suggest the cluster lifetime would be $\sim 10^8$ years, considerably less than the age of the Galaxy. In this paper, we wish to determine if, stability arguments aside, the spectrum of Sgr A\* could be due to a plasma trapped within the potential well of a dark cluster of arbitrary objects.
In addition to large scale gaseous features, there is ample evidence for the existence of rather strong stellar winds in and around Sgr A\* itself. The key wind sources appear to be the cluster of mass-losing, blue, luminous stars comprising the IRS 16 assemblage, located within several arcseconds of Sgr A\*. A variety of observations over the years (for a review see Morris & Serabyn, 1996) provide clear evidence of a hypersonic wind, with a velocity of $v_w \sim 500-1000
$ km s$^{-1}$, a number density $n_w \sim 10^{3-4}$ cm$^{-3}$, and a total mass loss rate $\dot M_w \sim 3-4\times 10^{-3} M_\odot~$yr$^{-1}$ pervading the inner parsec of the Galaxy. If the dark matter is distributed, it is likely that a portion of this wind is captured by the dark compact cluster and that it settles within the cluster’s potential well. Although the potential well of a cluster does not include a cusp such as that due to a black hole, the trapped plasma might conceivably still account for at least some of Sgr A\*’s radiative characteristics.
In §2 we discuss the simulation of the gas flow through a distributed dark mass cluster. In §3 we describe the resulting spectrum and present some preliminary semi-analytical spectral calculations based on Sgr A\* being a point mass while in §4 we summarize our analysis.
The 3D Hydrodynamical Model
===========================
In the classical Bondi-Hoyle scenario (Bondi & Hoyle, 1944), the mass accretion rate for a uniform hypersonic adiabatic gas flowing past a centralized mass is $$\label{mdot}
\dot M_{BH} = \pi {R_A}^2 m_H n_w v_w\;,$$ where $R_A \equiv 2 G M / {v_w}^2$ is the accretion radius and $M$ is the mass of the centralized object(s). At the GC, for the conditions described in the Introduction, we would therefore expect an accretion rate $\dot M_{BH} \sim
10^{21-22}$g sec$^{-1}$, with a capture radius $R_A \sim 0.01-0.02~$pc. Since this accretion rate is sub-Eddington for a $\sim$ one million solar mass concentration, the accreting gas is unimpeded by the escaping radiation field and is thus essentially in hydrodynamic free-fall starting at $R_A$. Our initial, simplistic, numerical simulations of this process, where we assume a point object and uniform flow (Ruffert & Melia, 1994; Coker & Melia, 1996) have verified these expectations.
The Stellar Wind Sources
------------------------
The GC wind, however, is unlikely to be uniform since the winds from many stars contribute to the mass ejection. We assume that the early-type stars enclosed (in projection) within the Western Arc, the Northern Arm, and the Bar produce the observed wind. Thus far, 25 such stars have been identified (Genzel, et. al., 1996), though the stellar wind characteristics of only 8 have been determined from their He I line emission (Najarro, et. al., 1997); the relavant characteristics of these 25 stars are summarized in Table 1. Two sources, IRS 13E1 and IRS 7W, seem to dominate the mass outflow with their high wind velocity ($\sim 1000 $ km sec$^{-1}$) and a mass loss rate of more than $2\times10^{-4}\;M_\odot$ yr$^{-1}$ each. Unfortunately, the temperature of the stellar winds is not well known, and so for simplicity we have assumed that all the winds are Mach 30; this corresponds to a temperature of $10^{4-5}$K. In addition, for the sources that are used in these calculations, their location in $z$ (i.e., along the line of sight) is determined randomly with the condition that the overall distribution in this direction approximately matches that in $x$ and $y$.
----------- ------------ ------------ ------------ ------------------- ----------------------------------------
Star x (arcsec) y (arcsec) z (arcsec) v (km sec$^{-1}$) $\dot M({10}^{-5}\;M_\odot$ yr$^{-1}$)
IRS 16NE -2.6 0.8 5.5 550 9.5
IRS 16NW 0.2 1.0 7.3 750 5.3
IRS 16C -1.0 0.2 -7.1 650 10.5
IRS 16SW -0.6 -1.3 4.9 650 15.5
IRS 13E1 3.4 -1.7 -1.5 1000 79.1
IRS 7W 4.1 4.8 -5.1 1000 20.7
AF 7.3 -6.7 8.5 700 8.7
IRS 15SW 1.5 10.1 700 16.5
IRS 15NE -1.6 11.4 750 18.0
IRS 29N 1.6 1.4 3.5 750 12.9
IRS 33E 0.0 -3.0 1.5 750 12.9
IRS 34W 3.9 1.6 -6.4 750 12.9
IRS 1W -5.3 0.3 7.8 750 12.9
IRS 9NW -2.5 -6.2 -3.8 750 12.9
IRS 6W 8.1 1.6 3.6 750 12.9
AF NW 8.3 -3.1 -2.1 750 12.9
BLUM 9.2 -5.0
IRS 9S -5.5 -9.2
Unnamed 1 1.3 -0.6
IRS 16SE -1.4 -1.4
IRS 29NE 1.1 1.8
IRS 7SE -2.7 3.0
Unnamed 2 3.8 -4.2
IRS 7E -4.2 4.9
AF NWW 10.2 -2.7
----------- ------------ ------------ ------------ ------------------- ----------------------------------------
: Parameters for Galactic Center Wind Sources
The sources are assumed to be stationary over the duration of the simulation. The stars without any observed He I line emission have been assigned a wind velocity of $750$ km sec$^{-1}$ and an equal mass loss rate chosen such that the total mass ejected by the 14 stars used here is equal to $3\times10^{-3}\;M_\odot$ yr$^{-1}$. Note that although we have matched the overall mass outflow rate to the observations, we have only used 14 of the 25 stars in the sample. There are two principal reasons for this: (1) stars further away than 10 arcsec (in projection) from Sgr A\* are outside of our volume of solution and therefore could not be included, and (2) due to our computational resolution limits, we needed to avoid excessively large local stellar densities.
In a complex flow, generated by many wind sources, the wind velocity and density are not uniform, so the accretion radius is not independent of angle. To set the length scale for the simulation, we shall therefore adopt the value $R_A = .018 $pc (for which 1$^{\prime\prime}$ = 2.3 $R_A$) as a reasonable mean representation of this quantity.
The Dark Cluster Potential
--------------------------
We wish to study the emission characteristics of a hot, magnetized plasma “trapped” within the dark cluster’s gravitational potential well. Following Haller & Melia (1996), we will represent the gravitational potential of the dark cluster with an “$\eta$-model” (Tremaine, et. al., 1994). This function represents an isotropic mass distribution with a single parameter. We here restrict our examination to the case $\eta=2.5$ since this provides the closest approximation to a King model that is physically realizable (i.e., a nonnegative distribution function). We scale the mass so that $2\times10^6\;M_\odot$ are enclosed within $0.01$ pc and the total integrated mass of the dark cluster is $2.7\times10^6\;M_\odot$. Thus, writing $r$ in units of $R_A$ we get for the enclosed mass as a function of $r$, $$M_\eta(r) = 2.7\times10^6 \left({{13.84r}\over{1+13.84r}}
\right)^{5/2} \;M_\odot\;.$$ A more recent assessment of the enclosed mass (Genzel, et. al., 1997) places a yet more rigorous constraint on the possibility of a distributed dark matter component. These newer observations may indeed invalidate the idea that [*any*]{} realistic stable distribution of objects can account for the observed gravitational potential.
The Hydrodynamics Code
----------------------
We use a modified version of the numerical finite difference algorithm ZEUS, a general purpose code for MHD fluids developed at NCSA (Stone & Norman, 1992; Norman, 1994). The code was run on the massively parallel Cray T3E at NASA’s Goddard Space Flight Center under the High Performance Computing Challenge program. Zone sizes are geometrically scaled by a factor of 1.02 so that the central zones are $\sim 20$ times smaller than the outermost zones, mimicking the “multiply nested grids” arrangement used by other researchers (e.g., Ruffert & Melia, 1994). This allows for maximal resolution of the central region (within the computer memory limits available) while sufficiently resolving the wind sources and minimizing zone-to-zone boundary effects. The total volume is $(40 R_A)^3$ or $\sim(0.7 $pc$)^3$ with the center of the spherically symmetric dark cluster distribution being located at the origin.
The density of the gas initially filling the volume of solution is set to a small value and the velocity is set to zero, while the internal energy density is chosen such that the initial gas temperature is $\sim10^2$ K. Free outflow conditions are imposed on the outermost zones and each time step is determined by the Courant condition with a Courant number of 0.5. The 14 stellar wind sources are modeled by forcing the velocity in 14 subregions of 125 zones each to be constant with time while the densities in these subvolumes are set so that the total mass flow into the volume of solution from each source is given by Table 1. Also, the magnetic field of the winds is assumed to always be at equipartition with the thermal energy density. The angular momentum and mass accretion rates are calculated by summing the relevant quantity in zones located within $0.1 R_A$ of the origin.
We assume the magnetic field is perfectly tangled and thus ignore the effects of the magnetic field on the large scale kinematics. We take the medium to be an adiabatic polytropic gas, with $\gamma = 5/3$. Building on previous work (Melia, 1994; Coker & Melia, 1997), we have included a first order approximation to magnetic dissipative heating as well as an accurate expression for the cooling due to magnetic bremsstrahlung, thermal bremsstrahlung, line emission, radiative recombination, and 2 photon continuum emission for a gas with cosmic abundance. For magnetic heating, we assume that the magnetic field never rises above equipartition. If compression and flux conservation would otherwise dictate a magnetic field larger than the equipartition value, the field lines are assumed to reconnect rapidly, converting the magnetic field energy into thermal energy, thereby re-establishing equipartition conditions. The cooling function includes a multiple-Gaussian fit to the relevant cooling emissivities provided by N. Gehrels (see Gehrels & Williams, 1993, and references cited therein), though with the thermal bremsstrahlung portion supplanted with more accurate expressions that are valid over a broader range of physical conditions and with the inclusion of magnetic bremsstrahlung. For details on the cooling expressions used in the hydrodynamics code as well as the spectrum calculations below, see Melia & Coker (1999). Note that cooling due to Comptonization and any pair production have not yet been included since they are not thought to be significant in the vicinity of Sgr A\*. Also, it is assumed that the optical depth is small throughout the volume of solution.
Results of the 3D Simulation
----------------------------
The integrated emissivity along the line-of-sight, at $\sim$ 1450 years after the start of the simulation, is shown in Figure 1. The grey scale is logarithmic with solid white corresponding to a frequency-integrated intensity of $\sim1.1\times10^{5}$ erg cm$^{-2}$ s$^{-1}$ steradian$^{-1}$, and black corresponding to $\sim1\times10^{-2}$ erg cm$^{-2}$ s$^{-1}$ steradian$^{-1}$. Sgr A\* is located at the center of the image. Of particular interest in this image is the appearance of streaks of high-velocity, high-density gas (“streamers”) that are very reminiscent of features, such as the so-called “Bullet” seen near the Galactic center (Yusef-Zadeh, et al. 1996). In our simulation, these structures are produced predominantly within the wind-wind collision regions, and in the future, we shall consider in greater detail the possibility that the observed high-velocity gas components near Sgr A\* are produced in this fashion. Note the dominant role played by IRS 13E1 to the lower right of the compact radio source.
After reaching equilibrium several sound crossing times ($\sim 1000$ years) after the start of the simulation, the enclosed mass and energy begin to fluctuate aperiodically on time scales of less than a few decades and with an amplitude of up to $50\%$, reflecting the turbulent cell nature of the flow in and out of the central region. Typically $2.7\times 10^{-3} \;M_\odot$ of gas is trapped within the cluster at any given time. Also, although the gas is generally supersonic, with most of the energy in kinetic form, the thermal energy can be boosted rather suddenly when the enclosed magnetic field energy is dissipated. This occurs when strong shocks pass through the central region; the shocks compress the field sufficiently to the point where it reaches, or even surpasses, equipartition and dissipation ensues.
The Spectrum of Sgr A\*
=======================
The Spectrum due to a Dark Cluster
----------------------------------
In order to calculate the observed continuum spectrum, we assume that the observer is positioned along the negative $z$-axis at infinity and we sum the emission from all zones that are located at a projected distance, $R_{xy}$, of less than $0.1\,R_A$. Since the size of Sgr A\* (at $\lambda$7 mm) is $\approx 10^{12-13}$ cm (Bower & Backer, 1998) and the smallest cell size in our simulation is $7\times10^{14}$ cm, in order to minimize the inaccuracy due to numerical fluctuations, we have calculated the spectrum from a central region roughly $10$ times this size ($0.1 R_A$), to include at least $100$ zones. Thus clearly our predicted spectrum constitutes an upper limit to the actual emission expected from Sgr A\*. At the temperature and density that we encounter here, the dominant components of the continuum emissivity are electron-ion and electron-electron bremsstrahlung; we ignore line emission in these spectral calculations.
As discussed in Melia (1998), the density in the central region reaches a peak value of roughly $10^8$ cm$^{-3}$ and the temperature is never greater than about $10^8$ K. Thus, since electrons begin to emit significant synchrotron radiation only above a few times $10^9$ K, the gas can only emit cyclotron radiation. However, the cyclotron emissivity is insignificant compared to bremsstrahlung so that the final spectrum, as shown in Figure 2, is a bremsstrahlung spectrum that,in the radio, falls more than 4 orders of magnitude short of the observed luminosity of Sgr A\* (compare with Figure 6b in Melia, 1998). Figure 2 shows spectra for 3 points in time during the 3D hydrodynamical simulation. The solid, dotted, and dashed curves are at approximately 1200, 1300, and 1400 years, respectively, after the start of the simulation. The flattening of the density, temperature, and magnetic field profiles is a direct consequence of the shallowness of the dark cluster potential compared to the steep potential gradient encountered by gas falling into a black hole. The high-energy shoulder of the emission shown in Figure 3 is characteristic of the highest temperature $T_{max}$ ($\sim 10^8$ K) attained by the gas and is consistent with the brightness temperature limit associated with Sgr A\*. Similarly, the X-ray and $\gamma$-ray emission is significantly below the observed upper limits (Predehl & Truemper, 1994; Goldwurm, et. al., 1994).
The Spectrum due to a Black Hole
--------------------------------
Although we have not yet run a 3D hydrodynamical simulation using the gravitational potential due to a point mass, we have undertaken semi-analytical calculations that improve upon earlier work (Melia, 1994). Although details will be presented elsewhere (Coker & Melia, 1999), we here present some preliminary results. We now explicitly solve for the velocity of the spherical flow using the relativistic Euler equation and integrate over $\mu$, the cosine of the angle between the line-of-sight and the flow. We also include effects due to inverse Compton scattering and the index of refraction.
In recent work (Lo, et. al., 1998) it has been suggested that, at mm wavelengths at least, the intrinsic size of Sgr A\* varies as $\nu^\alpha$ with $-1.9 < \alpha < -0.7$. It is difficult to compare the absolute observed size, which is usually the FWHM of a best-fit Gaussian (e.g. Krichbaum, et. al., 1998), to a theoretical size, which is usually given as the surface of last scatterring where $\tau(\nu) \equiv 1$; the observational size will tend to be larger than the theoretical size with these definitions. However, Figure 3 shows a plot of the predicited [*intrinsic*]{} size of Sgr A\* along with present observational values (see Lo, et. al. 1998 and references cited therein for details). The slope of the line is $\sim -0.6$, somewhat less steep than the observed value. This is not surprising since this particular model does not produce enough low frequency magnetic bremsstrahlung at large radii. The slope should steepen when a more optimal fit to the spectrum is found. Note that the lower size bound on the plot corresponds to the size of the black hole itself ($1 R_s \sim 6 \mu$as).
Summary
=======
It does not appear, based on these calculations, that the gravitational potential of a distributed dark mass, be it due to a compact cluster of stellar remnants or an even more exotic collection of objects (e.g. Tsiklauri & Viollier, 1998), can compress the gas from stellar winds to the point where the temperature, density, and magnetic field can produce an observationally significant cyclotron/synchrotron emissivity at GHz frequencies. Although the spatial resolution of the simulation near the origin can be improved over that used here, it is unlikely that it will alter the results; the lack of compression is due to the inherently flat potential of any distributed dark cluster. Thus, aside from the issue of whether any dark cluster can account for the observed GC potential and whether any such cluster is stable over a significant fraction of the age of the Galaxy, our 3D hydrodynamical simulation has shown that the radio emissivity from the gas trapped within any such cluster cannot reprodude the spectrum of Sgr A\*. We are left to conclude that either Sgr A\* is unrelated to the accreting or trapped GC gas or that Sgr A\* is the signature of an accreting, massive black hole. However, models of the structure and mechanism of radio emission from such a black hole are now faced with futher contraints due to recent observations of the intrinsic source structure of Sgr A\*. Coupled with the observed spectral energy distribution of Sgr A\*, we may be close to differentiating between the various accretion models.
This work was partially supported by NASA grant NGT-51637 and has made use of NASA’s Astrophysics Data System Abstract Service.
Backer, D.C. 1994, in The nuclei of normal galaxies, ed. R. Genzel & A.I. Harris (Dordrecht: Kluwer), 403. Bondi, H, & Hoyle, F. 1944, , 104, 273. Bower, G.C. & Backer, D.C. 1998, , 496, 97L. Coker, R., & Melia, F. 1996, in ASP Conf. Vol. 102, The Galactic Center: 4th ESO/CTIO Workshop, ed. R. Gredel (San Francisco: ASP), 403. Coker, R. & Melia, F., 1997, , 488, L149. Coker, R. & Melia, F., 1999, in preperation. Eckart, A. & Genzel, R., 1998, these proceedings. Gehrels, N. & Williams, E. D. 1993, , 418, L25. Genzel, R., Thatte, N., Krabbe, A., Kroker, H., & Tacconi-Garman, L.E., 1996, , 472, 153. Genzel, R., Eckart, A., Ott, T. & Eisenhauer, F. 1997, , 291, 219. Goldwurm, A., et. al., 1994, Nature, 371, 589. Haller, J.M., & Melia, F. 1996, , 464, 774. Haller, J.M., Rieke, M.J., Rieke, G.H., Tamblyn, P., Close, L., & Melia, F. 1996, , 456, 194. Herbst, T.M., Beckwith, S.V.W., Forrest, W.J., Pipher, J.L. 1993, , 105, 956. Lee, H.M. 1995, , 272, 605. Lo, K.Y., Shen, Z.-Q., Zhao, J.-H., & Ho, P. 1998, , 508, L61. Lutz, D., Krabbe, A. & Genzel, R. 1993, , 418, 244. Maoz, E. 1998, , 494, L181. Melia, F. 1994, , 426, 577. Melia, F. 1998, these proceedings. Melia, F. & Coker, R. 1999, , in press. Mezger, P.G., Duschl, W.J. & Zylka, R. 1996, , 7, 289. Morris, M. & Serabyn, E. 1996, , 34, 645. Najarro, F., Krabbe, A., Genzel, R., Lutz, D., Kudritzki, R., & Hillier, D. 1997, , 325, 700. Narayan, R., Yi, I. & Mahadevan, R. 1995, Nature, 374, 623. Norman, M. 1994, , 184, 50.01. Predehl, P. & Truemper, J. 1994, å, 290, 29. Roberts, D.A. & Goss, W.M. 1993, , 86, 133. Ruffert, M., & Melia, F. 1994, [*A.A.Letters*]{}, 288, L29. Stone, J.M., & Norman, M.L. 1992, , 80, 753. Tremaine, S., Richstone, D. O., Byun, Y., Dressler, A., Faber, S. M., Grillmair, C., Kormendy, J., & Lauer, T. R. 1994, , 107, 634. Yusef-Zadeh, F., Roberts, D.A., Goss, W.M., Frail, D. & Green, A. 1996, , 466, L25.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In systems with detected planets, hot-Jupiters and compact systems of multiple planets are nearly mutually exclusive. We compare the relative occurrence of these two architectures as a fraction of detected planetary systems to determine the role that metallicity plays in planet formation. We show that compact multi-planet systems occur more frequently around stars of increasingly lower metallicities using spectroscopically derived abundances for more than 700 planet hosts. At higher metallicities, compact multi-planet systems comprise a nearly constant fraction of the planet hosts despite the steep rise in the fraction of hosts containing hot and cool-Jupiters. Since metal poor stars have been underrepresented in planet searches, this implies that the occurrence rate of compact multis is higher than previously reported. Due to observational limits, radial velocity planet searches have focused mainly on high-metallicity stars where they have a higher chance of finding giant planets. New extreme-precision radial velocity instruments coming online that can detect these compact multi-planet systems can target lower metallicity stars to find them.'
author:
- 'John M. Brewer'
- Songhu Wang
- 'Debra A. Fischer'
- 'Daniel Foreman-Mackey'
title: 'Compact multi-planet systems are more common around metal poor hosts'
---
Introduction {#sec:intro}
============
The gas giant planet-metallicity correlation provided strong support for the core-accretion model [@1996Icar..124...62P] over gravitational instability: massive cores should form more rapidly in more metal rich (i.e. more massive) disks. Because the correlation is for giant planets on short period orbits, radial velocity searches made it possible to develop clean samples of stars with and without high-mass planets. As the number of discovered planets has exploded, we now know that almost all stars have planets , making it impossible to know if a given star is planet-free or simply has planets that evade our limited detection capabilities.
The search for extrasolar planets has identified two notable system architectures in the region close to the host star: multiple small planets on tight orbits, compact multi-planet systems [@2011ApJS..197....8L] and massive planets on short orbits, hot-Jupiters. These two system architectures are almost mutually exclusive, with few hot-Jupiters having close companions and almost no compact multi-planet systems having nearby massive planets. Hot-Jupiters are uncommon, but occur more frequently around stars with high amounts of heavy elements but small planets can occur around stars with a wide range of metallicities [@2012Natur.486..375B; @2014Natur.509..593B; @2015AJ....149...14W; @2018AJ....155...89P].
Compact multi-planet systems are mostly composed of planets near or below the current detection limits of radial velocity surveys. However, they tend to be very co-planar [@2014ApJ...790..146F], making them easy to detect in the *Kepler* transiting planet survey [@2010Sci...327..977B]. Some groups have looked at the average metallicity of these planets as a function of their radius and found that smaller planets are found around stars with a lower average metallicity [@2014Natur.509..593B; @2015AJ....149...14W; @2018MNRAS.480.2206O]. This hints that protoplanetary disks with less available solids may struggle to form larger planets, or even planets at all, but issues of selection bias and detection completeness still obscure a complete picture.
To circumvent these problems, we chose to look only at systems with detected planets and compare the properties of systems with unique architectures. Planet detection should not be biased toward a given type of system at a particular metallicity. We derived uniform stellar properties and elemental abundances from high resolution spectra for almost 3000 stars, including almost 1200 planet hosts [@Brewer:2016gf; @2018ApJS..237...38B] and compared the number of systems of a given architecture to all known hosts as a function of the host metallicity. As stars evolve, the measured surface abundances of heavy elements can decrease [@2017ApJ...840...99D; @2018ApJ...857...14S], which might bias any analysis of the influence of those elements on planet architecture. We limited our analysis to un-evolved stars ($\log g > 4.0$) and also looked at ratios of heavy elements to confirm our findings, since those ratios are relatively static over the main sequence lifetime of the stars.
Data and Analysis
=================
Stellar Properties and Abundances {#sec:abundances}
---------------------------------
The stellar properties and abundances used for this Letter were all derived from high resolution optical spectra taken with the Keck HIRES spectrograph and analyzed in a homogeneous manner [@Brewer:2016gf; @2018ApJS..237...38B]. The analysis procedure has been shown to recover surface gravities consistent with those from asteroseismology to within 0.05 dex [@2015ApJ...805..126B] in addition to accurate temperatures and precise abundances for the abundances of 15 elements (C, N, O, Na, Mg, Al, Si, Ca, Ti, V, Cr, Mn, Fe, Ni, and Y). Statistical uncertainties for the abundances range from $\sim 0.01$ dex (iron and silicon) to 0.04 dex (nitrogen), depending on the element.
The parameters and abundances are derived using forward modeling performed with the analysis package Spectroscopy Made Easy , fitting in an iterative fashion. After continuum normalizing the spectrum and extracting 20 short wavelength segments totaling 350 Å between 5160 and 7800 Å, the initial temperature and gravity are set using broadband colors and the abundance pattern is set to that of the Sun. We then fit for the global stellar properties (effective temperature, surface gravity, rotational and Doppler broadening, metallicity) and the abundances of three $\alpha$ elements (Ca, Si, Ti) to allow for departures from the solar abundance pattern. Using the derived parameters from this first fit, we perturb the temperature by $\pm 100$ K and re-fit, taking the $\chi^2$ weighted average as the global parameters from this stage. The global properties are then fixed, and we solve for the abundances of the 15 elements. This set of parameters and the abundance pattern are then used as a new starting point for a second iteration of the procedure.
The parameters and abundances are precise for dwarf stars of high signal-to-noise ratio (S/N), but trends in abundance with temperature have been identified for evolved stars and those with S/N $< 45$. To avoid potential contamination, we removed all stars with S/N $< 45$ and log surface gravities ($\log g$) $< 4.0$. The resulting combined catalog contains 1,148 planets around 716 stars.
Planets and System Architectures
--------------------------------
We cross-matched the stellar catalog with the confirmed planet catalog from the NASA Exoplanet Archive and adopted those planet parameters. We then defined three classes of exoplanet system architecture: hot-Jupiters, cool-Jupiters, and compact multi-planet systems. Hot-Jupiter systems are defined as having a planet with $M_{\mathrm{planet}} > 0.5 M_{\mathrm{Jupiter}}$ or $R_{\mathrm{planet}} > 0.75 R_{\mathrm{Jupiter}}$ and semi-major axis $<=0.3$ au resulting in 104 hot-Jupiter systems. Cool-Jupiters have the same mass or radius definition as hot-Jupiter planets, but have semi-major axes $>0.3$ au. This results in 87 cool-Jupiter systems. Finally, we defined compact multi-planet systems as having three or more planets orbiting at less than 1 au, resulting in 105 compact multi-planet systems. Only one hot-Jupiter system is also defined as a compact multi-Planet system, and nine cool-Jupiter systems are in the compact multi-Planet sample.
Planet Architecture Occurrence Analysis
---------------------------------------
To evaluate the relative occurrence rate of each system architecture as a fraction of known planet hosts at a given metallicity, we used Gaussian kernel density estimates (KDE) for the entire sample and each sub-population as a function of \[Fe/H\]. The optimum bandwidth was determined using Scott’s rule[@Anonymous:1979cw]. From the KDEs we generated probability density functions (PDF) and then divided the PDF of each architecture by that of all known hosts. This gives us an estimate of the fraction of planetary systems of a particular architecture given the overall occurrence of discovered planetary systems as a function of metallicity. To visualize the uncertainties in these occurrence rates, we drew 200 bootstrap realizations from each of the architecture samples and performed the same procedure for these samples. We then calculated the 68% and 95% confidence regions over the entire metallicity range based on those bootstrap realizations.
Results
=======
Comparing these planet architecture ratios, we find two opposing trends versus their log solar relative iron abundance, or \[Fe/H\] (Figure \[fig:arch\_frequency\]). For hot-Jupiters, we see the expected planet-metallicity correlation. The frequency of hot-Jupiters increases with increasing metallicity. At the highest metallicities, the frequency of compact multi-planet systems also increases, but the frequency is also consistent with being almost flat from $-0.3 < \mathrm{[Fe/H]} < 0.3$. However, at metallicities below -0.3 dex there is a sharp increase in the fraction of known planet hosts that are compact multi-planet systems, with a factor of three increase in the probability density over a range of just 0.2 dex.
{width="\textwidth"}
Recent studies have suggested that cool-Jupiters, giant planets residing more than 1 AU from their host star, may be companions to compact multi-planet systems where large mutual inclinations prevent us from seeing one or the other [@2014ApJ...796...47M; @2018ApJ...860..101Z]. Due to their distance from their host star, it is more difficult to detect cool-Jupiters through either radial velocities or transits and as a consequence we are much less complete. However, it is instructive to compare their distribution to the other hosts to see if they are clearly associated with one or the other population. As a function of metallicity, cool-Jupiters seem to closely trace the behavior of the hot-Jupiter systems and metallicities higher than $\sim -0.3$. The original planet-metallicity study included systems with planets on periods shorter than 4 years, which includes many cool-Jupiters, so this result is not too surprising. However, for systems with $\mathrm{[Fe/H]} < -0.3$, there is an increase in cool-Jupiter frequency similar to that of the compact multi-planet systems, although less strong and driven by the nine overlapping systems between the cool-Jupiters and compact multis.
{width="\textwidth"}
As a star ages, diffusion at the base of the convective zone can result in an apparent decrease in the amount of heavy elements at the stellar surface. The effect is more pronounced for more massive stars, but affects all elements heavier than helium roughly equally. Stars with lower initial metallicity should also have a higher ratio of $\alpha$-elements to iron , so we can use the Si/Fe ratio as a function of \[Fe/H\] to see if the increase in frequency of compact multis is due to age or inherently low metallicity (Figure \[fig:arch\_sife\_feh\]). At both low metallicity and high Si/Fe, none of the hosts are hot-Jupiters and an increasing fraction are compact multi-planet systems. Low-metallicity stars in our sample have a higher Si/Fe, as expected for stars with initially low metallicity. The metallicity relation we see is not related to an observational bias caused by diffusion.
Discussion
==========
Between $-0.3 < [Fe/H] < 0.4$, the fraction of systems that are compact multis stays relatively constant despite the steep increase in the fraction of hot and cool-Jupiters. Compact multis are already known to be common around hosts of solar composition [@2018ApJ...860..101Z]. The increasing fraction of stars hosting compact multi-planet systems at lower metallicities points to a previously unrecognized reservoir of small-planet hosts. This has implications for planet formation models and may suggest that the these dynamically cool systems may form after the gas disk dissipates [@2018MNRAS.480.2206O]. Previous studies have found evidence for two possible populations of planets, one with low mutual inclinations and low obliquities, and a second dynamically hotter one with fewer planets [@2013ApJ...771...11A; @2016ApJ...816...66B]. A separate study published while this Letter was being submitted also suggests that planet multiplicity may be tied to metallicity, with multi-planet systems more common around lower-metallicity stars [@2018arXiv180809451Z], supporting our result.
Stars of lower metallicity and higher Si/Fe ratios are generally older or members of the galactic thick-disk population . This could point to a changing mix of planet architectures based on formation time and location. In fact, one of the oldest verified and low-metallicity planet hosts, Kepler-444, is home to a compact multi-planet system [@Campante:2015ei]. New high precision radial velocity surveys looking for Earth-massed planets [@2017arXiv171105250G; @2016SPIE.9908E..6TJ] may find a much larger population of small planets around these lower metallicity stars.
D. A. Fischer and J. M. Brewer gratefully acknowledge support for this work funded under NSF 1616086. S. Wang thanks the Heising-Simons Foundation for their support. Data presented herein were obtained at the W. M. Keck Observatory from telescope time allocated to the National Aeronautics and Space Administration through the agency’s scientific partnership with the California Institute of Technology and the University of California as well as time from the Yale University TAC. The Observatory was made possible by the generous financial support of the W. M. Keck Foundation. We thank the many observers of the California Planet Search, who collected the majority of the spectra for more than a decade. This research has made use of the NASA Exoplanet Archive, which is operated by the California Institute of Technology, under contract with the National Aeronautics and Space Administration under the Exoplanet Exploration Program.
The authors wish to recognize and acknowledge the very significant cultural role and reverence that the summit of Maunakea has always had within the indigenous Hawaiian community. We are most fortunate to have the opportunity to conduct observations from this mountain.
natexlab\#1[\#1]{}
Albrecht, S., Winn, J. N., Marcy, G. W., [et al.]{} 2013, The Astrophysical Journal, 771, 11
Ballard, S., & Johnson, J. A. 2016, The Astrophysical Journal, 816, 66
Borucki, W. J., Koch, D., Basri, G., [et al.]{} 2010, Science, 327, 977
Brewer, J. M., & Fischer, D. A. 2018, The Astrophysical Journal Supplement Series, 237, 38
Brewer, J. M., Fischer, D. A., Basu, S., Valenti, J. A., & Piskunov, N. 2015, The Astrophysical Journal, 805, 126
Brewer, J. M., Fischer, D. A., Valenti, J. A., & Piskunov, N. 2016, The Astrophysical Journal Supplement Series, 225, 32
Buchhave, L. A., Latham, D. W., Johansen, A., [et al.]{} 2012, Nature, 486, 375
Buchhave, L. A., Bizzarro, M., Latham, D. W., [et al.]{} 2014, Nature, 509, 593
Burke, C. J., Christiansen, J. L., Mullally, F., [et al.]{} 2015, The Astrophysical Journal, 809, 8
Campante, T. L., Barclay, T., Swift, J. J., [et al.]{} 2015, The Astrophysical Journal, 799, 170
Dotter, A., Conroy, C., Cargile, P., & Asplund, M. 2017, The Astrophysical Journal, 840, 99
Fabrycky, D. C., Lissauer, J. J., Ragozzine, D., [et al.]{} 2014, Astrophysical Journal, 790, 146
Fischer, D. A., & Valenti, J. 2005, The Astrophysical Journal, 622, 1102
Gonz[á]{}lez Hern[á]{}ndez, J. I., Pepe, F., Molaro, P., & Santos, N. 2017, arXiv.org, arXiv:1711.05250
Howard, A. W., Marcy, G. W., Bryson, S. T., [et al.]{} 2012, The Astrophysical Journal Supplement, 201, 15
Jurgenson, C., Fischer, D., McCracken, T., [et al.]{} 2016, in Proceedings of the SPIE, Yale Univ. (United States) (International Society for Optics and Photonics), 99086T
Kordopatis, G., Wyse, R. F. G., Gilmore, G., [et al.]{} 2015, Astronomy and Astrophysics, 582, A122
Lissauer, J. J., Ragozzine, D., Fabrycky, D. C., [et al.]{} 2011, Astrophysical Journal, Supplement Series, 197, 8
Morton, T. D., & Winn, J. N. 2014, The Astrophysical Journal, 796, 47
Owen, J. E., & Murray-Clay, R. 2018, Monthly Notices of the Royal Astronomical Society, 480, 2206
Petigura, E. A., Marcy, G. W., Winn, J. N., [et al.]{} 2018, The Astronomical Journal, 155, 89
Piskunov, N., & Valenti, J. A. 2017, Astronomy and Astrophysics, 597, A16
Pollack, J. B., Hubickyj, O., Bodenheimer, P., [et al.]{} 1996, Icarus, 124, 62
Santos, N. C., Israelian, G., & Mayor, M. 2004, Astronomy and Astrophysics, 415, 1153
Scott, D. W. 1979, Biometrika, 66, 605
Souto, D., Cunha, K., Smith, V. V., [et al.]{} 2018, The Astrophysical Journal, 857, 14
Wang, J., & Fischer, D. A. 2015, The Astronomical Journal, 149, 14
Winn, J. N., & Fabrycky, D. C. 2015, Annual Review of Astronomy [&]{} Astrophysics, 53, 409
Zhu, W. 2018, arXiv.org, arXiv:1808.09451
Zhu, W., Petrovich, C., Wu, Y., Dong, S., & Xie, J. 2018, The Astrophysical Journal, 860, 101
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Humans comprehend a natural scene at a single glance; painters and other visual artists, through their abstract representations, stressed this capacity to the limit. The performance of computer vision solutions matched that of humans in many problems of visual recognition. In this paper we address the problem of recognizing the genre (subject) in digitized paintings using Convolutional Neural Networks (CNN) as part of the more general dealing with abstract and/or artistic representation of scenes. Initially we establish the state of the art performance by training a CNN from scratch. In the next level of evaluation, we identify aspects that hinder the CNNs’ recognition, such as artistic abstraction. Further, we test various domain adaptation methods that could enhance the subject recognition capabilities of the CNNs. The evaluation is performed on a database of 80,000 annotated digitized paintings, which is tentatively extended with artistic photographs, either original or stylized, in order to emulate artistic representations. Surprisingly, the most efficient domain adaptation is not the neural style transfer. Finally, the paper provides an experiment-based assessment of the abstraction level that CNNs are able to achieve.'
bibliography:
- 'SceneRecPaint.bib'
---
Introduction {#sec:introduction}
============
This paper aims to investigate the differences between the level of abstraction achieved by deep convolutional neural networks as compared to the human performance in the context of paining analysis. To synthesize the motivation, let us recall Pablo Picasso words: ”There is no abstract art. You must always start with something. Afterward you can remove all traces of reality”. Art historians and enthusiasts are able to note, while recalling major artistic works through the history, that the level of abstraction steadily increased.
In parallel, in the last period, works that use computer vision techniques to analyze visual art increased with respect to both the quantity and the quality of reported results. Two trends favored these developments. First, there were consistent efforts to digitize more and more paintings, such that modern systems may learn from large databases. Two of such popular efforts are Your Paintings (now Art UK[^1]) which contains more than 200,000 paintings tightly connected with historical British culture and WikiArt[^2] which contains around 100,000 paintings gathered from multiple national cultures. The databases come with multiple annotations. For this work we are particulary interested in annotations dealing with the painting’s subject or scene type. From this point of view, a more complete database is the WikiArt collection, where the labelling category is named *genre*. The second trend is purely technical and it deals with the development of the Deep Neural Networks, that allowed classification performances that were not imagined before. In this work, we will use the more popular Convolutional Neural Networks (CNN) to recognize the painting genre.
Let us, now, to establish the meaning of genre the relation with scene and with the image subject. A list of definitions for various paintings genres is presented in Table \[Tab:Genre\_Explain\]. To label a painting into a specific genre, in most of the cases, a user has to identify the subject of that painting. The exceptions are “Abstract Art”, “Design”, “Illustration” and “Sketch and Study”, where the main characteristic is related to the depiction mode. In this majority of cases, subject is related to the scene represented in the work of art. The term “genre” is typical for art domain, and is a more general, including, concept than mere “subject” or ”scene type”. In this work, while referring to paintings, we will use all three with the same meaning of ”genre”. In comparison, for a non-artistic photograph, as there is no artistic intervention in the depiction mode, the subject is more related to the scene, while the genre is hard to be defined. For artistic photos, the “genre” gets meaning again.
Starting from the idea that Deep Neural Networks share similarities with the human vision [@Cichy16] and the fact that such networks are already proven to be efficient in other perception-inspired areas, like object recognition or even in creating artistic images, we ask ourselves if they can pass the abstraction limit of artistic paintings and correctly recognize the scene type of such a work.
In this paper we will first work with the Residual Network (ResNet) on the standard WikiArt database so to obtain state of the art results. Afterwards, we will test different domain transfer augmentations to see if they can increase the recognition rate; also we will study if the network is capable to pass the abstraction limit and learn from different types of images that contain the same type of scenes. Furthermore, we introduce several alternatives for domain transfer to achieve a dual-task: improve the scene recognition performance and understand the abstraction capabilities of machine learning systems.
Regarding deep networks, multiple improvements have been proposed. In many situations, if the given task database is small, better performance is reachable if the network parameters are previously trained for a different task on a large database, such as ImageNet. Next, these values are updated to the given task. This is called fine–tuning and it is a case of transfer learning. As our investigation is related to a different domain transfer, we will avoid to use both of them simultaneuosly, in order to establish clearer conclusions. To compensate, we are relying on the recent architecture of the Residual Networks (Resnet [@He2016]) that was shown to be able to overcome the problem of vanishing gradients, reaching better accuracy for the same number of parameters, when compared to previous architectures.
Contribution and paper organization
-----------------------------------
This paper extends our previous works [@florea2017:Scia; @badea2017], being mostly developed from [@florea2017:Scia], where we had initiated the discussion about the efficiency of various methods to transfer information from the photographic domain to the paintings domain, such that the recognition by CNNs’ of paintings genre is improved. In this paper we significantly extend the discussion, by including other transfer methods and by adding more significant results that allow crisper conclusions. In the second work ([@badea2017]), we showed that the artistic style transfer remains as efficient even if a reduced number of iterations are performed while over–imposing the style of an artistic painting and the content from a photograph onto a new image, according to the neural style transfer introduced by Gatys et al. [-@Gatys:2015].
Overall, this paper claims several contributions along a number of directions. On one direction, we investigate which aspects, comprehensible by humans, hinder the CNNs while understanding a painting genre; subsequently by means of domain transfer, we retrieve information about the internal description and the organization of the painting clusters. In order to accomplish such a task, we annotate artistic photographic images with respect to the scene type related to genres and we stylize a large corpus of photographs using different style transfer methods. All this data will be made publicly available to be used in other research works.
On a second direction, this paper is the first to objectively evaluate the efficiency of the currently popular neural style transfer methods. Currently existing solutions [@Gatys:2015; @johnson2016; @ulyanov2016; @huang2017] compare themselves by speed, stability within video sequences or number of transferable styles. By quantifying the improvement while adapting photographs to the painting domain, we reach a surprising conclusion, namely that they are less or at most as efficient as non-neural style transfers solutions. Evermore, a CNN finds as informative the original photographs without any style transfer applied.
The remainder of the paper is organized as follows: section \[Sect:Related\] presents previous relevant works, section \[Sect:CNN\] summarizes the CNN choices made and section \[Sect:PaintingUnderstanding\] will discuss different aspects of painting understanding. Section \[Sect:Databases\] presents the used databases, while implementation details and results are presented in section \[Sect:Results\]. The paper ends with discussions about the impact of the results.
Related Work {#Sect:Related}
=============
This work investigates the capabilities of CNNs’ to recognize the subject of paintings as compared with the performance of humans. Thus, relevant prior work refers to solutions for object and scene recognition in paintings. As paintings are an abstraction of real images, scene recognition in photographs is also relevant. At last we aim to adapt information from photographs to paintings by means of style transfer.
**Object and scene recognition in paintings.** Computer based painting analysis has been in the focus of the computer vision community for a long period. A summary of various directions approached, algorithms and results for not-so-recent solutions can be found in the review of Bentowska and Coddington [-@Bentowska2010]. However, a plurality of works [@Agarwal2015; @Karayev2014; @Bar2015; @florea:17] addressed style (art movement) recognition as it is the main label associated with paintings. An intermediate topic is in the work of Monroy et al. [-@Monroy2014] which detected and extracted shapes (associated with objects) but as pre-processing for the final task which was that of restoration.
Object recognition has been in the focus of Crowley and Zisserman [-@Crowley2016] while searching the YourPaintings dataset with learning on photographic data; yet this dataset features “older” art, which is less abstracted than modern art.
Scene recognition in paintings is also named genre recognition following the labels from the WikiArt collection. This topic was approached by Condorovici et al. [-@Condorovici2013] and by Agarwal et al. [-@Agarwal2015]; both works, using the classical feature+classifier approach, tested smaller databases with a few (5) classes: 500 images - [@Condorovici2013] and 1500 images - [@Agarwal2015]. More extensive evaluation, using data from WikiArt, was performed by Saleh and Elgammal [-@saleh2015], which investigated an extensive list of visual features and metric learning to optimize the similarity measure between paintings and respectively by Tan et al. [-@Tan:16], which employed a fine tuned AlexNet architecture [@Krizhevsky:12] to recognize both style and genre of the paintings.
The process of transferring knowledge from natural photography to art objects has been previously addressed beyond the recent transfer from ImageNet to WikiArt [@Tan:16]; yet their solution is general as it is common to use ImageNet pre-trained CNNs on smaller databases. 3D object reconstruction can be augmented if information from old paintings is available [@Aubry2013]. Classifiers (deep CNNs) trained on real data are able to locate objects such as cars, cows and cathedrals [@Crowley2016] if the artistic rendering is not very abstract. The problem of detecting/recognizing objects in any type of data, regardless if it is real or artistic, was named cross-depiction by Hall et al. [@Hall2015]; however the problem is noted as being particular difficult and in the light of dedicated benchmarks [@Cai2015], the results show plenty of space for improvement.
A significant conclusion that arises is that all solutions that showed some degree of success did it for *older* artistic movements, where scene depiction was without particular abstraction. To our best knowledge there is no reported significant success for (more) modern art.
**Scene recognition in photographs.** Scene recognition in natural images is an intensively studied topic, but under the auspices of being significantly more difficult than object recognition or image classification [@Zhou:14]. We will refer the reader to a recent work [@Heranz2016] for the latest results on the topic. We will still note that the introduction of the SUN database [@xiao2010sun] (followed by the subsequent expansions) placed a significant landmark (and benchmark) on the issue. In this context, it was shown that using domain transfer (e.g from the Places database), the performance may be improved [@Zhou:14].
**Style transfer.** Any image claiming artistic representations adheres to some non-trivial ideas about content, color and composition; these ideas are often grouped in *styles*. It has been an important theme in computer vision [@bae2006] to develop methods that are able to transfer the style from an artistic image to a normal, common image. The methods are called *style transfer*. Currently, the methods may be classified as non-neural (such as the Laplacian style transfer [@Aubry2014]) and neural artistic style transfers, initiated by the method of Gatys et al. [-@Gatys:2015]. We will detail these algorithms in section \[Sect:PaintingUnderstanding\]. For the moment, we will note that the efficiency of the transfer was quantified, till now, only subjectively as that it was not rigourously defined the concept of the style so to evaluate objectively.
**Scene recognition by humans.** While it is beyond the purpose of this paper to discuss detailed aspects of the human neuro-mechanisms involved in scene recognition, following the integrating work of Sewards [-@Sewards2011] we stress only one important aspect: compared to object recognition, where localized structures are used, for scene recognition the process is significantly more tedious and complex. Object recognition “is solved in the brain via a cascade of reflexive, largely feedforward computations that culminate in a powerful neuronal representation in the inferior temporal cortex” [@DiCarlo12]. In contrast, scene recognition includes numerous and complex areas, as the process starts with peripheral object recognition, continues with central object recognition, activating areas such as the entorhinal cortex, hippocampus and subiculum [@Sewards2011].
Concluding, there is a consensus, from both the neuro-science and the computer vision communities that scene recognition is a particularly difficult task. This task becomes even harder when the subject images are heavily abstracted paintings produced within the movements of modern art.
CNNs: Architectures and Training {#Sect:CNN}
================================
Following AlexNet [@Krizhevsky:12] performance in the ImageNet challenge, in the recent years we have witnessed a steep increase in popularity of Convolutional Neural Networks (CNNs) when considering the task of image classification. Especially after Donahue et al. [-@donahue2014] showed that ImageNet pre-trained CNNs provide very good descriptors, it is very hard to find a large-enough database where state of the art performance is not related with CNNs. As such, the bulk of our experiments revolve, in one way or another, around the CNN algorithms, either as direct classification methods or as auxiliary tools.
The first network architecture tested in this work is AlexNet, proposed by Krizhevsky et al [-@Krizhevsky:12]. This architecture features 5 convolutional layers followed by 3 fully connected layers. The size of the receptive fields is large ($11 \times 11$), in comparison to more recent architectures; the design is optimized for the use of two parallel GPUs, although currently it easily fits on a single GPU. Even though it has been visibly surpassed by other networks, it still remains a main reference point in all architecture comparisons.
A more powerful architecture is represented by the VGG-type networks, proposed by Simonyan and Zisserman [-@Simonyan2014]. In contrast to AlexNet, this design features smaller receptive fields ($3 \times 3$), but with a stride between regions of only 1. These decisions have led to the possibility of increased depth, by stacking sequentially up to 5 convolutional layers, and, subsequently, to higher performance than AlexNet. Although we do not use this network for direct classification, the stacked convolutional layers of VGG–-19 trained on ImageNet produce very efficient feature descriptors, and, thus, are heavily used by the style transfer algorithms [@Gatys:2015; @johnson2016].
In the remainder of the paper, for the task of classification, we will use the Residual Network (ResNet) [@He2016] architecture, with 34 layers. All the hyper-parameters and the training procedure follow precisely the original ResNet [@He2016]. Nominally, the optimization algorithm is 1–bit Stochastic Gradient Descent, the initialization is random (i.e. from scratch) and when the recognition accuracy levels on the validation set, we decrease the learning rate by a factor of 10. The implementation is based on the CNTK library[^3].
Painting Understanding and Domain Transfer {#Sect:PaintingUnderstanding}
==========================================
The main task of this work is to generate results about the understanding of the machine learning systems (in our case deep CNN) grasp of art. In such a case, one needs tools to ease the comprehension of the system internal mechanisms.
For CNNs, the most popular visualization tool has been proposed by Zeiler and Fergus [-@Zeiler:14] by introducing deconvolutional layers and visualizing activations maps onto features. Attempts to visualize the CNNs, for scene recognition, using this technique, indicated that activations are related to objects. Thus it lead to the conclusion that multiple object detectors are incorporated in such a deep architecture [@Zhou:15]. This means that high activations were triggered simultaneously in multiple parts of an image, being thus rather spatially vague. In parallel, visualization of activations for genre [@Tan:16] have shown that, for instance, the landscape type of scene leads to activating almost the entire image, thus being less neat to draw any conclusion. Consequently, we tried a different approach to investigate the intrinsic mechanisms of deep CNNs. Our approach exploits domain transfer and simply correlates the final results with the training database.
Given the increased power of machine learning systems and the limited amount of data available to a specific task, a plethora of transfer learning techniques appeared [@Lu2015]. Transfer learning is particularly popular when associated with deep learning. First, let us recall that the lower layers of deep nets trained on large databases are extremely powerful features when coupled with a powerful classifier (such as SVM) and may act as feature selectors, no matter the task [@donahue2014]. Secondly, the process of fine tuning deep networks assumes using a network that has been pre-trained on another database, and, with a small learning rate, adapting it to the current task.
In contrast, the concept of domain transfer or domain adaptation appeared as an alternative to the increase of the amount of information over which a learner may be trained directly (without fine tuning) in order to improve its prediction capabilities. Many previous solutions and alternatives have been introduced. We will refer to the work of Ben-David et al. [-@Ben2010] for theoretical insights on the process.
It has been shown that domain transfer is feasible and the resulting learner has improved performance if the two domains are adapted. In this context one needs to mention the work of Saenko et al. [-@Saenko2010] that proved that using a trained transformation, the domain transfer is beneficial.
In our case, the target domain is the artistic painting domain, while for the source domain we consider several alternatives: consumer photographic image domain (represented by images from the SUN database), artistic photographic image domain (with instances from the Photo-Artist database), and a combination of features from the painting domain (taken from the “older”, less abstract styles of paintings); these will be detailed further in section \[Sect:Databases\] and, respectively, in \[Sect:Results\].
For the domain adaptation function we investigate several alternatives. As a visual distinctive characteristic of painting is the style, the investigated functions are in the category of style transfer methods. As they are a key concept to our work, in the next subsection we will detail them.
Style transfer
--------------
Following the seminal work of Reinhard et al. [-@Reinhard:2001] the idea to transfer elements from a reference image, with an artistic value, to a subject one increased in popularity. The main concepts introduced then still stand: first, both images are represented into a space that offers a meaningful description with respect to transferable elements; next, the subject image is altered such that it fits into the description of the reference image. Quite often the second step implies matching the estimated probability density functions (i.e histograms) of the two images. A simple version [-@Reinhard:2001; -@huang2017] is to assume the distribution in the feature space to be Gaussian, independent (or at least uncorrelated) with respect to the dimensions. Such an assumption is a sword with two edges: on one side, the feature space should be constrained to follow the decorrelation assumption, but, on the other hand, the matching process is very simple as only the mean and the variance on each feature dimension needs to be fitted.
While in the original work [@Reinhard:2001], the envisaged transferable element is color, later, the style of the image was approached [@bae2006]. Prior to popularization of CNNs, image descriptors were based on manually–set filter weights; a prominent is the case of the Laplacian style transfer [@Aubry2014]. In the last period, the spectacular results showed by Gatys et al. [-@Gatys:2015] helped the so called Neural Style Transfer methods to gain in popularity.
In the next subsections we will review the formalization of the main style transfer methods; $\mathbf{S}$ will denote the subject image, while $\mathbf{R}$ the style reference image. In some solutions, $\mathbf{S}$ is altered to mimic the style from $\mathbf{R}$, in others, a third image $\mathbf{X}$, is built from scratch to match the content from $\mathbf{S}$ and the style from $\mathbf{R}$.
### Laplacian style transfer
Following the findings of Sunkavalli et al. [-@sunkavalli2010], the complex transfer process produces better results if it is carried separately on different levels of details from the two images. The most simple form to have various details of an image is to use a pyramidal representation. The Laplacian style transfer [@Aubry2014] assumes a Laplacian pyramidal representation with multiple levels. On each level the style transfer is implemented as gradient transfer. Nominally, at each pixel, $\mathbf{p}$, in each level of the pyramid, given a neighborhood $\mathbf{q}$, the set of gradients in the two images is $\nabla
S_{\mathbf{p},\mathbf{q}}$ (for the subject image) and $\nabla R_{\mathbf{p},\mathbf{q}}$ (for the reference image). The transfer is implemented by iteratively correcting the local point $S(\mathbf{p})$:
$$\begin{array}{cl}
\mathbf{S}_n(\mathbf{p}) &= r(\mathbf{S}_{n-1}(\mathbf{p})); \enspace
r(i) = g + sign(i-g)t(|i - g|); \\
t(i) &= CDF^{-1}_{[\nabla \mathbf{R}(\mathbf{p})]}\left( CDF_{[\nabla \mathbf{S}(\mathbf{p})]}(\mathbf{p}) \right)
\end{array}$$
where CDF is the cumulative density function computed on the image, $r(\cdot)$ is the transfer function, while $\mathbf{S}_1(\mathbf{p})$ is the initial local pixel of the subject image and $\mathbf{S}_n(\mathbf{p})$ is the same value at iteration $n$. The mapping needs to be done iteratively, as in practice the CDF is approximated by the local cumulative histogram on a discrete support of values. Typically, 10 iterations suffice.
### Neural style transfer algorithm {#Sect:NeuralStyle}
The algorithm proposed by Gatys et al. [-@Gatys:2015] aims to transpose the style by representing an image into a space where content and style are separable. This is achieved by using the feature representation offered by the layers of a convolutional network architecture that do not have skip connections (e.g. VGG–19). The choice for this representation originates in the observation that higher layers (i.e. closer to the fully connected and to the classification layers) will focus on high-level features, thus onto content (the focus is on objects, not on the actual pixels), while the first layers can be used to reconstruct the image with accurate pixel values.
Differently with respect to previous works which adjust the subject (content) image, the original neural style algorithm [@Gatys:2015] starts from a white noise image, $\textbf{X}$, which is modified in order to match the over–imposed statistics.
Starting from the white noise image, the method employs gradient descent to smoothly adjust values in order to match the content and style of the desired source images. The global loss function is a linear combination between two other loss functions used for content and, respectively, the style matching. In the neural style transfer algorithm, the subject image will provide the content matching, $\mathbf{S}$, while the generated image is $\mathbf{X}$. Considering the activations $F_{ip}^l$ ($P_{ip}^l$) of the $i^{th}$ filter on the $l^{th}$ layer at position $p$ in a CNN representation of $\mathbf{S}$ (and respectively $\mathbf{X}$), the loss function with respect to content can be defined as:
$$L_{content} (\mathbf{S},\mathbf{X},l) = \frac{1}{2}\sum_{i,j} (F_{ip}^l-P_{ip}^l)^2$$
To underline the correlations between activations, the algorithm uses the Gram matrix; $G_{ip}^l$ is the inner product between the vector versions of the $i^{th}$ and $p^{th}$ feature maps of the $l^{th}$ layer: $$G_{ip}^l = \sum_k F_{ik}^lF_{pk}^l$$
With the Gram matrix defined, the contribution of each layer is brought into discussion. The $l^{th}$ layer is composed of $N_l$ feature maps of size $M_l$. Thus, considering, besides $\mathbf{X}$, the style source image $\mathbf{R}$, the Gram representations $A^l$ and $G^l$ are computed for layer $l$. The contribution of a certain layer and the total style loss function are:
$$E_l=\frac{1}{4N_l^2M_l^2}\sum_{i,j}(G_{ip}^l-A_{ip}^l)^2$$
$$L_{style}(\mathbf{A},\mathbf{X})=\sum_{l=0}^Lw_lE_l$$
The total loss function, which takes into account both content and style is a linear combination of the two aforementioned components:
$$L_{total}(\mathbf{X},\mathbf{S},\mathbf{R})=\alpha L_{content}(\mathbf{S},\mathbf{X})+\beta
L_{style}(\mathbf{A},\mathbf{X})$$
where $\alpha$ and $\beta$ are manually picked weights.
The dramatic visual impact of the original visual style transfer method encouraged other works to address the same subject. Li et al. [@li2017] showed that, in fact, this version of style transfer minimizes the Maximum Mean Discrepancy (MMD) with the second order polynomial kernel thus implementing another version of matching the feature distributions. Maximum Mean Discrepancy (MMD) is a test statistic for the two-sample testing problem, where acceptance or rejection decisions are made for a null hypothesis [@gretton2012]; thus, in other words, a minimum MMD enforces that two distributions, in this case that of the image $\mathbf{X}$ and that associated with the reference images ($\mathbf{S}$ for style and $\mathbf{R}$ for style) are similar enough, within a statistical relevance limit. The last work [@li2017] indicates that the relation of the neural style transfer abides to the original principle stated by Reinhard et al. [-@Reinhard:2001], namely of feature distribution matching. The neural style transfer [@Gatys:2015] implements an approximation of the exact solution, which would have been given by inverting and composing the Cumulative Density Functions (CDF).
In an attempt to accelerate the transfer, Ulyanov et al. [-@ulyanov2016] and Johnson et al. [@johnson2016] learned the inverse mapping by a deep CNN. Yet this approach has the disadvantage that only a specific reference image equivalent CDF may be learned by the CNN; another style from another reference image requires another CNN. Thus, this approximation restricts the number of transferable styles to the number of CNNs trained. The limitation was very recently approached by Huang and Belongie [-@huang2017], which assumes independent Gaussians and thus match only the mean and variance (to obtain the statistical match) following by training a CNN to approximate the multivariate Gaussian CDF and do the inverse mapping.
Although it is faster, the CNN learning of the inverse CDF as in [@johnson2016; @ulyanov2016] dramatically limits the number of transferable styles, making the concept ineffective as domain adaptation method. The Gaussian matching [@huang2017], while being able to transfer an infinity of styles, still imposes two approximations that have an effect on the final image quality. Concluding, the original neural style transfer [@Gatys:2015] remains the most viable candidate for domain adaptation between the photographic image domain and the paintings domain.
Comparison Between Neural and Laplacian Transfers
-------------------------------------------------
While the Laplacian and Neural style transfer are built on different paradigms (neural vs. non-neural), we stress that the two methods share similar concepts. First, let us note that the pooling paradigm from the CNN is similar with the downsampling used in the Laplacian pyramid. Noting that the best results for the neural style transfer are achieved using a VGG-19 architecture, this implies an input image of $224\times 224$; the VGG-19 has 4 pooling layers between convolutional layers, thus offering a 5 level pyramidal representation. For the same image input, the Laplacian transfer uses 7 levels, thus going for coarser style too. In each such pyramidal level, the VGG stacks 2 to 3 convolutions, while the Laplacian uses only one.
Another aspect is related to the used filters. For the Laplacian filters, the weights are preset, always having the same values, while for the neural style transfer, the weights are learned while trying to classify on the ImageNet task. It has been shown that the CNN lower level filters are similar to standard gradient filters [@fischer2014], yet those from upper levels have no such correspondence. In this point, one should note that CNN architectures use multiple filters in parallel (the VGG from 64 in the bottom layer to 512 in the upper ones), while the Laplacian only 1. This huge difference ensures that a much richer description is available for the neural transfer.
The last aspect taken into account for the comparison is the procedure used for matching the distributions. The original Neural Style transfer goes for minimizing the MMD, which is to ensure that the multi-dimensional feature distributions of the image to change and of the reference image do match from a statistical point of view. The Laplacian goes for the exact solution. Yet, the inverting of the CDF in the Laplacian is also imperfect, as the computed “cumulative density function” is only an approximation of the true CDF, aspect enhanced by the significant number of iterations, which are required to fill the discontinuities. Another point is that the full inversion is possible because the Laplacian uses 1D feature description, thus being less richer than the multi-dimensional one used for the neural version.
Concluding, just by looking at the richness of the two algorithms, the neural style transfer may be grossly perceived as a significant step ahead with respect to the Laplacian counterpart.
Databases {#Sect:Databases}
=========
For the various experiments undertaken, several databases have been employed. These are either a collection of digitized paintings, either a collection of digital photographs. In the next paragraphs we summarize their main characteristics.
WikiArt database
----------------
The WikiArt is a collection of digitized paintings used to popularize art. It is in continuous expansion, as every day new works of art are added. The version used is the same as in [@Karayev2014] and it contains approximately 80,000 digitized images of fine-art paintings. We note that this version is highly similar to ones used by prior art, to which we compare against.
The images are labelled within 27 different styles (cubism, rococo, realism, fauvism, etc.), 45 different genres (illustration, nude, abstract, portrait, landscape, marina, religious, literary, etc.) and belong to more than 1000 artists. To our knowledge this is the largest database currently available that contains genre annotations. For our tests we considered a set that contains 79434 images of paintings.
In WikiArt some of the genres were not well represented (i.e. less than 200 images) and we gathered them into a new class called “Others”. This led to a division of the database into 26 classes illustrated in Figure \[Fig:Database\]. The names of the classes and the number of training and testing images in each class can be seen in Table \[Tab:Genre\_Explain\]. Furthermore, as the genre label is not explained in prior works, we also contribute with an explanation of the genre meaning; these explanations are collected from an Internet encyclopedia[^4].
The list of genres that counted less than 200 examples and are included in the ”Others” class contains, among others: ”advertisement” – 94 examples – illustrate a work of art designed to market a product; ”capriccio”(194) – depict architectural fantasies, placing together buildings and/or elements in fantastical combinations; “caricature”(184) - variation of sketch, but with intended comical effect; “cloudscape” (162) - depicting clouds or sky; ”miniature” (79) - small paintings, typically originating in early middle age in the Byzantine or Central Asian empires; ”mosaic” (17) – images made from the assemblage of small pieces of colored materials; ”panorama”(17) – massive artworks that reveal a wide, all-encompassing view of a particular subject; ”pastorale” (80) - illustrates segments of life from shepherds herding livestock around open areas; “quadratura” (19) - an image based on techniques and spatial effects to create an illusion of three-dimensional space; “tessellation” (122) - images of aperiodically patterns from M. C. Escher; “vanitas”(34) - symbolic works of art, associated with the still life and “veduta” (188) – which are a highly detailed large-scale paintings.
We note that annotation is weak, as one may find arguable labels. For instance “literary” and “illustration” categories may in fact have “landscape” subjects. However, as this distribution matches practical situations, we used the database as it is, without altering the annotations.
After observing the main characteristics of paintings genres in Table \[Tab:Genre\_Explain\], one may conclude that the task is similar with the identification of a scene type in a photograph. Basically, in most cases, the genre is given by the topic (subject) rendered in the artwork. Under these circumstances, especially to specific genres, we consider that appropriate for domain transfer, are images labelled for scene content.
[|c|c|c|c|]{} & **Genre** & **No. imgs.** & **Description**\
------------------------------------------------------------------------
height 1pt @a xhline
1 & Abstract Art & 7201 & Uses shapes, forms etc. to replace accurate representations\
2 & Allegorical Painting & 809 & Expression of complex ideas using another subject\
3 & Animal Painting & 1233 & Paintings which depict animals\
4 & Battle Painting & 273 & The main subjects are battles and wars\
5 & Cityscape & 4089 & Works which contain cities or other large communities\
6 & Design & 1577 & Conceptual schemes of objects and structures\
7 & Figurative & 1782 & Forms inspired by objective sources, but altered\
8 & Flower Painting & 1270 & Paintings of flowers\
9 & Genre Painting & 10984 & Scenes of everyday life\
10 & History Painting& 656 & Depictions of historical events\
11 & Illustration & 2493 & Visual representations usually meant for books, magazines, etc.\
12 & Interior & 511 & Paintings depicting interiors of structures\
13 & Landscape & 11548 & Contains representations of land, or other natural scenes\
14 & Literary Painting& 418 & Subject taken from literary work\
15 & Marina & 1385 & These paintings show scenes from docks or ports\
16 & Mythological Painting& 1493 & Inspired by mythology\
17 & Nude Painting & 1758 & Paintings which contain nudes\
18 & Portrait & 12926 & Images of real individuals\
19 & Poster & 229 & Works which are usually intended for advertising\
20 & Religious Painting & 5703 & Inspiration is drawn from religious scenes\
21 & Self-Portrait & 1199 & The subject of the painting is the artist\
22 & Sketch and Study& 2778 & Drawings done for personal study or practice\
23 & Still Life & 2464 & Images which depict inanimate objects\
24 & Symbolic Painting& 1959 & Content suggested by symbols in the forms, lines, shapes, and colors\
25 & Wildlife Painting& 259 & Paintings of natural scenes, including animals in their habitats\
[c]{}
--------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------
{width="0.228\linewidth"} {width="0.249\linewidth"} {width="0.212\linewidth"} {width="0.212\linewidth"}
Allegorical Battle Cityscape Genre
--------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------
\
----------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------
{width="0.103\linewidth"} {width="0.177\linewidth"} {width="0.145\linewidth"} {width="0.1398\linewidth"} {width="0.13\linewidth"} {width="0.17\linewidth"}
Abstract Animal Design Figurative Flower History
----------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------
\
-------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------ ------------------------------------------------------------------------ ----------------------------------------------------------------------------------- -------------------------------------------------------------------------------------- ----------------------------------------------------------------------------
{width="0.146\linewidth"} {width="0.142\linewidth"} {width="0.14\linewidth"} {width="0.13\linewidth"} {width="0.145\linewidth"} {width="0.15\linewidth"}
Illustration Landscape Nude Portrait Poster Self-portrait
-------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------ ------------------------------------------------------------------------ ----------------------------------------------------------------------------------- -------------------------------------------------------------------------------------- ----------------------------------------------------------------------------
\
------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------ -------------------------------------------------------------------------------- ----------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------
{width="0.176\linewidth"} {width="0.169\linewidth"} {width="0.203\linewidth"} {width="0.167\linewidth"} {width="0.1545\linewidth"}
Interior Marina Religious Still Wildlife
------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------ -------------------------------------------------------------------------------- ----------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------
\
------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------- -------------------------------------------------------------------- --------------------------------------------------------------------------------
{width="0.248\linewidth"} {width="0.2025\linewidth"} {width="0.225\linewidth"} {width="0.252\linewidth"}
Literary Mythological Sketch Symbolic
------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------- -------------------------------------------------------------------- --------------------------------------------------------------------------------
\
Natural Scenes Databases
------------------------
As a first additional source of data we have considered images from the SUN (Scene Understanding) database [@xiao2010sun]. In its original form, it contains 899 classes and more than 130,000 images. Yet, only a few classes, which have a direct match with the content of the WikiArt genre database, were selected. These classes and the number of images in each one that were added in the training process at some step are: “Cityscape”–-2903 images, “Flower”–-229, “Landscape”-–7794, “Marina”–-1818.
Supplementary, since portraits are a very common type of paintings, we have retrieved 5993 images from the Labelled Faces in the Wild (LFW) database [@Huang2007a]. Again, these images are a mere rendering of one’s face without any artistic claim. In the next section, if it is not specified otherwise, these portraits will be associated with SUN photos. In total we have 20,075 normal photographs.
As a particular characteristic of the images from the SUN/LFW databases is that they were not selected for their artistic value, but strictly for their adherence to one of the scene classes tightly associated with studied genres. Artistic value, if there is any, is merely coincidental. Also there is no abstraction in this data.
Photo-Artist Data Set
---------------------
Thomas and Kovashka [-@thomas:16] have collected a large database of artistic photographs. Unlike the images from the SUN – LFW database, in this case, each image has an important artistic value, and its author is broadly accepted as an artist. The artistic value is encoded in the composition, framing, colors, or in other elements that are harder to objectively quantify. Yet this database was not labelled with respect to subject or scene composition.
From this large collection, we have selected and labelled a subset of 19,573 images distributed as follows: “Cityscape”–8068 images, “Interior”–3036 images, “Landscape”–4467 images, “Portrait”–4002 images.
Stylized Images
---------------
The most striking visual difference between photographs and paintings is the style of representing the subject within the paintings. In an effort to transfer knowledge between the two image types, we have stylized some sets of images as follows:
- images from the Photo-Artist database were stylized with Laplacian style transfer;
- images from the Photo-Artist database were stylized with neural style transfer.
- images of paintings with classical styles, that have a realistic depiction of reality, have been stylized using the neural transfer method and according to the styles of modern abstract art movements.
Example of stylized paintings may be seen in Figure \[Fig:Stylization\]. We have not stylized images from SUN–LFW database, as (one may see further in Table \[Tab:Transfer\]) their performance was not better than those of Artistic–Photo and the process is very tedious.
For the reasons previously mentioned, among neural solutions, we have used the original neural style transfer method [@Gatys:2015]. The motivation for trying to transfer the abstract style onto realistic paintings lies in the observation that, over time, the paintings adheres to the same favorite subjects.
[c]{}
---------------------------------------------------------------- ---------------------------------------------------------------------- -------------------------------------------------------------
{width="20.00000%"} {width="20.00000%"} {width="20.00000%"}
Artistic Photo Laplacian stylized image Reference painting
---------------------------------------------------------------- ---------------------------------------------------------------------- -------------------------------------------------------------
\
------------------------------------------------------------------ -------------------------------------------------------------------------- ----------------------------------------------------------------
{width="20.00000%"} {width="20.00000%"} {width="20.00000%"}
Classic style painting Neural stylized image Reference painting
------------------------------------------------------------------ -------------------------------------------------------------------------- ----------------------------------------------------------------
\
-------------------------------------------------------------------- ------------------------------------------------------------------- ---------------------------------------------------------
{width="20.00000%"} {width="20.00000%"} {width="20.00000%"}
Artistic photo Neural stylized image Reference painting
-------------------------------------------------------------------- ------------------------------------------------------------------- ---------------------------------------------------------
\
The process of neural stylization using the original neural solution is time consuming. For the initial tests [@florea2017:Scia] we have used a NVIDIA GTX 980 Ti and a NVIDIA K40; in this case, the stylization of an image took 2-10 minutes, which matches the more recent reports in [@ChenS16f]. Further acceleration [@badea2017] and the use the faster NVIDIA GTX 1080 Ti allows, currently, faster stylization with a duration scaled down to 30 sec; yet this is still slow considering the size of the databases, so that one will not produce an extremely large quantity of images.
Implementation and Results {#Sect:Results}
==========================
The main goal of this work is to study the performance of the CNNs in recognition of a paintings’ genre and, in parallel, to investigate various ways in which this performance can be increased. This includes experiments on the classification methods themselves, in order to establish both the state of the art performance and a baseline for further experimentation. Afterwards, we follow with experiments on various alterations brought to the database in the context of domain transfer.
Baseline and comparison with prior work
---------------------------------------
Genre recognition has been previously attempted using the WikiArt collection. We show that the used current procedure leads to slightly superior performance when compared to previously proposed solutions. Furthermore, as the list of prior art is not too rich, we have tested some classical solutions to establish more firmly the baseline performance.
Agarwal et al. [-@Agarwal2015], Tan et al. [-@Tan:16] and Saleh and Elgammal [-@saleh2015] used the WikiArt database for training and testing in order to classify paintings into different genres. While the first two used a very small subset, the later two solutions focused on 10 classes from the entire database, namely: Abstract, Cityscape, Genre, Illustration, Landscape, Nude, Portrait, Religious, Sketch and Study and Still life; in total their subset gathered $\sim$63,000 images.
We have adopted the division (training and testing) from Karayev et al. [-@Karayev2014] as working with a complete version of WikiArt. Furthermore, we stress that in our case the images from training and testing are completely different and are based on a random selection.
In order to compare our results to the ones reported by the mentioned previous works, we have selected the same classes of paintings for training and testing. While in the case from [@saleh2015] the test-to-train ratio is mentioned, in [@Tan:16], it is not. Under these circumstances and against our best efforts, the comparison with prior art is, maybe, less accurate.
The results from Table \[Table:Comp\_StateOfArt\] show that the proposed method gives slightly better results than previous method [@Tan:16], building upon the difference that they use a smaller fine–tuned (i.e. initialized on ImageNet) network (AlexNet), while we have used a larger one, based on the improved residual connections paradigm, but initialized from scratch. Also, the previous work used database augmentation while, for this result, we have not. The difference between here used ResNet-34 and the previously used AlexNet is above the stochastic margin, as we will discuss later. Furthermore, we report the average performance over 5 runs. Thus we may emphasize that our efforts produced state of the art genre recognition on the WikiArt collection.
On the 26-class problem, which will be furthered explored in the remainder of the paper, the residual CNN is able to retrieve better performance than classical solutions. In this test, the classical solutions coupled pyramidal Histogram of Oriented Gradients (pHog) [@Dalal2005] or pyramidal Local Binary Pattern (pLBP) [@Ojala:2002], as features, with a SVM for classification. The SVM uses radial basis function and we performed the grid searched for the best $\gamma,C$ parameters. Only features extracted from lower layers of a CNN (DeCAF) provide a description that is competitive with CNNs. Compared to our previous work, where we have reported 59.1% accuracy, the improved performance is due to better hyperparameter tuning. Currently, our effort produced state of the art performance on the 26 class problem.
[ |c|c|c| c|c|c | c|]{}
------------------------------------------------------------------------
height 1pt @a xhline
**Method** & **No. classes** & **No. images** & **Test ratio** & **Accuracy** (%)\
------------------------------------------------------------------------
height 1pt @a xhline
Agarwal et al. [-@Agarwal2015] - SIFT+BOW & & & & 82.53\
Agarwal et al. [-@Agarwal2015] - ensemble & & & & 84.46\
Saleh and Elgammal [-@saleh2015] - Classemes+Boost& & & & 57.87\
Saleh and Elgammal [-@saleh2015] - Classemes+ITML& & & & 60.28\
Saleh and Elgammal [-@saleh2015] - Classemes+Fusion& & & & 60.28\
Tan et al. [-@Tan:16] AlexNet - scratch & & & & 69.29\
Tan et al. [-@Tan:16] CNN- finetune & & & & *74.14*\
*Proposed - ResNet 34 - scratch* & & & 20% & **75.58**\
pHoG + SVM & & & & 44.37\
pLBP + SVM & & & & 39.58\
DeCAF + SVM & & & & 59.05\
AlexNet - scratch & & & & 53.02\
ResNet 34 - scratch [@florea2017:Scia] & & & & 59.1\
*Proposed - ResNet 34 - scratch* & & & & **61.64**\
*Proposed - ResNet 34 - scratch + augmentation* & & & & **63.58**\
------------------------------------------------------------------------
height 1pt @a xhline
**Database Basic Augmentation.** Besides using domain transfer and domain adaptation, one may improve the recognition performance by adding the original images modified by basic image processing techniques. The processing methods that have produced positive effects are flipping and slight rotation. The flipped versions are all horizontal flips of the original images. Regarding the rotations, all images have been rotated either clockwise or counterclockwise with $3^0$, $6^0$, $9^0$ or $12^0$. Adding noise, slight tone adjustment or small translations did not help.
When the size of the training database doubled, the improvement in recognition performance was of 2%, reaching $63.58\%$ while doubling the training time too. We consider the increase too small, so in the remainder of the experiments we have used only the original paintings without basic augmentation.
**K** **1** **2** **3** **4** **5**
-------------- ------- ------- ------- ------- -------
Recogn.\[%\] 61.64 76.24 82.14 88.46 90.09
: Top–$K$ accuracy for genre recognition rate in the 26-case scenario. Top–$K$ means that a positive recognition is achieved for an example, when any of the any the CNN model $K$ highest probability answers match the expected answer.\[Tab:TopN\]
**Confusion Matrix**. The confusion matrix for the best performer on the 26-class experiment is presented in Figure \[Fig:Confusion\]. We have marked classes that are particularly confused. It should be noted that from a human point of view, there is certain confusion between similar genres such as historical$\leftrightarrow$battle$\leftrightarrow$religious, portrait$\leftrightarrow$self portrait, poster$\leftrightarrow$illustration, animal$\leftrightarrow$wildlife etc. Some of these confusable images are, in fact, shown in Figure \[Fig:Database\].
Consequently, the Top–$K$ error may also show relevance, as in many cases there are multiple genre labels that can be truthfully associated with one image. The performance with respect to $K$ is presented in table \[Tab:TopN\]. One may note that greater improvement is from Top–1 to Top–2. For the best proposed alternative, ResNet with 34 layers, the Top-5 error is 9.91% - corresponding to **90.09**% accuracy. For the 10-class experiment the Top–5 accuracy is 96.75%.
![The confusion matrix for the 26 class. We have grouped and marked confusion classes.[]{data-label="Fig:Confusion"}](confuzie1.jpg){width="0.85\linewidth"}
**Stochastic Effect**. This test studied the effect of the stochastic nature of the CNN. Factors such as the random initialization of all the parameters can influence the results of any considered network. We ran the ResNet–34 several (5) times on the 26 classes and the accuracy results had a mean of 61.64% (Top-–1 accuracy) and a standard deviation of 0.33%. The results underline the fact that even though there is some variation caused by randomness, it does not influence the system significantly.
On the other hand, if alternative solutions show variation smaller than **0.33**% (“stochastic margin” as we named it), one may argue that these are not relevant to draw conclusions.
For the following experiments we will refer solely to the 26 classes scenario, as it is the most complete. We recall that the baseline performance (and also the best) is **61.64%**.
Influence of the artistic style
--------------------------------
Prior art [@Zhou:15] suggests that even in the case of the scene rendered in photographs, in fact, a deep network builds object detectors and can recognize objects presented in *a way seen before*. These conclusion are based on the viewing method of the CNN filters based on deconvolution [@Zeiler:14]. In the case of realistic scenes, Zhou et al. [-@Zhou:15] showed that filters activation is grouped on certain objects in association with a certain class. In the case of paintings’ genre, Tan et al. [-@Tan:16], using the same visualization technique, showed much more sparsity.
In the same, the deconvolution method has no procedure to identify the failure; in other words there is no way to prove that nothing in particular influences the decision on one class, but rather the small, easy–to–neglect weights, exploited in Distillation – Dark Knowledge [@hinton2015]. To avoid this potential uncertainty, our experiments focus directly on the network output, when the training set was adjusted into a specific direction.
With regard to the artistic style, we have performed two experiments. In the first one, we separate the training and testing based on style, thus asking the CNN to generalize across style, while in the second we have reviewed the baseline performance with respect to style.
For the first experiment, considering the full 79,434 images genre database, we selected all the images that are associated with *Cubist* and *Naive Art* styles and placed them in testing, resulting in 4,132 images for evaluation and 75,302 for training. Although numerically this is a weaker test than the baseline, as the training set is larger, the results are considerably worse: **50.82**% Top–1 accuracy and **82.10**% Top–5. We consider that this drop (from 61.64% - Top–1 and 90.99% Top–5) is due to the fact that these particular styles are rather different from the rest and the learner had no similar examples in the training database. Also these results argue for *a style oriented domain adaptation*.
The second experiment benefits from the fact that WikiArt images are annotated with multiple labels categories, and one of these category is the style. Thus given the baseline framework (i.e. keeping the same training and testing sets), we retrieve the genre recognition rate with respect to style (artistic movement). The results are showed in Figure \[Fig:Style\_recognition\].
{width="0.65\linewidth"}
One may note that better performance is achieved for older styles, from the early beginning of artistic painting (as we understand it today) to Post-Impressionism; these styles have a clear representation without too much abstraction. These results are in concordance with previous object detection in paintings [@Crowley2016], where good performance is achieved for paintings depicting scenes in a rather classical and realistic manner. Also, we achieve good performance for modern styles where the subject is rather unique; in cases such as Minimalism, Abstract Art or Color Field Painting the genre is in fact the style.
In contrast, for styles such as Surrealism, Naive Art, Cubism, Pop Art, the degree of abstraction is high and the CNN has difficulties in interpreting the scene; the scene subject is also variable. Based on this observation, coupled with the observation that in many cases modern paintings reinterpret classical compositions under the rules of new styles, we attempted to create new pseudo-paintings by using a realistic depiction of an old content and a modern style; this is the reason while we tried to neurally transfer abstract styles on older paintings content. Furthermore, this idea is in the same line with neural transfer methods [@Gatys:2015; @johnson2016; @ulyanov2016; @huang2017], where almost every time, the reference image had an abstract style, while content that easily comprehensible.
Domain transfer
---------------
The domain transfer experiments were performed iteratively, starting from small sets to larger ones. The motivation lies in the fact that producing relevant images for transfer requires non-negligible effort. We recall that we annotated $\sim 20,000$ artistic photos and we produced more that $30,000$ neurally adapted images.
Each iteration assumes $N$ ($N=250, 500 \dots$) maximum images in each of the 26 classes; on some classes, since they have too few examples, even in the first iteration this maximum is too generous from the begining. For a detailed number of images existing in each class, we kindly ask the reader to look into Table \[Tab:Genre\_Explain\] and to keep in mind that 80% is used for training while 20% is for testing. For iterations corresponding to given $N=250,\dots, 10000$, we have added, in all domain transfer scenarios, images from three classes, as follows: cityscape – 2903, landscape – 4467, portrait – 4002. At the end, we have added all images from all classes and all the images available in alternative, transferable, domains.
The obtained results may be numerically followed in Table \[Tab:Transfer\], while visually in Figure \[Fig:Domain\_Transfer\]. The baseline performance, where no domain transfer is involved, is on the row marked with “None”. The “Best-improvement” row marks the difference between the best achieving solution and the baseline. “Added image ratio” row represents the ratio between the number of added images and the number of paintings used in this training. Each cell from the column marked with “Avg. improv.” (average improvement) indicates the mean of the improvement for the transfer method.
[|c| c|c|c| c|c||c|]{}
------------------------------------------------------------------------
height 1pt @a xhline
& & **Avg.**\
& **250** & **500** & **1000** & **5000** & **All**& **Improv.**\
------------------------------------------------------------------------
height 1pt @a xhline
None & 19.12 & 27.95 & 35.66 & 54.61 & 61.64 & 0\
Normal Photos (SUN/LFW) & 25.28 & 32.04 & 38.21 & 55.31 & 61.67 & 2.71\
Artist Photos & 26.72 & **32.75** & 39.27 & 53.49 & 61.55 & 2.96\
Laplacian-on-Artist & 26.56 & 30.73 & [39.84]{} & **56.17** & **61.73** & 3.21\
Neural-on-Paints & **26.76** & 31.98 & 37.4 & 55.33 & 61.47 & 2.79\
Neural-on-Artist & 26.33 & 31.27 & **39.94** & 54.88 & 61.62 & 3.01\
Best-improvement & 7.64 & 4.8 & 4.18 & 1.56 & 0.09 & –\
Added image ratio\[%\] & 190.04& 103.38 & 58.98 & 27.33 & $\sim$26& –\
One should note that improvement (larger than stochastic margin) is achieved only when the number of added images is with at least 30% larger than the number of paintings in the training database and the number of paintings is not too large. Looking at the trends from the last two rows, one may estimate that in order to achieve some $5-10\%$ increase for the entire database, one will need about 3 times the number of paintings to be obtained from some other source, while the same procedure is assumed.
Another visible trend is that original paintings are better that any transferred and adapted image from other domain. We explain this behavior by the *content originality* of the paintings: artistic works should be considerably different in at least one aspect (e.g. content, local style, composition, subject, etc.) from its predecessors, as to be acknowledged as art. This produces a naturally *sparse* *domain* representation for paintings. The sparsity is not easily filled up by knowledge transfer from a similar representation.
At last, the performance of any domain transfer method is not too different from any other. Although by looking at overall improvement, the Laplacian style transfer on artistic photos, is marginally better, the difference is too small (especially given the stochastic margin) that one cannot conclude about a clear winner. This results is somehow surprising, as the highly praised neural style transfer (which produces visually pleasant images) does not show better performance than the Laplacian transfer or than the images themselves.
When comparing the normal photos with artistic photos, the artistic content and the way this content is represented, gives slightly better results. Yet a CNN learns too little from older artistic content to deal with new one.
When comparing the Laplacian transfer with the Neural style transfer, as introduced by Gatys et al. [-@Gatys:2015], the significant better representation (larger depth of the feature map) and the better fitting of the latter achieves nothing. At the end of this analysis, we must note that, although claimed, none of these methods implements a “style transfer”, where “style” has the meaning under which is used in art. Although this result has a negative connotation in the sense that expected hypothesis (“style transfer methods does not transfer styles“) does not hold, we feel that we should make it public, especially given the recent argumentation by Borji [-@Borji2018]. Furthermore, we see that has a strong positive connotation too: it shows that it is much easier to improve the performance of the CNN when dealing with understanding art by simply showing it very large quantities of relevant photographs.
![The genre recognition rate with respect to transferred images per class, as a function of the augmentation method. When the number of transferred images dominate over the number of painting in the database, the performance improves \[Fig:Domain\_Transfer\]](Domain_transfer_perfromance.png){width="0.88\linewidth"}
Discussion and Conclusions
==========================
In this paper we discussed the CNN capabilities to recognize the scene (genre) in a painting. The first contribution is that we clearly showed that machine learning systems (deep CNNs) are confused by the abstraction level from art. The experiment with abstract art showed that they cannot easily generalize with respect to style and that, the more abstract a style is, the lower is the recognition rate of the genre (painting subject). In this sense, the CNN is similar with humans, who also find the abstract representations of scenes to be more puzzling.
The secondary set of contributions results from the experimentation with domain transfer as an alternative to increase the overall performance and we have found that: (1) limited improvement is easily doable when the training set of paintings is small; (2) methods claiming to transfer style, either Laplacian based, either neural based, although produce visually very pleasant and intriguing images, are ineffective as domain adaptation methods with respect to style. One possible explanation is that “style” as understood by the transfer methods is not the same with “artistic style” in the sense of art movement.
Lastly, the third contribution is related to understanding the structure of the paintings domain. Due to the necessary criteria for some work to be accepted as a work of art, which implies significant artistic novelty with respect to its predecessors, the paintings domain is more sparse than that of the normal images. Given the problem to improve the performance, a CNN learns to better deal with some new work of art, when more works of art are presented to it in the training set. The CNNs are similar to humans in this behavior as well, as art expert do not learn their job looking at normal images.
Acknowledgment
==============
The work was supported by grants of the Romanian National Authority for Scientific Research and Innovation, CNCS UEFISCDI, number PN-II-RU-TE-2014-4-0733 and respectively, CCCDI-UEFISCDI, project number 96BM. The authors would like to thank NVIDIA Corporation for donating the Tesla K40c GPU that helped run the experimental setup for this research.
[^1]: <http://artuk.org/>
[^2]: <http://www.wikiart.org/>
[^3]: Available at <https://github.com/microsoft/cntk/>
[^4]: http://www.getty.edu
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Oleg A. Fonarev[^1]\
[*Racah Institute of Physics, The Hebrew University*]{}\
[*Jerusalem 91904, Israel*]{}\
[*E–mail: [email protected]*]{}
title: 'Wigner Functions in Curved Space–Time and Quantum Corrections to Thermal Equilibrium'
---
The Wigner function is known to be a very useful tool in exploring the semiclassical limit of the quantum theory. It is defined as a Fourier representation of the density matrix ${\bf \rho}$ with respect to the difference of coordinates$^{\ref{kn:wigner}}$: $${f}_{w}(q,p)=(\pi\hbar)^{-1}\:\int_{- \infty}^{\infty}
dQ\:e^{-2iQp/\hbar}\langle q+Q |{\bf \rho}|q-Q\rangle \; . \label{eq:wigo}$$ At a high temperature or after coarse graining, the Wigner function gives us directly quantum corrections to a distribution function in the classical phase space (the Boltzmann distribution if the system is in thermal equilibrium).\
In curved space–time the notion of the Fourier transformation becomes ambiguous because it is not invariant under general coordinate transformations. Different approaches to solve this problem have been proposed$^{\ref{kn:winter}-\ref{kn:fonarev1}}$. I will follow the approach by \[\[kn:fonarev1\],\[kn:fonarev11\]\] which uses the formalism of the tangent bundles. Consider any point $x$ on a space–time manifold ${\cal M}$. Let ${{\bf \varphi}}(x)$ be a (real) scalar field on the manifold. Let us introduce the tangent space ${\cal T}_{x}(\cal M)$ at point $x$ and define field ${{\bf
\Phi}}(x,y)$ on the tangent bundle as follows: $${{\bf \Phi}}(x,y) = \exp(y^{\alpha} \hat{\nabla}_{\alpha}) \, {{\bf
\varphi}}(x) \; . \label{eq:Phi}$$ Here the operator $\hat{\nabla}_{\alpha} = \nabla_{\alpha} - \Gamma_{\alpha\gamma}^{\beta} \:
y^{\gamma} \: \frac{\partial}{\partial y^{\beta}}$ is the horizontal lift of the derivative operator to the tangent bundle$^{\ref{kn:yano}}$. The generalized Wigner function is now defined as follows: $${f}(x,p) = (\pi\hbar)^{-4}\: \sqrt{-{g}(x)}\:\int_{{\cal T}_{x}(\cal M)}
d^{4}\,y\:e^{-2iy^{\alpha}p_{\alpha}/\hbar}\:\langle {{\bf \Phi}}(x,-y){{\bf
\Phi}^{\dagger}}(x,y)\rangle \; . \label{eq:wigf}$$ Here the brackets mean the averaging with a density matrix on an initial Cauchy hypersurface. In the Minkowski space–time, the tangent space ${\cal
T}_{x}(\cal M)$ coincides with the base space ${\cal M}$, the exponent in Eq.(\[eq:Phi\]) becomes the shift operator, i.e. ${{\bf \Phi}}(x,y) = {{\bf
\varphi}}(x+y)$ and we get the usuall definition of the relativistic Wigner function$^{\ref{kn:degroot}}$. In curved space–time, the definition (\[eq:wigf\]) is explicitly covariant and allows us to easily derive equations describing the evolution of the generalized Wigner function in the phase space, given a field equation for the field ${{\bf \varphi}}$.\
Suppose that ${{\bf \varphi}}$ obeys the covariant Klein–Gordon equation with the conformal coupling to gravity: $$\left(\nabla^{\alpha} \nabla_{\alpha} + m^{2}/\hbar^{2} -
\frac{1}{6} {R}\right){{\bf \varphi}} = 0 \; , \label{eq:Klein}$$ where $R$ is the scalar curvature. We then get two equations for the Wigner function$^{\ref{kn:winter}-\ref{kn:fonarev11}}$: $$\begin{aligned}
\left( m^{2} - p^{\alpha}p_{\alpha} \right) \:
f =
- \frac{\hbar^{2}}{4}\, D^{\alpha} D_{\alpha} \: f - \hspace*{1in} \nonumber \\
\mbox{} - \hbar^{2} \left( \frac{1}{6} R + \frac{1}{12} R_{\alpha\mu\beta\nu}
p^{\mu} p^{\nu} \frac{\partial^{2}}{\partial p_{\alpha} \partial p_{\beta}} +
\frac{1}{4} R_{\mu\nu} p^{\mu} \frac{\partial}{\partial p_{\nu}} \right) \: f +
{O}(\hbar^{4}) \; , \label{eq:mashf}\end{aligned}$$ $$\begin{aligned}
p^{\alpha} D_{\alpha} \, f =
\hbar^{2} \left(
\frac{1}{6} R_{\nu\beta\mu\alpha} p^{\mu} \frac{\partial^{2}}{\partial
p_{\alpha} \partial p_{\beta}} D^{\nu} -\frac{1}{24}
R_{\alpha\mu\beta\nu ; \sigma} p^{\mu} p^{\nu} \frac{\partial^{3}}{\partial
p_{\alpha} \partial p_{\beta} \partial p_{\sigma}} + \right. \nonumber \\
\left.
\mbox{} + \frac{1}{12} R_{\alpha}^{\nu} \frac{\partial}{\partial p_{\alpha}}
D_{\nu}
- \frac{1}{24} R_{\alpha\beta ; \nu} p^{\nu} \frac{\partial^{2}}{\partial
p_{\alpha} \partial p_{\beta}} - \frac{1}{24} R_{; \alpha}
\frac{\partial}{\partial p_{\alpha}} \right)\: f + {O}(\hbar^{4}) \; ,
\hspace*{.1in} \label{eq:tranf}\end{aligned}$$ where $
D_{\alpha} = \nabla_{\alpha}
+ \Gamma_{\alpha\beta}^{\gamma}\: p_{\gamma}\: \frac{\partial}{\partial
p_{\beta}}$ is the horizontal lift of the derivative operator to the [*cotangent*]{} bundle$^{\ref{kn:yano}}$.\
Eq.(\[eq:mashf\]) is a generalized mass–shell constraint, and it implies that in the semiclassical limit, the structure of the Wigner function must be as follows$^{\ref{kn:hu}}$: $$f = F_{0} \, {\delta}(m^{2}-p^{2}) +
\hbar^{2} F_{1} \, {\delta'}(m^{2}-p^{2}) +
\hbar^{2} F_{2} \, {\delta''}(m^{2}-p^{2}) + {O}(\hbar^{4}) \; .
\label{eq:expf}$$ All the $F_{n}$’s in (\[eq:expf\]) are easily found from eq.(\[eq:mashf\]) and they are expressed$^{\ref{kn:hu},\ref{kn:fonarev11}}$ in terms of certain local differential operators acting to the function $F_{0}$. The function $F_{0}$ satisfies the quantum corrected Vlasov equation (\[eq:tranf\]) on the mass–shell. $F_{0}$ is decomposed to classical and quantum parts: $$F_{0} = F_{cl} + \hbar^{2} F_{qu} + {O}(\hbar^{4}) \; , \label{eq:F0}$$ where $F_{qu}$ is found by integrating the classical distribution function $F_{cl}$ along classical trajectories and is in general nonlocal in the phase space.\
Thus, given a classical distribution function ${F_{cl}}(x,p)$ as a solution to the Vlasov equation, one is able, in principle, to evaluate the semiclassical expansion of the generalized Wigner function and of any physical observable expressible in terms of the Wigner function. For example, the number–flux ${N_{\alpha}}(x)$ and the stress–energy vector ${T_{\alpha\beta}}(x)$ are given by$^{\ref{kn:winter},\ref{kn:fonarev11}}$: $${N_{\alpha}}(x) = \int\frac{d^{4}\,p}{\sqrt{-{g}(x)}}\: p_{\alpha}\:f(x,p) \; ,
\label{eq:flux}$$ $$\begin{aligned}
{T_{\alpha\beta}}(x) =
\int\frac{d^{4}\,p}{\sqrt{-{g}(x)}}\: p_{\alpha}\,p_{\beta}\:{f}(x,p) +
\hspace*{1in} \nonumber \\
\mbox{} + \hbar^{2}\left(\frac{1}{6} {R_{\alpha\beta}}(x) + \frac{1}{12}
(\nabla_{\alpha}\nabla_{\beta} -
{g_{\alpha\beta}}(x)\nabla^{\nu}\nabla_{\nu})\right)
\int\frac{d^{4}\,p}{\sqrt{-{g}(x)}}\: {f}(x,p)
\; . \label{eq:ener}\end{aligned}$$ Let us now consider a stationary space–time admitting a global time–like Killing vector $\xi^{\alpha}$. In that case, any function, ${F_{cl}}(E)$, of the constant of motion, $E=\xi^{\alpha} p_{\alpha}$, is a solution of the Vlasov equation. The classical stress–energy tensor calculated with such distribution has a perfect fluid’s structure. The quantum corrections of the lowest adiabatic order to this distribution were found in Ref.\[\[kn:fonarev2\]\] and they possess the following properties: a) all the quantum corrections are local in the phase space and they are obtained by acting a certain differential operator to $F_{cl}$. This means that physical observables are expressed in terms of momenta of the classical distribution function; b) the stress–energy tensor has the following structure: $$T_{\alpha\beta} = (\varepsilon + p) v_{\alpha} v_{\beta} - p g_{\alpha\beta} -
\Pi_{\alpha\beta} \; , \label{eq:quen}$$ which is not longer a perfect fluid’s one : the anisotropic pressure tensor $\Pi_{\alpha\beta}$ appears ($\Pi^{\alpha}_{\alpha} = v^{\alpha}
\Pi_{\alpha\beta} = 0$); c) the eigenvector of the stress–energy tensor, $v_{\alpha}$, doesn’t coincide with the hydrodynamical velocity $u_{\alpha} = N_{\alpha}
(N^{\nu}N_{\nu})^{-1/2}$. This leads to the nonvanishing heat flow: $I_{\alpha} = (\varepsilon + p) (v_{\alpha} - u_{\alpha})$. $I_{\alpha}$ is proportional to the combination $N_{0}/6N_{2} - N_{1}/4N_{3}$, where $N_{k} = \int_{0}^{\infty} dE E^{k} {F_{cl}}(E)$. For the Boltzmann or Bose–Einstein distributions this combination doesn’t vanish.\
Only in static space–times (i.e. for nonrotating systems) the quantum corrections preserve the perfect fluid’s structure with the vanishing heat flow.\
Let us next consider a more general case of conformally stationary (or conformally static) space–times. The definition (\[eq:wigf\]) of the generalized Wigner function given allows one to easily link Wigner functions in two conformally related manifolds. Under the conformal transformation, ${g_{\alpha\beta}}(x) \mapsto {a}(x)^{2} {g_{\alpha\beta}}(x),
{\varphi}(x) \mapsto {\varphi}(x)/{a}(x)$, the Wigner function transforms as follows$^{\ref{kn:fonarev3}}$: $$f \mapsto a^{2} (1 + \hbar^{2} \hat{A} + {O}(\hbar^4)) f \; ,$$ where $\hat{A}$ is a certain differential operator involving derivatives of the conformal factor ${a}(x)$. If one knows a Wigner function in a stationary space–time, one is able to compute a Wigner function in a conformally stationary space–time. We have found$^{\ref{kn:fonarev3}}$ by this method the quantum corrections to a class of distribution functions isotropic in the momentum space$^{\ref{kn:ehlers}}$, $F_{cl}={F}(u^{\alpha}p_{\alpha},x)$.\
The Robertson–Walker Universes are examples of conformally static (and also conformally flat) space–times. The metric tensor takes the form: $g_{\alpha\beta}={a}(t)^2 \gamma_{\alpha\beta}$ where $\gamma_{\alpha\beta}$ is the metric tensor of a spherically symmetrical static space–time, and $t$ is the conformal time. The classical distribution function compatible with the symmetry of the space–time is any function of the three–momentum$^{\ref{kn:ehlers}}$. The quantum corrections to such distributions were first found by \[\[kn:pirk\]\] through a different method. The structure of the quantum–corrected stress–energy tensor is a perfect fluid’s one in this case. We also found$^{\ref{kn:fonarev3}}$ that the quantum corrected energy density takes the following form: $${\varepsilon}(t) = {\varepsilon_{cl}}(t) + \frac{1}{\kappa}
\frac{m^{2}}{M_{pl}^{2}} {\cal K}(\frac{m}{{\Theta}(t)}) {G_{0}^{0}}(t) +
{O}(\hbar^{4}) \label{eq:enRW}$$ where ${\cal K}$ is a certain function of the ratio of the particles mass $m$ to the local temperature ${\Theta}(t)=T/{a}(t)$, $M_{pl}$ is the Planck mass and $G_{0}^{0}$ is the zero–zero component of the Einstein tensor. If one substitutes Expr. (\[eq:enRW\]) into the Einstein equation, one finds that the effective gravitational “constant” appears which varies with time: $$\frac{1}{{\kappa_{eff}}(t)} = \frac{1}{\kappa} (1 - \frac{m^{2}}{M_{pl}^{2}}
{\cal K}(\frac{m}{{\Theta}(t)})) \; . \label{effgr}$$ The variation is negligible at present but it could lead to important physical effects in the early Universe.
[$^{\arabic{refer}}$]{}[ ]{}
\[kn:wigner\] E. P. Wigner, Phys. Rev. [**40**]{}, 749 (1932).
\[kn:winter\] J. Winter, Phys. Rev. D [**32**]{}, 1871 (1985).
\[kn:hu\] E. Calzetta, S. Habib, and B. L. Hu, Phys. Rev. D [**37**]{}, 2901 (1988).
\[kn:kandrup\] H. E. Kandrup, Phys. Rev. D [**37**]{}, 2165 (1988).
\[kn:fonarev1\] O. A. Fonarev, Izv. VUZov. Fizika (USSR) [**9**]{}, 47 (1990).
\[kn:fonarev11\] O. A. Fonarev, preprint gr–qc/9309005; to appear in J.Math.Phys. (USA).
\[kn:yano\] K. Yano and Sh. Ishihara, [*Tangent and Cotangent Bundles: Differential Geometry*]{} (Pure and Applied Mathematics: A Series of Monographs and Textbooks) (Marcel Dekker, Inc. , New York, 1973).
\[kn:degroot\] S. R. de Groot, W. A. van Leeuwen, and Ch. G. van Weert, [*Relativistic Kinetic Theory*]{} (North–Holland, Amsterdam, 1980).
\[kn:fonarev2\] O. A. Fonarev, Phys. Lett. A [**152**]{}, 153 (1991).
\[kn:fonarev3\] O. A. Fonarev, in preparation.
\[kn:ehlers\] J. Ehlers, P. Geren and R. K. Sachs, J. Math. Phys. [**9**]{}, 1344 (1968).
\[kn:pirk\] K. Pirk and G. B$\ddot{o}$rner, Class. Quantum Grav. [**6**]{}, 1855 (1989).
[^1]: talk given at the 3rd International Wigner Symposium, 5th-11th September 1993, Oxford, UK.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper we present some extensions of the Ailon-Rudnick Theorem, which says that if $f,g\in\C[T]$, then $\gcd(f^n-1,g^m-1)$ is bounded for all $n,m\ge 1$. More precisely, using a uniform bound for the number of torsion points on curves and results on the intersection of curves with algebraic subgroups of codimension at least $2$, we present two such extensions in the univariate case. We also give two multivariate analogues of the Ailon-Rudnick Theorem based on Hilbert’s irreducibility theorem and a result of Granville and Rudnick about torsion points on hypersurfaces.'
address: 'School of Mathematics and Statistics, University of New South Wales, Sydney NSW 2052, Australia'
author:
- Alina Ostafe
title: 'On some extensions of the Ailon-Rudnick Theorem'
---
\[theorem\][Lemma]{} \[theorem\][Example]{} \[theorem\][Corollary]{} \[theorem\][Proposition]{} \[theorem\][Definition]{} \[theorem\][Question]{} \[theorem\][Problem]{} \[theorem\][Remark]{} \[theorem\][Conjecture]{}
\#1[\#1]{}
${\left(}
\def$[)]{} $${\left[}
\def$$[\]]{}
\#1[[\#1]{}]{} \#1[[\#1]{}\_p]{}
${\left(}
\def$[)]{} \#1[\#1]{} \#1[\#1]{}
§ \#1 \#1
Introduction
============
Motivation
----------
Let $a,b$ be multiplicatively independent positive integers and $\varepsilon>0$. Bugeaud, Corvaja and Zannier [@BCZ] have proved that $$\gcd\(a^n-1,b^n-1\)\le \exp(\varepsilon n)$$ as $n$ tends to infinity. Corvaja and Zannier [@CZ05] have generalised this result and replaced $a^n,b^n$ with multiplicatively independent $S$-units $u,v\in\Z$.
In the function field case, Ailon and Rudnick [@AR Theorem 1] proved that if $f,g\in\C[T]$ are multiplicatively independent polynomials, then there exists $h \in \C[T]$ such that $$\label{eq:AR}
\gcd(f^n-1,g^n-1) \mid h$$ for all $n\ge 1$. Examining their argument one can easily see that the same statement holds in a larger generality; namely there exists $\widetilde h \in \C[T]$ such that $$\label{eq:AR Gen}
\gcd(f^n-1,g^m-1) \mid \widetilde h$$ for all $n,m\ge 1$.
In the case of finite fields $\F_q$ of characteristic $p$, Silverman [@Si] proves that even more restrictions on the polynomials $f,g\in\F_q[T]$ does not allow a similar conclusion as the result of [@AR]. In particular, Silverman proves that the analogue of is false in a very strong sense: there exists a constant $c(f,g;q)$, depending only on $f$, $g$ and $q$, such that $$\deg \gcd(f^n-1,g^n-1)\ge c(f,g;q) n$$ for infinitely many $n$.
More results in positive characteristic are obtained in [@CZ13; @Lau], as well as variants for elliptic divisibility sequences [@Si04; @Si05].
In this paper we present some extensions of the Ailon-Rudnick Theorem [@AR Theorem 1] over $\C$, both in the univariate and multivariate cases. Although the method of proof in the univariate case is similar to, or reduces to using, [@AR], we find these extensions exciting and we hope they will be of independent interest. Moreover, as we explain below, in certain situation we reduce our problem to applying [@AR Theorem 1], however for this we need a uniform bound for that depends only on the degree of the polynomials $f$ and $g$.
Besides the generality of results, the new ingredients of the paper are employing results [@BS; @BMZ99; @BMZ08; @Mau] on the number of points on intersections of curves in the $n$-dimensional multiplicative torus $\G_m^n$ with algebraic subgroups. We also present two multivariate generalisations that are based on the use of Hilbert’s irreducibility theorem [@Sch] and a transformation using the Kronecker substitution to reduce the problem to the univariate case, as well as a result of Granville and Rudnick [@GranRud] about torsion points on hypersurfaces.
Conventions and notation
------------------------
As usual, we denote $\C[X_1,\ldots,X_{\ell}]$ the polynomial ring in ${\ell}$ variables and $\C(X_1,\ldots,X_{\ell})$ the field of rational functions $F/G$, $F,G\in\C[X_1,\ldots,X_{\ell}]$. When working with univariate polynomials we reserve the variable $T$. All polynomials in $\C[T]$ are denoted with small letters $f,g,\ldots$, and for polynomials in $\C[X_1,\ldots,X_\ell]$ we use capital letters $F,G,\ldots$.
Throughout the paper, for a univariate polynomial $f\in\C[T]$, the notation $d_f$ will be used for the degree of $f$.
For a family of polynomials $F_1,\ldots,F_s \in \C[X_1, \ldots, X_{\ell}]$, we denote by $Z(F_1,\ldots,F_s)$ their zero set in $\C^{m}$.
Throughout the paper we assume that the greatest common divisor of two (or more) polynomials is monic, so it is well-defined.
We also define here the main concept of this paper.
The polynomials $F_1,\ldots,F_{s}\in\C[X_1,\ldots,X_{\ell}]$ are [*multiplicatively independent*]{} if there exists no nonzero vector $(\nu_1,\ldots,\nu_{s})\in\Z^{s}$ such that $$F_1^{\nu_1}\cdots F_s^{\nu_{s}}=1.$$ Similarly, we say that the polynomials $F_1,\ldots,F_{s}\in\C[X_1,\ldots,X_{\ell}]$ are [*multiplicatively independent in the group $\C(X_1,\ldots,X_{\ell})^*/\C^*$*]{} if there exists no nonzero vector $(\nu_1,\ldots,\nu_{s})\in\Z^{s}$ and $a\in\C^*$ such that $$F_1^{\nu_1}\cdots F_s^{\nu_{s}}=a.$$
We present now in more details the main results of this paper.
Our results: univariate case
----------------------------
Section \[sec:prel\] is dedicated to outlining the tools and results needed along the paper. In particular, in Section \[sec:ART\] we recall the result of [@AR Theorem 1] and, using a uniform bound for the number of points on a curve with coordinates roots of unity due to Beukers and Smyth [@BS], we derive in Lemma \[lem:univAR\] a version of that gives an upper bound on $\deg \gcd(f^n-1,g^m-1)$ that depends only the degrees of $f$ and $g$ (rather than on the polynomials themselves).
Such a uniform bound is crucial for some of our main results presented below and proved in Section \[sec:main\]. In particular, our first extension of [@AR Theorem 1], which is proved in Section \[sec:univar1\], is based on this uniform bound.
\[thm:genAR1\] Let $f,g,h_1,h_2\in\C[T]$. If $f$ and $g$ are multiplicatively independent in $\C(T)^*/\C^*$, then for all $n,m\ge 1$ we have $$\deg \gcd\(h_1\(f^n\),h_2\(g^m\)\)\le d_{h_1}d_{h_2} \(11(d_f+d_g)^{2}\)^{\min(d_f,d_g)}.$$
For the second extension of [@AR Theorem 1], which is proved in Section \[sec:univar2\], we apply the finiteness result of [@BMZ08; @Mau], see also [@BMZ99], for the number of points on the intersection of curves in $\G_m^n$ with algebraic subgroups, see Lemma \[lem:bmz\]. No uniform bounds are known so far for such finiteness results.
We recall that for a polynomial $f\in\C[T]$, we denote by $Z(f)$ the set of zeros of $f$ in $\C$.
\[thm:genAR2\] Let $f_1,\ldots,f_{\ell},\varphi_1,\ldots,\varphi_{k},g_1,\ldots,g_r,\psi_1,\ldots,\psi_s\in\C[T]$, $\ell,k,r,s\ge 1$, be multiplicatively independent polynomials such that $$\label{eq:cond}
Z(f_1\cdots f_{\ell})\cap Z(\varphi_1\cdots \varphi_k)=\emptyset,\quad Z(g_1\cdots g_r)\cap Z(\psi_1\cdots \psi_s)=\emptyset.$$ Then we have:
- For all $n_1,\ldots,n_{\ell},\nu_1,\ldots,\nu_k,m_1,\ldots,m_r,\mu_1,\ldots,\mu_s\ge 0$, there exists a polynomial $h\in \C[T]$ such that $$\gcd\(\prod_{i=1}^{\ell}f_i^{n_i}-\prod_{i=1}^k\varphi_i^{\nu_i},\prod_{i=1}^rg_i^{m_i}-\prod_{i=1}^s\psi_i^{\mu_i}\)\mid h.$$
- If in addition $$\gcd(f_1\cdots f_{\ell}-1,g_1\cdots g_r-1)=1,$$ then there exists a finite set $S$ and monoids $\cL_t \subseteq \N^{\ell+k+r+s}$, $t \in S$, such that the remaining set $$\cN = \N^{\ell+k+r+s}\setminus \cup_{t\in S} \cL_t$$ is of positive asymptotic density and for any vector $$(n_1,\ldots,n_{\ell},\nu_1,\ldots,\nu_k,m_1,\ldots,m_r,\mu_1,\ldots,\mu_s)\in\cN$$ we have $$\gcd\(\prod_{i=1}^{\ell}f_i^{n_i}-\prod_{i=1}^k\varphi_i^{\nu_i},\prod_{i=1}^rg_i^{m_i}-\prod_{i=1}^s\psi_i^{\mu_i}\)=1.$$
Although we prefer to keep the language of polynomials, one can easily see that Theorem \[thm:genAR2\] can be reformulated in terms of $S$-units in $\C[T]$ and implies that for any set of $S$-units, there exists a polynomial $h\in \C[T]$ such that for any multiplicatively independent $S$-units $U,V$ we have $$\gcd\(U-1,V-1\)\mid h.$$ In particular, this extension of [@AR] is fully analogous to the aforementioned extension of [@CZ05] over [@BCZ].
We also compare Theorem \[thm:genAR2\], which for multiplicatively independent $S$-units $U,V$, gives a uniform bound for $\deg\gcd(U-1,V-1)$, while the result of Corvaja and Zannier [@CZ1-08 Corollary 2.3] gives $$\deg \gcd(U-1,V-1)\ll \max(\deg U,\deg V)^{2/3}.$$ However, [@CZ1-08 Corollary 2.3] applies to more general situations.
It is interesting to unify Theorems \[thm:genAR1\] and \[thm:genAR2\] and obtain a similar result for $$\gcd\(h_1\(f_1^{n_1}\cdots f_{\ell}^{n_{\ell}}\),h_2\(g_1^{m_1}\cdots g_r^{m_r}\)\),$$ where $h_1,h_2\in\C[T]$. Similar ideas may work for this case however they require a uniform bound for the number of points on intersections of curves in $\G_{m}^{\ell+r}$ with algebraic subgroups of dimension $k \le \ell+r-2$ in Lemma \[lem:bmz\]. We note that for $\ell=r=1$ this was possible due to the uniform bounds of [@BS]. However, no such bounds are available in the more general case that we need.
Our results: multivariate case
------------------------------
For our first result in the multivariate case, we reduce the problem to the univariate case using Hilbert’s irreducibility theorem (see Section \[sec:Hilb\]), and to control the degree for such specialisation we also couple this approach with a transformation involving the Kronecker substitution. We obtain:
\[thm:multivAR\] Let $h_1,h_2\in\C[T]$ and $F,G\in\C[X_1,\ldots,X_{\ell}]$. We denote by $D=\max_{i=1\ldots,\ell}\(\deg_{X_i}F,\deg_{X_i}G\)$. If $F, G$ are multiplicatively independent in $\C(X_1,\ldots,X_{\ell})^*/\C^*$, then for all $n,m\ge 1$ we have $$\deg\gcd\(h_1\(F^n\),h_2\(G^m\)\)\le d_{h_1} d_{h_2} \(44(D+1)^{2{\ell}}\)^{(D+1)^{\ell}}.$$
We note that if $h_1=h_2=T-1$ as in [@AR Theorem 1], then in Theorem \[thm:multivAR\] we need $F,G$ to be just multiplicatively independent.
Theorem \[thm:multivAR\] is proved in Section \[sec:multivar1\].
Another natural extension of [@AR Theorem 1] to the multivariate case is related to the fact that the greatest common divisor of two univariate polynomials is given by their common zeros. Thus [@AR Theorem 1] says that the number of common zeros of $f^n-1$ and $g^m-1$, for two polynomials $f,g\in\C[T]$, is bounded by a constant depending only on $f$ and $g$ for all $n,m\ge 1$, and Lemma \[lem:univAR\] gives a uniform bound.
For positive integers $\ell,D\ge 1$, we denote $$\label{eq:gamma}
\gamma_{\ell}(D)=\binom{\ell+1+D^\ell}{\ell+1}.$$
We now obtain the following result proved in Section \[sec:multivar2\]. This multivariate generalisation is based on a result of Granville and Rudnick [@GranRud Corollary 3.1], which describes the structure of torsion points on hypersurfaces, see Lemma \[lem:GranRud\].
\[thm:CommonZeros\] Let $F_1,\ldots,F_{\ell+1}\in\C[X_1,\ldots,X_{\ell}]$ be multiplicatively independent polynomials of degree at most $D$. Then, $$\bigcup_{n_1,\ldots,n_{\ell+1}\in\N}Z\(F_1^{n_1}-1,\ldots,F_{\ell+1}^{n_{\ell+1}}-1\)$$ is contained in at most $$N \le (0.792\gamma_{\ell}(D)/\log \(\gamma_{\ell}(D)+1\))^{\gamma_{\ell}(D)}$$ algebraic varieties, each defined by at most $\ell+1$ polynomials of degree at most $(\ell+1)D^{\ell}\prod_{p\le \gamma_{\ell}(D)}p$, where the product runs over all primes $p\le \gamma_{\ell}(D)$.
We recall that by the prime number theorem, for an integer $k\ge 1$, $$\prod_{p\le k}p = \exp(k + o(k)).$$
We note that the bound on the number of algebraic subgroups that contain the points on $\cH$ with coordinates roots of unity may also follow from [@AS Theorem 1.1 and Lemma 2.6], which says that, for a hypersurface defined by $H\in\C[X_1,\ldots,X_s]$, $s\ge 2$, of degree $D$, the number of maximal torsion cosets contained in $\cH$ is at most $$c_1(s)D^{c_2(s)}$$ with $$\label{eq:c1c2}
c_1(s)=s^{\frac{3}{2}(2+s)5^{s}}\quad \textrm{and}\quad c_2(s)=\frac{1}{16}\(49 \cdot 5^{s-2}-4s-9\).$$
We also note that any argument that is based on the Bezout Theorem ultimately leads to bounds that depend on the exponents $n_1,\ldots,n_{\ell+1}$, while the bounds of Theorem \[thm:CommonZeros\] depend only on the initial data.
We conclude the paper with comments on future work.
Preliminaries {#sec:prel}
=============
The Ailon-Rudnick Theorem {#sec:ART}
-------------------------
The Ailon-Rudnick theorem is based on a well-known conjecture of Lang, proved by Ihara, Serre and Tate [@L], which says that a plane curve, which does not contain a translate of an algebraic subgroup of $\G_m^2$, contains only finitely many torsion points. In this case, Beukers and Smyth [@BS Section 4.1] give a uniform bound for the number of such points (see Lemma \[lem:BS\] below), and Corvaja and Zannier [@CZ08] give an upper bound (actually for curves in $\G_m^n$) for the maximal order of torsion points on the curve.
We now present the result of Ailon and Rudnick [@AR Theorem 1], coupled with the result of Beukers and Smyth [@BS Section 4.1], which we first mention separately.
\[lem:BS\] An algebraic curve $H(X,Y)=0$ has at most $11(\deg H)^2$ points which are roots of unity, unless $H$ has a factor of the form $X^i-\rho Y^j$ or $X^iY^j-\rho$ for some nonnegative integers $i,j$ not both zero and some root of unity $\rho$.
Using Lemma \[lem:BS\], we obtain the following more precise form of [@AR Theorem 1].
\[lem:univAR\] Let $f,g\in\C[T]$ be non constant polynomials. If $f$ and $g$ are multiplicatively independent, then for all $n,m\ge 1$, we have $$\deg\gcd\(f^n-1,g^m-1\)\le \(11(d_f+d_g)^2\)^{\min(d_f,d_g)}.$$
The proof, except the explicit bound for the degree, is given in [@AR Theorem 1]. To see the degree bound, we just apply Lemma \[lem:BS\].
Our curve is given in parametric form $\{\(f(t),g(t)\)~:~ t\in\C\}$ and we need to find the degree of the implicit form $H$ such that $H(f(t),g(t))=0$, $t\in\C$. This is obtained using resultants, that is $$H={{\operatorname{Res}}}_T\(f(T)-X_1,g(T)-X_2\),$$ which is a polynomial of degree $\deg g$ in $X_1$ and $\deg f$ in $X_2$. Thus, the total degree of $H$ is at most $\deg f+\deg g$.
Let $\widetilde{H}$ be an absolutely irreducible factor of $H$ and assume that $\widetilde{H}(f(t),g(t))=0$ for infinitely many $t\in\C$. As $\widetilde{H}(f(T),g(T))$ is a univariate polynomial , we must have the identity $\widetilde{H}(f(T)),g(T))=0$. Then, by Lemma \[lem:BS\] applied with the curve defined by the polynomial $\widetilde{H}$, we obtain that $\widetilde{H}$ is of the form $X_1^{n_1}X_2^{n_2}=\omega$, for some root of unity $\omega$ and integers $n_1,n_2$ not both zero. This implies that $f,g$ are multiplicatively dependent, which contradicts the hypothesis. Thus, there is no such absolutely irreducible divisor of $H$.
Therefore $\gcd\(f^n-1,g^m-1\)$ has at most $11(d_f+d_g)^2$ distinct zeros. As in the proof of [@AR Theorem 1], we see that the multiplicity of each zero is at most $\min(d_f,d_g)$. The bound now follows.
Intersection of curves with algebraic groups {#sec:intesect}
--------------------------------------------
We define $\G_m^k$ as the set of $k$-tuples of non-zero complex numbers equiped with the group law defined by component-wise multiplication. We refer to [@Sch Appendix by Umberto Zannier] for necessary definitions on algebraic subgroups.
One of the main tools in our paper is a result on the finiteness of the number of points on the intersection of a curve in $\G_m^k$ with algebraic subgroups of codimension at least $2$, initially obtained in [@BMZ99] for curves over $\ovQ$, and later on extended over $\C$, see [@BMZ08; @Mau] and references therein. We present it in the following form.
\[lem:bmz\] Let $C\subset \G_m^{k}$, $k\ge 2$, be an irreducible curve over $\C$. Assume that for every nonzero vector $(r_1,\ldots,r_{k})\in\Z^k$ the monomial $X_1^{r_1}\cdots X_{k}^{r_{k}}$ is not identically $1$ on $C$. Then there are finitely many points $(x_1,\ldots,x_{k})\in C(\C)$ for which there exist linearly independent vectors $(a_1,\ldots,a_{k}), (b_1,\ldots,b_{k})$ in $\Z^{k}$ such that $$x_1^{a_1}\cdots x_{k}^{a_{k}}=x_1^{b_1}\cdots x_{k}^{b_{k}}=1.$$
\[rem:translate\] As explained in [@BMZ99], the condition of Lemma \[lem:bmz\] that the monomial $X_1^{r_1}\cdots X_{k}^{r_{k}}$ is not identically $1$ on $C$ is equivalent with the curve not being contained in a proper subtorus of $\G_m^{k}$.
Torsion points on hypersurfaces
-------------------------------
Results regarding uniform bounds on the number of torsion points in subvarieties of $\G_m^k$ go back to work of Bombieri and Zannier [@BZ95], Schlickewei [@Schl] and Evertse [@Ever]. For example, Evertse [@Ever], improving bounds of Schlickewei [@Schl], shows that the number of non-degenerate solutions in roots of unity to the equation $a_1x_1+\cdots+a_kx_k=1$, $a_1,\ldots,a_k\in\C$, is at most $(k+1)^{3(k+1)^2}$.
For our results we use the following result of Granville and Rudnick, see [@GranRud Corollary 3.1], which describes the structure of the algebraic subgroups that contain the roots of unity on a hypersurface. Although the statement of their result does not contain the bound for the degree or the number of the polynomials defining the algebraic subgroups, this follows directly from or is explicitly stated in their proof. Moreover, we recall this result only for the case of hypersurfaces, however their result holds for any algebraic variety.
\[lem:GranRud\] Let $\cH=Z(H)$ be a hypersurface in $\C^k$ defined by a polynomial $H\in\C[X_1,\ldots,X_k]$ of degree $D$ and with $s(H)$ terms. There exists a finite list $\cB$ of at most $$N(H) \le (0.792s(H)/\log \(s(H)+1\))^{s(H)}$$ integer $k\times k$ matrices $B=(b_{j,i})$, $i,j=1,\ldots,k$, and $$b_{j,i}\le D\prod_{p\le s(H)}p,$$ where the product runs over all primes $p\le s(H)$, such that if $\xi\in\cH$ is a torsion point, then $\xi\in\cup_{B\in\cB}W_B$, where $$W_B=\bigcap_{j=1}^{k}Z\(X_1^{b_{j,1}}X_2^{b_{j,2}}\cdots X_k^{b_{j,k}}-1\).$$
The proof is essentially given in [@GranRud Corollary 3.1]. Indeed, each matrix $B$ corresponds to a partition of the set $\{1,2,\ldots,s(H)\}$, and thus, the number of matrices $B$ in the set $\cB$ is given by the number of such partitions, which, by [@BT Theorem 2.1], is at most $N(H)$.
The number of rows $n_B$ of a matrix $B \in \cB$ is not specified in [@GranRud Corollary 3.1]. However, we can choose the largest linear independent set of these vectors $\vec{b}_j$, which is of cardinality at most $k$, and all other varieties of the form $Z\(X_1^{b_{i,1}}X_2^{b_{i,2}}\cdots
X_k^{b_{i,k}}-1\)$ are defined by combinations of these vectors. Thus, we can consider $n_B\le k$. Repeating some rows if necessary we can take $n_B=k$ which concludes the proof.
Hilbertian fields and multiplicative independence {#sec:Hilb}
-------------------------------------------------
For the first multivariate generalisation of [@AR Theorem 1] we need a result which says that given $F_1,\ldots,F_s\in\C[X_1,\ldots,X_{\ell}]$ that are multiplicatively independent in $\C(X_1,\ldots,X_{\ell})^*/\C^*$, there exists a specialisation $(\alpha_2,\ldots,\alpha_{\ell})\in\C^{\ell-1}$ such that $F_i(X_1,\alpha_2,\ldots,\alpha_{\ell})$, $i=1,\ldots,s$, are multiplicatively independent in $\C(X_1)^*/\C^*$ (see Lemma \[lem:multindep\] below). Such a result follows directly from [@BMZ99 Theorem 1] which says that the points lying in the intersection of a curve $\cC$, not contained in any translate of a proper subtorus of $\G_m^{\ell}$, with the union of all proper algebraic subgroups is of bounded height (see also [@BMZ99 Theorem 1’]).
Furthermore, this also follows from previous work of N' eron [@Ne] (see also [@Serre Chapter 11]), Silverman [@Si83 Theorem C] and Masser [@Ma89] (see also [@Za Notes to Chapter 1] where Masser’s method is explained) on specialisations of finitely generated subgroups of abelian varieties. In particular, Masser’s result [@Ma89] gives explicit bounds for the least degree of a hypersurface containing the set of exceptional points, that is, points that lead to multiplicative dependence, of bounded degree and height.
Although the above results are sufficient for our purpose, for the sake of completeness we now give a simple self-contained proof that follows directly from Hilbert’s irreducibility theorem, see [@Sch Theorem 46]. Moreover, this proof does not appeal to the notion of height and applies to arbitrary Hilbertian field (see Definition \[def:HIT\] below), rather than to just finite extensions of $\Q$.
\[def:HIT\] We say that a field $\K$ is Hilbertian if for any irreducible polynomials $P_1,\ldots,P_r\in\K[X_1,\ldots,X_{\ell}]$ over $\K$ there exists a specialisation $(\alpha_2,\ldots,\alpha_{\ell})\in\K^{{\ell}-1}$ such that $P_i(X_1,\alpha_2,\ldots,\alpha_{\ell})$, $i=1,\ldots,r$, are all irreducible over $\K$.
In particular, by the famous Hilbert’s irreducibility theorem, any finite extension of $\Q$ is a Hilbertian field. Furthermore, by [@Sch Theorem 49] every finitely generated infinite field and every finitely generated transcendental extension of an arbitrary field are Hilbertian.
We also need the following simple fact.
\[lem:multconst\] Let $\K$ be an arbitrary field. The polynomials $F_1,\ldots,F_{s}\in\K[X_1,\ldots,X_{\ell}]$ are multiplicatively independent in $\K(X_1,\ldots,X_{\ell})^*/\K^*$ if and only if $$\frac{F_1}{F_1^*(\vec{0})},\ldots,\frac{F_{s}}{F_s^*(\vec{0})}$$ are multiplicatively independent, where $F_i^*(\vec{0}) =1$ if $F_i(\vec{0})=0$ and $F_i^*(\vec{0}) =F_i(\vec{0})$ otherwise.
Assume that $F_1,\ldots,F_{s}\in\K[X_1,\ldots,X_{\ell}]$ are multiplicatively independent in $\K(X_1,\ldots,X_{\ell})^*/\K^*$, but there exist integers $i_1,\ldots,i_{\ell}$ not all zero, such that $$\(\frac{F_1}{F_1^*(\vec{0})}\)^{i_1}\cdots\(\frac{F_{s}}{F_s^*(\vec{0})}\)^{i_s}=1.$$ Then, $F_1^{i_1}\cdots F_s^{i_s}=a$, where $a=\(F_1^*(\vec{0})\)^{i_1}\cdots\(F_s^*(\vec{0})\)^{i_s}$, and thus we obtain a contradiction.
For the other implication, assume $F_1^{i_1}\cdots F_s^{i_s}=a$ for some $a\in\K^*$ and integers $i_1,\ldots,i_s$ not all zero. From here we get again that $
a=\(F_1^*(\vec{0})\)^{i_1}\cdots\(F_s^*(\vec{0})\)^{i_s}$ (we note that if $F_k(\vec{0})=$ for some $k$, then we need to have $i_k=0$ as otherwise we get $a=0$, which is a contradiction). Thus we obtain again a contradiction with the fact that the polynomials $F_1/F_1^*(\vec{0}),\ldots, F_{s}/F_s^*(\vec{0})$ are multiplicatively independent.
We note that the conclusion of Lemma \[lem:multconst\] holds with any ${{\boldsymbol{\alpha}}}\in\K^{\ell}$, that is, $F_1,\ldots,F_{s}\in\K[X_1,\ldots,X_{\ell}]$ are multiplicatively independent in $\K(X_1,\ldots,X_{\ell})^*/\K^*$ if and only if $\frac{F_1}{F_1^*({{\boldsymbol{\alpha}}})},\ldots,\frac{F_{s}}{F_s^*({{\boldsymbol{\alpha}}})}$ are multiplicatively independent for any ${{\boldsymbol{\alpha}}}\in\K^{\ell}$, where as before $F_i^*({{\boldsymbol{\alpha}}}) =1$ if $F_i({{\boldsymbol{\alpha}}})=0$ and $F_i^*({{\boldsymbol{\alpha}}}) =F_i({{\boldsymbol{\alpha}}})$ otherwise.
The following result is easily derived from Lemma \[lem:multconst\].
\[lem:multindep\] Let $\K$ be a Hilbertian field and $F_1,\ldots,F_s\in\K[X_1,\ldots,X_{\ell}]$ multiplicatively independent polynomials in $\K(X_1,\ldots,X_{\ell})^*/\K^*$. Then, there exists a specialisation $(\alpha_2,\ldots,\alpha_{\ell})\in\K^{{\ell}-1}$ such that the polynomials $F_i(X_1,\alpha_2,\ldots,\alpha_{\ell})$, $i=1,\ldots,s$, are multiplicatively independent in $\K(X_1)^*/\K^*$.
By Lemma \[lem:multconst\] the polynomials $F_1,\ldots,F_s$ are multiplicatively independent in $\K(X_1,\ldots,X_{\ell})^*/\K^*$ if and only if $\frac{F_1}{F_1^*(\vec{0})},\ldots,\frac{F_{s}}{F_s^*(\vec{0})}$ are multiplicatively independent.
We denote $G_i=\frac{F_i}{F_i^*(\vec{0})}$, $i=1,\ldots,s$. Let $P_1,\ldots,P_r\in\K[X_1,\ldots,X_{\ell}]$ be the distinct irreducible factors of $G_1,\ldots,G_s$, that is, we have the factorisation $$G_i=P_1^{e_{i,1}}\cdots P_r^{e_{i,r}},\quad i=1,\ldots,s.$$
We note that the polynomials $G_1,\ldots,G_s$ being multiplicatively independent is equivalent with the matrix $(e_{i,j})_{\substack{1\le i\le s\\1\le j\le r}}$ having full rank.
Since $\K$ is Hilbertian, there exists a specialisation $(\alpha_2,\ldots,\alpha_{\ell})\in\K^{{\ell}-1}$ such that $P_j(X_1,\alpha_2,\ldots,\alpha_{\ell})$, $j=1,\ldots,r$, are all irreducible over $\K$ and for $i=1,\ldots,s$, we have the factorisation $$G_i(X_1,\alpha_2,\ldots,\alpha_{\ell})=P_1(X_1,\alpha_2,\ldots,\alpha_{\ell})^{e_{i,1}}\cdots P_r(X_1,\alpha_2,\ldots,\alpha_{\ell})^{e_{i,r}}.$$ If the polynomials $G_i(X_1,\alpha_2,\ldots,\alpha_{\ell})$, $i=1,\ldots,s$, are multiplicative dependent over $\K$, then there exist integers $\ell_1,\ldots,\ell_s$, not all zero such that $$G_1(X_1,\alpha_2,\ldots,\alpha_{\ell})^{\ell_1}\cdots G_s(X_1,\alpha_2,\ldots,\alpha_{\ell})^{\ell_s}=1.$$ This is equivalent to the fact that the matrix $(e_{i,j})_{\substack{1\le i\le r\\1\le j\le s}}$ does not have full rank, which contradicts the fact that the initial polynomials $G_1,\ldots,G_s\in\K[X_1,\ldots,X_{\ell}]$ are multiplicatively independent.
Recalling the definition of the polynomials $G_i$, $i=1,\ldots,s$, and applying again Lemma \[lem:multconst\], we get that the polynomials $F_i(X_1,\alpha_2,\ldots,\alpha_{\ell})$, $i=1,\ldots,s$, are multiplicatively independent in $\K(X_1)^*/\K^*$, which concludes the proof.
Multiplicities of zeroes
------------------------
To prove Theorem \[thm:genAR2\] we need a uniform bound for the multiplicities of zeros of polynomials of the form $f_1^{n_1}\cdots f_{\ell}^{n_{\ell}}-g_1^{m_1}\cdots g_r^{m_r}$. We present such a result below, as well as deduce as a consequence a similar uniform bound for rational functions, which we hope to be of independent interest.
For a rational function $h\in\C(T)$, we denote by $\Mult(h)$ the largest multiplicity and by $Z(h)$ the set of zeros of $h$ in $\C$, respectively. We also recall that for a polynomial $f\in\C[T]$, we use the notation $d_f$ for the degree of $f$.
\[lem:ABCpoly\] Let $f_1,\ldots,f_{\ell},g_1,\ldots,g_r\in\C[T]$ be polynomials satisfying $Z(f_1\cdots f_{\ell})\cap Z(g_1\cdots g_{\ell})=\emptyset$. Then, for all $n_1,\ldots,n_{\ell},m_1,\ldots,m_r\ge 0$, we have $$\Mult\(f_1^{n_1}\cdots f_{\ell}^{n_{\ell}}-g_1^{m_1}\cdots g_r^{m_r}\)\le \sum_{i=1}^{\ell}d_{f_i}+\sum_{j=1}^rd_{g_j}.$$
We denote $\vec{n}=(n_1,\ldots,n_{\ell})\in\N^{\ell}$ and $\vec{m}=(m_1,\ldots,m_r)\in\N^r$.
Writing the factorisation into linear factors, we have $$f_1^{n_1}\cdots f_{\ell}^{n_{\ell}}-g_1^{m_1}\cdots g_{r}^{m_r}=a_{\vec{n},\vec{m}}\prod_{t\in Z\(f_1^{n_1}\cdots f_{\ell}^{n_{\ell}}-g_1^{m_1}\cdots g_{r}^{m_r}\)}(T-t)^{e_t},$$ where $a_{\vec{n},\vec{m}}\in\C$ is the leading coefficient of $f_1^{n_1}\cdots f_{\ell}^{n_{\ell}}-g_1^{m_1}\cdots g_{r}^{m_r}$.
For simplicity we denote by $$S_{\vec{n},\vec{m}}=Z\(f_1^{n_1}\cdots f_{\ell}^{n_{\ell}}-g_1^{m_1}\cdots g_{r}^{m_r}\).$$
Let $\Mult=\max_{t\in S_{\vec{n},\vec{m}}} e_t$ be the largest multiplicity of the zeros of $f_1^{n_1}\cdots f_{\ell}^{n_{\ell}}-g_1^{m_1}\cdots g_{r}^{m_r}$.
The bound for $\Mult$ follows immediately from the polynomial $ABC$ theorem (proved first by Stothers [@St], and then independently by Mason [@Mas] and Silverman [@Si84]). Indeed, we apply the polynomial $ABC$ theorem with $A=a_{\vec{n},\vec{m}}\prod_{t\in S_{\vec{n},\vec{m}}}(T-t)^{e_t}$, $B=f_1^{n_1}\cdots f_{\ell}^{n_{\ell}}$ and $C=g_1^{m_1}\cdots g_{r}^{m_r}$, which are pairwise corpime. We get $$\label{eq:l}
\sum_{t\in S_{\vec{n},\vec{m}}}e_t\le \sum_{i=1}^{\ell}d_{f_i}+\sum_{j=1}^rd_{g_j}+\#S_{\vec{n},\vec{m}}-1.$$ Taking into account that $$\sum_{t\in S_{\vec{n},\vec{m}}}e_t\ge \Mult +\#S_{\vec{n},\vec{m}}-1,$$ from we obtain $$\Mult\le \sum_{i=1}^{\ell}d_{f_i}+\sum_{j=1}^rd_{g_j},$$ which concludes the proof.
We present now a similar result for rational functions.
\[cor:ABCrat\] Let $h_1,\ldots,h_{\ell}\in\C(T)$, $h_i=f_i/g_i$, $f_i,g_i\in\C[T]$, $i=1,\ldots,\ell$, with $Z(f_1\cdots f_{\ell})\cap Z(g_1\cdots g_{\ell})=\emptyset$. Then, for all $n_1,\ldots,n_{\ell}\ge 0$, we have $$\Mult\(h_1^{n_1}\cdots h_{\ell}^{n_{\ell}}-1\)\le \sum_{i=1}^{\ell} \(\deg f_i+\deg g_i\).$$
We note that $Z\(h_1^{n_1}\cdots h_{\ell}^{n_{\ell}}-1\)=Z\(f_1^{n_1}\cdots f_{\ell}^{n_{\ell}}-g_1^{n_1}\cdots g_{\ell}^{n_{\ell}}\)$. The result now follows directly from Lemma \[lem:ABCpoly\] applied with $r=\ell$ and $m_i=n_i$, $i=1,\ldots,\ell$.
Algebraic dependence
--------------------
We need the following result [@Plo Theorem 1.1] which gives a degree bound for the annihilating polynomial of algebraically dependent polynomials, which is always the case when the number of polynomials exceeds the number of variables. The result holds over any field, but we present it only over $\C$.
\[lem:perron\] Let $F_1,\ldots,F_{\ell+1}\in\C[X_1,\ldots,X_{\ell}]$ be of degree at most $D$. Then there exists a nonzero polynomial $R\in\C[Z_1,\ldots,Z_{\ell+1}]$ of degree at most $D^{\ell}$ such that $R(F_1,\ldots,F_{\ell+1})=0$.
Proofs of Main Results {#sec:main}
======================
Proof of Theorem \[thm:genAR1\] {#sec:univar1}
-------------------------------
We use the same idea as in the proof of [@AR Theorem 1] and Lemma \[lem:univAR\]. Indeed, we write the factorisation in linear factors, $$h_1=\prod_{i=1}^{d_{h_1}}(T-\omega_{1,i}),\quad h_2=\prod_{i=1}^{d_{h_2}}(T-\omega_{2,i}),$$ where $\omega_{1,i},\omega_{2,j}\in\C$, $i=1,\ldots,d_{h_1}$, $j=1,\ldots,d_{h_2}$.
Thus, we reduce the problem to estimating the degree of each $$\gcd \(f^n-\omega_{1,i},g^m-\omega_{2,j}\).$$ For simplicity we use the notation $\omega_1$ and $\omega_2$ for any two roots of $h_1$ and $h_2$, respectively, and we denote $$\cD_{n,m}(\omega_1,\omega_2)=\gcd \(f^n-\omega_{1},g^m-\omega_{2}\).$$
Thus, as in Lemma \[lem:univAR\], we need to bound the number of $t\in\C$ such that $f(t)^n=\omega_1$ and $g(t)^m=\omega_2$ for some positive integers $n$ and $m$.
For every $n,m\ge 1$, we fix an element $t_{n,m}\in\C$ such that $$\label{eq:root}
f(t_{n,m})^n=\omega_1,\quad g(t_{n,m})^m=\omega_2$$ (if no such $t_{n,m}$ exists then we immediately have $\deg \cD_{n,m}(\omega_1,\omega_2) =0$). We define new polynomials $$\widetilde{f}_{n,m}(T)=\frac{1}{f(t_{n,m})}f(T)\mand
\widetilde{g}_{n,m}(T)=\frac{1}{g(t_{n,m})}g(T).$$ As $f$ and $g$ are multiplicatively independent in $\C(T)^*/\C^*$, we obtain that $\widetilde{f}_{n,m}$ and $\widetilde{g}_{n,m}$ are multiplicatively independent for every $n,m$.
Thus, we can apply Lemma \[lem:univAR\] and conclude that $$\deg \gcd\(\widetilde{f}_{n,m}^n-1,\widetilde{g}_{n,m}^m-1\)\le \(11(d_f+d_g)^{2}\)^{\min(d_f,d_g)}.$$
From and the definition of $\widetilde{f}_{n,m}$ and $\widetilde{g}_{n,m}$, we have $$\deg \cD_{n,m}(\omega_1,\omega_2)=\deg \gcd\(\widetilde{f}_{n,m}^n-1,\widetilde{g}_{n,m}^m-1\),$$ and thus, for every $n,m\ge 1$, we get $$\deg\cD_{n,m}(\omega_1,\omega_2)\le \(11(d_f+d_g)^{2}\)^{\min(d_f,d_g)}.$$ As this holds for any roots $\omega_1,\omega_2$ of $h_1$ and $h_2$, respectively, we get $$\deg \gcd\(h_1\(f^n\),h_2\(g^m\)\)
\le d_{h_1}d_{h_2} \(11(d_f+d_g)^{2}\)^{\min(d_f,d_g)}$$ which concludes the proof.
Proof of Theorem \[thm:genAR2\] {#sec:univar2}
-------------------------------
We use the same idea as in the proof of [@AR Theorem1] combined with Lemma \[lem:bmz\].
First, we note that for any zero $t\in \C$ of $$\gcd\(\prod_{i=1}^{\ell}f_i^{n_i}-\prod_{i=1}^k\varphi_i^{\nu_i},\prod_{i=1}^rg_i^{m_i}-\prod_{i=1}^s\psi_i^{\mu_i}\)$$ the condition ensures that $\varphi_i(t),\psi_j(t)\ne0$, $i=1,\ldots,l$, $j=1,\ldots,k$. Therefore each such zero $t$ satisfies $$\label{eq:ffgg}
\prod_{i=1}^{\ell}f_i(t)^{n_i}\cdot\prod_{i=1}^k\varphi_i(t)^{-\nu_i}=\prod_{i=1}^rg_i(t)^{m_i}\cdot\prod_{i=1}^s\psi_i(t)^{-\mu_i}=1.$$
We apply Lemma \[lem:bmz\] with $k$ replaced by $L={\ell}+k+r+s$ and with the curve $$\begin{split}
C= \bigl\{(f_1(t),\ldots,f_{\ell}(t),\varphi_1(t)&,\ldots,\varphi_{k}(t),g_1(t),\ldots,g_r(t),\\
&\psi_1(t),\ldots,\psi_s(t))~:~ t\in\C\bigr\} \subseteq\G_m^{L}.
\end{split}$$ Indeed, we denote $$\vec{v}=(n_1,\ldots,n_{\ell},-\nu_1,\ldots,-\nu_k),\ \vec{w}=(m_1,\ldots,m_r,-\mu_1,\ldots,-\mu_s).$$ As the vectors $$(\vec{v},\vec{0}),\ (\vec{0},\vec{w})\in \Z^L$$ are linearly independent, by Lemma \[lem:bmz\] we obtain that there are finitely many $t\in\C$ such that holds for some vectors $\vec{v},\vec{w}$ as above.
We denote by $S$ the set of such $t\in\C$. For $\vec{v},\vec{w}$, we denote $$\cD_{\vec{v},\vec{w}}=\gcd\(\prod_{i=1}^{\ell}f_i^{n_i}-\prod_{i=1}^k\varphi_i^{\nu_i},\prod_{i=1}^rg_i^{m_i}-\prod_{i=1}^s\psi_i^{\mu_i}\).$$ We see from the above that set of zeros $Z(\cD_{\vec{v},\vec{w}})$ belongs to some fixed set that depends only on the above curve $C$ and thus only on the polynomials in the initial data. To give the upper bound for $\deg \cD_{\vec{v},\vec{w}}$ we only need to prove that the multiplicity of the roots $t\in S$ of $\cD_{\vec{v},\vec{w}}$ can be bounded uniformly for all integer vectors $\vec{v},\vec{w}$. This is given by Lemma \[lem:ABCpoly\] applied with the polynomials $\prod_{i=1}^{\ell}f_i^{n_i}-\prod_{i=1}^k\varphi_i^{\nu_i}$ and $\prod_{i=1}^rg_i^{m_i}-\prod_{i=1}^s\psi_i^{\mu_i}$.
Indeed, if we denote by $\Mult_1$ and $\Mult_2$ the largest multiplicity of roots in $S$ of the first and second polynomials, respectively, we get $$\Mult_1\le \sum_{i=1}^{\ell} d_{f_i}+\sum_{i=1}^kd_{\varphi_i},\quad \Mult_2\le \sum_{i=1}^{r} d_{g_i}+\sum_{i=1}^sd_{\psi_i}.$$
Thus, there exists a polynomial $h\in\C[T]$ defined by $$h=\prod_{t\in S} (T-t)^{d},\quad d=\min\(\sum_{i=1}^{\ell} d_{f_i}+\sum_{i=1}^kd_{\varphi_i},\sum_{i=1}^{r} d_{g_i}+\sum_{i=1}^sd_{\psi_i}\)$$ such that $\cD_{\vec{v},\vec{w}}\mid h$ for every vectors $\vec{v},\vec{w}$ as above. This concludes the proof of Part [**[i]{}**]{}.
For Part [**[ii]{}**]{}, for each $t\in S$, let $$\cL_t=\{\(\vec{v},\vec{w}\)\in \N^L
~:~(T-t)\mid \cD_{\vec{v},\vec{w}}\}.$$ We note that $\cL_s$ is actually a monoid as the sum of any two elements in $\cL_t$ is also an element of $\cL_t$. As the set $S$ is finite, there are finitely many such monoids $\cL_t$, $t\in S$, such that $\deg \cD_{\vec{v},\vec{w}}\ge 1$ for any $\(\vec{v},\vec{w}\)\in\cL_t$.
We are left to show that $\cup_{t\in S}\cL_t$ is not the entire space $\N^{L}$. Indeed this follows directly from [@AR Theorem 1] as for the diagonal case, that is $\vec{v}=n(\underbrace{1,\ldots,1}_{\ell},0,\ldots,0)\in\N^{\ell+k}$ and $\vec{w}=n(\underbrace{1,\ldots,1}_{r},0,\ldots,0)\in\N^{r+s}$, we have $$\gcd\(\(f_1\cdots f_{\ell}\)^n-1,\(g_1\cdots g_r\)^n-1\)=1$$ infinitely often.
Thus, for any $\(\vec{v},\vec{w}\)$ outside $\cup_{t\in S}\cL_t$, we have $\cD_{\vec{v},\vec{w}}=1$, and we conclude the proof.
Proof of Theorem \[thm:multivAR\] {#sec:multivar1}
---------------------------------
The idea of the proof lies in applying Hilbert’s irreducibility theorem, and in particular Lemma \[lem:multindep\], to reduce via specialisations to the univariate case and thus use Theorem \[thm:genAR1\].
We denote $d=D+1$, that is $d>\deg_{X_j}F,\deg_{X_j}G$ for any $j=1,\ldots,{\ell}$. We define the polynomials $$\begin{split}
&\widetilde{F}(X_1,\ldots,X_{\ell})=F\(X_1,X_2+X_1^d,\ldots,X_{\ell}+X_1^{d^{{\ell}-1}}\),\\
&\widetilde{G}(X_1,\ldots,X_{\ell})=G\(X_1,X_2+X_1^d,\ldots,X_{\ell}+X_1^{d^{{\ell}-1}}\).
\end{split}$$
The polynomials $\widetilde{F},\widetilde{G}$ have the property that $$\deg \widetilde{F},\deg \widetilde{G}\le D\frac{d^{{\ell}}-1}{d-1}<(D+1)^{\ell}$$ and $$\deg \widetilde{F}(X_1,\alpha_2,\ldots,\alpha_{\ell})=\deg \widetilde{F},\ \deg \widetilde{G}(X_1,\alpha_2,\ldots,\alpha_{\ell})=\deg \widetilde{G}$$ for any specialisation $(\alpha_2,\ldots,\alpha_{\ell})\in\C^{{\ell}-1}$.
Moreover, we note that the polynomials $\widetilde{F}, \widetilde{G}$ are also multiplicatively independent in $\C(X_1,\ldots,X_{\ell})^*/\C^*$. Indeed, if this would not be the case, then there exist $i_1,i_2$ not both zero and $a\in\C$ such that $$\widetilde{F}^{i_1}\widetilde{G}^{i_2}=a.$$ Composing this polynomial identity with the polynomial automorphism $$\label{eq:Kro}
(X_1,\ldots,X_{\ell})\to \(X_1,X_2-X_1^d,\ldots,X_{\ell}-X_1^{d^{{\ell}-1}}\)$$ we obtain that the polynomials $F,G$ are multiplicatively dependent in $\C(X_1,\ldots,X_{\ell})^*/\C^*$ and thus we get a contradiction.
Let $\K$ be the finite extension of $\Q$ by the coefficients of the polynomials $F,G$. By the Hilbert’s irreducibility theorem, see [@Sch Theorem 46], $\K$ is a Hilbertian field. We apply now Lemma \[lem:multindep\] with the polynomials $\widetilde{F},\widetilde{G}$, and thus there exists a specialisation $(\alpha_2,\ldots,\alpha_{\ell})\in\K^{{\ell}-1}$ such that $\widetilde{F}(X_1,\alpha_2,\ldots,\alpha_{\ell})$ and $\widetilde{G}(X_1,\alpha_2,\ldots,\alpha_{\ell})$ are multiplicatively independent in $\C(X_1)^*/\C^*$. For simplicity, we denote $f=\widetilde{F}(X_1,\alpha_2,\ldots,\alpha_{\ell})$ and $g=\widetilde{G}(X_1,\alpha_2,\ldots,\alpha_{\ell})$.
We denote $\cD_{n,m}=\gcd\(h_1\(F^n\),h_2\(G^m\)\)$. Moreover, we note that $$\cD_{n,m}\(X_1,X_2+X_1^d,\ldots,X_{\ell}+X_1^{d^{{\ell}-1}}\)=\gcd\(h_1\(\widetilde{F}^n\),h_2\(\widetilde{G}^m\)\).$$ We denote $E_{n,m}=\gcd\(h_1\(\widetilde{F}^n\),h_2\(\widetilde{G}^m\)\)$, and for the specialisation $(\alpha_2,\ldots,\alpha_{\ell})$ one has $$E_{n,m}(X_1,\alpha_2,\ldots,\alpha_{\ell})\mid \gcd\(h_1\(f^n\),h_2\(g^m\)\).$$ In particular, we have $$\deg \cD_{n,m}\le \deg E_{n,m}\le \deg \gcd\(h_1\(f^n\),h_2\(g^m\)\).$$ We make here the remark that using the automorphism was essential to have these degree inequalities, as if one just uses Hilbert’s irreducibility theorem applied directly with the polynomials $F$ and $G$, we cannot guarantee that when we make specialisations we get that $\deg \cD_{n,m}\le \deg \gcd\(h_1\(f^n\),h_2\(g^m\)\)$.
We apply now Theorem \[thm:genAR1\] and using the fact that $\deg f,\deg g< (D+1)^{\ell}$ we conclude that $$\deg \gcd\(h_1\(f^n\),h_2\(g^m\)\)\le d_{h_1} d_{h_2} (44(D+1)^{2\ell})^{(D+1)^{\ell}},$$ which finishes the proof.
Proof of Theorem \[thm:CommonZeros\] {#sec:multivar2}
------------------------------------
We define $$\cH=\{(F_1({{\boldsymbol{\alpha}}}),\ldots,F_{\ell+1}({{\boldsymbol{\alpha}}}))\mid {{\boldsymbol{\alpha}}}\in\C^{\ell}\}.$$ By Lemma \[lem:perron\] there exists a polynomial $R\in\C[Z_1,\ldots,Z_{\ell+1}]$ of degree at most $D^{\ell}$ such that $R(F_1,\ldots,F_{\ell+1})=0$. In other words, any point of $\cH$ is a point on the hypersurface defined by the zero set of $R$ in $\C^{{\ell+1}}$. In particular, any point ${{\boldsymbol{\alpha}}}\in\C^{\ell}$ such that $F_i({{\boldsymbol{\alpha}}})^{n_i}=1$, $i=1,\ldots,\ell+1$ gives a point on the hypersurface defined by the zero set of $R$ with coordinates roots of unity.
From Lemma \[lem:GranRud\] we get that there are at most $$N \le N(R) \le (0.792s(R)/\log \(s(R)+1\))^{s(R)}$$ algebraic subgroups, each defined by the zero set of at most $\ell+1$ polynomials of the form $$Z_1^{b_{j,1}}Z_2^{b_{j,2}}\cdots Z_{\ell+1}^{b_{j,\ell+1}}-1\in
\C[Z_1, \ldots Z_{\ell+1}]$$ with $$\sum_{i=1}^{\ell+1}b_{j,i}\le (\ell+1)D^{\ell}\prod_{p\le s(R)}p,
\qquad j = 1, \ldots, \ell+1,$$ where the product runs over all primes $p\le s(R)$, that contain all the points in $Z(R)$ with coordinates roots of unity. In particular, all points $\(F_1({{\boldsymbol{\alpha}}}),\ldots,F_{\ell+1}({{\boldsymbol{\alpha}}})\)$ such that $F_i({{\boldsymbol{\alpha}}})^{n_i}=1$, $i=1,\ldots,\ell+1$, lie in these algebraic subgroups. It remains to estimate $s(R)$.
As $R$ is a polynomial in $\ell+1$ variables and $\deg R\le D^{\ell}$, we have that $s(R)\le \gamma_{\ell}(D)$, where $ \gamma_{\ell}(D)$ is defined by .
Thus, the points ${{\boldsymbol{\alpha}}}$ such that $F_i({{\boldsymbol{\alpha}}})^{n_i}=1$, $i=1,\ldots,\ell+1$, lie in at most $N(H)$ algebraic varieties, each defined by at most $\ell+1$ polynomials of the form $F_1^{b_{j,1}}F_2^{b_{j,2}}\cdots F_{\ell+1}^{b_{j,\ell+1}}-1$ (note that these polynomials are non constant since $F_1,\ldots,F_{\ell+1}$ are multiplicatively independent) of degree at most $$\sum_{i=1}^{\ell+1}b_{j,i} \deg F_i
\le (\ell+1)D^{\ell+1}\prod_{p\le \gamma_{\ell}(D)}p,\qquad j = 1, \ldots, \ell+1,$$ where the product runs over all primes $p\le \gamma_{\ell}(D)$.
Final Comments and Questions
============================
Extensions over $\C$
--------------------
Lemma \[lem:bmz\] gives only the finiteness of the intersection of curves in $\G_m^{\ell}$ with algebraic subgroups. As already mentioned after Theorem \[thm:genAR2\], it is of high interest to have available uniform bounds for the size of this intersection. This implies uniform bounds on the degree of $h$ in Theorem \[thm:genAR2\].
More generally, one can ask for the number of solutions to $f(x,y) = 0$, with $x^n , y^m \in S$ for some nonzero integers $n$ and $m$, where $S$ is the group of $S$-units of some fixed number field. This leads again to further generalisations.
It is certainly interesting to obtain a similar result as Theorem \[thm:multivAR\] for $$\gcd\(H_1\(F_1^{n_1},\ldots,F_{s}^{n_{s}}\),H_2\(G_1^{m_1},\ldots,G_{r}^{m_{r}}\)\),$$ with polynomials $H_1\in\C[Y_1,\ldots,Y_s]$, $H_2\in\C[Z_1,\ldots,Z_r]$ and also $F_1,\ldots,F_s,G_1,\ldots,G_r\in\C[X_1,\ldots,X_{\ell}]$.
If one chooses $$H_1=Y_1\cdots Y_s-1,\quad H_2=Z_1\cdots Z_r-1,$$ then following the same proof as for Theorem \[thm:multivAR\], we reduce (via specialisations) the problem to Theorem \[thm:genAR2\], and thus get that $$\label{eq:degmultiv}
\deg \gcd\(H_1\(F_1^{n_1},\ldots,F_{s}^{n_{s}}\),H_2\(G_1^{m_1},\ldots,G_{r}^{m_{r}}\)\)$$ is bounded by a constant depending only on $F_1,\ldots,F_s,G_1,\ldots,G_r$.
However, the approach of Theorem \[thm:multivAR\] does not seem to work for more general multivariate polynomials $H_1,H_2$.
Dynamical analogues
-------------------
Another interesting direction of research is obtaining dynamical analogues of the results of Ailon-Rudnick [@AR] and Silverman [@Si]. That is, investigating the greatest common divisors of polynomials iterates.
More precisely, let $\K$ be a field and $f,g\in\K[X]$. We define $$f^{(0)}=T,\quad f^{(n)}=f(f^{(n-1)}),\quad n\ge 1,$$ and similarly for $g$.
Give, under some natural conditions, an upper bound for $$\deg\gcd\(f^{(n)},g^{(m)}\).$$
\[prob:gcd 1\] Show, under some natural conditions, that the iterates of $f$ and $g$ are coprime for infinitely many $n,m$.
We note that some conditions on $f,g$ are certainly needed in Problem \[prob:gcd 1\] as, for example, if $f$ and $g$ have $0$ as fixed point, that is, $f(0) = g(0) = 0$, then $f^{(n)}$ and $g^{(m)}$ are never coprime.
We note that there are many results regarding the arithmetic structure of polynomial iterates. For example in [@GNOS; @GOS; @Jon; @JB] and references therein, results regarding the irreducibility of iterates are given. Irreducible polynomials $f\in\Q[X]$ such that all the iterates $f^{(n)}$, $n\ge 1$, remain irreducible are called [*stable polynomials*]{}. For quadratic polynomials the stability is given by the presence of squares in the orbit of the critical point of the polynomial. Thus, if $f,g\in\K[X]$ are stable, then $f^{(n)}$ and $g^{(m)}$ are coprime for every $n,m\ge 1$.
For $h_1,h_2\in\K[X]$, one can also consider the more general case $$G_{n,m}=\gcd\(h_1\(f^{(n)}\),h_2\(g^{(m)}\)\).$$
We note that, following the ideas of [@AR Theorem 1] and of this paper, bounding the zeros of $G_{n,m}$ reduces to proving the finiteness (or even finding uniform bounds) of the number of $t\in\C$ such that $(f^{(n)}(t),g^{(m)}(t))\in V$, where $V$ is the set of zeros of $\{h_1(X_1),h_2(X_2)\}$.
This naturally leads to the question of counting the occurrences $$\(f^{(n)}(t),g^{(m)}(t)\) \in V, \qquad (n, m, t) \in [1, N] \times [1,N] \times \C,$$ for an arbitrary variety $V \subseteq \C^2$ and a sufficiently large integer $N\ge 1$.
For a fixed $t$ and the diagonal case $n=m$, this is of the same flavour as the uniform [*dynamical Mordell–Lang conjecture*]{}, which, for a fixed $(t_1,t_2)\in\C^2$ asserts that the iterates $n,m\ge 1$ such that $\(f^{(n)}(t_1),g^{(m)}(t_2)\)\in V$, see [@BGKT; @GT; @GTZ1] and references therein, lie in finitely many arithmetic progressions (which number does not depend on $t_1,t_2$).
Acknowledgements {#acknowledgements .unnumbered}
================
The author is very grateful to Joseph Silverman for drawing the attention on the Ailon-Rudnick Theorem and related results. The author would also like to thank Igor Shparlinski, Joseph Silverman, Thomas Tucker and Umberto Zannier for their valuable suggestions and stimulating discussions, and also for their comments on an early version of the paper. The research of A. O. was supported by the UNSW Vice Chancellor’s Fellowship.
[99]{} N. Ailon and Z. Rudnick, ‘Torsion points on curves and common divisors of $a^k-1$ and $b^k-1$’, [*Acta Arith.*]{}, [**113**]{} (2004), no. 1, 31–38.
I. Aliev and C. Smyth, ‘Solving algebraic equations in roots of unity’, [*Forum Math.*]{}, [**24**]{} (2012), 641–665.
R. Benedetto, D. Ghioca, P. Kurlberg and T. Tucker, ‘A case of the dynamical Mordell-Lang conjecture’, *Math. Ann.*, **352** (2012), 1–26.
D. Berend and T. Tassa, ‘Improved bounds on Bell numbers and on moments of sums of random variables’, *Probability and Mathematical Statistics*, **30** (2010), 185–205.
F. Beukers and C. J. Smyth, ‘Cyclotomic points on curves’, [*Number Theory for the Millenium*]{} (Urbana, Illinois, 2000), I, A K Peters, 2002, 67–85.
E. Bombieri, D. Masser and U. Zannier, ‘Intersecting a curve with algebraic subgroups of multiplicative groups’, [*Int. Math. Res. Not.*]{}, [**20**]{} (1999), 1119–1140.
E. Bombieri, D. Masser and U. Zannier, ‘On unlikely intersections of complex varieties with tori’, [*Acta Arith.*]{}, [**133**]{} (2008), 309–323.
E. Bombieri and U. Zannier, ‘Algebraic points on subvarieties of $\G_m^n$’, [*Internat. Math. Res. Notices*]{}, [**7**]{} (1995), 333–347.
Y. Bugeaud, P. Corvaja and U. Zannier, ‘An upper bound for the G.C.D. of $a^n-1$ and $b^n-1$’, [*Math. Z.*]{}, [**243**]{} (2003), 79–84.
P. Corvaja and U. Zannier, ‘A lower bound for the height of a rational function at $S$-unit points’, [*Monatsh. Math.*]{}, [**144**]{} (2005), 203–224.
P. Corvaja and U. Zannier, ‘On the maximal order of a torsion point on a curve in $\G_m^n$’, [*Rend. Lincei Mat. Appl.*]{}, [**19**]{} (2008), 73–78.
P. Corvaja, U. Zannier, ‘Some cases of VojtaÕs conjecture on integral points over function fields’, [*J. Alg. Geometry*]{}, [**17**]{} (2008), 295–333.
P. Corvaja and U. Zannier, ‘Greatest common divisors of $u-1$, $v-1$ in positive characteristic and rational points on curves over finite fields’, [*J. Eur. Math. Soc.*]{} (JEMS), [**15**]{} (2013), 1927–1942.
J.-H. Evertse, ‘The number of solutions of linear equations in roots of unity’, [*Acta Arith.*]{}, [**89**]{} (1999), 45–51.
L. Denis, ‘Facteurs communs et torsion en caract' eristique non nulle’, [*J. ThŽor. Nombres Bordeaux*]{}, [**23**]{} (2011), 347–352.
D. Ghioca and T.J. Tucker, ‘Periodic points, linearizing maps, and the dynamical Mordell-Lang problem’, *J. Number Theory*, **129** (2009), 1392–1403.
D. Ghioca, T. Tucker and M. Zieve, ‘Intersections of polynomial orbits, and a dynamical Mordell-Lang conjecture’, [*Invent. Math.*]{}, [**171**]{} (2008), 463–483.
D. Gomez-Perez, A. P. Nicol[á]{}s, A. Ostafe and D. Sadornil, ‘On the length of critical orbits of stable arbitrary polynomials over finite fields’, [*Rev. Matem. Iberoamer.*]{}, [**30**]{} (2014), 523-535.
D. Gomez-Perez, A. Ostafe and I. Shparlinski, ‘On irreducible divisors of iterated polynomials’, [*Rev. Matem. Iberoamer.*]{}, [**30**]{} (2014), 1123-1134.
A. Granville and Z. Rudnick, ‘Torsion points on curves’, [*NATO Sci. Ser. II Math. Phys. Chem.*]{}, [**237**]{} (2007), 85–92.
R. Jones, ‘The density of prime divisors in the arithmetic dynamics of quadratic polynomials’, [*J. Lond. Math. Soc.*]{}, [**78**]{} (2008), 523-544.
R. Jones and N. Boston, ‘Settled polynomials over finite fields’, [*Proc. Amer. Math. Soc.*]{}, [**140**]{} (2012), 1849-1863.
S. Lang, [*Fundamentals of Diophantine geometry*]{}, Springer, 1983.
R. C. Mason, [*Diophantine equations over function fields*]{}, London Mathematical Society Lecture Note Series [**96**]{} (1984), Cambridge University Press.
D. Masser, ‘Specializations of finitely generated subgroups of abelian varieties’, [*Trans. Amer. Math. Soc.*]{}, [**311**]{} (1989), 413–424.
D. Masser, ‘Unlikely intersections for curves in multiplicative groups over positive characteristic’, [*Quart. J. Math.*]{}, [**65**]{} (2014), 505–515.
G. Maurin, ‘Courbes alg' ebriques et ' equations multiplicatives’, [*Math. Ann.*]{}, [**341**]{} (2008), 789–824.
A. N' eron, ‘Probl\` emes arithm' etiques et g' eom' etriques rattach' es \` a la notion de rang d’une courbe alg' ebrique dans un corps’, [*Bull. Soc. Math. France*]{}, [**80**]{} (1952), 101–166.
A. Płoski, ‘Algebraic dependence of polynomials after O. Perron and some applications’, [*Computational Commutative and Non-Commutative Algebraic Geometry*]{}, IOS Press, 2005, 167–173.
A. Schinzel, [*Polynomials with special regard to reducibility. Appendix by Umberto Zannier*]{}, Encyclopedia of Mathematics and its Applications 77, Cambridge University Press, 2000.
H. P. Schlickewei, ‘Equations in roots of unity’, [*Acta Arith.*]{}, [**76**]{} (1996), 99–108.
J.-P. Serre, [*Lectures on the Mordell-Weil theorem*]{}, 2-nd ed., Vieweg, 1990.
J. H. Silverman, ‘Heights and the specialization map for families of abelian varieties’, [*J. Reine Angew. Math.*]{}, [**342**]{} (1983), 197–211.
J. H. Silverman, ‘The S-unit equation over function fields’, [*Proc. Camb. Philos. Soc.*]{}, [**95**]{} (1984), 3–4.
J. H. Silverman, ‘Common divisors of $a^n-1$ and $b^n-1$ over function fields’, [*New York Journal of Math.*]{} (electronic), [**10**]{} (2004), 37–43.
J. Silverman, ‘Common divisors of elliptic divisibility sequences over function fields’, [*Manuscripta Math*]{}, [**114**]{} (2004), 432–446.
J. H. Silverman, ‘Generalized greatest common divisors, divisibility sequences, and Vojta’s conjecture for blowups’, [*Monatsh. Math.*]{}, [**145**]{} (2005), 333–350.
W. W. Stothers, ‘Polynomial identities and Hauptmoduln’, [*Quarterly J. Math. Oxford*]{}, [**32**]{} (1981), 349–370.
U. Zannier, [*Some problems of unlikely intersections in arithmetic and geometry*]{}, Annals of Mathematics Studies 181, Princeton University Press, Princeton, NJ, 2012.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The Snake-in-the-Box problem is that of finding a longest induced path in an $n$-dimensional hypercube. We prove new lower bounds for the values $n\in \{11,12,13\}$. The Coil-in-the-Box problem is that of finding a longest induced cycle in an $n$-dimensional hypercube. We prove new lower bounds for the values $n\in \{10,11,12,13\}$.'
author:
- David Allison
- Daniël Paulusma
title: 'New Bounds for the Snake-in-the-Box Problem'
---
Introduction {#s-intro}
============
In 1958, Kautz [@Ka58] observed that the vertices of an induced path in a hypercube can be used as code words in Gray codes detecting single-bit errors. Applications of such codes date back to the 19th century, where they were used in telegraphy. They are nowadays used in many areas of mathematics, computer science and engineering, such as error correction in digital communication, disk sector encoding, clock domain crossing, computer network topologies and in the design of genetic algorithms, just to name a few; we refer to Drapela [@Dr15] for background information. The effectiveness of a Gray code depends on its length. Hence, the central research question is:
[*How can we find a longest induced path in a hypercube?*]{}
Kautz coined the name “Snake-in-the-Box” for this problem. The Snake-in-the-Box problem has been extensively studied in the literature, just as its variant “Coil-in-the-Box”, where the goal is to determine a longest induced cycle instead of a path. As we shall discuss, the lengths of a longest snake and a longest coil are not known for any hypercube of dimension larger than 8. The problem of finding a longest induced path is [[NP]{}]{}-hard even for bipartite graphs [@GJ79], but its complexity for hypercubes is still unknown. As exhaustive search is too time consuming, a variety of computational and mathematical techniques have been used to find lower bounds for these problems. See Drapela [@Dr15] for details on these techniques.
In this note the main aim is to show three new lower bounds for Snake-in-the-Box and four new lower bounds for Coil-in-the-Box. We do this in Section \[s-new\] after first giving the required terminology in Section \[s-terminology\]. Examples meeting our lower bounds are presented in the Appendix. Our methods are based on the stochastic beam search method, which is a general heuristic search method that was recently used by Meyerson et al. [@MDWP15; @MWDP14] to improve lower bounds for both Snake-in-the-Box and Coil-in-the-Box. We will explain our modifications to their algorithm in full detail in a future paper.
Terminology {#s-terminology}
===========
A graph $H$ is an [*induced subgraph*]{} of another graph $G$ if $H$ can be obtained from $G$ by a sequence of vertex deletions. For $r\geq 1$, the [*path*]{} $P_r$ on $r$ vertices is the graph with vertices $u_1,\ldots,u_r$ and edges $u_1u_2,\ldots,u_{r-1}u_r$. The [*cycle*]{} $C_r$ is the graph obtained from $P_r$ by adding the edge $u_ru_1$. We say that $P_r$ and $C_r$ are of [*length*]{} $r-1$ and $r$, respectively. For $n\geq 1$, the $n$-dimensional hypercube $Q_n$ is the graph that contains $2^n$ vertices, each of which is represented by a binary vector of length $n$, such that two vertices are adjacent if and only if their binary vectors differ by exactly one bit. A [*snake*]{} is an induced subgraph of $Q_n$ that is a path, and a [*coil*]{} is an induced subgraph of $Q_n$ that is a cycle. The [Snake-in-the-Box]{} problem is that of finding a longest snake in $Q_n$ for a given integer $n$ and the [Coil-in-the-Box]{} problem is that of finding a longest coil in $Q_n$ for a given integer $n$. Both problems have also been investigated for other graph classes (see, for example, [@Wo98]).
Known Results {#s-new}
=============
Besides snakes and coils, a number of variants and generalizations, such as symmetric coils [@AAC73] (also known as doubled coils), $k$-snakes and $k$-coils [@HRSW15; @Si66], and single-track circuit codes [@HPB96] have been studied. However, in this section we will only focus on the original notions of snakes and coils and we restrict our overview to exact or lower bounds for [Snake-in-the-Box]{} and [Coil-in-the-Box]{} for hypercubes of small dimensions, that is, we will consider $n$-dimensional hypercubes for $n\leq 20$. We refer to, for example, [@EL00; @Sn94; @Ze97] for upper bounds for these two problems.
For $n\leq 8$, both [Snake-in-the-Box]{} and [Coil-in-the-Box]{} have been solved exactly:
- For $n\leq 5$, Kautz [@Ka58] determined a longest coil. Using exhaustive search, Davies [@Da65] found a longest snake for $n\leq 6$ and a longest coil for $n=6$.
- For $n=7$, Potter, Robinson, Miller and Kochut [@PRMK94] used a genetic algorithm to find a longest snake and Kochut [@Ko96] used exhaustive search to find a longest coil.
- For $n=8$, Österg[å]{}rd and Pettersson used canonical augmentation to find a longest snake [@OP15] and a longest coil [@OP14].
For $n\geq 9$, only lower bounds for the length of a longest snake or coil are known. These lower bounds have been improved over the years and the state of the art results are as follows:
- For $n=9$, Wynn [@Wy12] proved that the length of a longest snake and coil is at least 190 and 188, respectively.
- For $n=10$, Kinny [@Ki12] showed that a longest snake has length at least 370 and Meyerson et al. [@MDWP15] showed that a longest coil has length at least 362.
- For $11\leq n\leq 12$, the results of Meyerson et al. [@MDWP15] from 2015 are known to give the best lower bounds for the lengths of a longest snake and coil.
- For $n=13$, Meyerson et al. [@MDWP15] showed that a longest snake has length at least 2520 and Abbott and Katchalski [@AK91] showed that a longest coil has length at least 2468.
- For $14\leq n\leq 20$, the results of Abbott and Katchalski [@AK91] from 1991 are still unbeaten. The same authors also showed a general lower bound by showing that for each $n\geq 21$, a longest coil has length at least $\frac{77}{256}2^n$.
We refer to Table \[t-survey\] for an overview of the currently best lower bounds on the length of a longest snake or coil for every $n\leq 20$. We note that the Snake-in-the-Box Records page http://ai1.ai.uga.edu/sib/sibwiki/ at the University of Georgia maintains a list of records for the Snake-in-the-Box and Coil-in-the-Box problems for every $n\leq 13$.
Dimension $\;\;$Snake Length $\;\;$Coil Length
----------- --------------------------------------- ------------------------------------
1 $\;$ $1^*$ $\;$ $0^*$
2 $\;$ 2$^*$ $\;$ [@Da65] $\;$ 4$^*$ $\;$ [@Ka58]
3 $\;$ 4$^*$ $\;$ [@Da65] $\;$ $6^*$ $\;$ [@Ka58]
4 $\;$ 7$^*$ $\;$ [@Da65] $\;$ $8^*$ $\;$ [@Ka58]
5 $\;\;$13$^*$ $\;$[@Da65] $\;$ $14^*$ [@Ka58]
6 $\;$ $26^*$$\;$ [@Da65] $\;$ $26^*$ [@Da65]
7 $\;$ 50$^*$ [@PRMK94] $\;$ 48$^*$ [@Ko96]
8 $\;$ 98$^*$ [@OP15] $\;$ 96$^*$ [@OP14]
9 $\;$ 190 [@Wy12] $\;$ 188 [@Wy12]
10 $\;$ 370 [@Ki12] $\;$ [**366**]{} (362 [@MDWP15])
11 $\;$ [**712**]{} $\;$ (707 [@MDWP15]) $\;$ [**692**]{} (668 [@MDWP15])
12 $\;$ [**1373**]{} (1302 [@MDWP15]) $\;$ [**1344**]{} (1276 [@MDWP15])
13 $\;$ [**2687**]{} (2520 [@MDWP15]) $\;$ [**2594**]{} (2468 [@AK91])
14 $\;$ 4932$\;\;\;$ [@AK91] $\;$ 4934$\;\;$ [@AK91]
15 $\;$ 9866$\;\;\;$ [@AK91] $\;$ 9868$\;\;$ [@AK91]
16 $\;$ 19738$\;$ [@AK91] $\;$ 19740 [@AK91]
17 $\;$ 39478$\;$ [@AK91] $\;$ 39480 [@AK91]
18 $\;$ 78958$\;$ [@AK91] $\;$ 78960 [@AK91]
19 $\;$ 157898 [@AK91] $\;$ 157900 [@AK91]
20 $\;$ 315798 [@AK91] $\;$ 315800 [@AK91]
: The lower bounds on the maximum length of a snake or coil in a hypercube of dimension $n=1,\ldots,20$. A \* indicates that the bound is optimal. The unreferenced results in bold are the new bounds proven in this paper; the previous records are placed between parentheses after our lower bound values. Note that the lower bounds on the provided snakes for $n\geq 14$ are deduced from the lower bound on the length of a longest coil.[]{data-label="t-survey"}
Our Results
===========
We prove that the length of a longest snake is at least 712 for $n=11$, at least 1373 for $n=12$ and at least 2687 for $n=13$ and that the length of a longest coil is at least 366 for $n=10$, at least 692 for $n=11$, at least 1344 for $n=12$ and at least 2594 for $n=13$; see also Table \[t-survey\]. We do this by giving examples of snakes and coils of these lengths in the Appendix. We checked the correctness of our solutions using the verifier provided at <http://ai1.ai.uga.edu/sib/sibwiki/doku.php/checker>.
Acknowledgements {#acknowledgements .unnumbered}
================
This work made use of the facilities of the Hamilton HPC Service of Durham University.
[100]{}
H. Abbott and M. Katchalski, On the construction of Snake In The Box codes, Utilitas Mathematica 40 (1991) 97–116.
L.E. Adelson, R. Alter and T.B, Curtz, Long snakes and a characterization of maximal snakes on the $d$-cube, Proc. 4th South-Eastern Conference on Combinatorics, Graph Theory and Computing, Congr. Numer. 8 (1973) 111–124.
D.W. Davies, Longest “separated” paths and loops in an $n$ cube, IEEE Trans. Electronic Computers 14 (1965) 261.
T.E. Drapela, The Snake-in-the-Box Problem: a primer, MSc Thesis, Athens, Georgia, 2015.
P.G. Emelyanova, and A. Lukito, On the maximal length of a snake in hypercubes of small dimension, Discrete Mathematics 218 (2000) 51–59.
M.R.Garey and D.S.Johnson, Computers and Intractability:A Guide to the Theory of [[NP]{}]{}-completeness, Freeman, San Francisco, 1979.
A. P. Hiltgen, K. G. Paterson, and M. Brandestini. Single-track Gray codes, IEEE Transactions on Information Theory 42 (1996) 1555–1561.
S. Hood, D. Recoskie, J. Sawada and D. Wong, Snakes, coils, and single-track circuit codes with spread $k$, Journal of Combinatorial Optimization 30 (2015) 42–62.
W.H. Kautz, Unit-distance error-checking codes, IRE Trans Electronic Computers EC-7 (1958) 179–180.
D. Kinny, A new approach to the Snake-In-The-Box-problem, Proc. ECAI 2012, 462–467.
K.J. Kochut, Snake-In-The-Box codes for dimension 7, Journal of Combinatorial Mathematics and Combinatorial Computing 20 (1996) 175–185.
S.J. Meyerson, T.E. Drapela, W.E. Whiteside and W.D. Potter, Finding longest paths in hypercubes: 11 new lower bounds for snakes, coils, and symmetrical coils, Proc. IEA/AIE 2015, 23–32.
S.J. Meyerson, W.E. Whiteside, T.E. Drapela and W.D. Potter, Finding longest paths in hypercubes, snakes and coils, Proc. IEEE CIES 2014, 103–109.
P.R.J. Österg[å]{}rd and V.H. Pettersson, Exhaustive search for snake-in-the-box codes, Graphs and Combinatorics 31 (2015) 1019–1028.
P.R.J. Österg[å]{}rd and V.H. Pettersson, On the maximum length of coil-in-the-box codes in dimension 8, Discrete Applied Mathematics 179 (2014) 193–200.
W. Potter, J. Robinson, J. Miller and K.J. Kochut, Using the genetic algorithm to find Snake- In-The-Box codes, Proc. IEA/AIE 1994, 421–426.
R.C. Singleton, Generalized snake-in-the-box codes. IEEE Transactions on Electronic Computers (1966) 596–602.
H.S. Snevily, The snake-in-a-box problem: a new upper bound, Discrete Mathematics 133 (1994) 307–314.
J. Wojciechowski, On constructing snakes in powers of complete graphs, Discrete Mathematics 181 (1998) 239–254
E. Wynn, Constructing circuit codes by permuting initial sequences, Manuscript, arXiv:1201.1647.
G. Zémor, An upper bound on the size of the snake-in-the-box, Combinatorica 17 (1997) 287–298.
The New Snakes and Coils
========================
Recall that two vertices in a hypercube are adjacent if and only if they differ in exactly one bit. A path $u_1u_2\cdots u_n$ in a hypercube $Q_n$ can be represented by a [*transition sequence*]{} $i_1,\ldots,i_{n-1}$ where for $j=1,\ldots,n-1$, $i_j$ denotes the vector entry of the bit in which vertices $u_{j-1}$ and $u_j$ differ. Similarly, a cycle in a hypercube can be represented by a transition sequence as well.
Snake of length 712 (Dimension: 11)
-----------------------------------
Snake of length 1373 (Dimension: 12)
------------------------------------
Snake of Length 2687 (Dimension: 13)
------------------------------------
Coil of length 366 (Dimension: 10)
----------------------------------
Coil of length 692 (Dimension: 11)
----------------------------------
A Coil of length 1344 (Dimension: 12)
-------------------------------------
A Coil of length 2594 (Dimension: 13)
-------------------------------------
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The parameters of inspiralling compact binaries can be estimated using matched filtering of gravitational-waveform templates against the output of laser-interferometric gravitational-wave detectors. The estimates are most sensitive to the accuracy with which the phases of the template and signal waveforms match over the many cycles received in the detector frequency bandwidth. Using a recently calculated formula, accurate to second post-Newtonian (2PN) order \[order $(v/c)^4$, where $v$ is the orbital velocity\], for the frequency sweep ($dF/dt$) induced by gravitational radiation damping, we study the statistical errors in the determination of such source parameters as the “chirp mass” $\cal M$, reduced mass $\mu$, and spin parameters $\beta$ and $\sigma$ (related to spin-orbit and spin-spin effects, respectively). We find that previous results using template phasing accurate to 1.5PN order actually underestimated the errors in $\cal M$, $\mu$, and $\beta$. Templates with 2PN phasing yield somewhat larger measurement errors because the 2PN corrections act to suppress slightly the importance of spin-orbit contributions to the phase, thereby increasing the measurement error on $\beta$. This, in turn, results in larger measurement errors on $\cal M$ and $\mu$ because of the strong correlations among the parameters. For two inspiralling neutron stars, the measurement errors increase by less than 16 percent.'
address: |
McDonnell Center for the Space Sciences, Department of Physics,\
Washington University, St. Louis, Missouri 63130
author:
- 'Eric Poisson and Clifford M. Will'
title: |
Gravitational waves from inspiralling compact binaries:\
Parameter estimation using second-post-Newtonian waveforms
---
Introduction
============
Inspiralling compact binary systems, composed of neutron stars and/or black holes, have been identified [@Thorne1987; @Schutz] as the most promising source of gravitational waves for interferometric detectors such as the American LIGO (Laser Interferometer Gravitational-wave Observatory [@LIGO]) and the European Virgo [@Virgo]. These systems evolve under the influence of gravitational radiation reaction, so that the gravitational-wave signal increases in amplitude as its frequency sweeps through the detector frequency bandwidth, from approximately 10 Hz to 1000 Hz. (This characteristic signal is often referred to as a “chirp.”) Inspiralling compact binaries are especially promising because recent estimates [@Narayan; @Phinney] indicate that their event rate could be as large as one hundred per year, for signals detectable out to hundreds of Mpc by the advanced version of LIGO, and because the signal can be accurately predicted using general relativity.
That the signal can be calculated with high accuracy is essential for the measurement of the source parameters [@Cutleretal], which include distance, position in the sky, orientation of the orbital plane, and the masses and spins of the companions. Loosely speaking, the measured signal is passed through a linear filter constructed from the expected signal $h(t;\bbox{\theta})$ and the spectral density of the detector noise [@Wainstein] (see below). The signal-to-noise ratio $S/N (\bbox{\theta})$ is then computed. The expected signal and the signal-to-noise ratio are expressed as functions of the vector $\bbox{\theta}$ which collectively represents the source parameters. The [*actual*]{} value of these parameters, which we denote $\bbox{\tilde{\theta}}$, is unknown prior to the measurement. When $\bbox{\theta} =
\bbox{\tilde{\theta}}$ the linear filter becomes the Wiener optimum filter which is well known to yield the largest possible signal-to-noise ratio [@Wainstein]. The source parameters can therefore be determined by maximizing $S/N(\bbox{\theta})$ over a broad collection of expected signals $h(t;\bbox{\theta})$, loosely referred to as “templates.”
The gravitational-wave signal can be characterized by a growing amplitude and a phase which accumulates nonlinearly with time [@Thorne1987]. The signal undergoes a number $N$ of oscillations, varying from 600 to 16 000 depending on the nature of the system (see below), as the frequency sweeps through the detector bandwidth. It has been established that it is the [*phasing*]{} of the signal which plays the largest role in parameter estimation [@Cutleretal; @FinnChernoff; @CutlerFlanagan]. This is because a slight variation in the parameters can quickly cause $h(t;\bbox{\theta})$ to get out of phase with respect to the true signal $h(t;\bbox{\tilde{\theta}})$, thus seriously reducing $S/N(\bbox{\theta})$ from its maximum possible value. Therefore a good match between the template’s phase and that of the measured signal, throughout the $N$ cycles, singles out, to a large extent, the value of the source parameters. (Clearly, this is only true for those parameters which affect the phasing of the waves, such as the masses and spins of the companions; see below.)
In principle, the gravitational-wave signal from an inspiralling compact binary can be calculated exactly using general relativity (this would require the numerical integration of Einstein’s equations). In practice, however, one must rely on some approximation scheme. It appears appropriate, in this context, to adopt a slow-motion approximation [@Will], and to solve the field equations using a combination of post-Newtonian and post-Minkowskian expansions [@BlanchetDamour]. To date, the waveform has been calculated accurately through order $(v/c)^4$ (where $v$ is the orbital velocity) beyond the leading-order, quadrupole-formula expression [@BDIWW; @BDI; @WW]. Leading-order expressions are referred to as “Newtonian;” the waveform is therefore known to second-post-Newtonian, or 2PN, order.
The detailed expression for the 2PN waveform is complicated: the dependence on the various angles (position of the source in the sky, orientation of the detector, orientation of the polarization axes) is not simple, and the waves have several frequency components given by the harmonics of the orbital frequency (assuming that the orbit is circular [@Peters; @LincolnWill]). For the purpose of this paper, and following Cutler and Flanagan [@CutlerFlanagan], we shall use a simplified expression for the waveform. We shall ignore all post-Newtonian corrections to the wave’s [*amplitude*]{}, and single out its dominant frequency component at twice the orbital frequency [@Thorne1987]. Thus, setting $G=c=1$, $$h(t;\bbox{\theta}) = r^{-1} Q(\mbox{angles})
{\cal M} (\pi {\cal M} F)^{2/3} \cos \Phi(t).
\label{1.1}$$ Here, $r$ is the distance to the source, $Q$ a function of the various angles mentioned above, $F(t)$ the gravitational-wave frequency, and $\Phi(t) = \int 2\pi F(t) dt$ the phase. We have also introduced the (so-called) [*chirp mass*]{} $\cal M$: If $\mu = m_1 m_2 / (m_1+m_2)$ is the reduced mass and $M=m_1+m_2$ the total mass, then $${\cal M} = \eta^{3/5} M, \qquad
\eta = \mu/M.
\label{1.2}$$
In Eq. (\[1.1\]) we use the most accurate expression available for the phase function $\Phi(t)$. It is determined by the 2PN expression for the frequency sweep [@BDIWW], $$\begin{aligned}
\frac{dF}{dt} &=& \frac{96}{5\pi {\cal M}^2} (\pi {\cal M} F)^{11/3}
\biggl[ 1 - \biggl( \frac{743}{336} + \frac{11}{4} \eta \biggr)
(\pi M F)^{2/3} + (4 \pi - \beta) (\pi M F)
\nonumber \\ & & \mbox{}
+ \biggl( \frac{34103}{18144} + \frac{13661}{2016}\eta +
\frac{59}{18} \eta^2 + \sigma \biggr) (\pi M F)^{4/3} \biggr].
\label{1.3}\end{aligned}$$ Apart from the parameters introduced previously, $dF/dt$ also depends on $\bbox{\hat{L}}$, the direction of orbital angular momentum, and on $\bbox{S}_1$ and $\bbox{S}_2$, the spin angular momentum of each companion. This dependence is hidden in the “spin-orbit” parameter [@KWW] $$\beta = \frac{1}{12} \sum_{i=1}^2
\bigl[ 113 (m_i/M)^2 + 75\eta \bigr]
\bbox{\hat{L}} \cdot \bbox{\chi}_i,
\label{1.4}$$ where $\bbox{\chi}_i = \bbox{S}_i/{m_i}^2$, and the “spin-spin” parameter [@KWW] $$\sigma = \frac{\eta}{48} \bigl(
-247 \bbox{\chi}_1 \cdot \bbox{\chi}_2 +
721 \bbox{\hat{L}} \cdot \bbox{\chi}_1
\bbox{\hat{L}} \cdot \bbox{\chi}_2 \bigr).
\label{1.5}$$
The purpose of this paper is to estimate the [*anticipated*]{} accuracy with which the various parameters (such as $\cal M$, $\eta$, $\beta$, and $\sigma$) can be determined during a gravitational-wave measurement. This analysis differs from that of Cutler and Flanagan [@CutlerFlanagan] in that it incorporates terms of 2PN order into the phasing \[terms of order $(\pi M F)^{4/3}$ in Eq. (\[1.2\])\]; their calculations were accurate only through 1.5PN order (terms of order $\pi M F$). Previous analyses also include Refs. [@FinnChernoff; @KrolakA; @KrolakB; @KrolakC].
Our main conclusion is that 1.5PN phasing [*underestimates*]{} the uncertainty in such parameters as $\cal M$, $\eta$, and $\beta$: 2PN phasing predicts somewhat larger measurement errors. This is true even when no attempt is made to determine the spin-spin parameter $\sigma$. If, however, $\sigma$ is also estimated, then the measurement errors become even larger. This is because the number of estimated parameters has increased with respect to the number contained in the 1.5PN waveform.
An independent analysis by Cutler and Flanagan [@CFprep] shows that 2PN waveforms are not sufficiently accurate for the purpose of parameter estimation: they produce systematic errors which are larger than the statistical errors inherent to the measurement process. This is because the 2PN waveform fails to remain in phase with the true general-relativistic signal, even when the source parameters are exactly matched [@paperII; @TN; @TS]. To construct templates such that the systematic errors will fall below the measurement errors will require an expression for the wave’s phasing accurate through at least 3PN order [@paperII; @TN; @TS]. To achieve such a high degree of accuracy is a major challenge for gravitational-wave theorists.
system A B C
------------ ------------ ------------ -----------
Newtonian 16 050 3 580 600
1PN 439 212 59
tail -208 -180 -51
spin-orbit $17\beta$ $14\beta$ $4\beta$
2PN 9 10 4
spin-spin $-2\sigma$ $-3\sigma$ $-\sigma$
: Contributions to the accumulated number of wave cycles measured in a LIGO/Virgo-type detector. The frequency entering the bandwidth is 10 Hz (seismic limit); the frequency leaving is 1000 Hz (system A) (shot noise), and 360 Hz (system B) and 190 Hz (system C) (innermost circular orbit). The various contributions correspond to various terms in Eq. (1.3). Newtonian: first term within the square brackets; 1PN: second term; tail: $4\pi (\pi M F)$; spin-orbit: $-\beta (\pi M F)$; 2PN: $(\pi M F)^{4/3}$ terms with $\sigma=0$; and spin-spin: $\sigma (\pi M F)^{4/3}$.
We shall consider the following three “canonical” binary systems:
> System A: two neutron stars, with $m_1=m_2=1.4 M_\odot$;
>
> System B: neutron star and black hole, with $m_1=1.4M_\odot$ (the neutron star) and $m_2=10 M_\odot$ (the black hole); and
>
> System C: two black holes, with $m_1=m_2=10 M_\odot$.
For each of these systems Table I summarizes the contribution from each term in Eq. (\[1.3\]) to the total number of gravitational-wave cycles received in a LIGO/Virgo-type detector.
The remainder of the paper is organized as follows. In Sec. II we review the theory of parameter estimation, as developed in previous papers by Finn [@Finn] and Cutler and Flanagan [@CutlerFlanagan]. In Sec. III we carry out the calculations for the waveform (\[1.1\]), and compute the anticipated uncertainty in the measured values of the source parameters. Our results are summarized and discussed in Sec. IV.
Parameter estimation: theory
============================
The theory of detection and measurement of gravitational-wave signals was put on a firm statistical foundation, rather similar to that underlying the theory of radar detection [@Wainstein; @Helstrom]. This was done by various authors, including Finn [@Finn] and Cutler and Flanagan [@CutlerFlanagan]. In this section we review the various aspects of the theory which are relevant for our purposes.
We assume that some criterion has been applied to conclude that a signal originating from an inspiralling compact binary has been received by a network of gravitational-wave detectors. It is therefore known that a signal of the form $h(t;\bbox{\theta})$ has passed through the detectors, and we seek to determine the value of the source parameters $\bbox{\theta}$ and the measurement error $\Delta \bbox{\theta} = \bbox{\theta}
- \bbox{\tilde{\theta}}$, where $\bbox{\tilde{\theta}}$ denotes the true value.
Finn [@Finn] has derived an expression for $p(\bbox{\theta}|s)$, the probability that the gravitational-wave signal is characterized by the parameters $\bbox{\theta}$, [*given*]{} that the detector output is $s(t)$ [*and*]{} that a signal $h(t;\bbox{\theta})$ — for [*any*]{} value of the parameters $\bbox{\theta}$ — is present. The detector output is given by $$s(t) = h(t;\bbox{\theta}) + n(t),
\label{2.1}$$ where $n(t)$ represents the detector noise, assumed to be a stationary, Gaussian random process. Finn shows that $$p(\bbox{\theta}|s) \propto p^{(0)}(\bbox{\theta})
\exp \Bigl[ -{\textstyle \frac{1}{2}}
\Bigl( h(\bbox{\theta}) - s
\Bigm| h(\bbox{\theta}) - s \Bigr) \Bigr],
\label{2.2}$$ where $p^{(0)}(\bbox{\theta})$ is the a priori probability that the signal is characterized by $\bbox{\theta}$ (this represents our prior information regarding the possible value of the parameters) and where the constant of proportionality is independent of $\bbox{\theta}$.
The inner product $(\cdot | \cdot)$ is defined as follows [@CutlerFlanagan]. The statistical properties of the detector noise can be summarized by its autocorrelation function $C_n(\tau) = \langle
n(t) n(t+\tau) \rangle$, where $\langle \cdot \rangle$ denotes a time average. (It is assumed that the noise has zero mean.) The Fourier transform of the autocorrelation function gives the noise spectral density $$S_n(f) = 2 \int_{-\infty}^\infty
C_n(\tau) e^{2\pi i f \tau} d\tau,
\label{2.3}$$ which is defined for $f>0$ only. The inner product is defined so that the probability for the noise $n(t)$ to have a particular realization $n_0(t)$ is given by $p(n=n_0) \propto
\exp[-(n_0|n_0)/2]$. It is given by $$(g | h) = 2 \int_0^\infty
\frac{\tilde{g}^*(f) \tilde{h}(f) +
\tilde{g}(f) \tilde{h}^*(f)}{S_n(f)} df,
\label{2.4}$$ where $\tilde{g}(f)$ is the Fourier transform of $g(t)$, $$\tilde{g}(f) = \int_{-\infty}^\infty
g(t) e^{2\pi i f t} dt;
\label{2.5}$$ an asterisk denotes complex conjugation.
We define $\rho$, the signal-to-noise ratio associated with the measurement, to be the norm of the signal $h(t;\bbox{\theta})$, $$\rho^2 = (h|h) = 4 \int_0^\infty
\frac{ |\tilde{h}(f)|^2 }{S_n(f)} df,
\label{2.6}$$ evaluated at $\bbox{\theta} = \bbox{\tilde{\theta}}$, the true value of the source parameters. This is the largest possible value of the signal-to-noise ratio, for $\tilde{h}^*(f;\bbox{\tilde{\theta}})/S_n(f)$ is just the Fourier transform of the Wiener optimum filter [@Wainstein].
In a given measurement, characterized by the particular detector output $s(t)$, the true value of the source parameters can be estimated by locating the value $\bbox{\hat{\theta}}$ at which the probability distribution function (\[2.2\]) is a maximum. This is the so-called maximum-likelihood estimator [@Wainstein]. In the limit of large signal-to-noise ratio, to which we henceforth specialize, $p(\bbox{\theta}|s)$ will be strongly peaked about that value. We now derive a simplified expression for $p(\bbox{\theta}|s)$ appropriate for this limiting case.
We first assume that $p^{(0)}(\bbox{\theta})$ is nearly uniform near $\bbox{\theta} = \bbox{\hat{\theta}}$. This indicates that the prior information is practically irrelevant to the determination of the source parameters; we shall relax this assumption below. Then, denoting $\xi(\bbox{\theta}) =
(h(\bbox{\theta}) -s|h(\bbox{\theta})-s)$, we have that $\xi$ is minimum at $\bbox{\theta} =
\bbox{\hat{\theta}}$. It follows that this can be expanded as $$\xi(\bbox{\theta}) = \xi(\bbox{\hat{\theta}}) +
{\textstyle \frac{1}{2}}
\xi_{,ab}(\bbox{\hat{\theta}})
\Delta \theta^a \Delta \theta^b + \cdots,
\label{2.7}$$ where “$,a$” denotes partial differentiation with respect to the parameter $\theta^a$, and $\Delta \theta^a = \theta^a - \hat{\theta}^a$; summation over repeated indices is understood. We assume that $\rho$ is sufficiently large that the higher-order terms can be neglected. Calculation yields $\xi_{,ab} = (h_{,ab}|h-s)
+ (h_{,a}|h_{,b})$, and we assume once more that $\rho$ is large enough that the first term can be neglected (see Cutler and Flanagan [@CutlerFlanagan] for details). We arrive at $$p(\bbox{\theta}|s) \propto p^{(0)}(\bbox{\theta})
\exp \Bigl[ - {\textstyle \frac{1}{2}}
\Gamma_{ab} \Delta \theta^a \Delta \theta^b \Bigr],
\label{2.8}$$ where $$\Gamma_{ab} = \bigl( h_{,a} \bigm| h_{,b} \bigr),
\label{2.9}$$ evaluated at $\bbox{\theta} = \bbox{\hat{\theta}}$, is the Fisher information matrix [@Helstrom].
We therefore see that in the limit of large signal-to-noise ratio, $p(\bbox{\theta}|s)$ takes a Gaussian form. From Eq. (\[2.8\]) it can be established that the variance-covariance matrix $\Sigma^{ab}$ is given by $$\Sigma^{ab} \equiv \langle \Delta \theta^a
\Delta \theta^b \rangle = (\bbox{\Gamma}^{-1})^{ab}.
\label{2.10}$$ Here, $\langle \cdot \rangle$ denotes an average over the probability distribution function (\[2.8\]), and $\bbox{\Gamma}^{-1}$ is the inverse of the Fisher matrix. We define the measurement error in the parameter $\theta^a$ to be $$\sigma_a =
\bigl\langle (\Delta \theta^a)^2 \bigr\rangle^{1/2}
= \sqrt{\Sigma^{aa}}
\label{2.11}$$ (no summation over repeated indices), and the correlation coefficient between parameters $\theta^a$ and $\theta^b$ as $$c^{ab} = \frac{\langle \Delta \theta^a
\Delta \theta^b \rangle}{\sigma_a \sigma_b} =
\frac{\Sigma^{ab}}{
\sqrt{\Sigma^{aa} \Sigma^{bb}}};
\label{2.12}$$ by definition each $c^{ab}$ must lie in the range $(-1,1)$.
Cutler and Flanagan [@CutlerFlanagan] have shown that in the limit of large signal-to-noise ratio, Eq. (\[2.8\]) is valid even when $p^{(0)}(\bbox{\theta})$ is not uniform near $\bbox{\theta} = \bbox{\hat{\theta}}$. In such cases the prior information plays an important role in the determination of the source parameters. The exponential factor is still peaked at $\bbox{\theta} = \bbox{\hat{\theta}}$, but $\bbox{\hat{\theta}}$ no longer represents the maximum-likelihood estimate, and the full probability distribution function $p(\bbox{\theta}|s)$ may not be a Gaussian.
For simplicity, and following Cutler and Flanagan [@CutlerFlanagan], we shall restrict attention to cases such that $p^{(0)}(\bbox{\theta})$ is a Gaussian, given by $$p^{(0)} (\bbox{\theta}) \propto
\exp \Bigl[ - {\textstyle \frac{1}{2}}
\Gamma_{ab}^{(0)} \bigl(\theta^a - \bar{\theta}^a\bigr)
\bigl(\theta^b - \bar{\theta}^b\bigr) \Bigl].
\label{2.13}$$ Then $p(\bbox{\theta}|s)$ will [*also*]{} take a Gaussian form, and the new variance-covariance matrix will be given by $$\bbox{\Sigma} = \bigl( \bbox{\Gamma} +
\bbox{\Gamma^{(0)}} \bigr)^{-1}.
\label{2.14}$$ It should be noted that in general, $p(\bbox{\theta}|s)$ will be peaked at a value $\langle \bbox{\theta} \rangle$ which differs both from $\bbox{\hat{\theta}}$ and $\bbox{\bar{\theta}}$.
Parameter estimation: calculations
==================================
We proceed with the calculation of the Fisher information matrix, Eq. (\[2.9\]), for gravitational-wave signals of the form (\[1.1\]), and for gravitational-wave detectors of the LIGO/Virgo type. For such detectors the anticipated noise spectral density can be approximated by the analytic expression [@CutlerFlanagan] $$S_n(f) = {\textstyle \frac{1}{5}} S_0
\bigl[ (f_0/f)^4 + 2 + 2(f/f_0)^2 \bigr],
\label{3.1}$$ where $S_0$ is a normalization constant irrelevant for our purposes, and $f_0$ the frequency at which $S_n(f)$ is minimum; we set $f_0 = 70$Hz, which is appropriate for advanced LIGO sensitivity [@LIGO]. To mimic seismic noise we assume that Eq. (\[3.1\]) is valid for $f > 10 {\rm Hz}$ only, and that $S_n(f) = \infty$ for $f < 10 {\rm Hz}$.
First, we integrate Eq. (\[1.3\]) to obtain expressions for $\Phi(F)$ and $t(F)$, respectively the phase and time as functions of gravitational-wave frequency. (Throughout this section we shall distinguish between $F$, the function of time describing the frequency sweep, and $f$, the Fourier-transform variable.) Expanding in powers of $(\pi M F)^{1/3}$ and truncating all expressions to 2PN order, we obtain $$\begin{aligned}
\Phi(F) &=& \phi_c - \frac{1}{16} (\pi {\cal M} F)^{-5/3}
\biggl[1 + \frac{5}{3} \biggl( \frac{743}{336} +
\frac{11}{4} \eta \biggr) (\pi M F)^{2/3}
- \frac{5}{2} (4\pi - \beta) (\pi M F)
\nonumber \\ & & \mbox{} +
5 \biggl( \frac{3058673}{1016064} + \frac{5429}{1008}\eta
+ \frac{617}{144} \eta^2 - \sigma \biggr) (\pi M F)^{4/3}
\biggr],
\label{3.2}\end{aligned}$$ where $\phi_c$ is (formally) the value of $\Phi$ at $F=\infty$, and $$\begin{aligned}
t(F) &=& t_c - \frac{5}{256} {\cal M} (\pi {\cal M} F)^{-8/3}
\biggl[1 + \frac{4}{3} \biggl( \frac{743}{336} +
\frac{11}{4} \eta \biggr) (\pi M F)^{2/3} -
\frac{8}{5} (4\pi - \beta) (\pi M F)
\nonumber \\ & & \mbox{} +
2 \biggl( \frac{3058673}{1016064} + \frac{5429}{1008}\eta
+ \frac{617}{144} \eta^2 - \sigma \biggr) (\pi M F)^{4/3}
\biggr],
\label{3.3}\end{aligned}$$ where (formally) $t_c = t(\infty)$. Of course, the signal cannot be allowed to reach arbitrarily high frequencies; it must be cut off at a frequency $F=F_i$ corresponding to the end of the inspiral. We put $\pi M F_i = (M/r_i)^{3/2} = 6^{-3/2}$; $r_i = 6M$ is the Schwarzschild radius of the innermost circular orbit for a test mass moving in the gravitational field of a mass $M$ [@foot1].
Next, we take the Fourier transform of Eq. (\[1.1\]) and calculate $\tilde{h}(f) = \int h(t) e^{2\pi i f t} dt$. It is sufficient to estimate $\tilde{h}(f)$ using the stationary phase approximation, according to which [@Jackson] $$\int g(t) e^{i \phi(t)} dt \simeq
\biggl[ \frac{2\pi i}{\phi''(t_0)} \biggr]^{1/2}
g(t_0) e^{i \phi(t_0)}
\label{3.4}$$ if $g(t)$ varies slowly near $t=t_0$ where the phase has a stationary point: $\phi'(t_0)=0$ (a prime denotes differentiation with respect to $t$). Substituting Eqs. (\[1.1\]) and (\[3.2\]) into (\[3.4\]), discarding the irrelevant negative-frequency component, and neglecting all post-Newtonian corrections to the amplitude of $\tilde{h}(f)$, we obtain $$\tilde{h}(f) = {\cal A} f^{-7/6} e^{i \psi(f)},
\label{3.5}$$ where ${\cal A} \propto {\cal M}^{5/6} Q(\mbox{angles})/r$, and $$\begin{aligned}
\psi(f) &=& 2\pi f t_c - \phi_c - \frac{\pi}{4} +
\frac{3}{128} (\pi {\cal M} f)^{-5/3}
\biggl[1 + \frac{20}{9} \biggl( \frac{743}{336} +
\frac{11}{4} \eta \biggr) (\pi M f)^{2/3}
\nonumber \\ & & \mbox{}
- 4(4\pi - \beta) (\pi M f) +
10 \epsilon \biggl( \frac{3058673}{1016064} + \frac{5429}{1008}\eta
+ \frac{617}{144} \eta^2 - \sigma \biggr) (\pi M f)^{4/3}
\biggr].
\label{3.6}\end{aligned}$$ We have introduced the parameter $\epsilon \equiv 1$. This gives us the freedom, for future use, of removing the 2PN terms from $\psi(f)$ by setting $\epsilon=0$.
We now substitute Eq. (\[3.5\]) into (\[2.6\]) and calculate the signal-to-noise ratio. We readily obtain $$\rho^2 = 20 {\cal A}^2 {S_0}^{-1} {f_0}^{-4/3} I(7),
\label{3.7}$$ where the integrals $I(q)$ represent various moments of the noise spectral density: $$I(q) \equiv \int_{1/7}^{x_i}
\frac{x^{-q/3}}{x^{-4} + 2 + 2x^2} dx,
\label{3.8}$$ where $x_i = f_i/f_0 = (6^{3/2} \pi M f_0)^{-1}$ is the frequency cutoff.
As the next step toward the computation of the Fisher matrix, we calculate the derivatives of $\tilde{h}(f)$ with respect to the seven parameters $$\bbox{\theta} = (\ln{\cal A},f_0 t_c, \phi_c, \ln {\cal M},
\ln \eta, \beta, \sigma).
\label{3.9}$$ We obtain $$\begin{aligned}
\tilde{h}_{,1} &=& \tilde{h}, \nonumber \\
\tilde{h}_{,2} &=& 2\pi i (f/f_0) \tilde{h}, \nonumber \\
\tilde{h}_{,3} &=& -i \tilde{h}, \nonumber \\
\tilde{h}_{,4} &=& - \frac{5 i}{128} (\pi {\cal M} f)^{-5/3}
(1 + A_4 v^2 - B_4 v^3 + C_4 v^4) \tilde{h}, \label{3.10} \\
\tilde{h}_{,5} &=& -\frac{i}{96} (\pi {\cal M} f)^{-5/3}
(A_5 v^2 - B_5 v^3 + C_5 v^4) \tilde{h}, \nonumber \\
\tilde{h}_{,6} &=& \frac{3i}{32} \eta^{-3/5}
(\pi {\cal M} f)^{-2/3} \tilde{h}, \nonumber \\
\tilde{h}_{,7} &=& -\frac{15 i}{64} \eta^{-4/5}
(\pi {\cal M} f)^{-1/3} \tilde{h}, \nonumber\end{aligned}$$ where $v\equiv(\pi M f)^{1/3}$. We also have defined $$\begin{aligned}
A_4 &=& \frac{4}{3} \biggl( \frac{743}{336} +
\frac{11}{4} \eta \biggr), \nonumber \\
B_4 &=& \frac{8}{5} (4\pi - \beta),
\label{3.11} \\
C_4 &=& 2 \epsilon \biggl(
\frac{3058673}{1016064} + \frac{5429}{1008} \eta
+ \frac{617}{144} \eta^2 - \sigma \biggr), \nonumber\end{aligned}$$ and $$\begin{aligned}
A_5 &=& \frac{743}{168} - \frac{33}{4} \eta, \nonumber \\
B_5 &=& \frac{27}{5} (4\pi - \beta),
\label{3.12} \\
C_5 &=& 18 \epsilon \biggl(
\frac{3058673}{1016064} - \frac{5429}{4032} \eta
- \frac{617}{96} \eta^2 - \sigma \biggr). \nonumber\end{aligned}$$
Finally, the components of $\bbox{\Gamma}$ can be obtained by evaluating the inner products $(h_{,a}|h_{,b})$ using Eq. (\[2.4\]). The $\Gamma_{ab}$’s can all be expressed in terms of the parameters $\bbox{\theta}$, the signal-to-noise ratio $\rho$, and the integrals $I(q)$. The expressions are too numerous and lengthy to be displayed here. As illustrating examples, we quote $$\Gamma_{1a} = \delta_{1a} \rho^2,
\label{3.13}$$ and $$\begin{aligned}
\Gamma_{46} &=& -\frac{15}{4096} \eta^{-3/5}
(\pi {\cal M} f_0)^{-7/3} \bigl[ J(14) + A_4 J(12)
\nonumber \\ & & \mbox{} -
B_4 J(11) + C_4 J(10) \bigr] \rho^2,
\label{3.14}\end{aligned}$$ where $J(q) \equiv I(q)/I(7)$. We note that even though the $\tilde{h}_{,a}$’s are expressed as truncated post-Newtonian expansions in Eq. (\[3.10\]), they must be treated as [*exact*]{} when computing $\bbox{\Gamma}$. This is to ensure that the eigenvalues of the Fisher matrix are always positive definite.
The variance-covariance matrix $\Sigma^{ab}$ can now be obtained from Eq. (\[2.14\]), and the measurement errors and correlation coefficients computed from Eqs. (\[2.11\]) and (\[2.12\]). Before doing so, however, we must first state our assumptions regarding the prior information available on the source parameters.
We take advantage of the fact that the dimensionless spin parameters, $\bbox{\chi}_1$ and $\bbox{\chi}_2$, must necessarily be smaller than unity. (This upper bound is strict for black holes, but only approximate for neutron stars.) It follows from Eqs. (\[1.4\]) and (\[1.5\]) that $|\beta|$ must be smaller than approximately 8.5, and that $|\sigma|$ must be smaller than approximately 5.0. Following Cutler and Flanagan [@CutlerFlanagan], we crudely incorporate this information into our calculations by taking $$p^{(0)} (\bbox{\theta}) \propto
\exp \bigl[ -{\textstyle \frac{1}{2}} (\beta/8.5)^2
- {\textstyle \frac{1}{2}} (\sigma/5.0)^2 \bigr].
\label{3.15}$$ We consider all other parameters to be unconstrained [@foot2].
Results and discussion
======================
Equation (\[3.12\]) implies that the Fisher matrix is block diagonal. The parameter $\theta^1 = \ln \cal A$ is therefore entirely uncorrelated with the other parameters, and we find $\sigma_1 = \Delta {\cal A} / {\cal A} =
1/\rho$, $c^{1a}=0$, in all cases. We shall no longer be concerned with this parameter.
The results concerning the other parameters are displayed in Tables II and III. All calculations were carried out assuming $\rho=10$, and that the companions are spinless, so that $\beta=\sigma=0$.
[lrrrrrrrrr]{} prior & $\epsilon$ & $\Delta t_c$ & $\Delta \phi_c$ & $\Delta {\cal M} / {\cal M}$ & $\Delta \eta / \eta$ & $\Delta \beta$ & $c^{{\cal M} \eta}$ & $c^{{\cal M} \beta}$ & $c^{\eta\beta}$\
\
\
yes & 1 & 1.07 & 2.94 & 0.036 % & 0.279 & 1.33 & -0.989 & 0.994 & -0.999\
no & 1 & 1.08 & 2.97 & 0.037 % & 0.282 & 1.35 & -0.989 & 0.994 & -0.999\
no & 0 & 1.13 & 4.09 & 0.034 % & 0.243 & 1.24 & -0.988 & 0.993 & -0.999\
no & -1 & 1.16 & 4.96 & 0.032 % & 0.213 & 1.15 & -0.986 & 0.992 & -0.999\
\
\
yes & 1 & 1.72 & 2.27 & 0.218 % & 0.503 & 2.29 & -0.993 & 0.996 & -0.999\
no & 1 & 1.76 & 2.32 & 0.226 % & 0.523 & 2.38 & -0.993 & 0.996 & -0.999\
no & 0 & 2.04 & 6.24 & 0.191 % & 0.386 & 1.99 & -0.990 & 0.994 & -0.999\
no & -1 & 2.20 & 8.68 & 0.171 % & 0.306 & 1.76 & -0.988 & 0.993 & -0.999\
\
\
yes & 1 & 1.50 & 2.19 & 0.54 % & 1.46 & 8.19 & -0.946 & 0.956 & -0.999\
no & 1 & 2.40 & 4.99 & 1.96 % & 5.50 & 30.8 & -0.996 & 0.997 & -0.999\
no & 0 & 3.53 & 9.27 & 1.42 % & 3.16 & 19.5 & -0.992 & 0.994 & -0.999\
no & -1 & 4.01 & 14.7 & 1.21 % & 2.22 & 14.9 & -0.989 & 0.992 & -0.999
To obtain the results of Table II we have estimated only six (including ${\cal A}$) of the seven parameters, leaving $\sigma$ out. In effect, we have truncated the original Fisher matrix to a smaller, $6\times 6$, matrix. This amounts to assuming before measurement that the spin-spin parameter must be very small; equivalently, this assumption can be implemented by taking $p^{(0)}(\sigma)$ to be very strongly peaked at $\sigma=0$.
For each of the three canonical systems, the first line of Table II displays the measurement errors and correlation coefficients as calculated using 2PN phasing ($\epsilon = 1$) and the prior probability distribution function (\[3.15\]). The second line shows the same quantities calculated without utilizing the prior information. We notice that the prior information makes virtually no difference for systems A and B, but is very significant for system C.
The third line of Table II displays the measurement errors and correlation coefficients assuming no prior information and 1.5PN phasing ($\epsilon=0$). Our values agree with those of Cutler and Flanagan [@CutlerFlanagan; @foot3]. We notice that the errors calculated using 1.5PN phasing are always [*larger*]{} for $t_c$ and $\phi_c$, and [*smaller*]{} for $\cal M$, $\eta$, and $\beta$, than those calculated using 2PN phasing. Thus, the measurement errors on the masses and spins are [*underestimated*]{} when evaluated using the less accurate 1.5PN phasing.
This can be explained with a simple argument. In Eq. (\[3.6\]), the 1PN and 2PN terms \[of order $(\pi M f)^{2/3}$ and $(\pi M f)^{4/3}$ respectively\] combine, when $\epsilon=1$, so as to reduce the relative importance of the $\pi M f$ term, when compared to the situation when $\epsilon=0$. In other words, the relative contribution to the total number of wave cycles coming from the $\pi M f$ term is less for 2PN phasing than it is for 1.5PN phasing (see Table I). It is therefore expected that 2PN phasing will produce larger measurement errors for $\beta$, since all information about $\beta$ comes from the $\pi M f$ term. But because $\beta$ is strongly correlated with both $\cal M$ and $\eta$, it follows that these parameters will [*also*]{} come with larger measurement errors. This is indeed what is observed. It is amusing to test this explanation by artificially setting $\epsilon=-1$ in our calculations, which we do in the fourth line of Table II. The argument suggests that the errors in $\cal M$, $\eta$, and $\beta$ should all decrease with respect to the values calculated using 1.5PN phasing, since the relative importance of the $\pi M f$ term is now increased. This is indeed what the results show.
[lrrrrrrrrrrrr]{} prior & $\Delta t_c$ & $\Delta \phi_c$ & $\Delta {\cal M} / {\cal M}$ & $\Delta \eta / \eta$ & $\Delta \beta$ & $\Delta \sigma$ & $c^{{\cal M} \eta}$ & $c^{{\cal M} \beta}$ & $c^{{\cal M} \sigma}$ & $c^{\eta\beta}$ & $c^{\eta \sigma}$ & $c^{\beta\sigma}$\
\
\
yes & 1.28 & 13.3 & 0.047 % & 0.507 & 1.77 & 4.79 & -0.956 & 0.996 & -0.648 & -0.964 & 0.835 & -0.660\
no & 2.72 & 46.9 & 0.120 % & 1.578 & 4.53 & 17.3 & -0.991 & 0.999 & -0.952 & -0.993 & 0.984 & -0.955\
\
\
yes & 2.54 & 23.6 & 0.280 % & 0.873 & 3.02 & 4.74 & -0.959 & 0.997 & -0.630 & -0.969 & 0.817 & -0.650\
no & 7.52 & 95.9 & 0.813 % & 3.10 & 8.98 & 19.4 & -0.993 & 0.999 & -0.961 & 0.995 & 0.986 & -0.964\
\
\
yes & 2.22 & 10.4 & 0.55 % & 1.51 & 8.22 & 4.81 & -0.849 & 0.920 & 0.191 & -0.984 & 0.257 & -0.081\
no & 17.0 & 179 & 7.23 % & 30.7 & 149 & 74.6 & -0.995 & 0.998 & -0.962 & -0.999 & 0.984 & -0.978
To obtain the results of Table III we have estimated all seven parameters, including both $\cal A$ and $\sigma$, and used 2PN phasing. For each of the three systems, the first line of Table III displays the measurement errors and correlation coefficients calculated using the prior probability distribution function (\[3.15\]). We notice that the measurement errors are all significantly larger than those displayed in Table II; this is expected from the fact that we are now estimating a larger number of parameters.
It is interesting to ask how the measurement errors increase as the number of estimated parameters increases. In Ref. [@CutlerFlanagan], Cutler and Flanagan initially estimate only five of their six parameters, leaving $\beta$ out. When they next include $\beta$ in their calculations, they find that the measurement errors on $\cal M$ and $\mu$ increase by a factor of order 10. In this paper, on the other hand, we have initially estimated only six of our seven parameters, leaving $\sigma$ out. When we next include $\sigma$ in our calculations, we find that the measurement errors on $\cal M$, $\mu$, and $\beta$ only increase by a factor of order unity. Thus, the inclusion of $\sigma$ in the calculation has less dramatic consequences than the inclusion of $\beta$. This confirms a conjecture formulated by Cutler and Flanagan [@CutlerFlanagan] at the end of their Sec. III. That this is so is largely due to the importance of prior information in the estimation of $\sigma$.
In the second line of Table III we display the results obtained when the prior information is [*not*]{} included in the calculations. We notice that for all systems, the prior information indeed plays a very important role. In fact, we see that Eq. (\[3.15\]) provides nearly [*all*]{} of the information regarding the spin-spin parameter $\sigma$. This explains why the measurement error on $\sigma$ is always nearly equal to 5.0, and its correlation coefficient with other parameters significantly smaller than unity. Of course, these results only apply to gravitational-wave measurements with $\rho=10$. To bring the error on $\sigma$ well below the a priori constraint $\sigma < |5.0|$, say $\Delta \sigma \lesssim 3$, the measurement would require a signal-to-noise ratio larger than approximately 45 for system A, 50 for system B, and 110 for system C.
We conclude with the following remark. It is clear that the results displayed in Tables II and III depend on a fairly large number of simplifying assumptions, and that a more careful treatment might produce somewhat different numbers. These assumptions include: (i) the simplified form (\[1.1\]) for the waveform; (ii) the neglect of (not yet calculated) higher-order terms in the post-Newtonian expansion (\[1.3\]); (iii) the neglect of $O(1\rho)$ corrections in the expression (\[2.8\]), (\[2.9\]) for $p(\bbox{\theta}|s)$; (iv) the analytic model (\[3.1\]) for the noise spectral density; and (v) our rather crude incorporation of the prior information. We shall leave for future work the difficult task of carefully examining the effect of these assumptions on our results.
Acknowledgments {#acknowledgments .unnumbered}
===============
Our warmest thanks to Eanna Flanagan for many useful conversations and his detailed comments on the manuscript. This work was supported by the National Science Foundation under Grant No. PHY 92-22902 and the National Aeronautics and Space Administration under Grant No. NAGW 3874.
K.S. Thorne, in [*300 Years of Gravitation*]{}, edited by S.W. Hawking and W. Israel (Cambridge University Press, Cambridge, England, 1987), p. 330. B.F. Schutz, Nature [**323**]{}, 310 (1986); Class. Quantum Grav. [**6**]{}, 1761 (1989). A. Abramovici, W.E. Althouse, R.W.P. Drever, Y. G[" u]{}rsel, S. Kawamura, F.J. Raab, D. Shoemaker, L. Siewers, R.E. Spero, K.S. Thorne, R.E. Vogt, R. Weiss, S.E. Whitcomb, and M.E. Zucker, Science [**256**]{}, 325 (1992). C. Bradaschia, E. Calloni, M. Cobal, R. Del Fasbro, A. Di Virgilio, A. Giazotto L.E. Holloway, H. Kautzky, B. Michelozzi, V. Montelatici, D. Pascuello, and W. Velloso, in [*Gravitation 1990*]{}, Proceedings of the Banff Summer Institute, edited by R. Mann and P. Wesson (World Scientific, Singapore, 1991). R. Narayan, T. Piran, and A. Shemi, Astrophys. J. [**379**]{}, L17 (1991). E.S. Phinney, Astrophys. J. [**380**]{}, L17 (1991). C. Cutler, T.A. Apostolatos, L. Bildsten, L.S. Finn, E.E. Flanagan, D. Kennefick, D.M. Markovic, A. Ori, E. Poisson, G.J. Sussman, and K.S. Thorne, Phys. Rev. Lett. [**70**]{}, 2984 (1993). L.A. Wainstein and V.D. Zubakov, [*Extraction of Signals from Noise*]{}, (Prentice-Hall, Englewood Cliffs, 1962). L.S. Finn and D.F. Chernoff Phys. Rev. D [**47**]{}, 2198 (1993). C. Cutler and E.E. Flanagan, Phys. Rev. D [**49**]{}, 2658 (1994). For an overview, see C.M. Will, in [*Relativistic Cosmology*]{}, Proceedings of the Eighth Nishinomiya-Yukawa Memorial Symposium, edited by M. Sasaki (Universal Academy Press, Kyoto, Japan, 1994), p. 83. L. Blanchet and T. Damour, Philos. Trans. R. Soc. London [**A320**]{}, 379 (1986); Phys. Rev. D [**37**]{}, 1410 (1988); Ann. Inst. H. Poincaré [**50**]{}, 377 (1989); Phys. Rev. D [**46**]{}, 4304 (1992). L. Blanchet, T. Damour, B.R. Iyer, C.M. Will, and A.G. Wiseman, [*Gravitational-radiation damping of compact binary systems to second post-Newtonian order*]{}, Phys. Rev. Lett. (in press). L. Blanchet, T. Damour, and B.R. Iyer, [ *Gravitational waves from inspiralling compact binaries: Energy loss and waveform to second-post-Newtonian order*]{} (unpublished). C.M. Will and A.G. Wiseman, in preparation. P.C. Peters, Phys. Rev. [**136**]{}, B1224 (1964). C.W. Lincoln and C.M. Will, Phys. Rev. D [**42**]{}, 1123 (1990). L.E. Kidder, C.M. Will, and A.G. Wiseman, Phys. Rev. D [**47**]{}, R4183 (1993). A. Krolak, J.A. Lobo, and B.J. Meers, Phys. Rev. D [**48**]{}, 3451 (1993). P. Joranowski and A. Krolak, Phys. Rev. D [**49**]{}, 1723 (1994). K. Kokkotas, A. Krolak, and G. Tsegas, Class. Quantum Grav. [**11**]{}, 1901 (1994); 2833 (1994). C. Cutler and E.E. Flanagan, in preparation. C. Cutler, L.S. Finn, E. Poisson, and G.J. Sussman, Phys. Rev. D [**47**]{}, 1511 (1993). H. Tagoshi and T. Nakamura, Phys. Rev. D [**49**]{} 4016 (1994). H. Tagoshi and M. Sasaki, Prog. Theor. Phys. [**92**]{}, 745 (1994). L.S. Finn, Phys. Rev. D [**46**]{}, 5236 (1992). C.W. Helstrom, [*Statistical Theory of Signal Detection*]{}, (Pergamon, Oxford, England, 1968). Strictly speaking, our expression for $F_i$ is only valid in the limit $\eta \to 0$. For simplicity, and because this will not affect our results significantly, we shall ignore the corrections to the innermost circular orbit which are due to the finite value of the mass ratio. These are computed by L.E. Kidder, C.M. Will, and A.G. Wiseman, Phys. Rev. D [**47**]{}, 3281 (1993) using post-Newtonian theory, and by S. Detweiler and J.K. Blackburn, Phys. Rev. D [**46**]{}, 2318 (1992) and G.B. Cook, Phys. Rev. D [**50**]{}, 5025 (1994) using numerical relativity. See, for example, J.D. Jackson, [*Classical Electrodynamics*]{}, (Wiley, New York, 1975), p. 316. Our calculations should take into account the fact that $\eta$ is restricted to the interval $(0,1/4]$. However, this information cannot be incorporated using a Gaussian probability distribution function such as Eq. (3.15). To keep the analysis simple we have preferred to leave $\eta$ unrestricted. Cutler and Flanagan use $\ln\mu$ as a parameter instead of $\ln\eta$. It is easy to show, by performing the coordinate transformation, that $\Delta \mu/ \mu =
(2/5) \Delta \eta/\eta$ and $c^{\mu a} = c^{\eta a}$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The world is long-tailed. What does this mean for computer vision and visual recognition? The main two implications are (1) the number of categories we need to consider in applications can be very large, and (2) the number of training examples for most categories can be very small. Current visual recognition algorithms have achieved excellent classification accuracy. However, they require many training examples to reach peak performance, which suggests that long-tailed distributions will not be dealt with well. We analyze this question in the context of eBird, a large fine-grained classification dataset, and a state-of-the-art deep network classification algorithm. We find that (a) peak classification performance on well-represented categories is excellent, (b) given enough data, classification performance suffers only minimally from an increase in the number of classes, (c) classification performance decays precipitously as the number of training examples decreases, (d) surprisingly, transfer learning is virtually absent in current methods. Our findings suggest that our community should come to grips with the question of long tails.'
bibliography:
- 'main.bib'
title: |
The Devil is in the Tails:\
Fine-grained Classification in the Wild
---
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Encryption has increasingly been used in all applications for various purposes, but it also brings big challenges to network security. In this paper, we take first steps towards addressing some of these challenges by introducing a novel system to identify key exchange protocols, which are usually required if encryption keys are not pre-shared. We observed that key exchange protocols yield certain patterns of high-entropy data blocks, e.g. as found in key material. We propose a multi-resolution approach of accurately detecting high-entropy data blocks and a method of generating scalable fingerprints for cryptographic protocols. We provide experimental evidence that our approach has great potential for identifying cryptographic protocols by their unique key exchanges, and furthermore for detecting malware traffic that includes customized key exchange protocols.'
author:
- Shoufu Luo
- Sven Dietrich
bibliography:
- 'entropy\_arxiv.bib'
title: Fingerprinting Cryptographic Protocols with Key Exchange using an Entropy Measure
---
Introduction
============
In the network security field, the use of encryption for malicious purposes brings new challenges to network security defense. For example, encryption has prevented botnet traffic from being inspected and detected by defense systems based on deep-packet inspection (DPI), which used to be very effective up to that point. For symmetric encryption and decryption, a secret key $k$ shared among two communicating parties is required, either pre-shared or negotiated on the fly using cryptographic key-exchange protocols. Most common cryptographic protocols [@ssl:rfc6101; @tls:rfc5246; @ssh:rfc4251] using symmetric encryption to secure the channel use a key exchange protocol, such as the Diffie-Hellman key exchange [@dhe:rfc2631]. Depending on the protocol design, key material is distributed differently along the traffic stream. As key material has high entropy compared to normal traffic, the traffic for the key exchange exhibits detectable characteristics, namely the uniqueness of the distribution of key material allowing for proper discriminating characteristics, as shown in Figure \[fig:demo\].
![Visualization of Entropy Distribution: dark portions are high-entropy blocks.[]{data-label="fig:demo"}](./blk_bar_demo.png)
Using an entropy metric, it may not be hard to test the hypothesis whether a byte string is “random,” if that byte string is sufficiently long. The problem becomes harder if the given string is relatively short, i.e. undersampled, or if the goal is to identify which part of the string contains random bytes, in particular, deciding the boundaries of those random bytes (also known as blocks of interest). It is therefore challenging to characterize a stream by the distribution of embedded random bytes, or so-called high-entropy blocks.
To avoid being treated as an anomaly, malware might try to use standard cryptographic protocols (e.g. SSL/TLS) for secure communication, effectively preventing DPI. However, standard protocols such as SSL can potentially be subject to a man-in-the-middle attack. However, malware in general tends to avoid using standard protocols and instead employs a customized variant. Only 10% of malware utilize TLS as a form of encryption, according to a recent study [@anderson2016deciphering]. To ensure fresh key material, a new key exchange is desirable for every new command-and-control (C&C) session of the malware [@dd2008:malware; @art:stormnugache].
Our work offers a systematic way to characterize network traffic through key exchange behaviors and generate scalable fingerprints based on detected high-entropy blocks. The system mainly consists of two parts: the high-entropy block detection and the fingerprint generation. First, we aim to identify high-entropy blocks from a traffic stream using sample entropy via a sliding window. Second, with all high-entropy blocks identified, entropy-based fingerprints for network flows will be generated by the distribution of high-entropy blocks. Our contribution also includes:
- A new method of identifying cryptographic protocols, raising the bar for malicious activities that abuse customizing cryptographic protocols to evade inspection.
- A voting mechanism that efficiently boosts the accuracy of entropy estimation when undersampled using a multi-resolution analysis.
- A statistical approach to estimate the range of high-entropy data blocks and build scalable entropy-based fingerprints for key exchange protocols in the form of regular expressions.
To the best of our knowledge, our work is the first attempt to fingerprint key exchange protocols by the distribution of key material and apply such a technique to malware detection. By design our approach can be implemented and deployed as a standalone system. However, it is not the intention to replace any existing detection techniques, but rather to complement them. This system can be built with existing systems as a plug-in component, in particular those relying on a certain degree of payload analysis, e.g. [@tamd]. Moreover, a component of our system can be a useful tool for the security community, e.g. for identifying high-entropy portions of a given data block, such as detection of packed malware binaries.
### Related Work
Olivain et al. [@netentropy] proposed to use cumulative entropy of network flows for detection of specific attacking behaviors targeted at known cryptographic protocols, i.e. SSL. Instead of an aggregation, our work aims to fingerprint the entropy distribution along the examined traffic. Our approach is still applicable for their purpose in a more precise way. Meanwhile, we adopt the technique they propose, *N-truncated entropy*, for entropy estimation, which is also used by Dorfinger et al. [@Dorfinger2011TMA] for classifying encrypted and unencrypted traffic. There is prior work [@white2013clear] that shows how entropy tests can be used to detect encrypted or compressed packets from network streams. Again, we provide a more reliable mechanism to detect high-entropy areas as one of our essential contributions.
Our work shares an interest from the field of protocol identification. Most of the work in that field is mainly learning-based, relying on network-observable features [@Kara:2005; @Wright:2006]. For example, Wright et al. [@Wright:2006] proposed to identify the cryptographic protocol of individual encrypted TCP connections using post-encryption observable features, such as timing, size, direction etc. To some extent, our approach can also be also applied for this purpose. However, there are known obfuscation techniques which could be used to evade this, such as obfsproxy[@obfsproxy] and FTE[@dyer2013protocol]. As discussed in [@wang2015seeing], obfuscation can be detected with entropy-based tests over the packet payloads. Our approach does the same by extracting entropy-based fingerprints.
Zhang et al.[@ZhangPM13] proposed to detect encrypted traffic by looking for *N* sequential high-entropy packets of all first *M* packets of one network flow adopting the cumulative entropy technique. In 2015, Zhang et al. [@zhang15] improved their previous work by detecting of high-entropy flows as an additional measure to score a host being a bot for BotHunter [@bothunter2007usenix]. Applicable to the same problem, our approach is different from theirs by fingerprinting malware with customized cryptographic protocols, such as Nugache, as will be shown. Unlike their work, our work does not rely on another system for detection.
The rest of this paper is organized as follows. We begin with background on entropy and its estimators. In section \[sec:method\], we discuss our methodology in detail, including how to identify high-entropy blocks, a voting mechanism as well as a filtering method for false positives reduction, etc. Following that, section \[sec:eval\] presents evaluation and analysis of our approach with three different dataset. Finally, we conclude this study by discussing limitations and directions of future work.
Background
==========
Entropy
-------
Introduced by Shannon [@shannon1948], entropy is used as a measurement of the amount of information that is missing before reception. In the context of cryptography, it is used as a measure of randomness (or uncertainty), equating higher entropy with higher randomness. Let $X$ be a discrete random variable under an arbitrary distribution $\mathcal{P}$ on a countable alphabet $\Sigma=\{x_1, ..., x_m\}$. The definition of Shannon entropy can be generally expressed by the equation (\[entropy\]), $$\label{entropy}
H(X) = -\sum^m\limits_{i=1} p(x_i) \log_2 p(x_i) % H(P) v.s. H(X), need to clarify
% H(P) v.s. H(X), need to clarify$$
The entropy $H(X)$ yields a maximum value when all $p(x _i)$ are equal to $\frac{1}{m}$, i.e. uniformly distributed. In cryptography, as a fundamental requirement of security, key material should have high entropy in order to be hard to predict.
Entropy Estimator
-----------------
Entropy can be easily obtained by the equation (\[entropy\]) if given a random variable whose probability distribution is known. However, in practice, $\mathcal{P}$ may remain unknown for most scenarios. Frequently, $p(x_i)$ could be still estimated by the relative frequencies of the outcome $x_i$ from a large number of trials. The probability of $x_i$ is thereby $\hat{p}(x_i) = \frac{n_i}{N}$, where $n_i$ is the number of times $x_i$ occurs and $N$ is the total number of trials or samples. Hereby, the *sample entropy*, a.k.a. maximum likelihood estimator (MLE) [@Antos2001], can be estimated as in the equation below.
$$\label{sample}
\hat{H}_N^{MLE}(X) \equiv -\sum^m\limits_{i=1} \hat{p}(x_i) \log_2 \hat{p}(x_i)$$
Even though MLE is an unbiased estimator of $H(X)$ when $N$ tends to infinity where $\hat{p}(x_i)$ approximates $p(x_i)$ and $\hat{H}_N^{MLE}(X)$ approximates real $H(X)$. When N is not sufficiently large, namely *undersampled*, $\hat{H}_N^{MLE}(X)$ highly bias, in particular, $N < m$ or $N \sim m$. There is no universal rate at which the error of MLE compared to $H(X)$ would be close to zero [@Antos2001]. There are attempts that aim to subtract the bias directly, such as the Miller-Madow corrector [@miller], the Jackknife corrector [@jackknife] and the Paninski corrector [@Paninski03]. However, the bias is still significantly high when $N < m$ or $N \sim m$. Moreover, it has been proven difficult to find an unbiased estimator [@Paninski03; @Schuermann04biasanalysis]. Unfortunately, the Paninski corrector is unbiased but if and only if $\mathcal{P}$ has a uniform distribution, which can not be guaranteed. Furthermore, according to this study [@netentropy], $\hat{H}_N^{MLE}(X) \sim H(X)$ is valid if and only if $N \gg m$, which typically means $N$ is of the order of roughly at least *10* times as large as $m$. In another word, if $\Sigma_0=\mathrm{\{0x00, ..., 0xff\}}$ (i.e. m=$|\Sigma_0|$=256), it would require around 2,000 samples to possibly obtain a reasonable estimated entropy. That makes it impractical for the purpose of profiling network traffic as key material usually is at most hundreds of bytes (256 bytes = 2048 bits). For example, in a typical TLS handshake, a client random number only contains 28 bytes.
N-truncated entropy $H_N(X)$
----------------------------
Similar to Olivain et al. [@netentropy], an accurate entropy value is not of our main focus, but rather the probability of a string being generated from a uniform distribution. The *N-truncated entropy* $H_N(X)$ proposed by Olivain et al. meets our needs, which is the average of the sample entropy $\hat{H}_N^{MLE}(X)$ over all strings of length of $N$ drawn at random from the distribution $\mathcal{P}$, as defined below. $$\label{netranc}
\small H_N(X)=\sum\limits_{ \Sigma_i n_i =N } \left[ {N \choose {n_0, ... , n_{m-1}}} \prod\limits_{i=0}^{m-1} p_i^{n_i} \left(- \sum^{m-1}\limits_{i=0}\frac{n_i}{N} \log_2 \frac{n_i}{N}\right) \right]$$
By construction, $\hat{H}_N^{MLE}(X)$ is an unbiased estimator of $H_N(X)$ for an arbitrary distribution $\mathcal{P}$. More importantly, $\hat{H}_N^{MLE}(X)$ gives a statistical indication that how close the distribution $\mathcal{P}$ is to being uniform by comparing to $\hat{H}_N^{MLE}(W)$ given that $W$ be a random variable under a uniform distribution $\mathcal{U}$. In section \[highent\], we describe how to obtain both values. Alternatively, if a string $s$ of length $N$ with each sample drawn from $\mathcal{P}$, we use $\hat{H}^{MLE}_N(s)$ instead of $\hat{H}_N^{MLE}(X)$. To differentiate this, $w$ is used instead if uniform distribution, $\mathcal{U}$. $H_N(X)$ has an upper bound of $log_2 \min\{m, N\}$ as it reaches its maximum value if all $\hat{p}_{x_i}$ are equal, either $\hat{p}_{x_i} = \frac{1}{N}$ if $N < m$ or $\hat{p}_{x_i} = \frac{1}{m}$ otherwise. In either case, uncertainty reaches its maximum.
Methodology {#sec:method}
===========
In this section, we discuss in detail the techniques we used and developed, accompanied by experimental evidence.
Sliding Window
--------------
To obtain entropy information of different portions within the traffic stream, a sliding window moves over the traffic with a step of one byte while sample entropy will be measured for each chunk of bytes in that window. Bytes in each window form a block.
The window size determines sample size, which directly impacts the accuracy of sample entropy. If the sample size is too small, the sample entropy might not be accurate enough to be meaningful. Equation (\[noiseprob\]) roughly estimates the probability of a N-byte string appearing to be “random”, i.e. each char in the alphabet only occurs once in the string. Fix $\Sigma$ to be $\Sigma_0$ and then let $m$=256. Let $N$=16 be a 16-byte sliding window. Pr\[X=e\]=0.6197. That is, there is a 40% probability that an arbitrary string appears random, i.e. a forty percent chance of a false positive. However, if N=32, Pr\[X=e\]=0.082. This confirms the discussion in Paninski et al. [@Paninski03] that one should never use less than 16 bytes for entropy estimation when $\Sigma_0$ is used. $$Pr[X=e]=1\cdot \frac{m-1}{m} \cdot ... \cdot \frac{m-N+1}{m} = \prod\limits_{i=0}^{N-1} \frac{m-i}{m}
\label{noiseprob}$$ If the sliding window grows to be too large, it is likely to mix high-entropy areas with low-entropy areas, confusing the difference between them. As shown in Figure \[fig:winsz\], when the window size is small, e.g. 16-byte, the curve is fuzzy and has too many valleys (low-entropy) and peaks (high-entropy), while as the window size goes larger, e.g. 1024 or 2048-byte, the curve becomes flatter and valleys or peaks are not distinctive anymore.
![Entropy plot of a TLS sample traffic using different sliding window sizes, from bottom to top (-byte): 16, 32, 64, 128, 256, 512, 1024 to 2048.[]{data-label="fig:winsz"}](./all_mix_8192.png)
A smaller window is more likely to mistakenly identify a non-random data area to be “random” (false positive), while a larger window possibly fails to identify real high-entropy area (false negative). The choice of window size will heavily depend on the minimum length of key materials of interest. In case of TLS, we choose a 32-byte sliding window as it is good for the minimum length of interests, i.e. 28-byte client random number. In summary, as the window slides over the data with a one-byte step, each block is labeled as either high-entropy or low-entropy. A list of consecutive either high-entropy blocks or low-entropy blocks then forms a *unit*, more precisely a high-entropy unit or a low-entropy unit respectively.
Baseline $H_N(\mathcal{U})$ {#highent}
---------------------------
To identify a high-entropy block, we follow the idea used by [@netentropy], i.e. the Monte-Carlo method, as it provides a level of confidence of a string being random. We first repeatedly generate strings of length of $N$ with each byte sampled from a random source, e.g. */dev/urandom* on MacOS X. Then, we calculate the mean $\mu$ and standard deviation $\sigma$ of sample entropy using all samples. Here, $\mu$ and $\sigma$ summarize the distribution of the sample entropy of random strings of length $N$. By a specific number $t$ of standard deviations, we can obtain the proportion of sample strings falling within the range of $\mu \pm t\times \sigma$. This proportion provides us with a confidence of a string being random if it falls within the given range. As exceeding the upper bound does not affect the randomness of the string, we ignore the upper bound and use the lower bound as a cutoff for a string being random, denoted by $\theta$, with a confidence by the proportion $\rho$: $$\theta = \mu\big(\hat{H}_{N}^{MLE}(w)\big) - t\times \sigma\big(\hat{H}_{N}^{MLE}(w)\big)
~~~~ ~~\rho = \frac{{number\ of\ samples\ above\ \theta}}{number\ of\ samples}$$ Consequently, any strings falling below the threshold are considered not random, i.e. low-entropy blocks. Similarly, any strings falling above the threshold will be considered random, i.e. high-entropy. Table \[tbl:HNmu\] shows thresholds ($\theta$) for $w$ using different window sizes (N) above a minimum level of confidence 99.0%.
$~~~N~~~$ $\mu$ $\sigma$ $t$ $\theta$ $\rho$
----------- ------------- --------------- -------------- ------------- ------------
16 3.94199 0.08290 2.8 3.7098 99.2%
32 4.88171 0.08134 2.7 4.6620 99.3%
64 5.76562 0.07664 2.6 5.5663 99.2%
128 6.55003 0.06733 2.5 6.3817 99.2%
256 7.17518 0.05240 2.5 7.0441 99.2%
512 7.59073 0.03364 2.4 7.5099 99.0%
1024 7.80894 0.01726 2.5 7.7658 99.1%
2048 7.90804 0.00814 2.5 7.8877 99.2%
: $\hat{H}_N^{MLE}(w)$ under Various Configurations[]{data-label="tbl:HNmu"}
The confidence measures the confidence of a string not being random when falling out of the range, rather than a confidence of a string being random when falling within the range. For example, let $N$ be 64 and $\Sigma$=$\Sigma_0$, then $\mu$=5.7656, $\sigma$=0.0766. With 99.4% of samples above $\theta$=$\mu$-3$\sigma$=5.53569 (i.e. $t = 3$), we would have *at least* 99.4% confidence that a string $s$ with $\hat{H}_N^{MLE}(s)$=5.5120 is not close to random, i.e. not a high-entropy block. Here, $t$ is our control variable. We can choose a smaller $t$ to tighten the range with a higher confidence or a larger $t$ to loosen the range, but with a lower confidence. In our study, we choose $t$ tightly to obtain a relatively high confidence, at least 99.0%. With the threshold, we could then transform sample entropy score to either one or zero. The plot turns to be to a square wave where *one* indicates high-entropy and *zero* for low-entropy as shown in figure \[fig:norm\]. The shadow in the upper plot shows the cutoff.
![Normalization: high-entropy blocks[]{data-label="fig:norm"}](./normalized.png)
The choice of $\Sigma$ {#sigma}
----------------------
\[voting\]
Due to statistical limitations, some data blocks may mistakenly be labeled as high-entropy blocks, i.e. a false positive, which will mislead the fingerprint and therefore must be avoided or minimized. In order to achieve this, we devised a voting mechanism using multi-resolution analysis, utilizing the choice of alphabet $\Sigma$. As will be shown, this mechanism dramatically reduces the rate of false positives.
Thus far we based our discussion on the choice of $\Sigma$ to be $\Sigma_0$ ($m$=256) with each char being a byte. In cryptography, however, the randomness of key material is defined at a more restrictive level, i.e. at a *bit* level, and thereby $\Sigma$={0, 1} ($m$=2). Let’s consider one experiment of tossing one coin that has two outcomes, and another experiment of tossing eight independent coins with two outcomes for each. According to basic probability theory, if each coin is uniformly drawn from $\Sigma$={0, 1}, the outcome of eight coins ($\Sigma_0$) will still follow a uniform distribution. In our estimation of $\hat{H}^{MLE}_N(w)$, we do generate each random byte by randomly sampling eight times over {0, 1} for all our sample strings. That being said, given that each bit is independently sampled uniformly from {0, 1}, we could choose a random variable of different number of $\tau$ bits (i.e. coins) and such a random variable will be guaranteed to have a uniform distribution.
As an extension to our previous computation of $\hat{H}^{MLE}_N(w)$, we outline the thresholds and their confidence levels for different $\tau$ while fixing $N$ to 32. We use the term $\tau$-bit measure, e.g. 2-bit measure. Previously, $N$ could be interpreted as either the window size and the sample size. In the case of $\tau$-bit measure, the sample size changes, i.e. $\frac{8}{\tau}$ N ($\tau \leq 8$). For convenience, we abuse the notation N, using it as the window size in the rest of this paper. The use of $\tau$-bit measure does not change the fundamentals of *N-truncated entropy* as it simply uses a larger sample size and a different alphabet.
$\tau$ m $\mu$ $\sigma$ $t$ $\theta$ $\rho$
--------- -------- ------------ -------------- ------------ -------------- ------------
1 2 0.9971 0.00399 4.18 0.9804 99.28%
2 4 1.9829 0.01387 3.59 1.9331 99.20%
4 16 3.8196 0.06715 3.02 3.6168 99.31%
8 256 4.8817 0.08135 3.0 4.6356 99.35%
: $\tau$-bit measure $\hat{H}_{32}^{MLE}(w)$[]{data-label="tbl:hmu2"}
Statistical methods such as sample entropy generally ignore potential structures or patterns occurring in the data. Therefore, a string with a high sample entropy score is not guaranteed to be random. For example, given a hexadecimal string $s$ be “55 55 bb bb”, i.e. $0101~ 0101~ 0101~ 0101~ 1010~ 1010~ 1010~ 1010$ in binary, we have $\hat{p}_0=\hat{p}_1=\frac{1}{2}$ if 1-bit measure ($\tau$=1) used, i.e. $\Sigma$={0, 1}, and then $\hat{H}^{MLE}_N(s)$ = 1. Consequently, $s$ will be labeled as high-entropy bytes in spite they are not at all. Taking another example from real world, a hexadecimal string from a TLS session: $16~ 03~ 01~ 0c~ 13~ 0b~ 00~ 0c~ 0f~ 00~ 0d~ 0e~ 10~ 04~ 7a~ 30~ 82$, which is a block of control information[^1] from the TLS handshake traffic. The two bytes *03 01* indicate the TLS version, i.e. TLS 1.0, *0c 13* for the length,*0b* for the protocol type, and another 3 bytes of length *00 0c 0f*. This block may not also appear “random“ if an 8-bit measure is used. Such cases are prone to false positives and mislead the process.
![A traffic sample from a TLS 1.2 session with a 1024-bit RSA public key.[]{data-label="fig:xsig"}](./xsig.png)
However, the idea is that if a string is random, no matter which $\tau$-bit measure is being used, its sample entropy $\hat{H}^{MLE}_N(s)$ should be always close to $\hat{H}_N(\mathcal{U})$. Thus, we propose to use a voting mechanism instead of using a sole $\tau$-measure. The voting rule is if any of chosen $\tau$-bit measure rejects the randomness of that block, the block will be labeled as non-random. It is a simple AND operation among the outcome of all measures. Figure \[fig:xsig\] shows the effectiveness of combining three $\tau$-measures, where the resulting signature by voting precisely outlines all high-entropy blocks in the TLS session. The last plot line, X-signature, is based on the voting over the three 1-bit, 4-bit and 8-bit measures.
Filtering Threshold
-------------------
Our voting mechanism effectively reduces false positives. However, in some scenarios, this approach may still not be sufficient to eliminate all false positives. There is still a chance that all $\tau$-bit measures falsely identify an ordinary block to be high-entropy because of accidentally some small actual randomness within the data. If there supposedly are no high-entropy data blocks, the length of a data block with randomness should be less than the minimum length of interest and the size of detected high-entropy units would appear to be relatively small compared to that, if there actually exists a high-entropy data block of interest. A filtering threshold denoted as $\xi$ is possibly chosen to eliminate those small high-entropy units. Our empirical study suggests $\xi=9$ to be a good choice when a 32-byte sliding window size chosen for detecting a minimum 20-byte high-entropy key material blocks. That means if there are only 9 consecutive high-entropy blocks detected between two low-entropy units, then a false positive is identified and filtered out in that case. Here, the “filter out” means labeling these blocks to be low-entropy instead of high-entropy.
Calibration {#cali}
-----------
Beyond identifying high-entropy blocks, it is also essential to describe the length of each unit in order to fingerprint the shape of the square wave as shown in \[fig:xsig\]. Due to its statistical inheritance and the way of measuring, the length of each unit (i.e. the number of detected consecutive high-entropy or low-entropy blocks) may vary because when the sliding window is partially over the target random bytes, it may still continue to yield high sample entropy blocks until the window moves sufficiently away from the target. For example, a TLS traffic stream contains a client random number as a chunk of 28 bytes. It is not difficult to anticipate that there will not be only exactly one high-entropy block detected in this case. The total number of high-entropy blocks detected around that chunk of data will not be fixed as well from case to case. However, our intention is not to determine an absolute value for each unit among all cases, but rather a certain reasonable range. Hereby, we resort to *Monte-Carlo* methods to empirically estimate the range. For example, to estimate the length of high-entropy unit around client random bytes, we sampled 100,000 *client hello* messages from TLS sessions.
![Distribution of length of detected high-entropy blocks (1) Left: over the TLS 28-byte client random string (2) Right: over the TLS 28-byte client random string and 32-byte session ID.[]{data-label="fig:range"}](./pjimage.png)
The result shown in figure \[fig:range\] indicates most of the length for the 28-byte client random string followed by the list of cipher suites fall within a range between six high-entropy blocks and twenty-four blocks. If a 32-byte TLS session ID (also random bytes) is present along with the client random bytes, adding up to 60 bytes, we obtain a range of $[38, 52]$ as shown in figure \[fig:range\]. A more conservative range would be $[20, 52]$.
Fingerprinting
--------------
Fingerprinting is a process to profile a key exchange protocol by its distribution of high-entropy blocks along traffic streams generated by such a protocol. A entropy-based fingerprint is a series of interleaving high-entropy units and low-entropy units with the length of each unit specified as a range. The reason that high-entropy blocks have to interleave with low-entropy ones is that otherwise two adjacent high-entropy or low-entropy blocks would be merged into one. Let $(s, l, r)$ represent one unit where $s \in \{1, 0\}, l, r \in \mathbf{Z^+}$, where $s$ be the sign indicating a high-entropy unit or low-entropy, $l$ be the minimum length and $r$ be the maximum length. An entropy-based fingerprint then is the concatenation of an ordered list of $(s, l, r)$ with $s$ alternating among one and zero. Alternatively, it can be concisely expressed as below, where $s_i \in \{1, 0\}, l_i, r_i \in \mathbf{Z^+}$. The benefit of such a representation is that this form aligns with standard regular expression and the matching process can be done very efficiently. The regular expression form will provide a flexible way of expressing the fingerprint, for instance, optional units, as will be shown in the experiment section. $$\bigparallel_{i=1}^n s_i \{l_i, r_i\}, s_i \neq s_{i+1}$$ The fingerprinting is straightforward in three steps: (1) identify high-entropy and low-entropy areas (units) of the anticipated traffic from a cryptographic protocol; (2) follow the technique described in section \[cali\] and estimate the range for each area; (3) formalize the units in a regular expression. Taking TLS using a cipher-suite of DHE-RSA-\* as an example, the fingerprint is as below: $$1\{8, 54\}0\{20,1024\}1\{8,54\}0\{30,800\}1\{80,260\} ....$$
During the detection phase, we have these steps: (1) scan the traffic stream by sliding a window over it and estimating sample entropy for each window using different $\tau$-bit measures; (2) normalize each block by its entropy score to either one or zero using the pre-calculated threshold $\theta$; (3) perform the voting (i.e. AND) of outcomes from each measures; (4) filter out the noises using filter threshold; (5) use regular expression to match the predefined fingerprint against the output (i.e. a string consisting of zeros and ones).
In our demonstration, we emphasize DHE-RSA-\* cipher-suite for TLS protocol as our approach aims to profile that a particular key exchange protocol and TLS is capable of using different key exchange protocols. SSL has evolved over time into the standard TLS protocol, which supports a long list of cipher suites with different key exchange protocols. To demonstrate, we choose to profile one set of key exchange protocol cipher suites, i.e. DHE-RSA-\*, (see \[tbl:cipher\]). By contrast, as an application of our system, most botnet C&C protocols are much simpler as most of them are designed for the sole purpose of performing a limited number of tasks.
Evaluation {#sec:eval}
==========
SSL/TLS is a well-known cryptographic protocol with fair complexity. The successful characterization of the TLS protocol provides the full ability to characterize other and simpler botnet C&C protocols. For evaluation, we first use TLS as our primary target and later extend it to the Nugache botnet. All streams are bidirectional and packets of a stream are correctly ordered with all TCP/IP headers removed. The *tshark* [@tshark] was used as a primary tool to process network traces in `pcap` [@pcap].
Datasets
--------
We obtained a data set of TLS network traffic from the ZMap project [@zmap13]. Initially, we extracted 16,240 TCP streams on standard port 443 from 800MB of raw traffic data and further reduced to 5,794 completed and validated TLS streams[^2]. Then, we extracted from those 5,794 streams the 1,378 streams that used one of the DHE-RSA-$\ast$ ciphersuites in Table \[tbl:cipher\]. We split 1,378 instances into two sets: the *d00200* set of 218 instances for parameter selection and signature refinement and the test set *d00300* of 1,160 instances for the testing of the final signature, denoted as the *d00015* set. We also extracted 1,204 TLS instances with other ciphersuites. We extracted 337 Nugache traffic streams from a set of raw Nugache traffic and divided instances into two groups: 162 instances of training set and 175 instances of testing set. Similar to TLS, we use the training set to tune the fingerprint and the testing set for validation.
**Cipher** ID **Name**
--------------- -----------------------------------------
0x00015 `TLS_DHE_RSA_WITH_DES_CBC_SHA`
0x00016 `TLS_DHE_RSA_WITH_3DES_EDE_CBC_SHA`
0x00033 `TLS_DHE_RSA_WITH_AES_128_CBC_SHA`
0x00039 `TLS_DHE_RSA_WITH_AES_256_CBC_SHA`
0x00045 `TLS_DHE_RSA_WITH_CAMELLIA_128_CBC_SHA`
0x00067 `TLS_DHE_RSA_WITH_AES_128_CBC_SHA256`
0x0006B `TLS_DHE_RSA_WITH_AES_256_CBC_SHA256`
0x00088 `TLS_DHE_RSA_WITH_CAMELLIA_256_CBC_SHA`
0x0009A `TLS_DHE_RSA_WITH_SEED_CBC_SHA`
0x0009E `TLS_DHE_RSA_WITH_AES_128_GCM_SHA256`
0x0009F `TLS_DHE_RSA_WITH_AES_256_GCM_SHA384`
: TLS Ciphersuites of Choice: DHE-RSA-\*[]{data-label="tbl:cipher"}
\
**Port** 80 25 22 143 21 111 179 139 110
-------------------- -------- --------- --------- --------- --------- ---------- --------- ---------- ---------
\# of Streams 582 189 168 125 96 44 18 5 5
\[tbl:unswnb\]
In addition, we used 3,412 non-TLS TCP streams from a data set generated by UNSW-NB15[@moustafa2015unsw]. This data set contains a variety of traffic types, but without any TLS traffic so we can use it as another dimension of negative cases for testing the fingerprints. Table above shows the traffic type of the majority by service ports, only including standard ports under 1024. The table does not show the whole spectrum of traffic types in this dataset, but rather provides a quick look. More details on this data set are available in the original paper.
TLS
---
We test the signature generated as previously described over the training set *dhe00200* with thresholds of the confidence $\rho$ above 99.2% for different measures. The results shown in table \[tbl:origsig\] do not seem promising at all, of all the best results from 1-4-8 and 1-2-4-8 only reach a recall rate, 62.84% and 64.22% with the confidence of 99.85% respectively, but it does confirm that the strategy of using multiple $\tau$-measures significantly improves the recall rate. Also, it is interesting to notice that the rate of multiple $\tau$-measures drops significantly below 10% with a confidence of 99.99%, which is reasonable because the threshold is too relaxed (with a higher proportion of high entropy blocks) to be accurate.
99.20% 99.85% 99.97% 99.99%
-------------- ------------ ------------- ------------- ------------- -- -- --
1-bit 8.72% 36.70% 26.15% 26.14%
2-bit 15.13% 10.55% 23.39% 23.39%
4-bit 47.25% 25.68% 8.26% 11.93%
8-bit 42.40% 28.44% 3.21% 10.09%
1-2-8 31.19% 17.43% 7.80% 4.58%
1-4-8 45.41% **62.84%** 38.99% 5.50%
1-2-4-8 39.44% **64.22%** 39.44% 5.05%
: Recall rate of the original signature for TLS[]{data-label="tbl:origsig"}
By manually checking those failures, we found three major issues of our original signature. One is the range of the server random bytes. It was a little bit tighter than it appeared, which is previously set to be (+, 8, 54) as we used the range estimated from client random bytes. It turns out to be inadequate as the bytes after the server random bytes appear more random than those after client random bytes, and therefore more likely produce a longer high-entropy block. Following the same method as we did for client random bytes, we increase the maximum length to 64. The second major issue is that we failed to consider optional random bytes such as key identifier fields for both issuer and subject of the certificate. The third one relies on the fact that two high-entropy areas might be adjacent to each other without a sufficient gap and get merged to a larger high-entropy area, e.g. the signature of certificate and the server key exchange parameters. For the later two cases, we introduce optional blocks to the signature making the signature scalable. In the regular expression, we can include optional strings. For instance, our TLS signature has been extended to include optional strings as below. This adjustment boosts the recall for most cases, as shown in Table \[tbl:newsig\]. For both cases of 1-4-8 and 1-2-4-8, the recall increases by around 20%. [$$`...1\{80,260\}(0\{20,1024\}|\{8,160\}(1\{8,70\}|\{8,70\}0\{0,300\}1\{8,70\})0\{0,500\})...'.$$]{}
99.20% 99.85% 99.97% 99.99%
-------------- ------------- ------------- -------------- -------------- -- -- --
1-bit 12.39% 6.88% 40.37% 40.37%
2-bit 21.10% 19.27% 38.53% 40.83%
4-bit 84.40% 73.39% 33.03% 13.30%
8-bit 67.43% 51.83% 11.01% 16.51%
1-2-8 41.28% 25.69% 18.34% 11.93%
1-4-8 55.05% **82.57%** 67.43% 12.39%
1-2-4-8 49.54% **87.61%** 67.43% 12.39%
: Recall using refined fingerprint[]{data-label="tbl:newsig"}
The noise threshold is used to remove false positives and make the fingerprint more reliable. As the threshold increases, the detection accuracy of high-entropy blocks will increase as we are eliminating those accidental “high-entropy” blocks. At a certain point, this elimination may hurt the effectiveness as true high-entropy blocks may be eliminated by such an excessively large threshold. We experimented with different filter thresholds $\xi$ using a 4-bit measure, as shown in figure \[fig:4bit\]. Given its initial purpose, this parameter should be kept as small as possible for effective filtering. Thus, $\xi=9$ is chosen based on the empirical results. As suggested by our test results, it appears to be a proper choice for other measures, e.g. 1-4-8 measure.
![Noise Threshold Selection over TLS traffic using a 4-bit measure[]{data-label="fig:4bit"}](./b4xi.png)
### Test Results
After two improvement procedures, i.e. signature refinement and parameter selection, the ultimate test over the testing sets, *d00300*, is shown in the table below. The multiple $\tau$-measure 1-4-8 now produces a good recall rate.
$\xi$=9 TP FN Recall
----------------------------------- -------- --------- --------------
4-bit measure ($\rho$=99.20%) 1056 104 91.03%
1-4-8 measure ($\rho$=99.85%) 1079 81 93.02%
\[tbl:newsigtest\]
Dataset Total Positive Negative
-------------------------------------- ----------- --------------- -------------- --
*d00200*: TLS w/ selected Cipher 1,160 1,079 81
*d00300*: TLS w/ other Cipher 1,204 61 1,143
*d00015*: non-TLS 3,412 0 3,412
: TLS signature over different datasets []{data-label="tbl:signug"}
Finally, we fixed our noise threshold $\xi=9$ and used the 1-4-8 measure. We summarize our results over three datasets as follows. Overall, the TLS signature has a precision of nearly 94.6% and its accuracy is around 94%, only including negative cases from *d00300* so as to have a equivalent size of positive cases. On the other hand, negative cases from non-TLS, i.e. *d00015*, turn out to be relatively trivial even though some of instances do contain high-entropy traffic, for example, SSH on port 22.
Application on Botnet Detection: Nugache
----------------------------------------
The Nugache botnet, was one of the first peer-to-peer botnets to use strong cryptography to protect its C&C channel, as the inter-peer communication was encrypted using individually negotiated session keys derived using a hybrid RSA/Rijndael scheme [@sok:p2pwned; @dd2008:malware; @art:stormnugache]. Specifically, Nugache uses a two-way RSA-like key exchange protocol for every session with a minimum length of 512 bits for the modulus. That is, one peer sends the length of the key to announce a peer key exchange, followed by an actual key [@dd2008:malware]; the other peer in turn replies with a message of the same length encrypted with that public key. Compared to TLS, signature extraction for Nugache is much easier because of the simplicity of its key exchange. Since there is little control information in key exchange messages, if consider the payload only, the signature can be simply defined as 1\*, meaning high-entropy blocks everywhere, which is also a strong detectable characteristic distinct from other cryptographic protocols. Following the same consideration, we choose $\xi=9$, which yields a fair recall rate.
99.20% 99.85% 99.97% 99.99%
--------- ------------ ------------ -------- -------- -- -- --
2-bit **92.21%** 67.90% 39.50% 17.28%
1-2-8 88.27% **90.12%** 73.46% 56.17%
1-4-8 89.51% **92.21%** 75.93% 56.17%
1-2-4-8 **90.12%** **95.06%** 77.16% 56.17%
: Recall on Nugache (N=32)[]{data-label="tbl:signug"}
The initial fingerprint we generated for Nugache includes two high-entropy areas, corresponding to the two-way key exchange. First, we test all $\tau$-bit measures with a fixed noise threshold value $\xi$=9. It shows the 2-bit measure produces good results (92.21% with $\rho=99.20\%$) but meanwhile our voting mechanism clearly outperforms a single $\tau$-bit measure given the same level of confidence. We conservatively choose the 1-4-8 measure as our metric in a general.
Dataset Desc Total Positive Negative
------------------------ --------------- ----------- -------------- --------------
- *Nugache* 175 162 13
*d00200, d00300* TLS 2,364 0 2,364
*d00015* non-TLS 3,412 0 3,412
: Nugache fingerprint over different datasets []{data-label="tbl:finalug"}
In Table \[tbl:finalug\], we summarize our testing results of the Nugache signature over three datasets as follows. It is encouraging that the Nugache signature generates no false positive and so has a precision of 100%. For obfuscation techniques, there still a portion of the traffic, although small, will appear to have low entropy.
Limitations & Future Work
=========================
One may argue that high entropy does not necessarily imply encryption, compressed data, or multimedia data. The critical point is the distribution of high-entropy data blocks not solely the presence of high-entropy data. A study [@ZhangPM13] provides evidence against such “common sense,” where it was shown that multimedia files could yield low entropy instead, although the authors also pointed out that in some cases compressed files do have high entropy. Such cases require a much closer look, which we left for future work. Furthermore, encodings, e.g. base64 [@base64:rfc4648], can significantly reduce the entropy of a string. For this case, we assume that a base64 detector as well as a decoder could be deployed to canonicalize the traffic data. It is also possible that one could easily inject arbitrary bytes to disturb the original distribution of high entropy and low entropy. In this case, we consider it to be a new protocol for which the traffic could be possibly fingerprinted, e.g. using optional units as we did for TLS. If the signature generation process is automated, then this approach would still be efficient. However, if more advanced obfuscation techniques [@obfsproxy; @ccs2013-fte] are applied, then our approach will fail at identifying the obfuscated protocol. Nevertheless, our proposed techniques may be still used to detect the obfuscation techniques themselves. To avoid being fingerprinted, malware could adopt plain TLS instead of customizing the protocol, running the risk of SSL inspection. It may explain why there only 10% of malware samples indeed utilize TLS. Nevertheless, the work [@anderson2016deciphering] also found that malware or botnets utilize TLS in a very customized way, i.e. advertising significantly much fewer cipher suites than enterprise TLS clients. A shorter list of cipher suites will reduce the control information (i.e. low-entropy blocks) and therefore may end with different fingerprints than enterprise-grade TLS clients. Investigating how effective our approach would be in such a scenario is left for future work. Under certain circumstances, it is possible that our approach may not be sufficient to rule out all possible false positives and we would recommend to coordinate with other tools for reducing false positives.
Last but not least, we are interested in looking at more diverse data, such as compressed data, SSH, and other malware traffic.
Conclusion
==========
In this paper, we proposed a novel voting-based method for accurately detecting high-entropy blocks, e.g. key material, in a network traffic stream, and a method based on regular expressions for generating a scalable fingerprint based on identified high-entropy blocks. Our approach can effectively put malware authors on the defense, as a longer key used for a more securely encrypted connection would make it more easily characterized and therefore more detectable. However, if a shorter key is used for making the connection less vulnerable to detection, then they would only achieve a less secure connection.
[^1]: Control information is commonly known to have low entropy.
[^2]: A large portion of hosts scanned by the ZMap client did not respond or reject connections for various reasons during TLS negotiation
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We discuss the existence and uniqueness of wavefunctions for inhomogenoeus boundary value problems associated to $x^2y^2$-type matrix model on a bounded domain of $\mathbb{R}^2$. Both properties involve a combination of the Cauchy-Kovalewski Theorem and a explicit calculations.'
author:
- 'L Boulton$^{1}$, M P Garcia del Moral$^{2}$ and A Restuccia$^{2}$'
title: Arguments towards the construction of a matrix model groundstate
---
Introduction
============
In this note we sketch a technique for determining the existence and uniqueness of groundstates of supersymmetric matrix models with global symmetry, subject to Dirichlet boundary conditions. We consider one of the simplest possible benchmark models available. This allows us to illustrate the actual procedure on a concrete setting. The hamiltonian, given by (\[hamil\]), is the canonical toy model which is often used to test spectral properties of the $D=11$ supermembrane. In the following we call it the $x^2y^2$-model.
We believe the approach we present below is new in the context of matrix models and its scope of applicability includes a wide range of physical settings. Wavefunctions of matrix models have been investigated by means of different methods in the past. They provide a better understanding of M-theory, either from the point of view of Supermembrane Theory [@dwhn; @hoppe; @fh] or from the point of view of other matrix models [@bfss]. They also play a role in Supersymmetric Yang-Mills theories in the slow-mode regime [@halpern; @yi; @porrati; @stern].
The $x^2y^2$-model
==================
The $x^2y^2$-model was first introduced in [@dwln]. In [@xy] it was shown that zero is not an eigenvalue. Numerical tests carried out in [@bgmr3] suggest the existence of more general Weyl-type wavefunctions. See also [@korcyl]. We now show analytically that the inhomogeneous boundary value problem associated to a zero eigenvalue for Dirchlet boundary conditions on a compact domain $\Omega$ has a solution which is unique. See (\[bvp\])-(\[bc\]). Without loss of generality we assume that $\Omega$ is a two-dimensional ball of finite radius $R>0$.
The supersymmetric hamiltonian to be considered is $$\label{hamil}
H=p_x^2+p_y^2+x^2y^2+x\sigma_3+y\sigma_1$$ where $\sigma_i$ are the Pauli matrices. The supersymmetric charges in this case are given by the expression $$Q=Q^{\dag}=
\left( \begin{array}{cc} -xy & i\partial_x-\partial_y\\
i\partial_x-\partial_y & xy\end{array} \right).$$ The wavefunctions are $\Psi=\left(\begin{array}{cc} \Psi_1 \\ \Psi_2 \end{array}\right)\in H^2(\Omega)\cap H^1_0(\Omega)$. Note they are subject to zero boundary conditions, $$\label{bc}
\Psi=0\quad\textrm{on}\quad \partial\Omega.$$
Existence and uniqueness of wavefunctions at zero
=================================================
The resolvent of $H$ is compact, so the inhomogeneous system $$\label{bvp}
H\Psi=\Phi$$ for regular enough right hand side $\Phi$ will have a unique solution if and only if zero is not an eigenvalue of $H$. We show that this is indeed the case.
Assume that $$\label{zero}
H\Psi=0.$$ Then $$\label{vanishingboth}
Q\Psi=Q^\dag\Psi=0 \quad \textrm{in}\quad \Omega.$$ According to the regularity properties of the elliptic operators, $H$ in this case, this condition holds true pointwise up to the boundary of $\Omega$.
Let $\mathbf{x}\in \partial\Omega$ and denote by $(\mathbf{n}_1,\mathbf{n}_2)$ the components of the normal to $\partial\Omega$ at $\mathbf{x}$. The tangent to $\partial\Omega$ at $\mathbf{x}$ is then $(\mathbf{n}_2,-\mathbf{n}_1)$ and we must have $$(\mathbf{n}_2\partial_x-\mathbf{n}_1\partial_y)\Psi(\mathbf{x})=0.$$
The solution of problem (\[zero\]) is regular, so we can extend it continuously up to the boundary. Then (\[vanishingboth\]) yields $$(i\partial_x+\partial_y)\Psi_2(\mathbf{x})=(i\partial_x-\partial_y)\Psi_1(\mathbf{x})=0$$ pointwise. Since $(\mathbf{n}_1,\mathbf{n}_2)\neq 0$, if $\mathbf{n}_2\ne 0$, $$\left(1+i\frac{\mathbf{n}_1}{\mathbf{n}_2}\right)\partial_y\Psi_2(\mathbf{x})=0 \quad \mathrm{implies}\quad \partial_y\Psi_2(\mathbf{x})=0\textrm{ and }\partial_x\Psi_2(\mathbf{x})=0,$$ and $$\left(-1+i\frac{\mathbf{n}_1}{\mathbf{n}_2}\right)\partial_y\Psi_1(\mathbf{x})=0 \quad \mathrm{implies}\quad \partial_y\Psi_1(\mathbf{x})=0\textrm{ and } \partial_x\Psi_1(\mathbf{x})=0.$$ A similar conclusion is obtained for $\mathbf{n}_1\ne 0$. Hence $\Psi$ and $\partial_{\mathbf{n}}\Psi$ must vanish on the boundary $\partial\Omega$.
Now, by virtue of the Cauchy-Kovalewski Theorem, cf. [@Folland], and from the fact that the potential is analytic, we conclude that $\Psi=0$ pointwise on the whole of $\overline{\Omega}$. That is, the only solution to (\[zero\]) is the zero wavefunction.
In a paper soon to be completed [@bgmr4], we will examine the existence and uniqueness of the solution of boundary value problems associated to the $SU(N)$ regularized model of the $D=11$ supermembrane on a bounded domain. This theory was introduced in [@dwhn] for an unbounded domain. We expect to being able to examine the existence of massless groundstates in supersymmetric matrix models with global and gauge symmetries.
Discussion
==========
The $x^2y^2$-model [@dwln] is a well-known benchmark for testing spectral properties of the $D=11$ supermembrane. The existence of a groundstate was analyzed in [@xy] by considering the whole space $\mathbb{R}^2$. In the latter case the model does not admit a solution [@xy]. We have considered this system confined to a ball of finite radius. In this case zero is also not an eigenvalue for the space of wavefunctions which on the boundary. Additionally we show that the inhomogeneous boundary value problem associated to zero eigenvalue has always a solution which is unique. We consider a viewpoint to determine the existence of groundstate wavefunctions which we believe is new in the context of matrix models. The conditions of supersymmetry and ellipticity of the hamiltonian operator are essential to this analysis. The results announced in this note can be further generalized to more interesting settings. A rigorous general framework in this respect is currently being consider [@bgmr4].
References {#references .unnumbered}
==========
[9]{}
B. de Wit, J. Hoppe, H. Nicolai, Nucl. Phys. [**B305**]{} (1988) 545. Jens Hoppe, [*On The Construction of Zero Energy States in Supersymmetric Matrix Models III*]{} arXiv:hep-th/9711033. J. Froehlich, J. Hoppe, [*On Zero-Mass Ground States in Super-Membrane Matrix Models*]{} arXiv:hep-th/9701119. T. Banks, W. Fischler, S.H. Shenker and L. Susskind, Phys.Rev.[**D55**]{} (1997) 5112-5128. M. Claudson, M. B. Halpern, Nucl.Phys. B [**250**]{} (1985) no. 4, 689. P. Yi, [*Nucl. Phys. B*]{} (1997) [**505**]{}, 307. M. Porrati, A. Rozenberg, [*Nucl. Phys. B*]{} (1998) [**515**]{}, 184-202. S. Sethi, M. Stern, [*Commun. Math. Phys.*]{} (1998) [**194**]{}, 675. B. de Wit, M. Luscher, H. Nicolai, Nucl.Phys. [**B320**]{} (1989) 135 G.M. Graf, D. Hasler, J. Hoppe, [*No zero energy states for the supersymmetric $x^2y^2$ potential*]{}, arXiv:math-ph/0109032. L. Boulton, M.P. Garcia del Moral, Alvaro Restuccia, [*Nucl.Phys. B*]{} (2012) [**856**]{}, 716-747. P. Korcyl, [*Phys.Rev.D*]{} (2006) [**74**]{},115012. G. Folland, Introduction to Partial Differential Equations. Second Edition. Princeton University Press, Princeton, 1995. L. Boulton, M.P. Garcia del Moral, A. Restuccia, In preparation.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
The formation and motion of lattice defects such as cracks, dislocations, or grain boundaries, occurs when the lattice configuration loses stability, that is, when an eigenvalue of the Hessian of the lattice energy functional becomes negative. When the atomistic energy is approximated by a hybrid energy that couples atomistic and continuum models, the accuracy of the approximation can only be guaranteed near deformations where both the atomistic energy as well as the hybrid energy are stable. We propose, therefore, that it is essential for the evaluation of the predictive capability of atomistic-to-continuum coupling methods near instabilities that a theoretical analysis be performed, at least for some representative model problems, that determines whether the hybrid energies remain stable [*up to the onset of instability of the atomistic energy*]{}.
We formulate a one-dimensional model problem with nearest and next-nearest neighbor interactions and use rigorous analysis, asymptotic methods, and numerical experiments to obtain such sharp stability estimates for the basic conservative quasicontinuum (QC) approximations. Our results show that the consistent quasi-nonlocal QC approximation correctly reproduces the stability of the atomistic system, whereas the inconsistent energy-based QC approximation incorrectly predicts instability at a significantly reduced applied load that we describe by an analytic criterion in terms of the derivatives of the atomistic potential.
address:
- 'School of Mathematics, 206 Church St. SE, University of Minnesota, Minneapolis, MN 55455, USA'
- 'Mathematical Institute, 24–29 St. Giles’, Oxford OX1 3LB, UK'
author:
- 'M. Dobson'
- 'M. Luskin'
- 'C. Ortner'
title: |
Accuracy of Quasicontinuum Approximations\
Near Instabilities
---
[^1]
Introduction
============
An important application of atomistic-to-continuum coupling methods is the study of the quasistatic deformation of a crystal in order to model instabilities such as dislocation formation during nanoindentation, crack growth, or the deformation of grain boundaries [@Miller:2003a]. In each of these applications, the quasistatic evolution provides an accurate approximation of the crystal deformation until the evolution approaches an unstable configuration. This occurs, for example, when a dislocation forms or moves or when a crack tip advances. The crystal will then typically undergo a dynamic process until it reaches a new stable configuration. In order to guarantee an accurate approximation of the entire quasistatic crystal deformation, up to the formation of an instability, it is crucial that the equilibrium in the atomistic/continuum hybrid method is stable whenever the corresponding atomistic equilibrium is. The purpose of this work is to investigate whether the quasicontinuum (QC) method has this property. In technical terms, this requires sharp estimates on the stability constant in the QC approximation.
The QC method is an atomistic-to-continuum coupling method that models the continuum region by using an energy density that exactly reproduces the lattice-based energy density at uniform strain (the Cauchy-Born rule) [@Miller:2003a; @Ortiz:1995a; @Shenoy:1999a]. Several variants of the QC approximation have been proposed that differ in how the atomistic and continuum regions are coupled [@Dobson:2008a; @E:2006; @Miller:2003a; @Shimokawa:2004]. In this paper, we present sharp stability analyses for the main examples of conservative QC approximations as a means to evaluate their relative predictive properties for defect formation and motion. Our sharp stability analyses compare the loads for which the atomistic energy is stable, that is, those loads where the Hessian of the atomistic energy is positive-definite, with the loads for which the QC energies are stable. It has previously been suggested and then observed in computational experiments that inconsistency at the atomistic-to-continuum interface can reduce the accuracy for computing a critical applied load [@Shenoy:1999a; @Miller:2003a; @Miller:2008; @E:2006]. In this paper, we give an analytical method to estimate the error in the critical applied load by deriving stability criteria in terms of the derivatives of the atomistic potential.
Although we present our techniques in a precise mathematical format, we believe that these techniques can be utilized in a more informal way by computational scientists to quantitatively evaluate the predictive capability of other atomistic-to-continuum or multiphysics models as they arise. For example, our quantitative approach has the potential to estimate the reduced critical applied load in QC approximations such as the quasi-nonlocal QC approximation (QNL), that are consistent for next-nearest interactions but not for longer range interactions. Since the longer range interactions are generally weak, such an estimate may give an analytical basis to judging that the reduced critical applied load for QNL with finite range interactions is within an acceptable error tolerance.
The accuracy of various QC approximations and other atomistic-to-continuum coupling methods is currently being investigated by both computational experiments and numerical analysis [@BadiaParksBochevGunzburgerLehoucq:2007; @Legoll:2005; @Dobson:2008c; @Dobson:2008b; @dobs-qcf2; @E:2005a; @Gunzburger:2008a; @Gunzburger:2008b; @LinP:2003a; @LinP:2006a; @luskin-cluster-2008; @mingyang; @Ortner:2008a; @PrudhommeBaumanOden:2005]. The main issue that has been studied to date in the mathematical analyses is the rate of convergence with respect to the smoothness of the continuum solution (however, see [@Legoll:2005; @dobs-qcf2; @Ortner:2008a] for analyses of the error of the QC solutions with respect to the atomistic solution, possibly containing defects). Some error estimates have been obtained that give theoretical justification for the accuracy of a QC approximation for all loads up to the critical atomistic load where the atomistic model loses stability [@Dobson:2008b; @Ortner:2008a], but other error estimates that have been presented do not hold near the atomistic limit loads. It is important to understand whether the break-down of these error estimates is an artifact of the analysis, or whether the particular QC approximation actually does incorrectly predict an instability before the applied load has reached the correct limit load of the atomistic model.
Two key ingredients in any approximation error analysis are the consistency and stability of the approximation scheme. For energy minimization problems, consistency means that the truncation error for the equilibrium equations is small in a suitably chosen norm, and stability is usually understood as the positivity of the Hessian of the functional. For the highly non-convex problems we consider here, stability must necessarily be a local property: The configuration space can be divided into stable and unstable regions, and the question we ask is whether the stability regions of different QC approximations approximate the stability region of the full atomistic model in a way that can be controlled in the setup of the method (for example, by a judicious choice of the atomistic region).
In this work, we initiate such a systematic study of the stability of QC approximations. In the present paper, we investigate conservative QC approximations, that is, QC approximations which are formulated in terms of the minimization of an energy functional. In a companion paper [@doblusort:qcf.stab], we study the stability of a force-based approach to atomistic-to-continuum coupling that is nonconservative.
In computational experiments, one often studies the evolution of a system under incremental loading. There, the critical load at which the system “jumps” from one energy well to another is often the goal of the computation. Thus, we will also study the effect of the “stability error” on the error in the critical load.
We will formulate a simple model problem, a one dimensional periodic atomistic chain with pairwise next-nearest neighbour interactions of Lennard-Jones type, for which we can analyze the issues layed out in the previous paragraphs. It is well known that the uniform configuration is stable only up to a critical value of the tensile strain (fracture). We use analytic, asymptotic, and numerical approaches to obtain sharp results for the stability of different QC approximations when applied to this simple model.
In Section \[sec:model\], we describe the model and the various QC approximations that we will analyze. In Section \[sec:analysis\], we study the stability of the atomistic model as well as two consistent QC approximations: the local QC approximation (QCL) and the quasi-nonlocal QC approximation (QNL). We prove that the critical applied strains for both of these approximations are equal to the critical applied strain for the atomistic model, up to second-order in the atomistic spacing.
A similar analysis for the inconsistent QCE approximation is more difficult because the uniform configuration is not an equilibrium. Thus, in Section \[sec:ana:qce\], we construct a first-order correction of the uniform configuration to approximate an equilibrium configuration, and we study the positive-definiteness of the Hessian for the linearization about this configuration. We explicitly construct a test function with strain concentrated in the atomistic-continuum interface that is unstable for applied strains bounded well away from the atomistic critical applied strain.
In Section \[sec:discus\], we analyze the accuracy in predicting the critical strain for onset of instability. For the QCL and QNL approximations, this involves comparing the effect of the difference between their modified stability criteria and that of the atomistic model. For QCE, since the solution to the nonlinear equilibrium equations are non-trivial, we provide computational results in addition to an analysis of the critical QCE strain predicted by the approximations derived in Section \[sec:ana:qce\].
The atomistic and quasicontinuum models {#sec:model}
=======================================
The atomistic model problem {#sec:intro:model_problem}
---------------------------
![\[fig:lj\]Lennard-Jones type interaction potential. The bond length $r_*$ is the turning point between the convex and concave regions of $\phi$.](lj2 "fig:"){height="5cm"}\
Suppose that the infinite lattice $\eps \Z$ is deformed uniformly into the lattice $y_F := F\eps \Z,$ where $F > 0$ is the macroscopic deformation gradient and where $\eps > 0$ scales the reference atomic spacing, that is, $$(y_F)_\ell:=F\ell\eps \quad\text{for }-\infty<\ell<\infty.$$ We admit $2N$-periodic perturbations $u = (u_\ell)_{\ell \in \Z}$ from the uniformly deformed lattice $y_F.$ More precisely, for fixed $N \in \N$, we admit deformations $y$ from the space $$\Ys_F := \big\{ y \in \R^\Z : y = y_F + u,\,
u \in \Us \big\},$$ where $\Us$ is the space of $2N$-periodic displacements with zero mean, $$\Us := \big\{ u \in \R^\Z : u_{\ell+2N} = u_\ell \text{ for }
\ell \in \Z, \text{ and } {\textstyle \sum_{\ell = -N+1}^N}
u_\ell = 0 \big\}.$$ We set $\eps = 1/N$ throughout so that the reference length of the periodic domain is fixed. Even though the energies and forces we will introduce are well-defined for all $2N$-periodic displacements, we require that they have zero mean in order to obtain locally unique solutions to the equilibrium equations. These zero mean constraints are an artifact of our periodic boundary conditions and are similarly used in the analysis of continuum problems with periodic boundary conditions.
We assume that the [*stored energy per period*]{} of a deformation $y
\in \Ys_F$ is given by a next-nearest neighbour pair interaction model, $$\label{bond}
\E_{\rm a}(y) := \eps \sum_{\ell = -N+1}^N \big(\phi(y_\ell')
+ \phi(y_\ell' + y_{\ell+1}')\big),$$ where $v_\ell'$ is the backward difference $$v_\ell' := \eps^{-1}(v_\ell - v_{\ell-1})\qquad\text{for }v \in \R^\Z,\
\ell \in \Z,$$ and where $\phi$ is a Lennard-Jones type interaction potential satisfying (see also Figure \[fig:lj\])
1. $\phi \in \CC^4((0, +\infty); \R)$,
2. there exists $r_* > 0$ such that $\phi$ is convex in $(0, r_*)$ and concave in $(r_*, +\infty)$, and
3. $\phi^{(k)}(r) \rightarrow 0$ rapidly as $r \nearrow \infty$, for $k = 0, \dots, 4$.
We have used the scaled interaction potential, [[$\eps\phi(r/\eps),$]{}]{} in the definition of the stored energy, $\E_{\rm a}(y),$ to obtain a continuum limit as $\eps\to 0.$ Assumptions (i) and (ii) are used throughout our analysis, while assumption (iii) serves primarily to motivate that next-nearest neighbour interaction terms are typically dominated by nearest-neighbour terms. Note, however, that even with assumption (iii), the relative size of next-nearest and nearest neighbour interactions is comparable when strains approach $r_*$.
We denote the first variation of the energy functional, $\E_{\rm
a}'(y)[u],$ at a deformation $y\in\Ys_F$ by $$\E_{\rm a}'(y)[u] :=\sum_{\ell = -N+1}^N \frac{\partial \E_{\rm a}(y)}{\partial y_\ell}u_{\ell}
=\eps \sum_{\ell = -N+1}^N \Big\{\phi'(y_\ell') u_\ell'
+\phi'(y_\ell'+y_{\ell+1}')(u_\ell' + u_{\ell+1}')\Big\},$$ for $u \in \Us$. In the absence of external forces, the uniformly deformed lattice $y = y_F$ is an equilibrium of the atomistic energy under perturbations from $\Us$, that is, $$\label{eq:DEa_yF}
\E_{\rm a}'(y_F)[u] = 0 \qquad \text{ for all } u \in \Us.$$
We identify the stability of $y_F$ with [*linear stability under perturbations from the space $\Us$*]{}. To make this precise, we denote the second variation of the energy functional, $\E_{\rm a}''(y)[u,v],$ evaluated at a deformation $y\in\Ys_F$, by $$\label{eq:defn_Ea_hess}
\begin{split}
\E_{\rm a}''(y)[u,v] &:=\sum_{\ell,\, m = -N+1}^N \frac{\partial^2 \E_{\rm a}(y)}{\partial y_\ell \partial y_m
}u_{\ell}v_m \\
&= \eps \sum_{\ell = -N+1}^N \Big\{
\phi''(y_\ell') u_\ell' v_\ell'
+ \phi''(y_\ell' + y_{\ell+1}') [u_\ell' + u_{\ell+1}'] [v_\ell' + v_{\ell+1}'] \Big\},
\end{split}$$ for all $u,v\in\Us$. The matrix $\Big(\frac{\partial^2 \E_{\rm
a}(y)}{\partial y_\ell \partial y_m }\Big)_{\ell, m = -N+1}^N$ is the Hessian for the energy functional. We say that the equilibrium $y_F$ is [*stable*]{} for the atomistic model if this Hessian, evaluated at $y = y_F$, is positive definite on the subspace $\Us$ of zero mean displacements, or equivalently, if $$\label{hess}
\E_{\rm a}''(y_F)[u,u] > 0 \qquad \text{ for all } u \in \Us \setminus\{0\}.$$ [[In Section \[sec:summary\], Definition \[def:stab\], we extend this definition of stability to the various QC approximation and their equilibria.]{}]{}
Note that if $y = y_F$, then $y_\ell' = F$ and $y_\ell' + y_{\ell+1}' =
2F$ for all $\ell$. Therefore, upon defining the quantities $$\phi_F'' := \phi''(F), \quad
\phi_{2F}'' := \phi''(2F), \quad \text{and} \quad
A_F = \phi_F'' + 4 \phi_{2F}'',$$ we can rewrite as follows $$\label{eq:Ea_hess_F}
\E_{\rm a}''(y_F)[u,u] = \eps \sum_{\ell = -N+1}^N \big\{ \phi_F'' |u_\ell'|^2
+ \phi_{2F}'' |u_\ell' + u_{\ell+1}'|^2 \big\} \quad \text{for } u\in\Us.$$ (We will use $A_F$ later.) The quantities $\phi_F''$ and $\phi_{2F}''$ will play a prominent role in the analysis of the stability of the atomistic model and its QC approximations and describe the strength of the nearest neighbor and next-nearest neighbor interactions, respectively. We similarly define the quantities $\phi^{(k)}_G$ for all $k \in \N$ and for all $G > 0$. For most realistic interaction potentials the second-nearest neighbour coefficient is non-positive, $\phi_{2F}''
\leq 0,$ except in the case of extreme compression (see Figure \[fig:lj\]). Therefore, in order to avoid having to distinguish several cases, we will assume throughout our analysis that $F \geq
r_*/2$. In this case, property (ii) of the interaction potential shows that $\phi_{2F}'' \leq 0$.
We also note that, for $u \in \Us$, both $u'$ and $u''$ are understood as $2N$-periodic chains, that is, $u', u'' \in \Us,$ where the centered second difference $u'' \in \Us$ is defined by $$u_\ell'' := \eps^{-2}(u_{\ell+1} - 2 u_\ell + u_{\ell-1})
\qquad\text{for }u \in \R^\Z,\
\ell \in \Z.$$ For $u, v \in \Us$, we also define the weighted $\ell^p$-norms $$\| v \|_{\ell^p_\eps} := \cases{
\displaystyle \Bigg( \sum_{\ell=-N+1}^N \eps |v_\ell|^p \Bigg)^{1/p}, &
1 \leq p < \infty, \\[10pt]
\displaystyle \max_{\ell = -N+1, \dots, N} |v_\ell|, & p = \infty,
}$$ as well as the weighted $\ell^2$-inner product $$\< u, v \> = \eps \sum_{\ell = -N+1}^N u_\ell v_\ell.$$
The local QC approximation (QCL) {#sec:model:qcl}
--------------------------------
Before we introduce different flavors of QC approximations, we note that we can rewrite the atomistic energy as a sum over the contributions from each atom, $$\begin{aligned}
\E_{\rm a}(y) =~& \eps \sum_{\ell = -N+1}^N E^a_\ell(y)
\qquad \text{where} \\
E^a_\ell(y) :=~& {{\textstyle \frac{1}{2}}} \big[ \phi(y_\ell')
+ \phi(y_{\ell+1}')
+ \phi(y_{\ell-1}' + y_\ell')
+ \phi(y_{\ell+1}'+y_{\ell+2}') \big].\end{aligned}$$
If $y$ is “smooth,” i.e., $y_\ell'$ varies slowly, then $E^a_\ell(y)
\approx E^c_\ell(y)$ where $$E^c_\ell(y) := {{\textstyle \frac{1}{2}}} \big[ \phi(y_\ell') + \phi(y_{\ell+1}')
+ \phi(2y_\ell') + \phi(2y_{\ell+1}') \big]
= {{\textstyle \frac{1}{2}}} \big[ \phi_{cb}(y_\ell') + \phi_{cb}(y_{\ell+1}') \big],$$ and where $\phi_{cb}(r) := \phi(r) + \phi(2r)$ is the so-called [ *Cauchy-Born stored energy density*]{}. In this case, we may expect that the atomistic model is accurately represented by the local QC (or continuum) model $$\label{qcldef}
\E_{\rm qcl}(y) := \eps \sum_{\ell = -N+1}^N E^c_\ell(y)
= \eps \sum_{\ell = -N+1}^N \phi_{cb}(y_\ell').$$ The main feature of this [*continuum*]{} model is that the next-nearest neighbour interactions have been replaced by nearest neighbour interactions, thus yielding a model with more [ *locality*]{}. Such a model can subsequently be coarse-grained (i.e., degrees of freedom are removed) which yields efficient numerical methods.
The energy-based QC approximation (QCE) {#sec:model:qce}
---------------------------------------
If $y_\ell'$ is “smooth” in the majority of the computational domain, but not in a small neighbourhood, say, $\{ -K, \dots, K \}$, where $K > 1$, then we can obtain sufficient accuracy and efficiency by coupling the atomistic model to the local QC model by simply choosing energy contributions $E^a_\ell$ in the [ *atomistic region*]{} $\As = \{-K, \dots, K\}$ and $E^c_\ell$ in the [*continuum region*]{} $\Cs = \{-N+1, \dots, N\} \setminus \As$. This approximation of the atomistic energy is often called the [*energy based QC approximation*]{} [@Ortiz:1995a] and yields the energy functional $$\begin{aligned}
\E_{\rm qce}(y) :=& \eps\sum_{\ell \in \Cs} E^c_\ell(y)
+ \eps \sum_{\ell \in \As} E^a_\ell(y).\end{aligned}$$
[[ It is now well-understood [@Dobson:2008a; @Dobson:2008c; @Dobson:2008b; @E:2006; @Shenoy:1999a] that the QCE approximation exhibits an inconsistency (“ghost force”) near the interface, which is displayed in the fact that $\E_{\rm qce}'(y_F)\ne 0 .$ The first remedy of this lack of consistency was the [*ghost force correction scheme*]{} [@Shenoy:1999a] which eventually led to the derivation of the force-based QC approximation [@Dobson:2008a] and which we analyze in [@doblusort:qcf.stab] and [@dobs-qcf2]. ]{}]{}
Quasi-nonlocal coupling (QNL) {#sec:model:qnl}
-----------------------------
An alternative approach was suggested in [@Shimokawa:2004], which requires a modification of the energy at the interface. This idea is best understood in terms of interactions rather than energy contributions of individual atoms (see also [@E:2006] where this has been extended to longer range interactions). The nearest neighbour interactions are left unchanged. A next-nearest neighbour interaction $\phi(\eps^{-1}(y_{\ell+1}-y_{\ell-1}))$ is left unchanged if at least one of the atoms $\ell+1, \ell-1$ belong to the atomistic region and is replaced by a Cauchy–Born approximation, $$\phi(\eps^{-1}(y_{\ell+1}-y_{\ell-1})) \approx
{{\textstyle \frac{1}{2}}} \big[ \phi(2y_\ell') + \phi(2y_{\ell+1}')]$$ if [*both*]{} atoms belong to the continuum region. This idea leads to the energy functional $$\begin{split}
\E_{\rm qnl}(y) :=\eps \sum_{\ell = -N+1}^N \phi(y_\ell')
+ \eps \sum_{\ell \in \As_{\rm qnl}} \phi(y_\ell' + y_{\ell+1}')
+ \eps \sum_{\ell \in \Cs_{\rm qnl}} {{\textstyle \frac{1}{2}}} \big[ \phi(2y_\ell')
+ \phi(2y_{\ell+1}')\big]
\end{split}$$ where $\As_{\rm qnl} = \{-K-1, \dots, K+1\}$ and $\Cs_{\rm qnl} =
\{-N+1,\dots, N\} \setminus \As_{\rm qnl}$ are modified atomistic and continuum regions. The QNL approximation is consistent, that is, $y = y_F$ is an equilibrium of the QNL energy functional. The label QNL comes from the original intuition of considering interfacial atoms as [ *quasi-nonlocal*]{}, i.e., they interact by different rules with atoms in the atomistic and continuum regions.
Stability of Quasicontinuum Approximations: Summary {#sec:summary}
===================================================
In this section, we briefly summarize the our main results.
We begin by giving a careful definition of a notion of stability. Our condition is slightly stronger than local minimality, which is the natural concept of stability in statics. However, an analysis of local minimality alone is usually not tractable. Moreover, for the deformations that we consider, our definition is in fact sufficiently general.
\[def:stab\] Let $\E : \Ys_F \to \R \cup \{+\infty\}$. We say that $y \in \Ys_F$ is a stable equilibrium of $\E$ if $\E$ is twice differentiable at $y$ and the following conditions hold:
- $\E'(y)[u] = 0$ for all $u \in \Us$,
- $\E''(y)[u,u] > 0$ for all $u \in \Us \setminus \{0\}$.
If only (i) holds, then we call $y$ a critical point of $\E$.
We focus on the deformation $y_F =
(F\ell\eps)_{\ell \in \Z}$ and ask for which macroscopic strains $F$ this deformation is a stable equilibrium. We know from that $y_F$ is a critical point of the atomistic energy $\E_{\rm a}$, and it is easy to see that $y_F$ is also a critical point of the QCL energy $\E_{\rm qcl}$ and of the QNL energy $\E_{\rm qnl}$. Our analysis in Section \[sec:analysis\] gives the following conditions under which $y_F$ is stable:
1. $y_F$ is a stable equilibrium of $\E_{\rm a}$ if and only if $A_F - \eps^2 \pi^2 \phi_{2F}'' + O(\eps^4) > 0$;
2. $y_F$ is a stable equilibrium of $\E_{\rm qcl}$ if and only if $A_F > 0$;
3. $y_F$ is a stable equilibrium of $\E_{\rm qnl}$ if and only if $A_F > 0$
where we recall that $A_F = \phi_F'' + 4 \phi_{2F}''$ is the continuum elastic modulus for the Cauchy–Born stored energy function $\phi_{cb}(r) = \phi(r) + \phi(2r)$. Points 1., 2., and 3. are established, respectively, in Propositions \[th:ana:stab\_a\], \[th:ana:qcl\_stab\], and \[th:ana:qnl\_stab\].
If we envision a quasistatic process in which $F$ is slowly increased, then we may wish to find the critical strain $F^*$ at which $y_F$ is no longer a stable equilibrium (fracture instability). If we denote the critical strains in the atomistic, QCL, and QNL models, respectively, by $F_{\rm a}^*$, $F_{\rm qcl}^*$ and $F_{\rm qnl}^*$, then 1.–3. imply that (cf. Section \[sec:discus\]) $$|F_{\rm a}^* - F_{\rm qcl}^*| = O(\eps^2) \quad \text{and} \quad
|F_{\rm a}^* - F_{\rm qnl}^*| = O(\eps^2).$$
For the QCE approximation defined in Section \[sec:model:qce\], the situation is more complicated. The occurrence of a “ghost force” in the QCE model implies that $y_F$ is [*not*]{} a critical point of $\E_{\rm qce}$, and consequently, we will need to analyze the stability of the second variation $\E_{\rm qce}''(y_{{\rm qce}, F})$ where $y_{{\rm qce}, F} \neq y_F$ is an appropriately chosen equilibrium of $\E_{\rm qce}$. Since $y_{{\rm qce}, F}$ solves a nonlinear equation, we will replace it by an approximate equilibrium in our analysis in Section \[sec:ana:qce\] where we obtain the following (simplified) result:
1. For $y_{{\rm qce}, F}$ to be a stable equilibrium of $\E_{\rm qce}$ it is necessary that $$1 + \frac{3 \phi_{2F}''}{2 \phi_F''}
+ \frac{\phi_F''' \phi_{2F}'}{2 |\phi_F''|^2} + O(\delta^2) > 0,$$ where $\delta = \max\{ |\phi^{(j)}(2F) / \phi''(F)|: j = 1,2,3
\}$ is assumed to be small.
We remark that 4. gives only a necessary but not a sufficient condition for stability of the QCE equilibrium $y_{{\rm qce}, F}$, which, moreover, depend on assumptions on the parameter $\delta$. We refer to Remark \[rem:delta\_discussion\] for a careful discussion of the role of $\delta$.
If we let $\tilde{F}_{\rm qce}^*$ denote the critical strain at which 4. fails (ignoring the $O(\delta^2)$ term), then we obtain $$| F_{\rm a}^* - \tilde{F}_{\rm qce}^* | = O(1),$$ which suggests that the QCE method is unable to predict the onset of fracture instability accurately. In Section \[sec:discus\], we confirm this asymptotic prediction with numerical experiments.
We have shown in [@qcf.iterative] that the stability properties of the ghost force correction scheme (GFC) can be understood for uniaxial tensile loading by considering the stability of the QC energy $$\label{gfc}
\E_{{\rm gfc},\, y_F}(y):=\E_{\rm qce}(y)-\E_{\rm qce}'(y_F)(y-y_F)
\quad\text{for all }y \in \Ys_F.$$ We note that $\E_{{\rm gfc},\, y_F}'(y_F)=0,$ so $y_F$ is an equilibrium of the $\E_{{\rm gfc},\, y_F}$ energy under perturbations from $\Us.$ We can therefore analyze the stability of $\E_{{\rm
gfc},\, y_F}(y)$ at $y_F$ by studying the Hessian $\E_{\rm
qce}''(y_F)=\E_{{\rm gfc},\, y_F}''(y_F).$ We show in Remark \[rmbounds\] in Section \[sec:model:qce\] that
1. $y_F$ is a stable equilibrium of $\E_{{\rm gfc},\, y_F}$ if and only if $A_F + \lambda_K
\phi_{2F}'' > 0$ where ${{\textstyle \frac{1}{2}}} \le \lambda_K \le 1.$
We give analytic and computation results in Sections \[sec:ana:qce\] and \[sec:discus\] showing that the ghost force correction scheme can be expected to improve the accuracy of the computation of the critical strain by the QCE method, that is, $$\tilde{F}_{\rm qce}^*<F_{\rm qce}^{y_F}<F_{\rm a}^*,$$ where $F_{\rm qce}^{y_F}$ is the critical strain at which $\E_{\rm qce}''(y_F)=\E_{{\rm gfc},\, y_F}''(y_F)$ is no longer positive definite, but the error in computing the critical strain by the GFC scheme is still $$| F_{\rm a}^* - F_{\rm qce}^{y_F} | = O(1).$$
Sharp Stability Analysis of Consistent QC Approximations {#sec:analysis}
========================================================
In this section, we analyze the stability of the atomistic model and two consistent QC approximations: the local QC approximation and the quasi-nonlocal QC approximation. In each case, we will give precise conditions on $F$ under which $y_F$ is stable in the respective approximation. The inconsistent energy-based QC approximation (QCE) is analyzed in Section \[sec:ana:qce\]. The corresponding result for QCE is less exact than for QCL and QNL, but shows that there is a much more significant loss of stability.
Atomistic model {#sec:ana:a}
---------------
Recalling the representation of $\E_{a}''(y_F)$ from and noting that $$\label{eq:rewrite_nnn}
|u_\ell'+u_{\ell+1}'|^2 = 2|u_\ell'|^2 + 2|u_{\ell+1}'|^2 - |u_{\ell+1}' - u_\ell'|^2,$$ we obtain $$\begin{aligned}
\notag
\E_{a}''(y_F)[u,u]
=~& \eps\sum_{\ell = -N+1}^N \phi_F'' |u_\ell'|^2
+ \eps \sum_{\ell = -N+1}^N \phi_{2F}'' \big(2|u_\ell'|^2 + 2|u_{\ell+1}'|^2 - |u_{\ell+1}' - u_\ell'|^2\big) \\
\notag
=~& \eps\sum_{\ell = -N+1}^N (\phi_{F}'' + 4 \phi_{2F}'') |u_\ell'|^2
+ \eps \sum_{\ell = -N+1}^N (-\eps^2\phi_{2F}'')|u_\ell''|^2 \\
\label{eq:Eq_second_diff_form}
=~& A_F \| u' \|_{\ell^2_\eps}^2 + (-\eps^2 \phi_{2F}'') \|u''\|_{\ell^2_\eps}^2.\end{aligned}$$
To quantify the influence of the strain gradient term, we define $$\mu_{\eps} := \inf_{\psi \in \Us \setminus\{0\}}
\frac{ \| \psi'' \|_{2} }{ \| \psi'\|_2 }.$$ Since $u$ is periodic, it follows that $u'$ has zero mean. In this case, the eigenvalue $\mu_{\eps}$ is known to be attained by the eigenfunction $\psi'_\ell=\sin(\eps \ell\pi)$ and is given by [@SuliMayers Exercise 13.9] $$\label{eq:defn_muaeps}
\mu_{\eps} = \frac{2 \sin(\pi\eps/2)}{\eps}.$$ Since $\sin(t) = t + O(t^3)$ as $t \searrow 0$, it follows that $\mu_\eps = \pi + O(\eps^2)$ as $\eps \searrow 0$. Thus, we obtain the following stability result for the atomistic model.
\[th:ana:stab\_a\] Suppose $\phi_{2F}'' \leq 0$. Then $y_F$ is stable in the atomistic model if and only if $A_F - \eps^2 \mu_{\eps}^{2} \phi_{2F}'' >
0$, where $\mu_{\eps}$ is the eigenvalue defined in .
By the definition of $\mu_{\eps}$, and using , we have $$\inf_{\substack{u \in \Us \\ \|u'\|_{\ell^2_\eps} = 1}} \E_{\rm a}''(y_F)[u,u]
= A_F - \eps^2 \phi_{2F}''
\inf_{\substack{u \in \Us \\ \|u'\|_{\ell^2_\eps} = 1}} \|u''\|_{\ell^2_\eps}^2
= A_F - \eps^2 \mu_\eps^2 \phi_{2F}''. \qedhere$$
The Local QC approximation
--------------------------
The equilibrium system, in variational form, for the QCL approximation is $$\E_{\rm qcl}'(y)[u] = \eps \sum_{\ell = -N+1}^N \big( \phi'(y_\ell') + 2\phi'(2 y_\ell') \big) u_\ell' = 0 \qquad \text{ for all } u \in \Us.$$ Since $u'$ has zero mean, it follows that $y = y_F$ is a critical point of $\E_{\rm qcl}$ for all $F$. The second variation of the local QC energy, evaluated at $y = y_F$, is given by $$\begin{aligned}
\E_{\rm qcl}''(y_F)[u,u] =~& \eps \sum_{\ell = -N+1}^N A_F |u_\ell'|^2
\qquad \text{ for } u \in \Us.\end{aligned}$$ Thus, recalling our definition of stability from Section \[sec:intro:model\_problem\], we obtain the following result.
\[th:ana:qcl\_stab\] The deformation $y_F$ is a stable equilibrium of the local QC approximation if and only if $A_F > 0$.
Comparing Proposition \[th:ana:qcl\_stab\] with Proposition \[th:ana:stab\_a\] we see a first discrepancy, albeit small, between the stability of the full atomistic model and the local QC approximation (or the Cauchy–Born approximation). In Section \[sec:discus\] we will show that this leads to a negligible error in the computed critical load.
Quasi-nonlocal coupling
-----------------------
By the construction of the QNL coupling rule at the interface, the deformation $y = y_F$ is an equilibrium of $\E_{\rm qnl}$ [@Shimokawa:2004]. The second variation of $\E_{\rm qnl}$ evaluated at $y =
y_F$ is given by $$\begin{split}
\E_{\rm qnl}''(y_F)[u,u] = \eps\sum_{\ell = -N+1}^N \phi_F'' |u_\ell'|^2
+~& \eps \sum_{\ell \in \As_{\rm qnl}} \phi_{2F}'' |u_\ell' + u_{\ell+1}'|^2 \\
+~& \eps \sum_{\ell \in \Cs_{\rm qnl}} 4 \phi_{2F}''
( {{\textstyle \frac{1}{2}}} |u_\ell'|^2 + {{\textstyle \frac{1}{2}}} |u_{\ell+1}'|^2 ).
\end{split}$$
We use to rewrite the second group on the right-hand side (the nonlocal interactions) in the form $$\eps \sum_{\ell = -K-1}^{K+1} \phi_{2F}'' |u_{\ell}'+u_{\ell+1}'|^2
= \eps \sum_{\ell = -K-1}^{K+1} \big(2\phi_{2F}'' (|u_\ell'|^2 + |u_{\ell+1}'|^2)
- \eps^2 \phi_{2F}'' |u_\ell''|^2\big),$$ to obtain $$\E_{\rm qnl}''(y_F)[u,u] = \eps \sum_{\ell = -N+1}^{N} A_F |u_\ell'|^2
+ \eps \sum_{\ell = -K-1}^{K+1} (-\eps^2 \phi_{2F}'') |u_\ell''|^2.$$ Except in the case $K \in \{N-1, N\}$, it now follows immediately that $y_F$ is stable in the QNL approximation if and only if $A_F > 0$.
\[th:ana:qnl\_stab\] Suppose that $K < N-1$ and that $\phi_{2F}'' \leq 0$, then $y_F$ is stable in the QNL approximation if and only if $A_F > 0$.
Stability Analysis of the Energy-based QC approximation {#sec:ana:qce}
=======================================================
We will explain in Remark \[rmbounds\] that there exists ${{\textstyle \frac{1}{2}}}
\le \lambda_K \le 1$ such that $$\E_{\rm qce}''(y_F) \quad
\text{is positive definite if and only if}\quad
A_F + \lambda_K \phi_{2F}'' > 0.$$ However, $y_F$ is not a critical point of $\E_{\rm qce},$ so we must be careful in extending the previous definition of stability to the QCE approximation. We cannot simply consider the positive-definiteness of $\E_{\rm qce}''(y_F).$ Instead, we analyze the second variation $\E_{\rm qce}''(y_{{\rm qce}, F})$ where $y_{{\rm qce}, F} \in \Ys_F$ solves the QCE equilibrium equations $$\frac{\partial \E_{\rm qce}}{\partial y_\ell} (y_{{\rm qce}, F}) = 0 \quad
\text{for } \ell = -N+1, \dots, N,$$ or equivalently $$\label{eq:qce:nonlin}
\E_{\rm qce}'(y_{{\rm qce}, F})[u] = 0 \qquad \text{ for all } u \in \Us.$$ We will see that, when the second-neighbour interactions are small compared with the first neighbour interactions (which we make precise in Lemma \[th:asymptotic\_expansion\]), there is a locally unique solution $y_{{\rm qce}, F}$ of the equilibrium equations, which is the correct QCE counterpart of $y_F$. We will then derive a stability criterion for the equilibrium deformation $y_{{\rm qce}, F}.$
In Proposition \[th:qce:final\_stab\_result\] below, we derive an upper bound for the coercivity of $\E_{\rm qce}''(y_{{\rm qce}, F})[u,\,u]$ with respect to the norm $\| u'\|_{\ell^2_\eps}.$ Even though the derivation of this upper bound is only rigorous for strains bounded away from the atomistic critical strain, it clearly identifies a source of instability that cannot be found by analyzing, for example, $\E_{\rm qce}''(y_F)$. Moreover, we will present numerical experiments in Section \[sec:discus\] showing that the critical strain predicted in our following analysis gives a remarkably accurate approximation to actual QCE critical strain.
In Section \[sec:discus\], we consider the critical strain for each approximation, namely the point at which the appropriate equilibrium deformation (either $y_F$ or $y_{{\rm qce},F}$) becomes unstable. We will see later in this section, as well as in Section \[sec:discus\] that predicting the loss of stability for the QCE approximation using $y_F$ greatly underestimates the error in approximating the atomistic critical strain by the QCE critical strain.
Due to the nonlinearity and nonlocality of the interaction law, we cannot compute $y_{{\rm qce}, F}$ explicitly. Instead, we will construct an approximation $\hat{y}_{{\rm qce},F}$ which is accurate whenever second-neighbour terms are dominated by first-neighbour terms. In the following paragraphs, we first present a semi-heuristic construction, motivated by the analysis in [@Dobson:2008c], and then a rigorous approximation result, the proof of which is given in Appendix \[sec:app\_qce\_eq\].
In (\[eq:app:DEqce\]) in the appendix, we provide an explicit representation of $\E_{\rm qce}'$. Inserting $y = y_F$, we obtain a variational representation of the atomistic-to-continuum interfacial truncation error terms that are often dubbed “ghost forces,” $$\label{eq:qce:ghost_force}
\begin{split}
\E_{\rm qce}'(y_F)[u] &~= \eps {{\textstyle \frac{1}{2}}} \phi_{2F}' \big\{
u_{-K-1}' - u_{-K+1}' - u_K' + u_{K+2}' \big\} \\
&~:= - \phi_{2F}' \< \hat g', u' \> \qquad \text{ for } u \in \Us,
\end{split}$$ where $$\label{eq:qce:defn_uhat}
\hat g_\ell' = \cases{
-{{\textstyle \frac{1}{2}}}, & \ell = -K-1, K+2, \\
{{\textstyle \frac{1}{2}}}, & \ell = -K+1, K, \\
0, & \text{otherwise}.
}$$ We note that (\[eq:qce:ghost\_force\]) makes our claim precise that $y_F$ is not a critical point of $\E_{\rm qce}$.
Motivated by property (iii) of the interaction potential $\phi$, we will assume that the parameters $$\delta_1 := \frac{\phi'(2F)}{\phi''(F)} \quad \text{and} \quad
\delta_2 := \frac{- \phi''(2F)}{\phi''(F)}$$ are small, and construct an approximation for $y_{{\rm qce}, F}$ which is asymptotically of second order as $\delta_1, \delta_2 \to 0$. Although such an approximation will not be valid near the critical strain for the QCE approximation, it will give us a rough impression how the inconsistency affects the stability of the system. [[We note that $\delta_1$ is scale invariant since we used a scaled interaction potential, $\eps\phi(r/\eps),$ in our definition of the stored energy .]{}]{}
A non-dimensionalization of (\[eq:qce:ghost\_force\]) shows that $y_{{\rm qce}, F} = y_{F} + O(\delta_1)$. If $\delta_1$ is small, then we can linearize (\[eq:qce:nonlin\]) about $y_F$ and find the first-order correction $y_{\rm lin} \in \Ys_F$, which is given by $$\label{eq:qce:lin_correction}
\E_{\rm qce}''(y_F)[ y_{\rm lin} - y_F, u] = - \E_{\rm qce}'(y_F)[u]
= \phi_{2F}' \< \hat g', u' \> \qquad \text{ for all } u \in \Us.$$ We note that this linear system is precisely the one analyzed in detail in [@Dobson:2008c]. However, instead of using the implicit representation of $y_{\rm lin} - y_F$ obtained there, we use the assumption that $\delta_2$ is small to simplify (\[eq:qce:lin\_correction\]) further and obtain a more explicit approximation.
Writing out the bilinear form $\E_{\rm qce}''(y_F)[u, u]$ explicitly (using (\[eq:app:qce\_hess\_2\]) as a starting point) gives $$\label{eq:Eqce_decomposition_final}
\begin{split}
\E_{\rm qce}''(y_F)[u,u] =~& \dots
+ \eps \sum_{\ell = 0}^N \phi_F'' |u_\ell'|^2
+ \eps \sum_{\ell = 0}^{K-1} \phi_{2F}'' |u_\ell' + u_{\ell+1}'|^2
+ \eps\sum_{\ell = K+2}^N 4 \phi_{2F}'' |u_\ell'|^2 \\
& + {{\textstyle \frac{\eps}{2}}} \phi_{2F}''|u_K'+u_{K+1}'|^2
+ {{\textstyle \frac{\eps}{2}}} \phi_{2F}'' |u_{K+1}'+u_{K+2}'|^2
+ {{\textstyle \frac{\eps}{2}}} 4 \phi_{2F}'' |u_{K+1}'|^2,
\end{split}$$ where we have only displayed the terms in the right half of the domain and indicated the terms in the left half by dots. Ignoring all terms involving $\phi_{2F}''$, which are of order $\delta_2$ relative to the remaining terms, we arrive at the following approximation of : $$\phi_F'' \big\< (\hat y_{{\rm qce}, F} - y_F)', u' \big\> = \phi_{2F}' \< \hat g', u' \>
\qquad \text{ for all } u \in \Us,$$ the solution of which is given by $$\hat y_{{\rm qce}, F} = y_F + \delta_1 \hat g.$$
The following lemma makes this approximation rigorous. A complete proof is given in Appendix \[sec:app\_qce\_eq\].
\[th:asymptotic\_expansion\] If $\delta_1$ and $\delta_2$ are sufficiently small, then there exists a (locally unique) solution $y_{{\rm qce}, F}$ of such that $$\| (y_{{\rm qce}, F} - \hat y_{{\rm qce}, F})' \|_{\ell^\infty}
\leq C (\delta_1^2 + \delta_1\delta_2),$$ where $C$ may depend on $\phi$ (and its derivatives) and on $F$, but is independent of $\eps$.
From now on, we will also assume that $\delta_3 :=
\phi_{2F}''' / \phi_{F}''$ is small, and combine the three small parameters into a single parameter $$\delta := \max(|\delta_1|, |\delta_2|, |\delta_3|).$$ We will neglect all terms which are of order $O(\delta^2)$. [[A careful discussion of the parameter $\delta$ and the validity of the asymptotic analysis is given in Remark \[rem:delta\_discussion\].]{}]{}
In the following, we will again only show terms appearing on the right half of the domain. Our goal in the remainder of this section is to obtain an estimate for the smallest eigenvalue of $\E_{\rm qce}''(\hat
y_{{\rm qce}, F})$. Using (\[eq:app:qce\_hess\_2\]), we can represent $\E_{\rm qce}''(\hat y_{{\rm qce}, F})$ as $$\begin{aligned}
\E_{\rm qce}''(\hat y_{{\rm qce}, F})[u,u] = \dots
+ \eps \sum_{\ell = 0}^{K-2} \big\{ A_F |u_\ell'|^2
- \eps^2 \phi_{2F}'' |u_\ell''|^2 \big\}
+ \eps \sum_{\ell = K+3}^N A_F &\,|u_\ell'|^2 \\
+ \eps \big\{\phi_F'' + 2 \phi_{2F}'' + 2 \phi''(2F+{{\textstyle \frac{1}{2}}}\delta_1)\big\}
&\,|u_{K-1}'|^2 \\
+ \eps \big\{\phi''(F+{{\textstyle \frac{1}{2}}}\delta_1)+ 3\phi''(2F+{{\textstyle \frac{1}{2}}}\delta_1)\big\}
&\, |u_{K}'|^2 \\
+ \eps \big\{ \phi_F'' + \phi''(2F-{{\textstyle \frac{1}{2}}}\delta_1)
+ \phi''(2F + {{\textstyle \frac{1}{2}}}\delta_1)
+ 2 \phi_{2F}'' \big\} &\, |u_{K+1}'|^2 \\
+ \eps \big\{ \phi''(F-{{\textstyle \frac{1}{2}}}\delta_1) + \phi''(2F-{{\textstyle \frac{1}{2}}}\delta_1)
+ 4 \phi''(2F-\delta_1) \big\} &\, |u_{K+2}'|^2 \\
- \eps^3 \big\{
\phi''(2F+{{\textstyle \frac{1}{2}}}\delta_1) |u_{K-1}''|^2
+ {{\textstyle \frac{1}{2}}} \phi''(2F+{{\textstyle \frac{1}{2}}}\delta_1) |u_K''|^2
+ {{\textstyle \frac{1}{2}}} \phi''(2F-{{\textstyle \frac{1}{2}}}\delta_1) |u_{K+1}''|^2 \big\}. \hspace{-1.2cm} &\end{aligned}$$ We expand all terms containing $\delta_1$ and neglect all terms which are of order $O(\delta^2)$ relative to $\phi''_F,$ which is the order of magnitude of the coefficient of the diagonal term of $\E_{\rm
qce}''(\hat{y}_{{\rm qce}, F}).$ For example, we have, for some $\vartheta
\in (0, 1)$, $$\frac{\phi''(2F+{{\textstyle \frac{1}{2}}}\delta_1)}{\phi''_F}
= \frac{\phi_{2F}''}{\phi_F''} +
\frac{\phi'''(2F + \vartheta {{\textstyle \frac{1}{2}}}\delta_1)}{\phi''_F} ({{\textstyle \frac{1}{2}}}\delta_1)
= \frac{\phi_{2F}''}{\phi_F''} + O(\delta_3\delta_1),$$ as $\delta_1,\delta_3 \to 0$. Thus, the $O(\delta_1)$ perturbation of a second-neighbour term will not affect our final result. On the other hand, expanding a nearest neighbour term gives $$\frac{\phi''(F + {{\textstyle \frac{1}{2}}}\delta_1)}{\phi''_F}
= 1 + \frac{\phi_F'''}{\phi''_F} ({{\textstyle \frac{1}{2}}}\delta_1) + O(\delta_1^2),$$ as $\delta_1 \to 0$. Proceeding in the same fashion for the remaining terms, we arrive at $$\begin{aligned}
\notag \E_{\rm qce}''(\hat y_{{\rm qce}, F})[u,u] =~& \dots
+ \eps \sum_{\ell = 0}^{K-1} A_F |u_\ell'|^2
- \eps^3 \sum_{\ell = 0}^{K-1} \phi_{2F}'' |u_\ell''|^2
+ \eps \sum_{\ell = K+3}^N A_F |u_\ell'|^2 \\
\label{eq:qce:hatH_asymp}
& + \eps \big\{ A_F + ({{\textstyle \frac{1}{2}}} \delta_1 \phi_F''' - \phi_{2F}'') \big\}
|u_{K}'|^2 + \eps A_F |u_{K+1}'|^2 \\
\notag &
+ \eps \big\{ A_F - ({{\textstyle \frac{1}{2}}}\delta_1\phi_F''' - \phi_{2F}'') \big\} |u_{K+2}'|^2
- \eps^3 {{\textstyle \frac{1}{2}}} \phi_{2F}'' \big\{ |u_{K}''|^2 + |u_{K+1}''|^2 \big\} \\
\notag & + O\big(\phi_F'' \delta^2 \|u'\|_{\ell^2_\eps}^2\big).\end{aligned}$$
Clearly, our focus must be the coefficients of the terms $|u_K'|^2$ and $|u_{K+2}'|^2$, and in particular, on the quantity $$\label{sign}
{{\textstyle \frac{1}{2}}}\delta_1\phi_F''' - \phi_{2F}'' = \frac{\phi_F''' \phi_{2F}'
- 2 \phi_F'' \phi_{2F}''}{2 \phi_F''}.$$ Depending on the sign of ${{\textstyle \frac{1}{2}}}\delta_1\phi_F''' - \phi_{2F}''<0,$ we see that the “weakest bonds” are either between atoms $K-1$ and $K$ (as well as $-K+1$ and $-K$) or between atoms $K+1$ and $K+2$ (as well as $-K-1$ and $-K-2$).
If ${{\textstyle \frac{1}{2}}}\delta_1\phi_F''' - \phi_{2F}''<0,$ we insert the test function $w \in \Us$, defined by $$w_\ell' = \cases{
({{\textstyle \frac{1}{2}}} \eps^{-1})^{1/2}, & \ell = K, \\
-({{\textstyle \frac{1}{2}}} \eps^{-1})^{1/2}, & \ell = -K+1, \\
0, & \text{otherwise},}$$ into (\[eq:qce:hatH\_asymp\]) to obtain $$\label{eq:qce:stab_final}
\begin{split}
\inf_{\substack{u \in \Us \\ \| u'\|_{\ell^2_\eps} = 1} }
\E_{\rm qce}''(\hat{y}_{{\rm qce}, F})[u,u]
\leq~& \E_{\rm qce}''(\hat y_{{\rm qce}, F})[w,w] \\[-4mm]
=~& A_F \Big\{ 1 + \frac{\phi_F''' \phi_{2F}'
- 5\phi_F'' \phi_{2F}''}{2 A_F \phi_F''} + O(\delta^2) \Big\}.
\end{split}$$ Note that the constant $2$ in front of $\phi_F'' \phi_{2F}''$ was replaced by $5$ due to the strain gradient terms in (\[eq:qce:hatH\_asymp\]) which slightly stabilize the system.
If ${{\textstyle \frac{1}{2}}}\delta_1\phi_F''' - \phi_{2F}''>0,$ we use the alternative test function $w \in \Us$, defined by $$\label{alt}
w_\ell' = \cases{
({{\textstyle \frac{1}{2}}} \eps^{-1})^{1/2}, & \ell = K+2, \\
-({{\textstyle \frac{1}{2}}} \eps^{-1})^{1/2}, & \ell = -K-1, \\
0, & \text{otherwise},}$$ to test (\[eq:qce:hatH\_asymp\]), which gives $$\label{eq:qce:stab_final3}
\begin{split}
\inf_{\substack{u \in \Us \\ \| u'\|_{\ell^2_\eps} = 1} }
\E_{\rm qce}''(\hat{y}_{{\rm qce}, F})[u,u]
\leq~& \E_{\rm qce}''(\hat y_{{\rm qce}, F})[w,w] \\[-4mm]
=~& A_F \Big\{ 1 - \frac{ \phi_F''' \phi_{2F}'-\phi_F'' \phi_{2F}''
}{2 A_F \phi_F''} + O(\delta^2) \Big\}.
\end{split}$$ In this case, only a single strain gradient term affects the final result, and therefore this correction is only small.
Due to the stabilizing effect of the strain gradient terms for our perturbation, the right hand sides of and might both be bounded below by $A_F,$ so our estimate will involve a $\min$ over three terms. Recalling that $y_{{\rm qce}, F} = \hat{y}_{{\rm qce}, F} + O(\delta^2),$ we obtain the following result:
\[th:qce:final\_stab\_result\] There exist constants $\hat{\delta}$ and $\hat{C}$, which may depend on $\phi$ and its derivatives and on $F$ but not on $\eps$, such that, if $\delta \leq \hat{\delta}$, then $$\label{eq:qce:stab_final2}
\begin{split}
\inf_{\substack{u \in \Us \\ \| u'\|_{\ell^2_\eps} = 1} }
\E_{\rm qce}''(y_{{\rm qce}, F})[u,u] \leq \phi_F'' \Big( \min\!\Big\{\,&
1+\frac{3\phi_{2F}''}{\phi_F''}
\pm \Big( \frac{\phi_F'''\phi_{2F}'}{2|\phi_{F}''|^2}
- \frac{3}{2} \frac{\phi_{2F}''}{\phi_F''} \Big),
\frac{A_F}{\phi_F''} \Big\}
+ \hat{C} \delta^2 \, \Big).
\end{split}$$
The bounds (\[eq:qce:stab\_final\]) and (\[eq:qce:stab\_final3\]) are rigorous provided $\delta$ is sufficiently small so that $F -
{{\textstyle \frac{1}{2}}} \delta_1$ is bounded away from zero. Moreover, if $\delta$ is sufficiently small, then Lemma \[th:asymptotic\_expansion\] gives a rigorous bound for the error $\| (y_{{\rm qce}, F} - \hat{y}_{\rm
qce})' \|_{\ell^\infty}$ which only adds an additional $O(\delta^2)$ error to the estimate.
For typical interaction potentials, we would expect that $\phi_F''' <
0$ (as $\phi_F''$ is decreasing), that $\phi_{2F}' > 0$, and we have already postulated that $\phi_F'' > 0$ and $\phi_{2F}'' < 0$. Thus, the two terms in the numerator of the right hand side of have opposing sign and may, in principle even cancel each other. However, we have found in numerical tests that for typical potentials such as the Morse or Lennard–Jones potentials the first term is dominant, that is, $ {{\textstyle \frac{1}{2}}}\delta_1\phi_F''' -
{{\textstyle \frac{3}{2}}} \phi_{2F}''< \phi''_{2F} $ and $$\min\!\Big\{\,
1+\frac{3\phi_{2F}''}{\phi_F''}
\pm \left( \frac{\phi_F'''\phi_{2F}'}{2|\phi_{F}''|^2}
- \frac{3}{2} \frac{\phi_{2F}''}{\phi_F''} \right),
\frac{A_F}{\phi_F''} \Big\}=
1 + \frac{3}{2} \frac{\phi_{2F}''}{\phi_F''}
+ \frac{\phi_F'''\phi_{2F}'}{2|\phi_{F}''|^2}$$ in Proposition \[th:qce:final\_stab\_result\].
Proposition \[th:qce:final\_stab\_result\] as well as the subsequent discussion clearly shows that the spurious QCE instability is due to a combination of the effect of the“ghost force” error and of the anharmonicity of the atomistic potential.
\[rmbounds\] A variant of the analysis presented above shows that $\E_{\rm
qce}''(y_F)$ is positive definite if and only if $A_F + \lambda_K
\phi_{2F}'' > 0$ where ${{\textstyle \frac{1}{2}}} \le \lambda_K \le 1$. The lower bound can be obtained using the test function in the bilinear form $\E_{\rm qce}''(y_F)[u,\,u]$ given explicitly by , while the upper bound can be obtained from the estimate $$\E_{\rm qce}''(y_F)[u,\,u]\ge (A_F+\phi_{2F}'')\| u'\|_{\ell^2_\eps}^2
\qquad\text{ for all } u \in \Us,$$ which also follows from (see also [@Dobson:2008b Lemma 2.1]). Thus, the lower bound is related to the second term in which we have noted above is generally greater than the first term, and we can conclude that the critical strain for QCE obtained by linearizing about $y_F,$ rather than the equilibrium solution $y_{{\rm qce}, F},$ significantly underestimates the loss of stability (see also Figure \[fig:crit\_strains\_a\]).
The study of the positive-definiteness of $\E_{\rm qce}''(y_F)$ is relevant to the stability of the ghost-force correction iteration and is discussed in more detail in [@qcf.iterative].
\[rem:delta\_discussion\] While our rigorous results, Lemma \[th:asymptotic\_expansion\] and Proposition \[th:qce:final\_stab\_result\], are proven only for sufficiently small $\delta$, one usually expects that such asymptotic expansions have a wider range of validity than that predicted by the analysis. For this reason, we have neglected to give more explicit bounds on how small $\delta$ needs to be.
Nevertheless, a relatively simple asymptotic analysis such as the one we have presented cannot usually give complete information near the onset of instability. Our aim was mainly to demonstrate that the inconsistency at the interface leads to a decreased stability of the QCE approximation when compared to the full atomistic model or the consistent QC approximations. We will see in Section \[sec:discus\] that, if we use to predict the onset of instability for QCE, then we observe a fairly significant loss of stability of the QCE approximation when compared to the full atomistic model. In numerical experiments, we will also see that the prediction given by (\[eq:qce:stab\_final2\]) is qualitatively fairly accurate for the Morse potential [[for a range of parameters that explores the dependence of our results on $\delta.$]{}]{}
Prediction of the Limit Strain for Fracture Instability {#sec:discus}
=======================================================
The deformation $y_F \in \Ys_F$ is an equilibrium of the atomistic energy for [*all $F > 0$*]{}. However, it is established in Proposition \[th:ana:stab\_a\] that $y_F$ is [*stable*]{} if and only if ${F < F_{\rm a}^*}$ where $F_{\rm a}^*$ is the solution of the equation $$\label{eq:defn_Fa_crit}
\psi_{\rm a}(F_{\rm a}^*) := \phi''(F_{\rm a}^*) + (4 - \eps^2 \mu_\eps^2) \phi''(2F_{\rm a}^*) = 0.$$ We call $F_{\rm a}^*$ the critical strain for the atomistic model. The goal of the present section is to use the stability analyses of the different QC approximations in Sections \[sec:analysis\] and \[sec:ana:qce\] to investigate how well the critical strains for the different QC approximations approximate that of the atomistic model.
In order to test our predictions against numerical values, we will use the Morse potential $$\label{eq:defn_morse}
\phi_\alpha(r) = e^{-2 \alpha (r - 1)} - 2e^{-\alpha(r-1)}=(e^{-\alpha(r-1)}-1)^2-1,$$ where $\alpha \geq 1$ is a fixed parameter, and the Lennard–Jones potential $$\phi_{\rm lj}(r) = \frac{1}{r^{12}} - \frac{2}{r^6}.$$
Limit strain for the QCL and QNL approximations
-----------------------------------------------
The critical strain $F_{\rm c}^*$ for the local QC approximation as well as the QNL approximation (cf. Propositions \[th:ana:qcl\_stab\] and \[th:ana:qnl\_stab\]) is the solution to the equation $$\psi_{\rm c}(F_{\rm c}^*) := \phi''(F_{\rm c}^*) + 4 \phi''(2F_{\rm c}^*) = 0.$$
We note that the critical strain $F_{\rm c}^*$ for the QCL and QNL models is independent of $N$ which is convenient for the following analysis. Inserting $F_{\rm c}^*$ into (\[eq:defn\_Fa\_crit\]) gives $$\psi_{\rm a}(F_{\rm c}^*)
= \psi_{\rm c}(F_{\rm c}^*) - \eps^2\mu_\eps^2 \phi''(2 F_{\rm c}^*)
= - \eps^2\mu_\eps^2\phi''(2 F_{\rm c}^*),$$ and hence $$\psi_{\rm a}(F_{\rm a}^*) - \psi_{\rm a}(F_{\rm c}^*) = \eps^2\mu_\eps^2\phi''(2F_{\rm c}^*).$$ A linearization of the left-hand side gives $$\psi_{\rm a}'(F_{\rm c}^*) ( F_{\rm a}^* - F_{\rm c}^*)
= \eps^2\mu_\eps^2 \phi''(2 F_{\rm c}^*)
+ O( |F_{\rm a}^* - F_{\rm c}^*|^2 ).$$ Noting that $\psi_{\rm a}'(F_{\rm c}^*) = \psi_{\rm c}'(F_{\rm c}^*) + O(\eps^2)$, we find that the [*relative error*]{} satisfies $$\label{eq:rel_err_ac}
\begin{split}
\left| \frac{F_{\rm a}^* - F_{\rm c}^*}{F_{0} - F_{\rm c}^*} \right|
=~& \eps^2 \left| \frac{\pi^2 \phi''(2F_{\rm c}^*)}{
(\phi'''(F_{\rm c}^*)+8\phi'''(2F_{\rm c}^*))(F_{0} - F_{\rm c}^*)} \right|
+ O(\eps^4) \\
:=~& \eps^2 C_{\rm err}(\phi) + O(\eps^4),
\end{split}$$ where $F_{0}$ is the energy-minimizing macroscopic deformation gradient which satisfies $$\frac{d\E_{\rm a}(y_F)}{dF}(F_0) =\phi'(F_0) + 2 \phi'(2F_0) = 0.$$
$\phi$ $\phi_2$ $\phi_3$ $\phi_4$ $\phi_5$ $\phi_6$ $\phi_7$ $\phi_{\rm lj}$
--------------------- ---------- ---------- ---------- ---------- ---------- ---------- -----------------
$C_{\rm err}(\phi)$ 1.0877 0.3796 0.1339 0.0485 0.0177 0.0065 0.0635
: \[tbl:errconsts\] Numerical values of the error constant $C_{\rm err}(\phi)$ defined in , for various choices of $\phi$.
In Table \[tbl:errconsts\] we display numerical values of $C_{\rm
err}(\phi)$ for the Morse potential $\phi = \phi_\alpha$, with $\alpha = 2, \dots, 7$, and for the Lennard–Jones potential $\phi =
\phi_{\rm lj}$. We observe that the constant decays exponentially as the stiffness increases, and that it is moderate even for very soft interaction potentials ($C_{\rm err}(\phi_2) \approx 1.0877$).
Limit strain for the QCE approximation
--------------------------------------
In Section \[sec:ana:qce\], we have computed a rough estimate for the coercivity constant of the QCE approximation. We argued that, for as long as the second neighbour interaction is small in comparison to the nearest neighbour interaction, we have the bound $$\inf_{\substack{u \in \Us \\ \|u'\|_{\ell^2_\eps} = 1}}
\E_{\rm qce}''(y_{{\rm qce}, F})[u,u] \leq \phi_F'' \Big\{
1 + \frac{3}{2} \frac{\phi_{2F}''}{\phi_F''} + \frac{\phi_F'''\phi_{2F}'}{2 |\phi_{F}''|^2} + O(\delta^2) \Big\}.$$ Even though this bound will, in all likelihood, become invalid near the critical strain, it is nevertheless reasonable to expect that solving $$\label{eq:discus:critF_qce}
\tilde\psi_{\rm qce}(\tilde{F}_{\rm qce}^*) :=
\phi_F'' + {{\textstyle \frac{3}{2}}} \phi_{2F}'' + \frac{\phi_F'''\phi_{2F}'}{2 \phi_{F}''} = 0,$$ will give a good approximation for the exact critical strain, $F_{\rm qce}^*$. The latter is, loosely speaking, defined as the maximal strain $F > 0$ for which a stable “elastic” equilibrium of $\E_{\rm qce}$ exists in $\Ys_F$. A deformation $y$ can be called elastic if $y_\ell'= O(1)$ for all $\ell$, as opposed to fractured if $y_{\ell_0}' = O(N)$ for some $\ell_0$.
We could use the same argument as in the previous subsection to obtain a representation of the error; however, since $\tilde F_{\rm qce}^*$ depends only on $F$ but not on $\eps$ we can simply solve for $\tilde
F_{\rm qce}^*$ directly.
For the Morse potential , with stiffness parameter $2 \leq \alpha \leq 7$, we have computed both $F_{\rm
qce}^*$ (for $N = 40, K= 10$ as well as for $N = 100, K = 20$) and $\tilde F_{\rm qce}^*$ numerically and have plotted these critical strains in Figure \[fig:crit\_strains\_a\], comparing them against $F_0$ and $F_{\rm c}^*$. We have also included the critical strain $F_{\rm qce}^{y_F}$, below which $\E_{\rm qce}''(y_F)$ is positive definite, to demonstrate that it bears no relation to the stability or instability of the QCE approximation. We discuss $F_{\rm qce}^{y_F}$ in detail in [@qcf.iterative] where we argue that it describes the stability of the ghost-force correction scheme.
In Figure \[fig:crit\_strains\_b\], we plot the relative errors $$\alpha \mapsto \left| \frac{F_{\rm qce}^*(\alpha) - F_{\rm c}^*(\alpha)}{
F_{\rm c}^*(\alpha) - F_0(\alpha)} \right|
\quad \text{and} \quad
\alpha \mapsto \left| \frac{\tilde F_{\rm qce}^*(\alpha) - F_{\rm c}^*(\alpha)}{
F_{\rm c}^*(\alpha) - F_0(\alpha)} \right|.$$
![\[fig:crit\_strains\_a\] Critical strains $F_{\rm
qce}^*,\ \tilde F_{\rm qce}^*,\ F_{\rm qce}^{y_F},\ F_{\rm c}^*$ and the equilibrium strain $F_0$, computed for the Morse potential with varying $\alpha$. The critical strains for the QCE Hessian, $F_{\rm qce}^*,$ are computed with $N = 40$ and $K = 10$. The approximation, $\tilde F_{\rm qce}^*,$ is computed using the asymptotic approximation . The strain $F_{\rm qce}^{y_F}$ is the critical strain at which $\E_{\rm qce}''(y_F)$ is no longer positive definite.](crit_strains_a){height="6.6cm"}
![\[fig:crit\_strains\_b\] Relative errors of the critical strains (computed and predicted) for the QCE approximation against the critical strains of the QCL/QNL approximation. The errors are computed explicitly for $N = 40, K = 10$ as well as for $N =
100, K = 20$, using the Morse potential with varying $\alpha$. These two curves are very close and may be hard to distinguish. Additionally, we show the critical strain for loss of positive definiteness of $\E_{\rm
qce}''(y_F),$ which does not predict the loss of stability that the QCE experiences correctly for any parameter value.](crit_strains_b){height="6.6cm"}
We observe that the prediction for the critical strain, as well as the prediction for the relative error, obtained from our asymptotic analysis is insufficient for very soft potentials but becomes fairly accurate with increasing stiffness. In particular, it provides a good prediction of the relative errors for the critical strains for $\alpha
\geq 3.5$.
For a correct interpretation of our results, we must first of all note that the relative errors for the critical strains decay exponentially with increasing stiffness $\alpha$. While, for small $\alpha$ (soft potentials) the error is quite severe, one could argue that it is insignificant (i.e., well below 10%) for moderately large $\alpha$ (stiff potentials). However, our point of view is that, by a careful choice of the atomistic region one should be able to control this error, as is the case for consistent QC approximations such as QNL. For the QCE approximation, this is impossible: the error in the critical strain is [*uncontrolled*]{}.
Conclusion {#conclusion .unnumbered}
==========
We propose sharp stability analysis as a theoretical criterion for evaluating the predictive capability of atomistic-to-continuum coupling methods. Our results show that a sharp stability analysis is as important as a sharp truncation error (consistency) analysis for the evaluation of atomistic-to-continuum coupling methods, and provides a new means to distinguish the relative merits of the various methods. Our results also provide an approach to establish a theoretical basis for the conclusions of the benchmark numerical tests reported in [@Miller:2008], in particular for the poor performance of the QCE approximation in predicting the movement of a dipole of Lomer dislocations under applied shear.
[[ Of course, the simple one-dimensional situation that we have considered here cannot nearly capture the complexity of atomistic-to-continuum coupling methods in 2D/3D. Even the much simpler question of whether QCL (the Cauchy-Born continuum model without coupling) can correctly predict bifurcation points becomes much more difficult since it is possible, in general, that the stability region for the Cauchy–Born model is much larger than that for atomistic model [@E:2007a]. However, in many interesting situations this effect does not occur [@HudsOrt:a], and it is an interesting question to characterize these. Concerning the stability of the coupling mechanism, no rigorous results are available in 2D/3D. Until such an analysis is available, we propose that careful numerical experiments should be performed, which experimentally investigate the stability properties of atomistic-to-continuum coupling methods. ]{}]{}
Representations of $\E_{\rm qce}'$ and $\E_{\rm qce}''$ {#sec:app_qce1}
=======================================================
Our aim in this section is to derive useful representations for the first and second variations $\E_{\rm qce}'(y)$ and $\E_{\rm qce}''(y)$ of the QCE energy functional. For notational convenience, we will only write out terms in the right half of the domain $\{-N+1, \dots, N\}$, indicating the remaining terms (which can be obtained from symmetry considerations) by dots. For example, we write $$\begin{aligned}
\E_{\rm qce}(y) =~& \dots + \eps \sum_{\ell = 0}^{N} \phi(y_\ell')
+ \eps \sum_{\ell = 0}^{K-1} \phi(y_\ell' + y_{\ell+1}')
+ \eps \sum_{\ell = K+2}^N \phi(2 y_\ell') \\
& + {{\textstyle \frac{\eps}{2}}} \phi(y_K' + y_{K+1}')
+ {{\textstyle \frac{\eps}{2}}} \phi(y_{K+1}' + y_{K+2}')
+ {{\textstyle \frac{\eps}{2}}} \phi(2y_{K+1}').\end{aligned}$$
The first variation is a linear form on $\Us$, given by $$\begin{aligned}
\E_{\rm qce}'(y)[u] = \dots + \eps \sum_{\ell = 0}^N \phi'(y_\ell') u_\ell'
+ \eps \sum_{\ell = 0}^{K-1} \phi'(y_\ell'+y_{\ell+1}')(u_\ell' + u_{\ell+1}')& \\
+ {{\textstyle \frac{\eps}{2}}} \phi'(y_K' + y_{K+1}')(u_K'+u_{K+1}')
+ {{\textstyle \frac{\eps}{2}}} \phi'(y_{K+1}' + y_{K+2}')(u_{K+1}'+u_{K+2}') & \\
+ {{\textstyle \frac{\eps}{2}}} \phi'(2y_{K+1}') (2 u_{K+1}')
+ \eps \sum_{\ell = K+2}^N \phi'(2y_\ell')(2u_\ell')&.\end{aligned}$$ Collecting terms related to element strains $u_\ell'$, we obtain $$\begin{aligned}
\label{eq:app:DEqce}
\begin{split}
\E_{\rm qce}'(y)[u] = \dots +
\eps \sum_{\ell = 0}^{K-1} \big\{ \phi'(y_\ell')
+ \phi'(y_{\ell-1}'+y_\ell')+\phi'(y_\ell'+y_{\ell+1}')\big\}~& u_\ell' \\
+ \eps \big\{ \phi'(y_K') + \phi'(y_{K-1}'+y_K')
+ {{\textstyle \frac{1}{2}}} \phi'(y_{K}'+y_{K+1}')\big\}~& u_K' \\
+ \eps \big\{ \phi'(y_{K+1}') + {{\textstyle \frac{1}{2}}} \phi'(y_{K}'+y_{K+1}')
+{{\textstyle \frac{1}{2}}} \phi'(y_{K+1}'+y_{K+2}') + \phi'(2y_{K+1}') \big\} ~& u_{K+1}'\\
+ \eps \big\{ \phi'(y_{K+2}' + {{\textstyle \frac{1}{2}}} \phi'(y_{K+1}'+y_{K+2}')
+ 2 \phi'(2y_{K+2}') \big\} ~& u_{K+2}' \\
+ \eps \sum_{\ell = K+3}^N \big\{ \phi'(y_\ell')
+ 2 \phi'(2y_\ell') \big\} ~& u_{\ell}'.
\end{split}\end{aligned}$$
Similarly, the second variation can be written in the form $$\label{eq:app:qce_hess_1}
\begin{split}
\E_{\rm qce}''(y)[u,u] = \dots
+ \eps \sum_{\ell = 0}^N \phi''(y_\ell') |u_\ell'|^2
+ \eps \sum_{\ell = 0}^{K-1} \phi''(y_\ell'+y_{\ell+1}')
|u_\ell' + u_{\ell+1}'|^2& \\
+ {{\textstyle \frac{\eps}{2}}} \phi''(y_K'+y_{K+1}')|u_K'+u_{K+1}'|^2
+ {{\textstyle \frac{\eps}{2}}} \phi''(y_{K+1}'+y_{K+2}') |u_{K+1}'+u_{K+2}'|^2 & \\
+ {{\textstyle \frac{\eps}{2}}} \phi''(2 y_{K+1}') |2u_{K+1}'|^2
+ \eps\sum_{\ell = K+2}^N \phi''(2y_\ell') |2u_\ell'|^2&.
\end{split}$$ Using (\[eq:rewrite\_nnn\]) to replace all second-neighbour terms in (\[eq:app:qce\_hess\_1\]), we obtain the alternative representation $$\begin{aligned}
\label{eq:app:qce_hess_2}
\E_{\rm qce}''(y)[u,u] = \dots
+ \eps \sum_{\ell = 0}^{K-1} \big[
\phi''(y_\ell') + 2 \phi''(y_{\ell-1}'+y_\ell')
+ 2\phi''(y_{\ell}'+y_{\ell+1}') \big]& |u_\ell'|^2 \\
\notag
+ \eps \big[ \phi''(y_K') + 2 \phi''(y_{K-1}'+y_K')
+ \phi''(y_K'+y_{K+1}') \big]& |u_K'|^2 \\
\notag
+ \eps \big[ \phi''(y_{K+1}') + \phi''(y_{K}'+y_{K+1}') +
\phi''(y_{K+1}'+y_{K+2}') + 2\phi''(2y_{K+1}') \big]& |u_{K+1}'|^2 \\
\notag
+ \eps \big[ \phi''(y_{K+2}') + \phi''(y_{K+1}'+y_{K+2}')
+ 4 \phi''(2y_{K+2}') \big]& |u_{K+2}'|^2 \\[-1mm]
\notag
+ \eps \sum_{\ell = K+3}^{N} \big[ \phi''(y_\ell')
+ 4 \phi''(2 y_\ell')\big]& |u_\ell'|^2 \\[-4mm]
\notag
- \eps^3 \sum_{\ell = 0}^{K-1} \phi''(y_\ell'+u_{\ell+1}') |u_\ell''|^2
- {{\textstyle \frac{1}{2}}} \eps^3 \big\{ \phi''(y_K'+y_{K+1}') |u_K''|^2
+ \phi''(y_{K+1}'+y_{K+2}') |u_{K+1}''|^2 \big\}. \hspace{-2cm} &\end{aligned}$$ While somewhat unwieldy at first glance, this representation is particularly useful for the stability analysis in Section \[sec:ana:qce\].
Proof of Lemma \[th:asymptotic\_expansion\] {#sec:app_qce_eq}
===========================================
In this section, we complete the proof of Lemma \[th:asymptotic\_expansion\] which was merely hinted at in the main text of Section \[sec:ana:qce\]. Recall that $\hat{y}_{{\rm qce}, F} =
y_F + \delta_1 \hat{g}$ where $\hat{g}$ is given by , and recall, moreover, that $\hat{y}_{\rm
qce}$ solves the linear system $$\label{eq:app_qce:defnyhat}
\phi_F'' \< (\hat y_{{\rm qce}, F} - y_F)', u' \> = \phi_{2F}' \< \hat g', u' \>
= - \E_{\rm qce}'(y_F)[u] \qquad \text{ for all } u \in \Us.$$ Our strategy is to prove that $\hat{y}_{{\rm qce}, F}$ has a residual of order O($\delta_1^2 + \delta_1\delta_2$) and that $\E_{\rm qce}''(\hat y_{{\rm qce}, F})$ is an isomorphism between suitable function spaces. We will then apply a quantitative inverse function theorem to prove the existence of a solution $y_{{\rm qce}, F}$ of the QCE criticality condition which is “close” to $\hat y_{{\rm qce}, F}$. Before we embark on this analysis, we make several comments and introduce some notation that will be helpful later on.
To ensure that $\E_{\rm qce}$ is sufficiently differentiable in a neighbourhood of $\hat y_{{\rm qce}, F}$ we only need to assume that $F > 0$ and that $\delta_1$ is sufficiently small, e.g., $\delta_1 \leq F$. In that case, $\E_{\rm qce}$ is three times differentiable at $y$ for any $y
\in \Ys_F$ such that $\|y' - \hat y_{{\rm qce}, F}' \|_{\ell^\infty} <
{{\textstyle \frac{1}{2}}} \delta_1$.
We will interpret $\E_{\rm qce}'$ as a nonlinear operator from $\Us^{1,\infty}$ to $\Us^{-1,\infty}$ which are, respectively, the spaces $\Us$ and $\Us^*$ endowed with the Sobolev-type norms, $$\| u \|_{\Us^{1,\infty}} = \|u'\|_{\ell^\infty} \quad \text{for } u \in \Us,
\quad \text{and} \quad
\| T \|_{\Us^{-1,\infty}} = \sup_{\substack{v \in \Us \\\|v'\|_{\ell^1_\eps} = 1}} T[v]
\quad \text{for } T \in \Us^*.$$ Consequently, for $y \in \Ys_F$, $\E_{\rm qce}''(y)$ can be understood as a linear operator from $\Us^{1,\infty}$ to $\Us^{-1,\infty}$.
Our justification for defining $\hat{y}_{{\rm qce}, F}$ as we did in is the bound $$\label{eq:D2EyF_expansion}
\big| \E_{\rm qce}''(y_F)[u,v] - \phi_F'' \<u', v'\> \big|
\leq \phi_F'' c_1 \delta_2 \|u'\|_{\ell^\infty_\eps} \|v'\|_{\ell^1_\eps}
\quad \text{ for all } u, v \in \Us,$$ where $c_1 = 5$, which follows from . We can formulate this bound equivalently as $$\label{eq:D2EyF_expansion_2}
\| \E_{\rm qce}''(y_F) - \phi_F'' L_1 \|_{L(\Us^{1,\infty},\ \Us^{-1,\infty})}
\leq \phi_F'' c_1 \delta_2,$$ where $L_1 : \Us \to \Us^*$ is given by $$L_1 (u)[ v ] = \< u', v' \> \qquad \text{ for all } u, v \in \Us.$$ We also remark that $L_1 : \Us^{1,\infty} \to \Us^{-1,\infty}$ is an isomorphism, uniformly bounded in $N$, more precisely, $$\label{eq:qceapp:L1_iso}
\| L_1^{-1} \|_{L(\Us^{-1,\infty},\ \Us^{1,\infty})} \leq 2.$$ This result follows, for example, as a special case of [@Ortner:2008a Eq. (36)] or [@dobs-qcf2 Eq. (5.2)], and is also contained in [@doblusort:qcf.stab].
We are now ready to estimate the residual of $\hat{y}_{{\rm qce}, F}$. Expanding $\E_{\rm qce}'(\hat y_{{\rm qce}, F})$ to first order gives $$\label{eq:DEqce(haty)}
\begin{split}
\E_{\rm qce}'(\hat y_{{\rm qce}, F})[v] =~& \big\{ \E_{\rm qce}'(y_F)[v]
+ \delta_1 \E_{\rm qce}''(y_F)[\hat g, v] \big\} \\
& + \delta_1 \int_0^1 \big\{
\E_{\rm qce}''(y_F + t \delta_1 \hat g)[\hat g, v]
- \E_{\rm qce}''(y_F)[\hat g, v] \big\} \dt.
\end{split}$$ We will estimate the two groups on the right-hand side of separately. Using and , we obtain $$\label{app:qce:resest_1}
\begin{split}
\big|\E_{\rm qce}'(y_F)[v] + \delta_1 \E_{\rm qce}''(y_F)[\hat g, v]\big|
=~& \delta_1 \big| - \phi_F'' \< \hat g', v'\> + \E_{\rm qce}''(y_F)[\hat g, v]
\big| \\
\leq~& \phi_F'' c_1 \delta_1 \delta_2 \| \hat g'\|_{\ell^\infty_\eps}
\|v'\|_{\ell^1_\eps} \qquad \text{ for all } v \in \Us.
\end{split}$$
To estimate the second group in we simply use the regularity of the interaction potential (we assumed that $\phi \in
\CC^3(0, +\infty)$) and Hölder’s inequality to obtain $$\label{app:qce:resest_2}
\big| \E_{\rm qce}''(y_F + t \delta_1 \hat g)[\hat g, v]
- \E_{\rm qce}''(y_F)[\hat g, v] \big| \leq \phi_F'' c_2 t \delta_1
\| \hat g'\|_{\ell^\infty}^2 \|v'\|_{\ell^1_\eps},$$ where $(\phi_F'' c_2)$ is a local Lipschitz constant for $\phi''$, that is, there exists a universal constant $\hat{c}_2$ such that $$c_2 = \hat c_2 \sup_{|r| \leq {{\textstyle \frac{1}{2}}}\delta_1} \frac{\max(|\phi'''(F+r)|,
|\phi'''(2(F+r))|)}{\phi_F''}.$$ In particular, if $\delta_1$ is sufficiently small then we may assume that $$c_2 = 2 \hat{c}_2 \frac{\max( |\phi_F'''|, |\phi_{2F}'''|)}{\phi_F''}.$$
Inserting and into (\[eq:DEqce(haty)\]), and using the fact that $\|\hat
g'\|_{\ell^\infty} = {{\textstyle \frac{1}{2}}}$, we obtain the $\Us^{-1,\infty}$-residual estimate $$\| \E_{\rm qce}'(\hat y_{{\rm qce}, F}) \|_{\Us^{-1,\infty}}
\leq \phi_F''({{\textstyle \frac{1}{2}}} c_1 \delta_1 \delta_2 + {{\textstyle \frac{1}{8}}} c_2 \delta_1^2).$$
Next, we estimate $\| \E_{\rm qce}''(\hat y_{{\rm qce}, F})^{-1} \|_{L(\Us^{-1,\infty},\
\Us^{1,\infty})}$. Using and a similar argument as for gives $$\begin{aligned}
\| \E_{\rm qce}''(\hat y_{{\rm qce}, F}) - \phi_F'' L_1 \|_{L(\Us^{1,\infty},\ \Us^{-1\infty})}
\leq~&
\| \E_{\rm qce}''(\hat y_{{\rm qce}, F}) - \E_{\rm qce}''(y_F) \|_{L(\Us^{1,\infty},\ \Us^{-1\infty})} \\
& + \| \E_{\rm qce}''(y_F) - \phi_F'' L_1 \|_{L(\Us^{1,\infty},\ \Us^{-1\infty})} \\
\leq~& \phi_F'' ({{\textstyle \frac{1}{2}}} c_2 \delta_1 + c_1 \delta_2).\end{aligned}$$ Moreover, from , we deduce that $$\| (\phi_F'' L_1)^{-1} \|_{L(\Us^{-1,\infty},\ \Us^{1,\infty})} \leq \frac{2}{\phi_F''}.$$
A standard result of operator theory states that if $X, Y$ are Banach spaces and $T, S : X \to Y$ are bounded linear operators with $T$ being invertible and satisfying $\|S - T\| < 1 / \|T^{-1}\|,$ then $S$ is invertible and $$\|S^{-1} \| \leq \frac{ \|T^{-1}\|}{1 - \|T^{-1}\| \|S-T\|}.$$ In our case, setting $T = \phi_F'' L_1$ and $S = \E_{\rm qce}''(\hat
y_{{\rm qce}, F})$, this translates to $$\| \E_{\rm qce}''(\hat y_{{\rm qce}, F})^{-1} \|_{L(\Us^{-1,\infty},\ \Us^{1,\infty})}
\leq \frac{2}{\phi_F''(1 - {{\textstyle \frac{1}{2}}} c_2 \delta_1 - c_1 \delta_2)},$$ provided that the denominator is positive. Thus, for $\delta_1,
\delta_2$ sufficiently small, we obtain the bound $$\| \E_{\rm qce}''(\hat y_{{\rm qce}, F})^{-1} \|_{L(\Us^{-1,\infty},\ \Us^{1,\infty})}
\leq \frac{4}{\phi_F''}.$$
We now apply the following version of the inverse function theorem.
\[th:inverse\_fcn\_thm\] Let $X, Y$ be Banach spaces, $U$ an open subset of $X$, and let $F :
U \to Y$ be Fréchet differentiable. Suppose that $x_0 \in U$ satisfies the conditions $$\begin{aligned}
& \| F(x_0) \|_{Y} \leq \eta, \quad
\| F'(x_0)^{-1} \|_{L(Y,X)} \leq \sigma^{-1}, \\
& \overline{B_X(x_0, 2\eta\sigma^{-1})} \subset U, \\
& \| F'(x_1) - F'(x_2) \|_{L(X,Y)} \leq L \| x_1 - x_2 \|_X \quad
\text{for} \quad \|x_j - x_0\|_X \leq 2 \eta \sigma^{-1}, \\
& \text{ and } 2 L \sigma^{-2} \eta < 1,
\end{aligned}$$ then there exists $x \in X$ such that $F(x) = 0$ and $\|x - x_0\|_X
\leq 2\eta\sigma^{-1}$.
The result follows, for example, by applying Theorem 2.1 in [@ortner_apostex] with the choices $R = 2 \eta\sigma^{-1}$, $\omega(x_0, R) = LR$ and $\bar\omega(x_0, R) = \frac12 L
R^2$. Similar results can be obtained by tracking the constants in most proofs of the inverse function theorem, and assuming local Lipschitz continuity of $F'$.
For our purposes, we set $X = \Us^{1,\infty}$, $Y =
\Us^{-1,\infty}$, $F(u) = \E_{\rm qce}'(\hat y_{{\rm qce}, F} + u)$, and $x_0 =
0$. Assuming that $\delta_1, \delta_2$ are sufficiently small, our previous analysis gives the residual and stability estimates $$\eta = \phi_F'' ({{\textstyle \frac{1}{2}}} c_1\delta_1\delta_2 + {{\textstyle \frac{1}{8}}} c_2 \delta_1^2)
\quad \text{and} \quad \sigma = {{\textstyle \frac{1}{4}}} \phi_F'',$$ and, in particular, $$2\eta\sigma^{-1} = 4 c_1 \delta_1\delta_2 + c_2 \delta_1^2.$$
To ensure that $\overline{B_{\Us^{1,\infty}}(0, 2\eta\sigma^{-1})}$ remains within the region of differentiability of $F$, that is, to ensure that $(\hat y_{{\rm qce}, F} + u)_\ell' > 0$ for $\|u'\|_{\ell^\infty} \leq
2\eta\sigma^{-1}$, it is clearly enough to assume that $\delta_1$ and $\delta_2$ are sufficiently small.
A modification of then allows the choice $L =
2 \phi_F'' c_2$ for the local Lipschitz constant.
Thus, the condition ensuring the existence of a solution $y_{{\rm qce}, F}$ of becomes $$4 L \sigma^{-2} \eta
= 64 c_1c_2 \delta_1\delta_2 + 16 c_2^2 \delta_1^2
< 1,$$ which is satisfied, once again, if we assume that $\delta_1$ and $\delta_2$ are sufficiently small. An application of Lemma \[th:inverse\_fcn\_thm\] concludes the proof of Lemma \[th:asymptotic\_expansion\].
[99]{}
S. Badia, M. L. Parks, P. B. Bochev, M. Gunzburger, and R. B. Lehoucq. On atomistic-to-continuum coupling by blending. , 7(1):381–406, 2008.
X. Blanc, C. Le Bris, and F. Legoll. Analysis of a prototypical multiscale method coupling atomistic and continuum mechanics. , 39(4):797–826, 2005.
M. Dobson and M. Luskin. Analysis of a force-based quasicontinuum approximation. , 42(1):113–139, 2008.
M. Dobson and M. Luskin. . , 43:591–604, 2009.
M. Dobson and M. Luskin. . , 47:2455–2475, 2009.
M. Dobson, M. Luskin, and C. Ortner. Sharp stability estimates for force-based quasicontinuum approximation of homogeneous tensile deformation. Multiscale Modeling and Simulation, to appear.
M. Dobson, M. Luskin, and C. Ortner. Stability, instability, and error of the force-based quasicontinuum approximation. Arch. Rat. Mech. Anal., to appear.
W. E, J. Lu, and J. Yang. . , 74(21):214115, 2004.
W. E and P. Ming. Analysis of the local quasicontinuum method. In [*Frontiers and prospects of contemporary applied mathematics*]{}, volume 6 of [*Ser. Contemp. Appl. Math. CAM*]{}, pages 18–32. Higher Ed. Press, Beijing, 2005.
W. E and P. Ming. Cauchy-[B]{}orn rule and the stability of crystalline solids: static problems. , 183(2):241–297, 2007.
M. Gunzburger and Y. Zhang. A quadrature-rule type approximation for the quasicontinuum method. manuscript, 2008.
M. Gunzburger and Y. Zhang. Quadrature-rule type approximations to the quasicontinuum method for short and long-range interatomic interactions. manuscript, 2008.
T. Hudson and C. Ortner. On the stability of atomistic models and their cauchy–born approximations. in preparation.
P. Lin. Theoretical and numerical analysis for the quasi-continuum approximation of a material particle model. , 72(242):657–675, 2003.
P. Lin. Convergence analysis of a quasi-continuum approximation for a two-dimensional material without defects. , 45(1):313–332 (electronic), 2007.
M. Luskin and C. Ortner. An analysis of node-based cluster summation rules in the quasicontinuum method. , 47(4):3070–3086, 2009.
D. M, M. Luskin, and C. Ortner. Iterative methods for the force-based quasicontinuum approximation. arXiv:0910.2013.
R. Miller and E. Tadmor. . , 9:203–239, 2003.
R. Miller and E. Tadmor. Benchmarking multiscale methods. , 17:053001 (51pp), 2009.
P. Ming and J. Z. Yang. Analysis of a one-dimensional nonlocal quasi-continuum method. , 7(4):1838–1875, 2009.
M. Ortiz, R. Phillips, and E. B. Tadmor. . , 73(6):1529–1563, 1996.
C. Ortner. A posteriori existence in numerical computations. , 47(4):2550–2577, 2009.
C. Ortner and E. S[ü]{}li. Analysis of a quasicontinuum method in one dimension. , 42(1):57–91, 2008.
S. Prudhomme, P. T. Bauman, and J. T. Oden. Error control for molecular statics problems. , 4(5-6):647–662, 2006.
V. B. Shenoy, R. Miller, E. B. Tadmor, D. Rodney, R. Phillips, and M. Ortiz. . , 47(3):611–642, 1999.
T. Shimokawa, J. Mortensen, J. Schiotz, and K. Jacobsen. . , 69(21):214104, 2004.
E. S[ü]{}li and D. F. Mayers. . Cambridge University Press, Cambridge, 2003.
[^1]: This work was supported in part by DMS-0757355, DMS-0811039, the Department of Energy under Award Number DE-FG02-05ER25706, the Institute for Mathematics and Its Applications, the University of Minnesota Supercomputing Institute, the University of Minnesota Doctoral Dissertation Fellowship, and the EPSRC critical mass programme “New Frontier in the Mathematics of Solids” (OxMoS). The project was initiated during a visit of ML to OxMoS.\
[ .]{} To appear in J. Mech. Phys. Solids.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We demonstrate a spatially resolved autocorrelation measurement with a Bose-Einstein condensate (BEC) and measure the evolution of the spatial profile of its quantum mechanical phase. Upon release of the BEC from the magnetic trap, its phase develops a form that we measure to be quadratic in the spatial coordinate. Our experiments also reveal the effects of the repulsive interaction between two overlapping BEC wavepackets and we measure the small momentum they impart to each other.'
address: |
$^{1}$National Institute of Standards and Technology, Gaithersburg, MD 20899\
$^{2}$Georgia Southern University, Statesboro, GA 30460-8031
author:
- 'J. E. Simsarian,$^{1}$ J. Denschlag,$^{1}$ Mark Edwards,$^{1,2}$ Charles W. Clark,$^{1}$ L. Deng,$^{1}$ E. W. Hagley,$^{1}$ K. Helmerson,$^{1}$ S. L. Rolston,$^{1}$ and W.D. Phillips$^{1}$'
title: ' Imaging the Phase of an Evolving Bose-Einstein Condensate Wavefunction'
---
= 10000
A trapped Bose-Einstein condensate [@anderson0] has unique value as a source for atom lasers [@mewes] and matter-wave interferometry [@berman] because its atoms occupy the same quantum state, with uniform spatial phase. However, when released from the trapping potential, a BEC with repulsive atom-atom interactions expands, developing a non-uniform phase profile. Understanding this phase evolution will be important for applications of coherent matter waves. We have developed a new interferometric technique using spatially resolved autocorrelation to measure the functional form and time evolution of the phase of a BEC wavepacket expanding under the influence of its mean field repulsion.
In 1997, the coherence of weakly interacting BECs was demonstrated by releasing two spatially separated condensates and observing their interference [@andrews]. Subsequent experiments have further investigated condensate coherence properties. One [@stenger] used velocity-resolved Bragg diffraction [@kozuma] to probe the momentum spectrum of trapped and released BECs. A complementary experiment [@hagley2] that used matter-wave interferometry can be interpreted as a measurement of the spatial correlation function, whose Fourier transform is the momentum spectrum. These experiments showed that a trapped condensate has a uniform phase, and a released condensate develops a non-uniform phase profile. (Recently the influence of non-zero temperature on coherence properties was also investigated [@bloch]). The experiments reported in this Letter combine spatial resolution and interferometry to measure the functional form of the time-dependent phase profile of a released condensate. We also make the first measurement of the velocity imparted to two equal BEC wavepackets from their mutual mean-field repulsion [@ketterle].
We perform our experiments with a condensate of $1.8(4) \times
10^6$ [@uncertainty] sodium atoms in the $3S_{1/2}$, $F=1$, $m_{F}=-1$ state. The sample has no discernable non-condensed (i.e. thermal) component. The condensate is prepared following the method of Ref. [@kozuma] and is held in a magnetic trap with trapping frequencies $\omega_x = \sqrt{2}\omega_y = 2 \omega_z
= 2\pi \times $27 Hz. Using a scattering length of $a = 2.8$ nm, the calculated Thomas-Fermi diameters [@dalfovo] are 47 $\mu$m, 66 $\mu$m, and 94 $\mu$m, respectively.
We release the BEC from the magnetic trap and it expands, driven mostly by the mean-field repulsion of the atoms. This expansion implies the development of a nonuniform spatial phase profile (recall that the velocity field is proportional to the gradient of the quantum phase). After an expansion time $T_{0}$, we probe the phase profile with matter-wave Bragg interferometry [@giltner; @torii; @denschlag]. Our interferometer splits the BEC into two wavepackets and recombines them with a chosen overlap, producing interference fringes, which we measure with absorption imaging [@imaging]. From the dependence of the fringe spacing on the overlap, we extract the phase profile of the wavepackets.
=3.2in
Our atom interferometer [@denschlag] consists of three optically-induced Bragg-diffraction pulses applied successively in time (Fig. 1). Each pulse consists of two counter-propagating laser beams whose frequencies differ by 100 kHz. They are detuned by about $-2$ GHz from atomic resonance ($\lambda = 2\pi/k = 589$ nm) so that spontaneous emission is negligible. The first pulse has a duration of 6 $\mu$s and intensity sufficient to provide a $\pi/2$ pulse, which coherently splits the BEC into two wavepackets, $\psi_{{ A}}$ and $\psi_{{
B}}$. The wavepackets have about the same number of atoms and only differ in their momenta: $p=0$ and $p=2\hbar k$. At a time $T_{1}=1$ ms after the first Bragg pulse, the two wavepackets are completely separated and a second Bragg pulse (a $\pi$ pulse) of 12 $\mu$s duration transfers $\psi_{{ B}}$ to a state with $p\approx 0$ and $\psi_{{
A}}$ to $p\approx 2\hbar k$ [@almost]. After a variable time $T_{2}$ the wavepackets partially overlap again and we apply a third pulse, of 6 $\mu$s duration (a $\pi/2$ pulse). This last pulse splits each wavepacket into the two momentum states. The interference of the overlapping wavepackets in each of the two momentum states allows the determination of the local phase difference between them. By changing the time $T_{2}$ we vary $\delta x = x_{ A} - x_{ B}$, the separation of $\psi_{{A}}$ and $\psi_{{B}}$ at the time of the final Bragg pulse. The set of data at different $\delta x$ constitutes a new type of spatial autocorrelation measurement that is similar to the “FROG” technique [@frog] used to measure the complete field of ultrafast laser pulses. From these measurements we obtain the phase profile of the wavepackets in the $x$ direction.
=3.3in
Figure \[fig2\]a-e shows one interferometer output port for different $\delta x$ (different $T_2$) after an expansion time $T_{0}$ = 4 ms. In general, we observe straight, evenly spaced fringes (although for small $T_{0}$ and $T_{2}$ the fringes may be somewhat curved). There is a value of $\delta x = x_{0} \neq 0$ where we observe no fringes (Fig. \[fig2\]c) and the fringe spacing decreases as $|\delta x - x_{0}|$ increases. Figure \[fig2\]f, a cut through Fig. \[fig2\]d, shows the high-contrast fringes [@contrast]. Our data analysis uses the average fringe period $d$, obtained from plots like Fig. \[fig2\]f.
The fringes come from two different effects: the interference of two wavepackets with quadratic phase profile, and a relative velocity between the wavepackets’ centers. The data can be understood by calculating the fringe spacing along $x$ at output port 1 [@separable]. We assume that the phase $\phi$ of the wavefunction $f {\rm
e}^{i\phi}$ can be written as $\phi = \frac{\alpha}{2}x^{2}+\beta x$. The equal spacing of the fringes implies, as predicted in the Thomas-Fermi limit [@castin], that $\phi$ has no significant higher-order terms [@polynomials]. The curvature coefficient $\alpha$ describes the mean-field expansion of the wavepackets and $\beta$ describes a relative repulsion velocity. The velocity arises because the wavepackets experience a repulsive push as they first separate and again as they recombine. The density at port 1 (see Fig. 1) just after the final interferometer pulse is the interference pattern $|{\psi_{A1} + \psi_{B1}}|^{2}$ of the wavepackets $\psi_{{A1}}$ and $\psi_{{B1}}$: $$|f(x-\delta x)e^{i(\frac{\alpha}{2}(x-\delta x)^{2}
- \beta(x-\delta x))} + f (x)e^{i(\frac{\alpha}{2} x^{2} +
\beta x)}|^{2},
\label{wavefunctions}$$ where we assume that the amplitudes and curvatures of the wavepackets are equal and their velocities have equal magnitude and opposite direction. The cross term of (\[wavefunctions\]) is $$2f(x-\delta x)f(x){\rm cos}\left[ \left(\alpha \, \delta x +
\frac{{M}\, \delta v}{\hbar}\right)x + C\right],
\label{fringes}$$ where $M$ is the sodium mass, ${M}\, \delta v/\hbar \equiv
2 \beta$, and $C$ is independent of $x$ [@constant]. $\delta v = v_{{B}} - v_{{A}}$ is the relative repulsion velocity between the wavepackets $\psi_{{A1}}$ and $\psi_{{B1}}$. Expression (\[fringes\]) predicts fringes with spatial frequency, $$\kappa = \alpha \, \delta x + \frac{{M}\, \delta v}{\hbar},
\label{line}$$ where $|\kappa| = 2 \pi/d$. When there are no fringes, $\kappa$ = 0 and the wavepacket separation $\delta x =
x_{0} \equiv -{M}\, \delta v/ \alpha \hbar$.
=3.3in
Figure \[fig3\] plots the measured $\kappa $ vs. $\delta x$ [@separation] for $T_{0}$ = 1 and 4 ms. The data are well fit by a straight line as expected from Eq. (3) in the approximation that $\alpha$ and $\delta v$ are independent of $\delta x$. The slopes of the lines are the phase curvatures $\alpha$, and the $\kappa$ intercepts give the relative velocities $\delta v$.
We checked the validity of the data analysis procedure by analyzing data simulated with a 1-D Gross-Pitaevskii (GP) treatment. Despite variations of $\delta v$ and $\alpha$ with $\delta x$ (due to their continued evolution during the variable time $T_{2}$), we find that $\kappa$ is still linear in $\delta x$. The slopes and intercepts in general are averages over the range of $\delta x$ used in the experiment.
The interference fringes used to determine $\alpha$ and $\delta v$ are created at the time of the final interferometer pulse. Because the two outputs overlap at that moment, we wait a time $T_{3}$ for them to separate before imaging. During this time, the wavepackets continue to expand. The 1-D simulations show that the fringe spacings and the wavepackets expand in the same proportion. We correct $\kappa$ (by typically 15 $\%$) for this, using the calculated expansion from a 3-D solution of the GP equation described below.
=3.3in
The different slopes and intercepts of the two lines in Fig. \[fig3\] show that the curvature $\alpha$ and relative velocity $\delta v$ of the wavepackets depend on the release time $T_{0}$ before the first interferometer pulse. Figure \[fig4\] plots the dependence of $\alpha$ and $\delta v$ on various release times $T_{0}$. The condensate initially has a uniform phase so that immediately after its release from the trap $\alpha = 0$. We nevertheless measure a nonzero $\alpha$ for $T_{0}$ = 0 ms because the BEC expands during $T_{1}$ and $T_{2}$. As a function of time, $\alpha$ behaves as $\dot D$/$D$ where $D$ is the wavepacket diameter and $\dot D$ is its rate of change [@castin]. At early times when the mean-field energy is being converted to kinetic energy, $\dot D$ increases rapidly, [*increasing*]{} $\alpha$. At late times, after the mean-field energy has been converted, $D$ increases while $\dot D$ is nearly constant, [*decreasing*]{} $\alpha$.
We predict the time evolution of $\alpha$ using the Lagrangian Variational Method (LVM) [@perez]. The LVM uses trial wavefunctions with time dependent parameters to provide approximate solutions of the 3-D time-dependent GP equation. In the model, the effect of the interferometer pulses is to replace the original wavepacket with a superposition of wavepackets having different momenta; e.g., the action of our first interferometer pulse is $\psi_{0} \rightarrow \left(\psi_{0} + e^{i2kx}
\psi_{0}\right)/\sqrt{2}$. We use Gaussian trial wavefunctions in the LVM and, for simplicity, neglect the interaction between the wavepackets, to calculate the phase curvature $\alpha$ at the time of the last interferometer pulse. This result, with $T_{1} = T_{2}$, is the solid line of Fig. 4a.
We use energy conservation to calculate the relative repulsion velocity $\delta v$ between $\psi_{A1}$ and $\psi_{B1}$ because we neglect wavepacket interactions in the LVM. In the Thomas-Fermi approximation, we can calculate the amount of energy available for repulsion when $T_0$ = 0. A trapped condensate has $\frac{5}{7}\mu$ average total energy per particle, where $\mu$ is the chemical potential [@dalfovo]. After release from the trap, it has $\frac{2}{7}
\mu$ average mean-field energy per particle. Applying a $\pi/2$ Bragg pulse to the BEC causes a density corrugation, which increases the mean-field energy to $\frac{3}{7}\mu$ per particle. In the approximation that the wavepackets do not deform as they separate and recombine, one can show that 1/3 of the total mean-field energy goes into expansion of the wavepackets, and 2/3 is available for kinetic energy of center-of-mass motion. Therefore $\frac{2}{7}\mu$ of mean-field energy per particle is available for repulsion. The corresponding repulsion velocity is only about $10^{-2}$ of a photon recoil velocity. The repulsion energy and $\delta v$ decrease for larger $T_0$ because both are inversely proportional to the condensate volume, which we calculate with the LVM. The two curves shown in Fig. \[fig4\]b are the calculated $\delta v$ when $\delta x = 0$ (solid curve) and $\delta v$ averaged over the different $\delta x$ used in the experiment (dashed curve). The 1-D GP simulations suggest that for small $T_{0}$, the results of the experiment should be closer to the solid curve; and for large $T_0$, closer to the dashed curve. The data is consistent with this trend.
=3.4in
In a related set of experiments we performed interferometry in the trap. This differs from the experiments on a released BEC because there is no expansion before the first interferometer pulse [@contraction] and the magnetic trap changes the relative velocity of the wavepackets between the interferometer pulses (Fig. 5a). To better reveal the velocity differences, we choose $T_{1}$ = $T_{2} = T$ to suppress fringes arising from the phase curvature. As with the released BEC measurements, we observe equally spaced fringes at the output of the interferometer, although the fringes are almost entirely due to a relative velocity $v$ between the wavepackets $\psi_{{A1}}$ and $\psi_{{B1}}$ at the time of the third interferometer pulse. We obtain $v$ from the fringe periodicity after a small correction for residual phase curvature [@velcorrection].
Two effects contribute to $v$: the mutual repulsion between the wavepackets $\psi_{{A}}$ and $\psi_{{B}}$ and the different action of the trapping potential on the two wavepackets in the interferometer. The latter effect occurs because after the first Bragg pulse, $\psi_{{A}}$ remains at the minimum of the magnetic potential while $\psi_{{B}}$ is displaced. Wavepacket $\psi_{{B}}$ therefore spends more time away from the center of the trap and experiences more acceleration than $\psi_{{A}}$.
Following the last Bragg pulse, $\psi_{{A1}}$ and $\psi_{{B1}}$ have a velocity difference which for our parameters can be approximated by $v \approx -\frac{2\hbar k}{M}{\rm sin^{2}}(\omega_{x} {T}) + \delta v$ [@approx]. Figure \[fig5\]b plots $v$ versus $T$, and the curve is a fit to the above expression. We obtain the trap frequency $\omega_x /2\pi = 26.7(15)$ Hz, in excellent agreement with an independent measurement. We also obtain the relative velocity from the mean-field repulsion $\delta v = 0.49(12)$ mm/s, which we expect to be somewhat larger than for the released measurements because the wavepackets contract, producing a larger mean field.
In conclusion, we demonstrate an autocorrelating matter-wave interferometer and use it to study the evolution of a BEC phase profile by analyzing spatial images of interference patterns. We study how the phase curvature of the condensate develops in time and measure the repulsion velocity between two BEC wavepackets. Our interferometric method should be useful for characterizing other interesting condensate phase profiles. For example, it can be applied to detect excitations of a BEC with characteristic phase patterns, such as vortices and solitons [@denschlag; @burger; @matthews; @madison; @jackson]. The method should be useful for further studies of the interaction of coherent wavepackets and to study the coherence of atom lasers.
We thank T. Busch, D. Feder, and L. Collins for helpful discussions. This work was supported in part by the US Office of Naval Research and NASA. J.D. acknowledges support from the Alexander von Humboldt foundation. M.E. and C.W.C. acknowledge partial support from NSF grant numbers 9802547 and 9803377.
M. H. Anderson [*et al.*]{}, Science [**269**]{}, 198 (1995); K. B. Davis [*et al.*]{}, Phys. Rev. Lett. [**75**]{}, 3969 (1995); C. C. Bradley, C. A. Sackett, and R. G. Hulet, Phys. Rev. Lett. [**78**]{}, 985 (1997); see also C. C. Bradley [*et al.*]{}, Phys. Rev. Lett. [**75**]{}, 1687 (1995). M.-O. Mewes [*et al.*]{}, Phys. Rev. Lett. [**78**]{}, 582 (1997); E. W. Hagley [*et al.*]{}, Science [**283**]{}, 1706 (1999); I. Bloch, T. W. Hänsch, and T. Esslinger, Phys. Rev. Lett. [**82**]{}, 3008 (1999); B. P. Anderson and M. A. Kasevich, Science [**282**]{}, 1686 (1998). P. R. Berman, Ed., [*Atom Interferometry*]{} (Academic Press, Cambridge, 1997). M. R. Andrews [*et al.*]{}, Science [**275**]{}, 637 (1997). J. Stenger [*et al.*]{}, Phys. Rev. Lett. [**82**]{}, 4569 (1999). M. Kozuma [*et al.*]{}, Phys. Rev. Lett. [**82**]{}, 871 (1999). E. W. Hagley [*et al.*]{}, Phys. Rev. Lett. [**83**]{}, 3112 (1999); M. Trippenbach [*et al.*]{}, J. Phys. B [**33**]{}, 47 (2000). I. Bloch, T. W. H[ä]{}nsch, and T. Esslinger, Nature [**403**]{}, 166 (2000). The mean-field energy shift of the out-coupled wavepacket in [@stenger] can be used to infer the repulsion velocity. Here we measure the velocity directly. All uncertainties reported here are 1 standard deviation combined statistical and systematic uncertainties. F. Dalfovo [*et al.*]{}, Rev. Mod. Phys. [**71**]{}, 463 (1999). D. M. Giltner, R. W. Mc Gowan, and S. A. Lee, Phys. Rev. Lett. [**75**]{}, 2638 (1995). Y. Torii [*et al.*]{}, Phys. Rev. A [**61**]{}, 041602(R) (2000). J. Denschlag [*et al.*]{}, Science [**287**]{}, 97 (2000). The condensate was imaged by first optically pumping the atoms to the $F=2$ ground state and then imaging the absorption of a probe beam on the $F = 2 \rightarrow F' = 3$ transition. The pulse had a 5 $\mu$s duration, $\approx$ 170 mW/cm$^{2}$ intensity, and was detuned 15 MHz from resonance. The momenta are not exactly $p=0$ and $p=
2 \hbar k$ because of repulsion effects that will be discussed. R. Trebino [*et al.*]{}, Rev. Sci. Instrum. [**68**]{}, 3277 (1997). The observation is consistent with the full predicted fringe contrast when we include the finite imaging resolution. We can treat the $x$ direction independently of $y$ and $z$ when the wavefunction $\psi= f e^{i\phi(x,y,z)}$ is separable, i.e. $\phi(x,y,z) = \varphi(x) + \eta(y,z)$. Straight fringes imply this separability. Y. Castin and R. Dum, Phys. Rev. Lett. [**77**]{}, 5315 (1996). For $T_{0} =
6$ ms and an interference pattern with many fringes, we found that the coefficients of third and fourth order terms are smaller than 1$\times 10^{-5}$ $\mu$m$^{-3}$ and 1$\times 10^{-7}$ $\mu$m$^{-4}$ respectively. In practice, $C$ includes a random phase from mirror vibrations. We use $\delta x =\frac{2\hbar k}{M}{(T_{2} -
T_{1}})-x_{\epsilon}$ where $\frac{2\hbar k}{M}$ = 5.9 cm/s is the two photon recoil velocity of sodium and $x_{\epsilon}$ is a small correction of the order $\delta v \, {T_{1}}$ due to the repulsion of the wavepackets. We include the correction in our data analysis in a self-consistent manner. The correction modifies $\alpha$ insignificantly, but increases the final values of $\delta v$ by $\approx$ 0.05 mm/s. V. M. Pérez–Garcia [*et al.*]{}, Phys. Rev. Lett. [**77**]{}, 5320 (1996). In fact, after the first pulse the wavepackets contract because the reduced mean-field energy can no longer support the wavepacket size. We also correct the fringe spacings for the contraction of the wavepackets between the final interferometer pulse and when the image is taken. We assume $\delta v<<\frac{\hbar k}{M}$, $T\omega_{x}<<1$, and harmonic motion in the trap. S. Burger [*et al.*]{}, Phys. Rev. Lett. [**83**]{}, 5198 (1999). M. R. Matthews [*et al.*]{}, Phys. Rev. Lett. [**83**]{}, 2498 (1999). K. W. Madison [*et al.*]{}, Phys. Rev. Lett. [**84**]{}, 806 (2000). A. D. Jackson, G.M. Kavoulakis, and C. J. Pethick, Phys. Rev. A [**58**]{}, 2417 (1998).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Let $K$ be a finite, connected, abstract simplicial complex. The Morse complex of $K$, first introduced by Chari and Joswig, is the simplicial complex constructed from all gradient vector fields on $K$. We show that if $K$ is neither the boundary of the $n$-simplex nor a cycle, then ${\mathrm{Aut}}({\mathcal{M}}(K))\cong {\mathrm{Aut}}(K)$. In the case where $K= C_n$, a cycle of length $n$, we show that ${\mathrm{Aut}}({\mathcal{M}}(C_n))\cong {\mathrm{Aut}}(C_{2n})$. In the case where $K={\partial\Delta}^n$, we prove that ${\mathrm{Aut}}({\mathcal{M}}({\partial\Delta}^n))\cong {\mathrm{Aut}}({\partial\Delta}^n)\times \mathbb{Z}_2$. These results are based on recent work of Capitelli and Minian.'
author:
- 'Maxwell Lin, Nicholas A. Scoville'
date: April 2019
title: On the automorphism group of the Morse complex
---
*Keywords:* Discrete Morse theory, automorphism group, Morse complex, gradient vector field\
*2000 MSC:*Primary: 55U05, 08A35 Secondary: 52B05, 57Q05
Introduction
============
In 2005, Chari and Joswig [@CJ-2005] introduced the Morse complex of a simplicial complex. The Morse complex is based on Forman’s discrete Morse theory [@F-95; @F-02] where after fixing a simplicial complex $K$, one builds a new simplicial complex ${\mathcal{M}}(K)$ from the collection of all gradient vector fields or arrows on $K$. Chari and Joswig computed the homotopy type of the Morse complex when $K$ is the $n$-simplex. Ayala et al. have shown that the pure Morse complex of a tree is collapsible and some other results on the pure Morse complex of an arbitrary graph [@A-F-Q-V-08]. Kozlov studied shellability and other properties for trees [@Kozlov99], although the language of the Morse complex was not available to him at the time. Recently, Capitelli and Minian showed that the isomorphism type of the Morse complex completely determines the isomorphism type of the corresponding simplicial complex [@CM-17]. Other than these results, very little is known about the Morse complex. Its sheer size alone makes it a notoriously complex object of study.
The goal of this paper is to compute the automorphism group of the Morse complex ${\mathcal{M}}(K)$. We derive a formula relating ${\mathrm{Aut}}({\mathcal{M}}(K))$ to ${\mathrm{Aut}}(K)$ for $K$ any finite, connected, abstract simplicial complex. Our main result is the following:
\[thm: main\] Let $K$ be a finite, connected, abstract simplicial complex. Then $${\mathrm{Aut}}({\mathcal{M}}(K))\cong
\begin{cases}
{\mathrm{Aut}}(K) & \text{if } K\neq {\partial\Delta}^n,C_n \\
{\mathrm{Aut}}(C_{2n}) & \text{if } K= C_n \\
{\mathrm{Aut}}(K)\times \mathbb{Z}_2 & \text{if } K= {\partial\Delta}^n. \\
\end{cases}$$
Here ${\partial\Delta}^n$ is the boundary of the $n$-simplex and $C_n$ is the cycle of length $n$. Theorem \[thm: main\] is proved in three parts. The first is Proposition \[prop: aut(M(K))=aut(K)\] where we prove that ${\mathrm{Aut}}({\mathcal{M}}(K))\cong {\mathrm{Aut}}(K)$ for $K\neq C_n, {\partial\Delta}^n.$ In Section \[Induced Maps\], we show that one can induce an automorphism on ${\mathcal{M}}(K)$ from an automorphism of $K$ . We then show that there is an injection of ${\mathrm{Aut}}(K)$ into ${\mathrm{Aut}}({\mathcal{M}}(K))$ in Proposition \[prop: autK subgroup\]. We are then able to show that ${\mathrm{Aut}}({\mathcal{M}}(K))\cong {\mathrm{Aut}}(K)$ for $K\neq {\partial\Delta}^n, C_n$ by utilizing results of Capitelli and Minian [@CM-17]. These results concern when we may pull an automorphism of the Morse complex back to an automorphism of the original complex. The case where $K=C_n$ follows as a corollary of the more general fact that ${\mathrm{Aut}}({\mathcal{M}}(K))\cong {\mathrm{Aut}}({\mathcal{H}}(K))$ where ${\mathcal{H}}(K)$ is the Hasse diagram of $K$. We establish this later isomorphism in Proposition \[prop: aut(M(K))=aut(H(K))\] and prove that ${\mathrm{Aut}}({\mathcal{M}}(C_n))\cong {\mathrm{Aut}}(C_{2n})$ in Proposition \[prop: aut C\_n\].
The case where $K={\partial\Delta}^n$ is Proposition \[prop: aut(M(K)) pt2\]. In this case, there are automorphisms of the Morse complex which are not induced by a simplicial map on the original complex, called ghost automorphisms. We define what we call the reflection map $\pi$ which is a cosimplicial map in the sense that if $\alpha\subseteq \beta$, then $\pi(\alpha) \supseteq \pi(\beta)$. This cosimplicial map induces and then generates all the ghost automorphisms of the Morse complex of ${\partial\Delta}^n$. By studying these ghost automorphisms, we account for all automorphisms of ${\mathcal{M}}({\partial\Delta}^n)$.
The outline of this paper is as follows: Section \[sec: background\] gives necessary background, terminology, and notation. Section \[sec: The Automorphism group\] is the heart of the paper where we compute the automorphism group of ${\mathcal{M}}(K)$. In Section \[Induced Maps\], we show that any automorphism $K$ induces an automorphisms of ${\mathcal{M}}(K)$ so that there is an injective homomorphism ${\mathrm{Aut}}(K) \to {\mathrm{Aut}}({\mathcal{M}}(K))$. This injection turns out to be an isomorphism in the case where in the case where $K\neq C_n$ or ${\partial\Delta}^n$ by Proposition \[prop: aut(M(K))=aut(K)\]. We then turn to the cases $K=C_n$ and $K={\partial\Delta}^n$. We show that ${\mathrm{Aut}}({\mathcal{M}}(C_n))\cong {\mathrm{Aut}}(C_{2n})$ in Section \[sec: The Morse complex of the Hasse diagram\]. This follows fairly easily from the fact that ${\mathrm{Aut}}({\mathcal{M}}(K))\cong {\mathrm{Aut}}({\mathcal{H}}(K))$ (Proposition \[prop: aut(M(K))=aut(H(K))\]). Finally, Section \[sec: Morse complex of pd\] is devoted to computing the automorphism group of ${\partial\Delta}^n$ via the ghost automorphisms mentioned above.
Background {#sec: background}
==========
All our simplicial complexes are assumed to be abstract, finite, and connected simplicial complexes. Our reference for the basics of simplicial complexes is [@F-P11] or [@J-11].
Let $n\geq 1$ be an integer, and write $[v_n]:=\{v_0, v_1, \ldots, v_n\}$. We use $K$ to denote a simplicial complex and $\alpha, \sigma$ etc. to denote a simplex of $K$. If $K$ is a simplicial complex on $n+1$ vertices, the set $V(K):=[v_n]$ is the **vertex set** of $K$ or the set of $0$-simplices of $K$. We use $\sigma^{(i)}$ to denote a simplex of dimension $i$, and we write $\tau < \sigma^{(i)}$ to denote any face of $\sigma$ of dimension strictly less than $i$. The number $\dim(\sigma)-\dim(\tau)$ is called the **codimension of $\tau$ with respect to $\sigma$**.
A **simplicial map** $f\colon K \to L$ is a function induced by a map on the vertex sets $f_V\colon V(K)\to V(L)$ with the property that if $\sigma=v_{i_0}v_{i_1}\ldots v_{i_m}$ is a simplex in $K$, then $f(\sigma):=f_V(v_{i_0})f_V(v_{i_1})\ldots f_V(v_{i_m})$ is a simplex of $L$.
If $v_0, v_1, \ldots, v_n$ are all the vertices of a simplex $\sigma$, we will often use the notation $\sigma:=\prod\limits_{i=0}^n v_i$.
\[lem: comp\] If $f\colon A\to B$ and $g\colon B\to C$ are simplicial maps, then $(g\circ f)_V = g_V \circ f_V$.
Consider any $\sigma \in V(A)$. Then, $(g\circ f)_V(\sigma) = (g\circ f)(\sigma)$. Since $f,g$ are simplicial maps, we also know that $\dim f(\sigma) = \dim \sigma =0$, so $f(\sigma) \in V(B)$. Likewise, we know $\dim f(\sigma) = \dim g(f(\sigma))$, so $g(f(\sigma)) \in V(C)$. Therefore, $(g\circ f)_V(\sigma) = (g\circ f)(\sigma) = (g \circ f_V)(\sigma) = (g_V\circ f_V)(\sigma),$ as desired.
A simplicial map which is a bijection is a **simplicial isomorphism**, and if $f\colon K \to K$ is a simplicial isomorphism, we say that $f$ is a **simplicial automorphism**. The **automorphism group** of $K$ is defined by $${\mathrm{Aut}}(K):=\{f\colon K \to K \mid f \text{ is an automorphism}\}.$$
Because we need to refer to them below, we define a cycle $C_n$ and the boundary of the $n$-simplex ${\partial\Delta}^n.$
Let $n \geq 3$ be an integer. Define the **cycle of length $n$**, denoted $C_n$, to be the $1$-dimensional simplicial complex (graph) with vertex set $V(C_n):=\{v_0, v_1, \ldots, v_{n-1}\}$ and edge set given by $$\{v_0,v_1\}, \{v_1,v_2\}, \{v_2, v_3\}\ldots, \{v_{n-2}, v_{n-1}\}, \{v_{n-1},v_0\}.$$
Let $n\geq 1$ be an integer. The **boundary of the $n$-simplex**, denotes ${\partial\Delta}^n$ is the simplicial complex given by ${\partial\Delta}^n:=\mathcal{P}([v_n])- \{\emptyset, \{v_0,\ldots, v_{n}\}\}.$
The Morse complex
-----------------
In this section, we recall the basics of the Morse complex. Our references for discrete Morse theory in general are [@F-02; @KnudsonBook; @KozlovBook] and the Morse complex in particular are [@CJ-2005; @CM-17].
Let $K$ be a simplicial complex. A **discrete vector field** $V$ on $K$ is defined by $$V:=\{(\sigma^{(p)}, \tau^{(p+1)}) : \sigma< \tau, \text{ each simplex of } K \text{ is in at most one pair}\}.$$ Any pair in $(\sigma,\tau)\in V$ is called a **regular pair**, and $\sigma, \tau$ are called **regular simplices** or just **regular**. If $(\sigma^{(p)},\tau^{(p+1)})\in V$, we say that $p+1$ is the **index** of the regular pair. Any simplex in $K$ which is not in $V$ is called **critical**.
Let $V$ be a discrete vector field on a simplicial complex $K$. A **$V$-path** is a sequence of simplices $$\alpha^{(p)}_0, \beta^{(p+1)}_0, \alpha^{(p)}_1, \beta^{(p+1)}_1, \alpha^{(p)}_2\ldots , \beta^{(p+1)}_{k-1}, \alpha^{(p)}_{k}$$ of $K$ such that $(\alpha^{(p)}_i,\beta^{(p+1)}_i)\in V$ and $\beta^{(p+1)}_i>\alpha_{i+1}^{(p)}\neq \alpha_{i}^{(p)}$ for $0\leq i\leq k-1$. If $k\neq 0$, then the $V$-path is called **non-trivial.** A $V$-path is said to be **closed** if $\alpha_{k}^{(p)}=\alpha_0^{(p)}$. A discrete vector field $V$ which contains no non-trivial closed $V$-paths is called a **gradient vector field**.
\[thatExampleV\] An example of a gradient vector field is given on a triangulation of the Möbius band below:
$$\begin{tikzpicture}[
decoration={markings,mark=at position 0.6 with {\arrow{triangle 60}}},
]
\filldraw[fill=black!30, draw=black] (0,0)--(6,0)--(6,2)--(0,2)--cycle;
\node[inner sep=1pt, circle, fill=black] (a) at (0,0) [draw] {};
\node[inner sep=1pt, circle, fill=black] (b) at (2,0) [draw] {};
\node[inner sep=1pt, circle, fill=black] (c) at (4,0) [draw] {};
\node[inner sep=1pt, circle, fill=black] (d) at (6,0) [draw] {};
\node[inner sep=1pt, circle, fill=black] (e) at (0,2) [draw] {};
\node[inner sep=1pt, circle, fill=black] (f) at (2,2) [draw] {};
\node[inner sep=1pt, circle, fill=black] (g) at (4,2) [draw] {};
\node[inner sep=1pt, circle, fill=black] (h) at (6,2) [draw] {};
\draw[->-] (b)--(a) node[midway] {};
\draw[->-] (a)--(e) node[midway] {};
\draw[-] (a)--(f) node[midway] {};
\draw[->-] (c)--(b) node[midway] {};
\draw[-] (b)--(g) node[midway] {};
\draw[-] (b)--(f) node[midway] {};
\draw[-] (c)--(d) node[midway] {};
\draw[-] (c)--(h) node[midway] {};
\draw[-] (c)--(g) node[midway] {};
\draw[->-] (h)--(d) node[midway] {};
\draw[-] (e)--(f) node[midway] {};
\draw[->-] (f)--(g) node[midway] {};
\draw[->-] (g)--(h) node[midway] {};
\draw[-triangle 60] (1,1)--(.5,1.5);
\draw[-triangle 60] (2,.5)--(1,.5);
\draw[-triangle 60] (3,1)--(2.5,1.5);
\draw[-triangle 60] (4,.5)--(3,.5);
\draw[-triangle 60] (5,1)--(4.5,1.5);
\draw[-triangle 60] (5.5,0)--(5.5,1);
\node[anchor = north ] at (a) {};
\node[anchor = north ] at (b) {};
\node[anchor = north ] at (c) {};
\node[anchor = north ] at (d) {};
\node[anchor = south ] at (e) {};
\node[anchor = south ] at (f) {};
\node[anchor = south ] at (g) {};
\node[anchor = south ] at (h) {};
\node[anchor = east ] at (a) {{$v_0$}};
\node[anchor = west ] at (d) {{$v_3$}};
\node[anchor = east ] at (e) {{$v_3$}};
\node[anchor = west ] at (h) {{$v_0$}};
\end{tikzpicture}$$
Another gradient vector field on the Möbius band is
$$\begin{tikzpicture}[
decoration={markings,mark=at position 0.6 with {\arrow{triangle 60}}},
]
\filldraw[fill=black!30, draw=black] (0,0)--(6,0)--(6,2)--(0,2)--cycle;
\node[inner sep=1pt, circle, fill=black] (a) at (0,0) [draw] {};
\node[inner sep=1pt, circle, fill=black] (b) at (2,0) [draw] {};
\node[inner sep=1pt, circle, fill=black] (c) at (4,0) [draw] {};
\node[inner sep=1pt, circle, fill=black] (d) at (6,0) [draw] {};
\node[inner sep=1pt, circle, fill=black] (e) at (0,2) [draw] {};
\node[inner sep=1pt, circle, fill=black] (f) at (2,2) [draw] {};
\node[inner sep=1pt, circle, fill=black] (g) at (4,2) [draw] {};
\node[inner sep=1pt, circle, fill=black] (h) at (6,2) [draw] {};
\draw[->-] (b)--(a) node[midway] {};
\draw[-] (a)--(e) node[midway] {};
\draw[-] (a)--(f) node[midway] {};
\draw[->-] (c)--(b) node[midway] {};
\draw[-] (b)--(g) node[midway] {};
\draw[-] (b)--(f) node[midway] {};
\draw[-] (c)--(d) node[midway] {};
\draw[-] (c)--(h) node[midway] {};
\draw[-] (c)--(g) node[midway] {};
\draw[-] (h)--(d) node[midway] {};
\draw[->-] (e)--(f) node[midway] {};
\draw[-] (f)--(g) node[midway] {};
\draw[->-] (g)--(h) node[midway] {};
\draw[-triangle 60] (1,1)--(.5,1.5);
\draw[-triangle 60] (3,1)--(2.5,1.5);
\draw[-triangle 60] (3.5,0)--(3.5,1);
\draw[-triangle 60] (5,1)--(5.5,.5);
\node[anchor = north ] at (a) {};
\node[anchor = north ] at (b) {};
\node[anchor = north ] at (c) {};
\node[anchor = north ] at (d) {};
\node[anchor = south ] at (e) {};
\node[anchor = south ] at (f) {};
\node[anchor = south ] at (g) {};
\node[anchor = south ] at (h) {};
\node[anchor = east ] at (a) {{$v_0$}};
\node[anchor = west ] at (d) {{$v_3$}};
\node[anchor = east ] at (e) {{$v_3$}};
\node[anchor = west ] at (h) {{$v_0$}};
\end{tikzpicture}$$
If a gradient vector field $V$ has only one regular pair, we call $V$ **primitive**. Given multiple primitive gradient vector fields, we may sometimes combine them to form a new gradient vector field. This will be accomplished “overlaying" all the arrows of each primitive gradient vector field. Clearly such a construction may or may not yield a gradient vector field.
\[ex: s1\] Let primitive gradient vector fields $V_0,V_1,V_2$ be given by
$$\begin{tikzpicture}[
decoration={markings,mark=at position 0.6 with {\arrow{triangle 60}}},
]
\node[inner sep=1pt, circle, fill=black] (1) at (0,0) [draw] {};
\node[inner sep=1pt, circle, fill=black] (2) at (1,0) [draw] {};
\node[inner sep=1pt, circle, fill=black] (4) at (0,1) [draw] {};
\node[inner sep=1pt, circle, fill=black] (3) at (1,1) [draw] {};
\node[inner sep=1pt, circle, fill=black] (a) at (3,0) [draw] {};
\node[inner sep=1pt, circle, fill=black] (b) at (4,0) [draw] {};
\node[inner sep=1pt, circle, fill=black] (d) at (3,1) [draw] {};
\node[inner sep=1pt, circle, fill=black] (c) at (4,1) [draw] {};
\node[inner sep=1pt, circle, fill=black] (e) at (6,0) [draw] {};
\node[inner sep=1pt, circle, fill=black] (f) at (7,0) [draw] {};
\node[inner sep=1pt, circle, fill=black] (h) at (6,1) [draw] {};
\node[inner sep=1pt, circle, fill=black] (g) at (7,1) [draw] {};
\draw[-] (1)--(2);
\draw[-] (2)--(3);
\draw[->-] (4)--(3);
\draw[-] (4)--(1);
\draw[->-] (a)--(b);
\draw[-] (b)--(c);
\draw[-] (c)--(d);
\draw[-] (d)--(a);
\draw[-] (e)--(f);
\draw[-] (f)--(g);
\draw[-] (g)--(h);
\draw[->-] (e)--(h);
\end{tikzpicture}$$ respectively. Then $V_0,V_1$ combine to form a new gradient vector field $V$ $$\begin{tikzpicture}[
decoration={markings,mark=at position 0.6 with {\arrow{triangle 60}}},
]
\node[inner sep=1pt, circle, fill=black] (1) at (0,0) [draw] {};
\node[inner sep=1pt, circle, fill=black] (2) at (2,0) [draw] {};
\node[inner sep=1pt, circle, fill=black] (4) at (0,2) [draw] {};
\node[inner sep=1pt, circle, fill=black] (3) at (2,2) [draw] {};
\draw[->-] (1)--(2);
\draw[-] (2)--(3);
\draw[->-] (4)--(3);
\draw[-] (4)--(1);
\end{tikzpicture}$$ but clearly combining $V_1$ and $V_2$ $$\begin{tikzpicture}[
decoration={markings,mark=at position 0.6 with {\arrow{triangle 60}}},
]
\node[inner sep=1pt, circle, fill=black] (1) at (0,0) [draw] {};
\node[inner sep=1pt, circle, fill=black] (2) at (2,0) [draw] {};
\node[inner sep=1pt, circle, fill=black] (4) at (0,2) [draw] {};
\node[inner sep=1pt, circle, fill=black] (3) at (2,2) [draw] {};
\draw[->-] (1)--(2);
\draw[-] (2)--(3);
\draw[-] (4)--(3);
\draw[->-] (1)--(4);
\end{tikzpicture}$$ is not a gradient vector field since the bottom left vertex is in two pairs of $V$.
If $V,W$ are two gradient vector fields, write $V\leq W$ whenever the regular pairs of $V$ are also regular pairs of $W$. In general, we say that a collection of primitive vector fields $V_0, V_1, \ldots, V_n$ is **compatible** if there exists a gradient vector field $V$ such that $V_i\leq V$ for all $0\leq i\leq n$.
\[MorseComplexDef2\] The **Morse complex** of $K$, denote $\mathcal{M}(K)$, is the simplicial complex whose vertices are given by primitive gradient vector fields and whose $n$-simplices are given by gradient vector fields with $n+1$ regular pairs. A gradient vector field $V$ is then associated with all primitive gradient vector fields $V:=\{V_0, \ldots, V_n\}$ with $V_i\leq V$ for all $0\leq i\leq n$.
Let $K={\partial\Delta}^1=C_3$ be the simplicial complex given by
$$\begin{tikzpicture}[scale=1]
\node[inner sep=1pt, circle, fill=black](a) at (0,2) {};
\node[inner sep=1pt, circle, fill=black](b) at (1,0) {};
\node[inner sep=1pt, circle, fill=black](c) at (-1,0) {};
\draw[-] (a)--(b);
\draw[-] (a)--(c);
\draw[-] (b)--(c);
\node[anchor = south] at (a) {\small{$a$}};
\node[anchor = west] at (b) {\small{$b$}};
\node[anchor = east] at (c) {\small{$c$}};
\end{tikzpicture}$$
Here we adopt the convention that if $(x,xy)$ is a primitive vector field consisting of a vertex and edge, we denote this by $xy$. Note that this notation only works for a primitive vector of index $1$. Then one checks that the Morse complex ${\mathcal{M}}(K)$ is given by:
$$\begin{tikzpicture}[scale=1]
\node[inner sep=1pt, circle, fill=black](a) at (0,2) {};
\node[inner sep=1pt, circle, fill=black](b) at (1,0) {};
\node[inner sep=1pt, circle, fill=black](c) at (-1,0) {};
\node[inner sep=1pt, circle, fill=black](d) at (0,-1) {};
\node[inner sep=1pt, circle, fill=black](e) at (1,-3) {};
\node[inner sep=1pt, circle, fill=black](f) at (-1,-3) {};
\draw[-] (a)--(b);
\draw[-] (a)--(c);
\draw[-] (a)--(d);
\draw [white, line width=1.5mm] (b) -- (c) node[midway, left] {};
\draw[-] (b)--(c);
\draw[-] (d)--(e);
\draw[-] (d)--(f);
\draw[-] (e)--(f);
\draw[-] (b)--(e);
\draw[-] (c)--(f);
\node[anchor = east] at (a) {\small{$ab$}};
\node[anchor = west] at (b) {\small{$bc$}};
\node[anchor = east] at (c) {\small{$ca$}};
\node[anchor = east] at (d) {\small{$cb$}};
\node[anchor = west] at (e) {\small{$ac$}};
\node[anchor = east] at (f) {\small{$ba$}};
\end{tikzpicture}$$
The Automorphism group of ${\mathcal{M}}(K)$ {#sec: The Automorphism group}
============================================
This section is devoted to computing the automorphism group of ${\mathcal{M}}(K)$. We first show in section \[Induced Maps\] that certain automorphisms of $K$ induce automorphisms of ${\mathcal{M}}(K)$. It will then follow by Proposition \[prop: autK subgroup\] that ${\mathrm{Aut}}(K)$ is isomorphic to a subgroup of ${\mathrm{Aut}}({\mathcal{M}}(K))$. We show in Proposition \[prop: aut(M(K))=aut(K)\] that ${\mathrm{Aut}}(K)\cong {\mathrm{Aut}}({\mathcal{M}}(K))$ in the case where $K\neq C_n$ or ${\partial\Delta}^n$. The next two sections compute ${\mathrm{Aut}}({\mathcal{M}}(K))$ in the two excluded cases.
Induced maps on ${\mathcal{M}}(K)$ {#Induced Maps}
----------------------------------
If $f_V\colon V(K) \to V(L)$ is a bijection on the vertex sets and induces a simplicial map $f\colon K\to L$, then $f$ is an isomorphism. We are interested in isomorphisms from $K$ to $K$, i.e., automorphisms of $K$. Given an automorphism of $K$, we now define an induced automorphism on ${\mathcal{M}}(K)$.
Let $f\colon K\to K$ be an automorphism. Define the **induced automorphism on the Morse complex** $f_{\ast_V}\colon V(\mathcal{M}(K)) \to V(\mathcal{M}(K))$ by $f_{\ast_V}(v) = (f(\sigma), f(\tau))$ where $v = (\sigma^{(p)}, \tau^{(p+1)}) \in V(\mathcal{M}(K))$.
We then extend $f_{\ast_V}$ to a simplicial map on all of ${\mathcal{M}}(K)$. Below we will justify our claims that this yields a well-defined automorphism of ${\mathcal{M}}(K)$.
\[prop: primitive to primitive\] If $v =(\sigma^{(p)}, \tau^{(p+1)}) \in V(\mathcal{M}(K))$, then $f_{\ast_V}(v) \in V(\mathcal{M}(K))$.
We seek to show that $(f(\sigma), f(\tau))$ is a primitive vector of $K$. Since $f_V$ is a bijection, we know that $\dim f(\sigma) = \dim \sigma$ and $\dim f(\tau) = \dim \tau$. Hence we have $\dim f(\sigma) = \dim\sigma = \dim\tau-1 =\dim f(\tau) - 1$. Since $\sigma \subseteq \tau$, $f(\sigma)\subseteq f(\tau)$ so that $(f(\sigma), f(\tau))$ is primitive and $(f(\sigma), f(\tau))\in V({\mathcal{M}}(K)).$
\[prop: f\_\* simplicial\] Let $f\colon K \to K$ be a simplicial automorphism. Then the induced map $f_{\ast}$ is simplicial.
Suppose for the sake of contradiction that $f_{\ast}\colon \mathcal{M}(K)\to \mathcal{M}(K)$ is not a simplicial map. Then there is some $\alpha = v_{i_0}v_{i_1}\cdots v_{i_m}\in \mathcal{M}(K)$ such that $f_{\ast} (\alpha) = f_{\ast_V}(v_{i_0}) f_{\ast_V}(v_{i_1})\cdots f_{\ast_V}(v_{i_m})\not\in \mathcal{M}(K)$. This implies that the induced function $f_\ast(\alpha)$ either does not induce a gradient vector field (i.e. two vertices in $f_\ast(\alpha)$ are not compatible), or it induces a discrete vector field that contains a nontrivial closed $V$-path.\
**Case 1:** Suppose for the sake of contradiction that there are at least two distinct simplices, say $v_{i_j} = (\sigma_j, \tau_j)$ and $v_{i_k} = (\sigma_k, \tau_k)$, which are incompatible. This implies that exactly one of the following holds:
1. $f(\sigma_i) = f(\sigma_j)$
2. $f(\tau_i) = f(\tau_j)$
3. $f(\tau_j) = f(\sigma_k)$
4. $f(\sigma_j) = f(\tau_k)$.
Suppose $(1)$ is true. Since $f$ is an automorphism, this implies $\sigma_i = \sigma_j$. However, since $\tau_i \neq \tau_j$, we cannot simultaneously have both $(\sigma_i, \tau_i)$ and $(\sigma_i, \tau_j)$, so this would imply $\alpha \not\in \mathcal{M}(K)$, a contradiction. The cases in which $(2)$, $(3)$, or $(4)$ are true lead to similar contradictions, again due to the fact that $f$ is a simplicial automorphism.\
**Case 2:** Suppose for the sake of contradiction that $f_\ast(\alpha)$ contains a nontrivial closed V-path. Then, $f_\ast(\alpha)$ contains some vertices $$f_\ast(\sigma_0, \tau_0), f_\ast(\sigma_1, \tau_1), f_\ast(\sigma_2, \tau_2), \dots , f_\ast(\sigma_{k}, \tau_{k})$$ $$= (f(\sigma_0), f(\tau_0)), (f(\sigma_1), f(\tau_1)), (f(\sigma_2), f(\tau_2)), \dots , (f(\sigma_{k-1}), f(\tau_{k-1}))$$ such that $f(\sigma_0), f(\tau_0), f(\sigma_1), f(\tau_1), \dots , f(\sigma_{k-1}), f(\tau_{k-1}), f(\sigma_0)$ is a closed V-path, in which $f(\sigma_i) < f(\tau_{i-1})$ for $0 \leq i \leq k-1$ (where the indices are taken mod $k$). Since $f$ is an automorphism, it follows that $\sigma_0, \tau_0, \sigma_1, \tau_1, \dots , \sigma_{k-1}, \tau_{k-1}, \sigma_0$ must also be a nontrivial closed V-path. Since $(\sigma_0, \tau_0), (\sigma_1, \tau_1), \dots , (\sigma_{k-1},\tau_{k-1})$ are vertices in $\alpha$, it follows that $\alpha \not\in {\mathcal{M}}(K)$, a contradiction. Therefore, $f_\ast(\alpha) \in {\mathcal{M}}(K)$, so $f_{\ast}$ must be a simplicial map.
We give an example to show that if the simplicial map $f\colon K \to K$ is not an automorphism, then $f_*$ is not necessarily a simplicial map. Indeed, consider the simplicial complex $K = \Delta^2$ with vertex set $\{a,b,c\}$. Define $f_V\colon V(K)\to V(K)$ by $f_V(v) = a$ for all vertices $v$. It is easy to verify that $f$ is a simplicial map. Let $f_\ast\colon {\mathcal{M}}(K) \to {\mathcal{M}}(K)$ be induced by $f$. Notice that $(a,ab) \in {\mathcal{M}}(K)$, but $f_\ast((a,ab)) = (f(a), f(ab)) = (a,a) \not\in {\mathcal{M}}(K)$. In order to avoid these “degenerate" primitive vectors, we must impose the constraint that $f$ be an isomorphism.
We now show that a simplicial automorphism on a simplicial complex $K$ gives rise to a simplicial automorphism on ${\mathcal{M}}(K)$.
\[prop: induced auto\] Let $f\colon K\to K$ be a simplicial automorphism. Then the induced map $f_{\ast}\colon \mathcal{M}(K) \to \mathcal{M}(K)$ is a simplicial automorphism.
We know by Proposition \[prop: f\_\* simplicial\] that $f_*$ is simplicial. It thus suffices to show that $f_{\ast_V}\colon V(\mathcal{M}(K)) \to V(\mathcal{M}(K))$ is a bijection. Consider any $(\sigma, \tau) \in V(\mathcal{M}(K))$. We know that $f$ is a simplicial isomorphism, so it has an inverse $g\colon K \to K$. Thus, $(g(\sigma), g(\tau)) \in V(\mathcal{M}(K))$, as an argument similar to that in Proposition \[prop: primitive to primitive\] verifies that $(g(\sigma), g(\tau))$ is primitive. Hence, $f_{\ast_V}((g(\sigma), g(\tau))) = (f\circ g(\sigma), f\circ g(\tau)) = (\sigma, \tau)$. Therefore, $f_{\ast_V}$ is surjective. Since $V(\mathcal{M}(K))$ is finite, this implies $f_{\ast_V}$ is bijective.
We can then show that the induced map respects composition.
\[lem: induced comp\] Let $f,g\colon K\to K$ be simplicial automorphisms with induced automorphisms $f_\ast, g_\ast \colon {\mathcal{M}}(K)\to {\mathcal{M}}(K)$, respectively. Then $(f\circ g)_\ast = f_\ast \circ g_\ast$.
It suffices to show that $(f\circ g)_{\ast_V} = (f_{\ast} \circ g_{\ast})_V$. Consider any $(\sigma, \tau) \in {\mathcal{M}}(K)$. We have $$\begin{aligned}
(f\circ g)_{\ast_V}((\sigma,\tau)) &= (f\circ g(\sigma), f\circ g(\tau))\\
&= f_{\ast_V}(g(\sigma), g(\tau))\\
&= (f_{\ast_V} \circ g_{\ast_V})((\sigma, \tau)).\end{aligned}$$ Hence, $(f\circ g)_{\ast_V} = f_{\ast_V}\circ g_{\ast_V}$. By Lemma \[lem: comp\], we know that $f_{\ast_V} \circ g_{\ast_V} = (f_{\ast} \circ g_{\ast})_V $, so $(f\circ g)_{\ast_V} = (f_{\ast} \circ g_{\ast})_V$.
Because automorphisms of $K$ induce automorphisms of ${\mathcal{M}}(K)$, we next show that we are able to obtain not only elements of ${\mathrm{Aut}}({\mathcal{M}}(K))$, but that ${\mathrm{Aut}}(K)$ actually corresponds to a subgroup of ${\mathrm{Aut}}({\mathcal{M}}(K)).$
\[prop: autK subgroup\] The group $\text{Aut}(K)$ is isomorphic to a subgroup of $\text{Aut}(\mathcal{M}(K))$. In particular, there is an injective homomorphism $\phi\colon {\mathrm{Aut}}(K)\to {\mathrm{Aut}}({\mathcal{M}}(K)).$
Consider the function $\phi\colon \text{Aut}(K) \to \text{Aut}(\mathcal{M}(K))$ that sends each simplicial automorphism $f \in \text{Aut}(K)$ to its induced simplicial automorphism $f_{\ast}\in \text{Aut}(\mathcal{M}(K))$. We first show that $\phi$ is a homomorphism. Consider any $f,g \in \text{Aut}(K)$, and an arbitrary $\sigma = \prod_{i=0}^m(\sigma_i, \tau_i) \in \mathcal{M}(K)$ where each $(\sigma_i,\tau_i)$ is a vertex of $\sigma$. Since $f,g\in {\mathrm{Aut}}(K)$, $f\circ g\in {\mathrm{Aut}}(K)$. This induces $(f\circ g)_{\ast}\colon \mathcal{M}(K) \to \mathcal{M}(K)$. Notice that
$$\begin{aligned}
\phi(f\circ g)(\sigma) &=(f\circ g)_{\ast}(\sigma) \\
&= \prod_{i=0}^m (f\circ g)_{\ast_V}((\sigma_i, \tau_i)) \\
&= \prod_{i=0}^m((f\circ g)(\sigma_i), (f\circ g)(\tau_i)).\end{aligned}$$
On the other hand, we have
$$\begin{aligned}
(\phi(f) \circ \phi(g))(\sigma) &= (f_\ast \circ g_\ast)(\sigma) \\
&= f_\ast\left(\prod_{i=0}^mg_{\ast_V}((\sigma_i, \tau_i)) \right) \\
&= f_\ast\left(\prod_{i=0}^m(g(\sigma_i), g(\tau_i))\right) \\
&= \prod_{i=0}^m((f\circ g)(\sigma_i), (f\circ g)(\tau_i))\end{aligned}$$
Thus $\phi(f \circ g) = \phi(f)\circ \phi(g)$.\
We now need to show that $\phi$ is injective. Consider any $f \in {\mathrm{Ker}}(\phi)$. We claim that $f = {\mathrm{id}}_K$. Again, consider an arbitrary $\sigma = \prod_{i=0}^m(\sigma_i, \tau_i) \in \mathcal{M}(K)$. Notice $f$ induces a simplicial map $f_\ast \colon {\mathcal{M}}(K) \to {\mathcal{M}}(K)$. Since $f \in {\mathrm{Ker}}(\phi)$, we have $$\begin{aligned}
\phi(f)(\sigma) &= \sigma\\
f_\ast (\sigma) &= \sigma\\
\prod_{i=0}^m f_{\ast_V}((\sigma_i,\tau_i)) &= \prod_{i=0}^m (\sigma_i,\tau_i)\\
\prod_{i=0}^m (f(\sigma_i),f(\tau_i)) &= \prod_{i=0}^m (\sigma_i,\tau_i).\end{aligned}$$ As this holds for any choice of $\sigma$, we conclude that $f = {\mathrm{id}}_K$. Since ${\mathrm{Ker}}(\phi)$ is trivial, it follows that $\phi$ is injective.
Proposition \[prop: autK subgroup\] guarantees that ${\mathrm{Aut}}(K)\leq {\mathrm{Aut}}({\mathcal{M}}(K))$ by inducing an automorphism of ${\mathcal{M}}(K)$ from an automorphism of $K$. Thus if every automorphism of ${\mathcal{M}}(K)$ is induced by an automorphism of $K$, then ${\mathrm{Aut}}({\mathcal{M}}(K))={\mathrm{Aut}}(K).$ We will first show that this is indeed the case for $K\neq C_n, {\partial\Delta}^n$.
\[prop: aut(M(K))=aut(K)\] If $K$ is a simplicial complex other than $C_n$ or ${\partial\Delta}^n$, then ${\mathrm{Aut}}({\mathcal{M}}(K)) \cong {\mathrm{Aut}}(K)$ .
As in Proposition \[prop: autK subgroup\], define a function $\phi\colon \text{Aut}(K) \to \text{Aut}(\mathcal{M}(K))$ that sends each simplicial automorphism $f \in \text{Aut}(K)$ to its induced simplicial automorphism $f_{\ast}\in \text{Aut}(\mathcal{M}(K))$. We know that $\phi$ is an injective homomorphism by Proposition \[prop: autK subgroup\].
We now show that there is a surjection onto ${\mathrm{Aut}}({\mathcal{M}}(K))$. If so, since ${\mathcal{M}}(K)$ and $K$ are finite, this implies that $\phi$ is an isomorphism. Let $F\in {\mathrm{Aut}}({\mathcal{M}}(K)).$ If $K$ is a 1-dimensional simplicial complex (i.e., a graph) other than $C_n$, then $F$ is induced by a simplicial isomorphism $f\colon K\to K$ by [@CM-17 Theorem 3.5]. Thus the result for $K$ a graph other than $C_n$.
Now suppose that $K\neq {\partial\Delta}^n$ is a simplicial complex of dimension greater than or equal to $2$, and let $F\in {\mathrm{Aut}}({\mathcal{M}}(K)).$ In the proof of Theorem A [@CM-17], Capitelli and Minian construct a simplicial isomorphism $f\colon K \to K$ that induces the given $F\colon {\mathcal{M}}(K)\to {\mathcal{M}}(K).$ Their construction of $f$ relies on a condition that is not satisfied for $K={\partial\Delta}^n$ according to the contrapositive of [@CM-17 Theorem 4.2]. Thus the result for $K\neq {\partial\Delta}^n$ a simplicial complex of dimension greater than $2$.
The Morse complex of the Hasse diagram {#sec: The Morse complex of the Hasse diagram}
--------------------------------------
In this section, we will show that computing the automorphism group of ${\mathcal{M}}(K)$ is equivalent to computing the automorphism group of ${\mathcal{H}}(K)$. We will then immediately be able to compute the Morse complex of $C_n$, one of the two cases excluded in Proposition \[prop: aut(M(K))=aut(K)\]. Section \[sec: Morse complex of pd\] is then devoted to computing the Morse complex of our final special case, $K={\partial\Delta}^n.$ We briefly recall here the definition of the Hasse diagram.
The **Hasse diagram** of $K$, denoted $\mathcal{H}_K$ or $\mathcal{H}$, is defined as the partially ordered set of simplices of $K$ ordered by the face relations. We view ${\mathcal{H}}$ as a graph.
We adopt the convention that if $\sigma^{(p)}<\tau^{(p+1)}$ are two nodes of the Hasse diagram, the edge joining them is denoted $\sigma\tau$.
\[prop: aut(M(K))=aut(H(K))\] For any simplicial complex $K$, ${\mathrm{Aut}}({\mathcal{M}}(K)) \cong {\mathrm{Aut}}({\mathcal{H}}(K))$.
We will construct an isomorphism $\phi\colon {\mathrm{Aut}}({\mathcal{H}}(K)) \to {\mathrm{Aut}}({\mathcal{M}}(K))$. Consider an arbitrary automorphism $f \in {\mathrm{Aut}}({\mathcal{H}}(K)) $. Define a function\
$m \colon E({\mathcal{H}}(K)) \to V({\mathcal{M}}(K))$ by $m(\sigma\tau) = (\sigma, \tau)$, where $\sigma$ is a codimension $1$ face of $\tau$. Notice that $m$ has an inverse $m^{-1} \colon V({\mathcal{M}}(K)) \to E({\mathcal{H}}(K))$ given by $m^{-1}(\sigma, \tau)=\sigma\tau$. Define a function $g_V\colon V({\mathcal{M}}(K)) \to V({\mathcal{M}}(K))$ by $g_V := m\circ f \circ m^{-1}$. Clearly $g_V((\sigma,\tau))$ is another vertex of $V({\mathcal{M}}(K))$. Since $f$ is a simplicial isomorphism, it has an inverse $f^{-1} \colon {\mathcal{H}}(K) \to {\mathcal{H}}(K)$. It is then clear that $g_V$ has an inverse $g_V^{-1} := m^{-1} \circ f^{-1} \circ m$. Hence $g_V$ is a bijection.\
We now show that $g$ is simplicial. Consider any $\sigma := (\sigma_1, \tau_1)(\sigma_2,\tau_2)\cdots (\sigma_k,\tau_k)\in {\mathcal{M}}(K)$. Observe that $$\begin{aligned}
g(\sigma) &= (m\circ f\circ m^{-1})(\sigma) \\
&= (m\circ f)\left( \prod_{i=1}^k \sigma_i\tau_i\right)\\
&= m\left( \prod_{i=1}^k f_V(\sigma_i) f_V(\tau_i)\right)\\
&= \prod_{i=1}^k m(f_V(\sigma_i)f_V(\tau_i)).\end{aligned}$$
Suppose that $\prod_{i=1}^k m(f_V(\sigma_i)f_V(\tau_i))= \prod_{i=1}^k (f_V(\sigma_i), f_V(\tau_i))$ as the case $\prod_{i=1}^k m(f_V(\sigma_i)f_V(\tau_i))= \prod_{i=1}^k (f_V(\tau_i), f_V(\sigma_i))$ is identical. We claim that each $(f_V(\sigma_i), f_V(\tau_i))$ is a primitive vector in ${\mathcal{M}}(K)$. Since $\sigma_i\tau_i \in {\mathcal{H}}(K)$, by the construction of ${\mathcal{H}}(K)$, we know that either $\sigma_i$ is a codimension $1$ face of $\tau_i$, or vice versa, and since $f$ is an automorphism, $f_V(\sigma_i)f_V(\tau_i) \in {\mathcal{H}}(K)$. Again, by construction of ${\mathcal{H}}(K)$, we know that either $f_V(\sigma_i)$ is a codimension $1$ face of $f_V(\tau_i)$, or vice versa. Thus, applying $m$ to $f_V(\sigma_i)f_V(\tau_i)$ will form a primitive vector $(f_V(\sigma_i)f_V(\tau_i))\in {\mathcal{M}}(K)$.\
We now must show that all primitive vectors in $g(\sigma) = \prod_{i=1}^k (f_V(\sigma_i), f_V(\tau_i))$ are compatible. Suppose for the sake of contradiction they are not all compatible. The first possibility is that there are two compatible primitive vectors $(\sigma_k,\tau_k),(\sigma_{\ell}, \tau_{\ell}) \in V({\mathcal{M}}(K))$ such that
$$(f_V(\sigma_k),f_V(\tau_k))(f_V(\sigma_{\ell}),f_V(\tau_{\ell})) \not\in {\mathcal{M}}(K).$$ However, this implies that $\{f_V(\sigma_k),f_V(\tau_k)\} \cap \{f_V(\sigma_{\ell}),f_V(\tau_{\ell})\} \neq \emptyset$. Since $f_V$ is bijective, this implies that $\{\sigma_k,\tau_k\} \cap \{\sigma_{\ell}, \tau_{\ell}\} \neq \emptyset$, so that $(\sigma_k,\tau_k),(\sigma_{\ell}, \tau_{\ell})$ are not compatible, which is a contradiction. The second possibility is that $g(\sigma)$ contains a nontrivial closed $V$-path. In a similar manner to the first case, this nontrivial closed $V$-path in the image of $g$ can be pulled back to obtain a nontrivial closed $V$-path in $\sigma$, a contradiction. We conclude that $g(\sigma)\in {\mathcal{M}}(K)$ so that $g$ is a simplicial map.
Since $g_V$ is a bijection and $g$ is simplicial, it follows that $g$ is a simplicial automorphism on ${\mathcal{M}}(K)$. Define $\phi(f) := g$. We first show that $\phi$ is a homomorphism. Suppose we have $a, b \in {\mathrm{Aut}}({\mathcal{H}}(K))$. We seek to show that $\phi(a\circ b) = \phi(a)\circ \phi(b)$. By the definition of a simplicial map, it suffices to show that $\phi(a\circ b)_V = (\phi(a)\circ \phi(b))_V$. We have $$\begin{aligned}
\phi(a\circ b)_V&=& m\circ a\circ b \circ m^{-1}\\
&=& (m \circ a \circ m^{-1}) \circ (m \circ b \circ m^{-1})\\
&=& \phi(a)_V \circ \phi(b)_V.\end{aligned}$$ By Lemma \[lem: comp\], we know that $\phi(a\circ b)_V = \phi(a)_V \circ \phi(b)_V = (\phi(a)\circ \phi(b))_V$, as desired.\
To see that $\phi$ is injective, suppose that $\phi(a) = \phi(b)$. Then we have $m \circ a \circ m^{-1} = m\circ b \circ m^{-1}$ which implies that $a=b$.\
Finally, we show that $\phi$ is surjective. For any $g \in {\mathrm{Aut}}({\mathcal{M}}(K))$, $g$ must be induced by some $g_V: V({\mathcal{M}}(K)) \to V({\mathcal{M}}(K))$. Then construct an $f \colon {\mathcal{H}}(K)\to {\mathcal{H}}(K)$ defined by $f = m^{-1} \circ g \circ m$. We see that $\phi(f) = m\circ (m^{-1} \circ g \circ m) \circ m^{-1} = g$, as desired. Therefore, we conclude that ${\mathrm{Aut}}({\mathcal{H}}(K)) \cong {\mathrm{Aut}}({\mathcal{M}}(K))$.
We now are able to easily compute the automorphism group of $C_n$.
\[prop: aut C\_n\] If $K = C_n$, then then ${\mathrm{Aut}}({\mathcal{M}}(C_n)) \cong {\mathrm{Aut}}(C_{2n}) $.
It suffices to show ${\mathrm{Aut}}({\mathcal{H}}(C_n)) \cong {\mathrm{Aut}}(C_{2n})$. We construct the Hasse diagram of $C_n$:
$$\begin{tikzpicture}[scale=1.7]
\node[inner sep=1pt, circle, fill=black](a1) at (0,0) {};
\node[inner sep=1pt, circle, fill=black](a2) at (1,0) {};
\node[inner sep=1pt, circle, fill=black](a3) at (2,0) {};
\node[inner sep=1pt, circle, fill=black](a4) at (4,0) {};
\node[inner sep=1pt, circle, fill=black](ak) at (5,0) {};
\node at (2.7,0) {\ldots};
\node[inner sep=1pt, circle, fill=black](a1a2) at (0,1) {};
\node[inner sep=1pt, circle, fill=black](a2a3) at (1,1) {};
\node[inner sep=1pt, circle, fill=black](a3a4) at (2,1) {};
\node[inner sep=1pt, circle, fill=black](a4ak) at (4,1) {};
\node[inner sep=1pt, circle, fill=black](a1ak) at (5,1) {};
\node at (2.6,1) {\ldots};
\node at (2.9,0.6) {$\ddots$};
\draw[-] (a1)--(a1a2); \draw[-] (a1)--(2.6, 0.45); \draw[-] (3.2, 0.65)--(a1ak);
\draw[-] (a2)--(a1a2); \draw[-] (a2)--(a2a3);
\draw[-] (a3)--(a2a3); \draw[-] (a3)--(a3a4);
\draw[-] (ak)--(a1ak);
\draw[-] (a4)--(a4ak);
\draw[-] (ak)--(a4ak);
\draw[-] (a3a4)--(2.4,0.7);
\draw[-] (a4)--(3.5,0.3);
\node[anchor = east] at (a1) {\small{$a_1$}};
\node[anchor = east] at (a2) {\small{$a_2$}};
\node[anchor = east] at (a3) {\small{$a_3$}};
\node[anchor = east] at (a4) {\small{$a_{n-1}$}};
\node[anchor = east] at (ak) {\small{$a_n$}};
\node[anchor = east] at (a1a2) {\small{$a_1a_2$}};
\node[anchor = south] at (a1ak) {\small{$a_1a_n$}};
\node[anchor = east] at (a2a3) {\small{$a_2a_3$}};
\node[anchor = east] at (a3a4) {\small{$a_3a_4$}};
\node[anchor = east] at (a4ak) {\small{$a_{n-1}a_n$}};
\end{tikzpicture}$$
It is clear that this Hasse Diagram can be redrawn as: $$\begin{tikzpicture}
\node[regular polygon,regular polygon sides=13,minimum size=5cm,draw] (a){};
\foreach \x in {9,...,13}{\node[circle,fill,inner sep=1pt] at (a.corner \x) {};}
\foreach \x in {1,...,5}{\node[circle,fill,inner sep=1pt] at (a.corner \x) {};}
\draw [line width=1.1cm, white] (1.1,-2) -- (-2.3,-2);
\node at (-1.71,-1.55) {$\ddots$};
\node at (0.85,-2.17) {$\dots$};
\node[anchor = south] at (a.corner 1) {\small{$a_1$}};
\node[anchor = east] at (a.corner 2) {\small{$a_1a_n$}};
\node[anchor = east] at (a.corner 3) {\small{$a_n$}};
\node[anchor = east] at (a.corner 4) {\small{$a_{n-1}a_n$}};
\node[anchor = east] at (a.corner 5) {\small{$a_{n-1}$}};
\node[anchor = west] at (a.corner 12) {\small{$a_2$}};
\node[anchor = west] at (a.corner 10) {\small{$a_3$}};
\node[anchor = west] at (a.corner 13) {\small{$a_1a_2$}};
\node[anchor = west] at (a.corner 11) {\small{$a_2a_3$}};
\node[anchor = west] at (a.corner 9) {\small{$a_3a_4$}};
\end{tikzpicture}$$ which is $C_{2n}$. Therefore, ${\mathrm{Aut}}({\mathcal{H}}(C_n)) \cong {\mathrm{Aut}}(C_{2n})$.
Let $D_{n}$ be the dihedral group of order $n$. It is well known that $D_{2n}\cong D_n \times \mathbb{Z}_2$ for $n$ odd. Since ${\mathrm{Aut}}(C_n) \cong D_{2n}$, we have ${\mathrm{Aut}}({\mathcal{M}}(C_n)) \cong {\mathrm{Aut}}(C_{2n}) \cong {\mathrm{Aut}}(C_n) \times \mathbb{Z}_2$ whenever $n$ is odd, yielding the same formula as the automorphism group of ${\partial\Delta}^n$ that we show in Section \[sec: Morse complex of pd\].
The Morse complex of $\partial\Delta^n$ {#sec: Morse complex of pd}
---------------------------------------
We now investigate the case where $K={\partial\Delta}^n$. In this case, as in the $K=C_n$ case, there are automorphisms of the Morse complex which are not induced by an automorphism of the original complex. While these automorphisms of the Morse complex are not induced by simplicial maps, we will show below that they are induced by what we are calling the reflection map. This is not a simplicial map, but rather a “cosimplicial map," a term we define in Definition \[defn: cosimplicial\]. The automorphisms induced by the simplicial maps and those induced by this cosimplicial map will then be shown to generate all possible automorphisms of the Morse complex, allowing us to compute and ${\mathrm{Aut}}({\mathcal{M}}({\partial\Delta}^n)$ in Theorem \[prop: aut(M(K)) pt2\]. To illustrate, we first look at an example.
\[ex: delta1\] Let $K={\partial\Delta}^1=C_3$, and recall that we computed the Morse complex of $K$ in Example \[ex: s1\]. For reference, we give the Morse complex here, noting again the convention that $ab$ is shorthand for the primitive vector $(a,ab)$.
$$\begin{tikzpicture}[scale=1]
\node[inner sep=1pt, circle, fill=black](a) at (0,2) {};
\node[inner sep=1pt, circle, fill=black](b) at (1,0) {};
\node[inner sep=1pt, circle, fill=black](c) at (-1,0) {};
\node[inner sep=1pt, circle, fill=black](d) at (0,-1) {};
\node[inner sep=1pt, circle, fill=black](e) at (1,-3) {};
\node[inner sep=1pt, circle, fill=black](f) at (-1,-3) {};
\draw[-] (a)--(b);
\draw[-] (a)--(c);
\draw[-] (a)--(d);
\draw [white, line width=1.5mm] (b) -- (c) node[midway, left] {};
\draw[-] (b)--(c);
\draw[-] (d)--(e);
\draw[-] (d)--(f);
\draw[-] (e)--(f);
\draw[-] (b)--(e);
\draw[-] (c)--(f);
\node[anchor = east] at (a) {\small{$ab$}};
\node[anchor = west] at (b) {\small{$bc$}};
\node[anchor = east] at (c) {\small{$ca$}};
\node[anchor = east] at (d) {\small{$cb$}};
\node[anchor = west] at (e) {\small{$ac$}};
\node[anchor = east] at (f) {\small{$ba$}};
\end{tikzpicture}$$ We see that ${\mathrm{Aut}}({\partial\Delta}^1)$ is the symmetries of a triangle, and has six automorphisms. Meanwhile, ${\mathrm{Aut}}({\mathcal{M}}({\partial\Delta}^1))$ has twelve automorphisms. Six of the automorphisms of ${\mathcal{M}}({\partial\Delta}^1)$ arise from automorphisms of ${\partial\Delta}^n$, but the other six do not. For example, define $F\colon {\mathcal{M}}({\partial\Delta}^1)\to {\mathcal{M}}({\partial\Delta}^1)$ by $$\begin{aligned}
F(ab)&=&cb\\
F(ba)&=&ca\\
F(ca)&=&ba\\
F(cb)&=&ab\\
F(ac)&=&bc\\
F(bc)&=&ac.\end{aligned}$$ Then it is easy to see that $F$ is a simplicial automorphism. However, it is not induced by any automorphism of ${\partial\Delta}^1$. Furthermore, composition of $F$ with any automorphism induced by an automorphism of ${\partial\Delta}^1$ yields a new automorphism of ${\mathcal{M}}({\partial\Delta}^1)$ that is not induced by an automorphism of ${\partial\Delta}^1$. For example, if $f\colon {\partial\Delta}^1\to {\partial\Delta}^1$ is given by $f(a)=a, f(b)=c$, and $f(c)=b$, then $F\circ f_*$ is an automorphism of ${\mathcal{M}}({\partial\Delta}^1)$ which is not equal to any automorphism induced from ${\partial\Delta}^1$.
The map $F$ in Example \[ex: delta1\] is what in general we call $\pi_*$, the induced map of the reflection map $\pi_n \colon {\partial\Delta}^n\to {\partial\Delta}^n$ given below in Definition \[defn: reflection map\]. We will show in Lemma \[lem: ghost2\] that $\pi_*$ is a simplicial automorphism of ${\mathcal{M}}({\partial\Delta}^n)$ and that it generates all the “missing" automorphisms of ${\mathcal{M}}({\partial\Delta}^n)$ in Theorem \[prop: aut(M(K)) pt2\].
\[defn: reflection map\] Let ${\partial\Delta}^{n}$ be the boundary of the $n$-simplex on the vertices $\{v_0, v_1, \ldots, v_n\}$ and write $\delta:=v_0v_1\cdots v_{n}$. Define the **reflection map** $\pi_n=\pi\colon {\partial\Delta}^n \to {\partial\Delta}^n$ by $\pi(\sigma) := \delta - \sigma $.
The reflection map is a cosimplicial map in the following sense:
\[defn: cosimplicial\] Let $K$ be a simplicial complex, $f\colon K \to K$ a function such that if $\sigma$ is a simplex in $K$, $f(\sigma)$ is a simplex of $K$. Then $f$ is called a **cosimplicial map** if whenever $\tau\subseteq \sigma$, then $f(\tau)\supseteq f(\sigma).$ If in addition $f$ is a bijection on the simplices of $K$, we say that $f$ is a **cosimplicial automorphism**.
We now illustrate how the reflection map $K={\partial\Delta}^2$ gives rise to an automorphism on the Hasse diagram of $\partial\Delta^2$.
$$\begin{tikzpicture}[scale=1.7]
\node[inner sep=1pt, circle, fill=black](a) at (0,0) {};
\node[inner sep=1pt, circle, fill=black](b) at (1,0) {};
\node[inner sep=1pt, circle, fill=black](c) at (2,0) {};
\node[inner sep=1pt, circle, fill=black](d) at (3,0) {};
\node[inner sep=1pt, circle, fill=black](ab) at (-1,1) {};
\node[inner sep=1pt, circle, fill=black](ac) at (0,1) {};
\node[inner sep=1pt, circle, fill=black](ad) at (1,1) {};
\node[inner sep=1pt, circle, fill=black](bc) at (2,1) {};
\node[inner sep=1pt, circle, fill=black](bd) at (3,1) {};
\node[inner sep=1pt, circle, fill=black](cd) at (4,1) {};
\node[inner sep=1pt, circle, fill=black](bcd) at (3,2) {};
\node[inner sep=1pt, circle, fill=black](acd) at (2,2) {};
\node[inner sep=1pt, circle, fill=black](abd) at (1,2) {};
\node[inner sep=1pt, circle, fill=black](abc) at (0,2) {};
\draw[-] (a)--(ab); \draw[-] (a)--(ac); \draw[-] (a)--(ad);
\draw[-] (b)--(ab); \draw[-] (b)--(bc); \draw[-] (b)--(bd);
\draw[-] (c)--(ac); \draw[-] (c)--(bc); \draw[-] (c)--(cd);
\draw[-] (d)--(ad); \draw[-] (d)--(bd); \draw[-] (d)--(cd);
\draw[-] (bcd)--(bc); \draw[-] (bcd)--(bd); \draw[-] (bcd)--(cd);
\draw[-] (acd)--(ac); \draw[-] (acd)--(ad); \draw[-] (acd)--(cd);
\draw[-] (abd)--(ab); \draw[-] (abd)--(ad); \draw[-] (abd)--(bd);
\draw[-] (abc)--(ab); \draw[-] (abc)--(ac); \draw[-] (abc)--(bc);
\node[anchor = east] at (a) {\small{$a$}};
\node[anchor = east] at (b) {\small{$b$}};
\node[anchor = east] at (c) {\small{$c$}};
\node[anchor = east] at (d) {\small{$d$}};
\node[anchor = east] at (ab) {\small{$ab$}};
\node[anchor = east] at (ad) {\small{$ad$}};
\node[anchor = east] at (ac) {\small{$ac$}};
\node[anchor = east] at (bc) {\small{$bc$}};
\node[anchor = east] at (bd) {\small{$bd$}};
\node[anchor = east] at (cd) {\small{$cd$}};
\node[anchor = east] at (abc) {\small{$abc$}};
\node[anchor = east] at (bcd) {\small{$bcd$}};
\node[anchor = east] at (acd) {\small{$acd$}};
\node[anchor = east] at (abd) {\small{$abd$}};
\end{tikzpicture}$$
Observe that $\partial\Delta^2$ demonstrates rotational symmetry about the center of the diagram. (Notice that $K=C_n$ also demonstrates rotational symmetry, as can be seen in the Hasse diagram in Proposition \[prop: aut C\_n\].) An automorphism of ${\partial\Delta}^1$ induces an automorphism of the Hasse diagram which permutes the $0-$simplices, but the rotational symmetry of the Hasse diagram arises from the reflection map. For instance, the map induced by the reflection map sends $a\leftrightarrow bcd$, $b\leftrightarrow acd$, and so on, which visually rotates the Hasse Diagram upside down. For simplicial complexes other than $K=\partial\Delta^n$ and $C_n$, the Hasse diagram does not exhibit this rotational symmetry. Hence all automorphisms of the Hasse diagrams correspond to permutations of the $0-$simplices, which in turn correspond to automorphisms on the original $K$. (This follows from Proposition \[prop: aut(M(K))=aut(K)\].)
We now give several basic properties of the reflection map.
\[lem: ghost1\] The reflection map is a cosimplicial automorphism that commutes with all members of ${\mathrm{Aut}}(K)$.
Clearly if $\sigma$ is a simplex in ${\partial\Delta}^n$, then $\delta-\sigma$ is a simplex in ${\partial\Delta}^n$ since ${\partial\Delta}^n$ by definition is made up of all proper, nonempty subsets of $\delta$. If $\sigma\subseteq \tau$, then $\pi(\sigma)=\delta-\sigma \supseteq \delta-\tau=\pi(\tau)$. Hence $\pi$ is a cosimplicial map. Next notice that we have $\pi(\pi(\sigma)) = \sigma$, so $\pi$ is its own right and left inverse. It follows that $\pi$ is a cosimplicial automorphism.
To see that $\pi$ commutes with all simplicial automorphisms of ${\partial\Delta}^n$, let $f \in {\mathrm{Aut}}({\partial\Delta}^n)$. Then $f$ is induced by a bijection $f_V\colon V(K)\to V(K)$. Consider any simplex $\sigma \in K$. We seek to show that $\pi(f(\sigma))= f ( \pi (\sigma))$. It thus suffice to show that $\delta - f(\sigma) = f(\delta - \sigma)$. We proceed by subset inclusion. Consider any vertex $v \in \delta - f(\sigma)$. Since $f$ is an automorphism, we can express $ v = f(w)$ for some vertex $w$. Then we have $f(w) \not \in f(\sigma)$, so $w \not \in \sigma$. Thus, $w \in \delta - \sigma$, so it follows that $v= f(w) \in f(\delta - \sigma)$. Hence $\delta - f(\sigma) \subseteq f(\delta - \sigma)$. For the other direction, consider any $v \in f(\delta - \sigma)$. Again, we can express $v = f(w)$ for some vertex $w$, hence $w \in \delta - \sigma$. Then $w \not\in \sigma$, so $f(w) \not\in f(\sigma)$. So $v = f(w) \in \delta - f(\sigma)$. Therefore, $f(\delta - \sigma) \subseteq \delta - f(\sigma) $. We conclude that $\delta - f(\sigma) = f(\delta - \sigma)$ so that the reflection map commutes with all of ${\mathrm{Aut}}({\partial\Delta}^n).$
As in Proposition \[prop: f\_\* simplicial\], the reflection map induces a function on the Morse complex $\pi_{\ast_V} \colon V({\mathcal{M}}({\partial\Delta}^n)) \to V({\mathcal{M}}({\partial\Delta}^n))$ defined by $\pi_{\ast_V}((\sigma,\tau)) = (\pi(\tau), \pi(\sigma))$. Even though $\pi$ is not a simplicial map, the induced map is a map on the vertex set of ${\mathcal{M}}({\partial\Delta}^n)$ which induces a simplicial map on ${\mathcal{M}}({\partial\Delta}^n)$. The following lemma shows that this induced map on the Morse complex behaves in a similar way to the cosimplicial automorphism on ${\partial\Delta}^n.$
\[lem: ghost2\] Let $\pi_n=\pi\colon {\partial\Delta}^n\to {\partial\Delta}^n$ be the reflection map, and $\pi_{\ast_V} \colon V({\mathcal{M}}({\partial\Delta}^n)) \to V({\mathcal{M}}({\partial\Delta}^n))$ the induced function on the Morse complex. Then $\pi_{\ast_V}$ is a bijection that commutes with all bijections $g_{\ast_V}\colon V({\mathcal{M}}({\partial\Delta}^n)) \to V({\mathcal{M}}({\partial\Delta}^n))$ that are induced by some $g \in {\mathrm{Aut}}({\partial\Delta}^n)$ . Moreover, the induced function $\pi_\ast$ is a simplicial map that commutes with all members of ${\mathrm{Aut}}({\mathcal{M}}({\partial\Delta}^n))$.
That $\pi_{\ast_V}$ is a bijection follows from the fact that $\pi_{\ast_V}$ is its own inverse; that is, $$\pi_{\ast_V}\circ \pi_{\ast_V}((\sigma,\tau)) = \pi_{\ast_V}((\pi(\tau), \pi(\sigma))) = (\pi(\pi(\sigma)),\pi(\pi(\tau))) = (\sigma,\tau).$$
To see that $\pi_{\ast_V}$ commutes with any induced bijection $g_{\ast_V}$ observe that
$$\begin{aligned}
\pi_{\ast_V} \circ g_{\ast_V} ((\sigma,\tau)) &=& \pi_{\ast_V} ((g(\sigma), g(\tau)) \\
&=& (\pi\circ g(\tau), \pi\circ g(\sigma))\\
&=&(g\circ \pi(\tau), g\circ \pi(\sigma)) \\
&=& g_{\ast_V}(\pi(\tau), \pi(\sigma))\\
&=& g_{\ast_V}\circ \pi((\sigma,\tau))\end{aligned}$$
where Lemma \[lem: ghost1\] justifies the fact that $\pi$ and $g$ commute.
Since $\pi_{\ast_V} \circ g_{\ast_V}=g_{\ast_V}\circ \pi_{\ast_V}$, they induce the same function on ${\mathcal{M}}(K)$. Then $\pi_{\ast_V}\circ g_{\ast_V}$ induces $\pi_{\ast} \circ g_\ast$, and $g_{\ast_V}\circ \pi_{\ast_V} $ induces $g_\ast\circ\pi_\ast$, thus $\pi_{\ast} \circ g_\ast = g_\ast\circ\pi_\ast$.
It remains to verify that this is a simplicial map. Consider any $\sigma = \prod_{i=0}^k (\alpha_i, \beta_i)\in {\mathcal{M}}(K)$. We seek to show that $\pi_\ast(\sigma) \in {\mathcal{M}}(K)$. As we already know that $\pi_{\ast_V}$ is a bijection, it remains to show that if $\pi_\ast(\sigma)$ contains incompatible primitive vectors, so does $\sigma$.\
**Case 1:** There exists two primitive vectors $\pi_\ast((\alpha_i, \beta_i)) = (\pi(\beta_i), \pi(\alpha_i))$ and $\pi_\ast((\alpha_j, \beta_j)) = (\pi(\beta_j),\pi(\alpha_j))$ of $\pi_\ast(\sigma)$ that are not compatible. Then, $\{\pi(\beta_i), \pi(\alpha_i)\} \cap \{\pi(\beta_j), \pi(\alpha_j)\} \neq \emptyset$. As $\pi$ is bijective, it follows that $\{\beta_i, \alpha_i\} \cap \{\beta_j,\alpha_j\} \neq \emptyset$, so $(\alpha_i,\beta_i)$ and $(\alpha_j, \beta_j)$ are not compatible.\
**Case 2:** There exists a nontrivial $V$-path $$(\pi(\beta_{i_0}), \pi(\alpha_{i_0})),(\pi(\beta_{i_1}), \pi(\alpha_{i_1})) , (\pi(\beta_{i_2}), \pi(\alpha_{i_2})), \dots , (\pi(\beta_{i_m}), \pi(\alpha_{i_m}))$$ where each $\pi(\beta_{i_j})$ is a codimension $1$ face of $\pi(\alpha_{i_{j-1}})$ for each $1\leq j \leq m$, and $\pi(\beta_{i_0})$ is a codimension $1$ face of $\pi(\alpha_{i_m})$. However, since $\pi$ is a cosimplicial automorphism, we must have that each $\alpha_{i_{j-1}}$ is a codimension $1$ face of $\beta_{i_j}$ for $1\leq j \leq m$, and $\alpha_{i_m}$. is a codimension $1$ face of $\beta_{i_0}$. However, this would imply that $(\alpha_{i_m}, \beta_{i_m}), (\alpha_{i_{m-1}}, \beta_{i_{m-1}}), (\alpha_{i_{m-2}}, \beta_{i_{m-2}}), \dots , (\alpha_{i_{0}}, \beta_{i_{0}}) \in \sigma $ is a nontrivial closed V-path, and are therefore not compatible.
Thus $\pi_*$ is a simplicial map, and we therefore conclude that $\pi_\ast \in {\mathrm{Aut}}({\mathcal{M}}(K))$.
We will refer to the simplicial automorphism $\pi_\ast \in {\mathrm{Aut}}({\mathcal{M}}(K))$ induced by $\pi_{\ast_V}$ (and in general, any automorphism of ${\mathcal{M}}({\partial\Delta}^n)$ that is not induced by a simplicial automorphism of ${\partial\Delta}^n$) as a **ghost automorphism** on ${\mathcal{M}}({\partial\Delta}^n)$. In the proof of Theorem \[prop: aut(M(K)) pt2\], we will see that this ghost automorphism generates all the other ghost automorphisms of ${\mathcal{M}}({\partial\Delta}^n)$.
Next we will compute the cardinality of ${\mathrm{Aut}}({\mathcal{M}}({\partial\Delta}^n))$. We first fix some notation and terminology. Let ${\mathcal{H}}_i$ denote the set of nodes of the Hasse diagram that correspond to simplices in ${\partial\Delta}^n$ of dimension $i$, and let $H = \{{\mathcal{H}}_0, {\mathcal{H}}_1, \dots , {\mathcal{H}}_{n-1}\}$. We also define ${\mathcal{H}}_{-1} = {\mathcal{H}}_{n} = \emptyset$ for convenience. Call ${\mathcal{H}}_i$ the **$i$th layer** of the Hasse diagram of ${\partial\Delta}^n$. Observe that $|{\mathcal{H}}_i| = \dbinom{n+1}{i+1}$ for indices $0\leq i \leq n-1$. Abusing language, we will use simplex to mean both a simplex of $\partial\Delta^n$ and the corresponding vertex of ${\mathcal{H}}(\partial\Delta^n)$. We also define the **degree** of layer ${\mathcal{H}}_i$, denoted as ${\textrm{Deg }}{\mathcal{H}}_i$, as ${\textrm{Deg }}\sigma$ where $\sigma \in {\mathcal{H}}_i$. Note that this is well-defined since for ${\partial\Delta}^n$, it is clear that if $\sigma, \tau\in {\mathcal{H}}_i$, then ${\textrm{Deg }}\sigma={\textrm{Deg }}\tau$. We say that two layers ${\mathcal{H}}_i, {\mathcal{H}}_j$ are **connected** if there exist $\sigma \in {\mathcal{H}}_i, \tau \in {\mathcal{H}}_j$ that are connected in ${\mathcal{H}}(\partial\Delta^n)$. It is clear by the construction of the Hasse diagram that two layers are connected if and only if $i$ and $j$ are consecutive. It is also clear that connectivity of layers is preserved under automorphisms of ${\mathcal{H}}(K)$.\
\[lem: order\] $|{\mathrm{Aut}}({\mathcal{M}}(\partial\Delta^n))| = 2|{\mathrm{Aut}}(\partial\Delta^n)|$.
By Proposition \[prop: aut(M(K))=aut(H(K))\], it suffices to show that $|{\mathrm{Aut}}({\mathcal{H}}({\partial\Delta}^n))| = 2|{\mathrm{Aut}}({\partial\Delta}^n)|$. Consider an arbitrary automorphism $f \in {\mathrm{Aut}}({\mathcal{H}}(\partial\Delta^n))$. We claim that the image of any ${\mathcal{H}}_i$ under $f$ will either be ${\mathcal{H}}_i$ or ${\mathcal{H}}_{n-i-1}$. We first establish that $f$ takes ${\mathcal{H}}_0$ to ${\mathcal{H}}_0$ or ${\mathcal{H}}_{n-1}$, and then proceed by induction on $n$. To that end, observe that if $\sigma\in {\mathcal{H}}_0$ or ${\mathcal{H}}_{n-1}$, then ${\textrm{Deg }}\sigma =n$. If $\sigma\in {\mathcal{H}}_j$ for $1\leq j\leq n-2$, then $\sigma$ has $j+1$ faces of dimension $(j-1)$ and $(n+1)-(j+1)$ cofaces of dimension $j+1$. Hence ${\textrm{Deg }}\sigma=(j+1)+(n+1)-(j+1)=n+1.$ Therefore, the only layer with the same degree as ${\mathcal{H}}_0$ is ${\mathcal{H}}_{n-1}$, so that any automorphism must send a node of ${\mathcal{H}}_0$ into either ${\mathcal{H}}_0$ or ${\mathcal{H}}_{n-1}$.
Next we establish that $f({\mathcal{H}}_0)={\mathcal{H}}_0$ or ${\mathcal{H}}_{n-1}$. Let $ab$ be a $1$-simplex of ${\partial\Delta}^n$ and suppose for the sake of contradiction that $f$ sends $a, b \in {\mathcal{H}}_0$ to two different layers, say $f(a) \in {\mathcal{H}}_0$ and $f(b) \in {\mathcal{H}}_{n-1}$. Now $ab$ is connected to both $a$ and $b$, and since $f$ is an automorphism, $f(ab)$ must be connected to both $f(a)$ and $f(b)$. Since $f(a) \in {\mathcal{H}}_0$ and $f(ab)$ is connected to $f(a)$, it follows that $f(ab) \in {\mathcal{H}}_1$. Similarly, since $f(b)$ and $f(ab)$ are connected, $f(ab) \in {\mathcal{H}}_{n-2}$. Thus, ${\mathcal{H}}_{n-2} = {\mathcal{H}}_1$, so we must have $n=3$. But it is easily seen by inspection of the $n=3$ case that such an automorphism is impossible. Thus $f({\mathcal{H}}_0) = {\mathcal{H}}_0$ or ${\mathcal{H}}_{n-1}$.
Having established that $f({\mathcal{H}}_0)={\mathcal{H}}_0$ or ${\mathcal{H}}_{n-1}$ for any automorphism $f$, we now show by induction that $f({\mathcal{H}}_i)={\mathcal{H}}_i$ or ${\mathcal{H}}_{n-i-1}$ for all $1\leq i \leq n-1$. For the first case, suppose that $f({\mathcal{H}}_0) = {\mathcal{H}}_0$, and suppose the inductive hypothesis that for some integer $0< k<n-1$, we have $f({\mathcal{H}}_j) = {\mathcal{H}}_j$ for all integers $0\leq j\leq k$. We seek to show that $f({\mathcal{H}}_{k+1}) = {\mathcal{H}}_{k+1}$. Notice that ${\mathcal{H}}_{k+1}$ is connected to ${\mathcal{H}}_k$. Thus, $f({\mathcal{H}}_{k+1})$ is connected to $f({\mathcal{H}}_k)$. Additionally, the only layers connected to ${\mathcal{H}}_k$ are ${\mathcal{H}}_{k-1}$ and ${\mathcal{H}}_{k+1}$. Since $f({\mathcal{H}}_k) = {\mathcal{H}}_k$ by the inductive hypothesis, we know $f({\mathcal{H}}_{k+1}) \subseteq {\mathcal{H}}_{k-1}\cup {\mathcal{H}}_{k+1}$. However, by the inductive hypothesis, we know that $f({\mathcal{H}}_{k-1}) = {\mathcal{H}}_{k-1}$. Since $f$ is an isomorphism, this means we cannot send any simplices of ${\mathcal{H}}_{k+1}$ to ${\mathcal{H}}_{k-1}$ under $f$. Therefore, $f({\mathcal{H}}_{k+1}) \subseteq {\mathcal{H}}_{k+1} \implies f({\mathcal{H}}_{k+1}) = {\mathcal{H}}_{k+1}$.
The case where $f({\mathcal{H}}_0) = {\mathcal{H}}_{n-1}$ is similar. Suppose that for some integer $0< k<n-1$, we have $f({\mathcal{H}}_j) = {\mathcal{H}}_{n-j-1}$ for all integers $0\leq j\leq k$. We seek to show that $f({\mathcal{H}}_{k+1}) = {\mathcal{H}}_{n-k-2}$. Since ${\mathcal{H}}_{k+1}$ is connected to ${\mathcal{H}}_k$, $f({\mathcal{H}}_{k+1})$ is connected to $f({\mathcal{H}}_{k}) = {\mathcal{H}}_{n-k-1}$. The only layers connected to ${\mathcal{H}}_{n-k-1}$ are ${\mathcal{H}}_{n-k-2}, {\mathcal{H}}_{n-k}$, so $f({\mathcal{H}}_{k+1}) \subseteq {\mathcal{H}}_{n-k-2}\cup {\mathcal{H}}_{n-k}$. By the inductive hypothesis, $f({\mathcal{H}}_{k-1}) = {\mathcal{H}}_{n-k}$. Since $f$ is injective, we cannot send any simplices of ${\mathcal{H}}_{k+1}$ to ${\mathcal{H}}_{n-k}$. Thus, $f({\mathcal{H}}_{k+1}) \subseteq {\mathcal{H}}_{n-k-2}$. We know that $$\begin{aligned}
|f({\mathcal{H}}_{k+1})| &=& |{\mathcal{H}}_{k+1}|\\
&=& \binom{n+1}{k+2}\\
&=& \binom{n+1}{n-k-1} \\
&=& |{\mathcal{H}}_{n-k-2}|.\end{aligned}$$ Hence $f({\mathcal{H}}_{k+1}) = {\mathcal{H}}_{n-k-2}$, as desired.
Having established the behaviour of each layer under automorphism, we now establish a group action to count $|{\mathrm{Aut}}({\mathcal{H}}({\partial\Delta}^n))|$. Define an action of ${\mathrm{Aut}}({\mathcal{H}}({\partial\Delta}^n))$ on $H = \{{\mathcal{H}}_0, {\mathcal{H}}_1, \dots , {\mathcal{H}}_{n-1}\}$ by
$$(g, {\mathcal{H}}_i) \mapsto g{\mathcal{H}}_i := \{g(\sigma)\colon \sigma \in {\mathcal{H}}_i\}.$$
We verify this is indeed a group action by noting that, ${\mathrm{id}}_{{\mathcal{H}}({\partial\Delta}^n)} ({\mathcal{H}}_i) = {\mathcal{H}}_i$ for all ${\mathcal{H}}_i$, and that if $g,h \in {\mathrm{Aut}}({\mathcal{H}}({\partial\Delta}^n))$, we have $$\begin{aligned}
(gh)({\mathcal{H}}_i) &=& \{(g\circ h)(\sigma): \sigma\in {\mathcal{H}}_i\}\\
&=& g(\{h(\sigma): \sigma \in {\mathcal{H}}_i\} )\\
&=& g(h({\mathcal{H}}_i)).\end{aligned}$$
By the Orbit Stabilizer theorem, we have $|{\mathrm{Aut}}({\mathcal{H}}({\partial\Delta}^n))| = |\mathrm{Orb}({\mathcal{H}}_0)||\mathrm{Stab}({\mathcal{H}}_0)|$. Suppose $f \in |{\mathrm{Aut}}({\mathcal{H}}({\partial\Delta}^n))|$ fixes ${\mathcal{H}}_0$. Then, $f$ is bijective on the set of vertices of ${\partial\Delta}^n$ so that it corresponds with an automorphism of ${\partial\Delta}^n$. Likewise, any automorphism of ${\partial\Delta}^n$ is induced by a bijective map $V({\partial\Delta}^n)\to V({\partial\Delta}^n)$, so must correspond with an automorphism on ${\mathcal{H}}({\partial\Delta}^n)$ that fixes ${\mathcal{H}}_0$. Therefore, $|\mathrm{Stab}({\mathcal{H}}_0)| = |{\mathrm{Aut}}({\partial\Delta}^n)|$. We also know that $\mathrm{Orb}({\mathcal{H}}_0) = 2$ since automorphisms of the Hasse Diagram send ${\mathcal{H}}_0$ to either ${\mathcal{H}}_0$ or ${\mathcal{H}}_{n-1}$, as we showed earlier in this proof. We conclude that $|{\mathrm{Aut}}({\mathcal{H}}({\partial\Delta}^n))| = 2|{\mathrm{Aut}}({\partial\Delta}^n)|$, as desired.
We are now able to compute the automorphism group of the Morse complex in the case where $K={\partial\Delta}^n$.
\[prop: aut(M(K)) pt2\] If $K ={\partial\Delta}^n$, with $n\geq 2$, then ${\mathrm{Aut}}({\mathcal{M}}({\partial\Delta}^n)) \cong {\mathrm{Aut}}({\partial\Delta}^n) \times \mathbb{Z}_2$.
For any function $f$, write $f^n := \underbrace{f \circ f \circ \cdots \circ f}_{n\text{ } f\text{'s}}$. We construct an isomorphism $\phi \colon {\mathrm{Aut}}({\partial\Delta}^n)\times \mathbb{Z}_2 \to {\mathrm{Aut}}({\mathcal{M}}({\partial\Delta}^n))$. For each $(f, i) \in {\mathrm{Aut}}(K)\times \mathbb{Z}_2$, $i=0,1$, define $\phi((f,i)) := f_\ast \circ \pi_\ast^i$, where $f_\ast$ is the automorphism of ${\mathcal{M}}({\partial\Delta}^n)$ induced by $f$, and $\pi_\ast$ is the ghost automorphism induced by the reflection map. We first show that $\phi$ is a homomorphism. Suppose we have $(f, i), (g, j) \in {\mathrm{Aut}}({\partial\Delta}^n)\times \mathbb{Z}_2$. By Lemma \[lem: induced comp\] and the fact that $\pi_\ast$ commutes with all $f_*$, we have $$\begin{aligned}
\phi((f,i)(g,j)) &= \phi( (f\circ g, i+j) ) \\ &= (f\circ g)_\ast \circ \pi_\ast^{i+j} \\ &= f_\ast \circ g_\ast \circ \pi_\ast^i\circ \pi_\ast^j \\&= (f_\ast \circ \pi_\ast^i) \circ (g_\ast \circ \pi_\ast^j )\\& = \phi((f,i))\phi((g,j)),\end{aligned}$$ as desired.
Next, we will show that $\phi$ is a bijection. We first show that $\phi$ is injective. Suppose we have $\phi((f,i)) = \phi((g,j))$ for some $(f,i),(g,j) \in {\mathrm{Aut}}({\partial\Delta}^n)\times \mathbb{Z}_2$. Then, we have $f_\ast \circ \pi_\ast^i = g_\ast \circ \pi_\ast^j$. If $i=j$, then, we have $f_\ast = g_\ast$, so $f=g$ by Proposition \[prop: autK subgroup\]. We claim that $i\neq j$ is impossible. Suppose by contradiction that $i=0, j=1$, so that $f_\ast = g_\ast\circ \pi_\ast$. However, this implies that $g_\ast \circ \pi_\ast$ is induced by some simplicial automorphism on ${\partial\Delta}^n$. Consider any $(\sigma,\tau)\in V({\mathcal{M}}({\partial\Delta}^n))$ with $\dim\sigma=0$. We have $f_{\ast}((\sigma,\tau)) = (f(\sigma), f(\tau))$. We know that $\dim \sigma = \dim f(\sigma) = 0$. However, $g_\ast \circ \pi_\ast((\sigma,\tau)) = g_\ast(\pi(\tau),\pi(\sigma)) = (g(\pi(\tau)), g(\pi(\sigma)))$. We know that $\dim \pi(\tau) = n- \dim\tau = n - 1$. Then, $\dim g(\pi(\tau)) = n-1 > 0$, so $\dim f(\sigma) \neq \dim (g(\pi(\tau)))$, a contradiction. Thus, $i\neq j$ is not possible. Hence $\phi$ is injective.
Finally, by Proposition \[prop: aut(M(K))=aut(H(K))\] and Lemma \[lem: order\], we see that $|{\mathrm{Aut}}({\mathcal{M}}({\partial\Delta}^n))| = |{\mathrm{Aut}}({\mathcal{H}}({\partial\Delta}^n))| = 2|{\mathrm{Aut}}({\partial\Delta}^n)| = |{\mathrm{Aut}}({\partial\Delta}^n)\times \mathbb{Z}_2|$. Since these groups are finite, $\phi$ is a bijection. We conclude that $\phi$ is an isomorphism.
Combining Propositions \[prop: aut(M(K))=aut(K)\], \[prop: aut C\_n\], and Theorem \[prop: aut(M(K)) pt2\] thus yields Theorem \[thm: main\] as promised.
[10]{}
R. Ayala, L. M. Fern[á]{}ndez, A. Quintero, and J. A. Vilches, *A note on the pure [M]{}orse complex of a graph*, Topology Appl. **155** (2008), no. 17-18, 2084–2089.
N. A. Capitelli and E. G. Minian, *A simplicial complex is uniquely determined by its set of discrete [M]{}orse functions*, Discrete Comput. Geom. **58** (2017), no. 1, 144–157.
M. K. Chari and M. Joswig, *Complexes of discrete [M]{}orse functions*, Discrete Math. **302** (2005), no. 1-3, 39–51.
D. L. Ferrario and R. A. Piccinini, *Simplicial structures in topology*, CMS Books in Mathematics/Ouvrages de Mathématiques de la SMC, Springer, New York, 2011, Translated from the 2009 Italian original by Maria Nair Piccinini.
R. Forman, *Morse theory for cell complexes*, Adv. Math. **134** (1998), no. 1, 90–145.
[to3em]{}, *A user’s guide to discrete [M]{}orse theory*, Sém. Lothar. Combin. **48** (2002), Art. B48c, 35.
J. Jonsson, *Introduction to simplicial homology*, 2011, Available at the website [[ https://people.kth.se/\~jakobj/homology.html]( https://people.kth.se/~jakobj/homology.html)]{}.
K. Knudson, *Morse theory: Smooth and discrete*, World Scientific Publishing Co. Pte. Ltd., Hackensack, NJ, 2015.
D. N. Kozlov, *Complexes of directed trees*, J. Combin. Theory Ser. A **88** (1999), no. 1, 112–122.
[to3em]{}, *Combinatorial algebraic topology*, Algorithms and Computation in Mathematics, vol. 21, Springer, Berlin, 2008.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Multiple Object Tracking (MOT) is a long-standing task in computer vision. Current approaches based on the tracking by detection paradigm either require some sort of domain knowledge or supervision to associate data correctly into tracks. In this work, we present an unsupervised multiple object tracking approach based on visual features and minimum cost lifted multicuts. Our method is based on straight-forward spatio-temporal cues that can be extracted from neighboring frames in an image sequences without superivison. Clustering based on these cues enables us to learn the required appearance invariances for the tracking task at hand and train an autoencoder to generate suitable latent representation. Thus, the resulting latent representations can serve as robust appearance cues for tracking even over large temporal distances where no reliable spatio-temporal features could be extracted. We show that, despite being trained without using the provided annotations, our model provides competitive results on the challenging MOT Benchmark for pedestrian tracking.'
author:
-
-
-
bibliography:
- 'egbib.bib'
- 'paper.bib'
title: |
Unsupervised Multiple Person Tracking\
using AutoEncoder-Based Lifted Multicuts
---
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Although the long-term numerical integrations of planetary orbits indicate that our planetary system is dynamically stable at least $\pm$4 Gyr, the dynamics of our Solar System includes both chaotic and stable motions: the large planets exhibit remarkable stability on gigayear timescales, while the subsystem of the terrestrial planets is weekly chaotic with a maximum Lyapunov exponent reaching the value of 1/5 Myr$^{-1}$. In this paper the dynamics of the Sun–Venus–Earth–Mars-Jupiter–Saturn model is studied, where the mass of Earth was magnified via a mass factor $\kappa_E$. The resulting systems dominated by a massive Earth may serve also as models for exoplanetary systems that are similar to our one. This work is a continuation of our previous study, where the same model was used and the masses of the inner planets were uniformly magnified. That model was found to be substantially stable against the mass growth. Our simulations were undertaken for more then 100 different values of $\kappa_E$ for a time of 20, in some cases for 100 Myrs. A major result was the appearance of an instability window at $\kappa_E \approx 5$, where Mars escaped. This new result has important implications for the theories of the planetary system formation process and mechanism. It is shown that with increasing $\kappa_E$ the system splits into two, well separated subsystems: one consists of the inner, the other one consists of the outer planets. According to the results the model became more stable as $\kappa_E$ increases and only when $\kappa_E \ge$ 540 Mars escaped, on a Myr timescale. We found an interesting protection mechanism for Venus. These results give insights also to the stability of the habitable zone of exoplanetary systems, which harbour planets with relatively small eccentricities and inclinations.'
author:
- |
Áron Süli$^{1}$[^1], Rudolf Dvorak$^{2}$ and Florian Freistetter$^{3}$\
$^{1}$Eötvös Loránd University, Department of Astronomy, Pázmány Péter sétány 1/A, Budapest, H-1518, Hungary\
$^{2}$University of Vienna, Institute of Astronomy, Türkenschanzstrasse 17, A-1180, Vienna, Austria\
$^{3}$Astrophysikalisches Institut und Universitäts-Sternwarte, Schillergäßchen 2-3, D-07745 Jena, Deutschland
title: 'The stability of the terrestrial planets with a more massive “Earth”'
---
\[firstpage\]
celestial mechanics – Solar System: general.
Introduction
============
The determination of the stability of our Solar System is one of the oldest problems in astronomy. The question has been debated over more than 300 years, and has attracted the attention of many famous mathematicians over the course of history. The problem played a central role in the development of non–linear dynamics and chaos theory. Despite the considerable efforts, we do not possess a definite answer to the question of whether our Solar System is stable or not. This is partly a result of the fact that the definition of the term stability is not unambiguous when it is used in relation to the problem of planetary motion. In addition to the vagueness of the concept of stability, the planets in our planetary system show a character typical of dynamical chaos. The physical basis of this chaotic behaviour is now partly understood as a consequence of resonance overlapping and three body resonances [@Lecar2001; @Murray99; @Nesvorny99] which can manifest themselves in dramatic and relatively sudden changes in an orbit. In the last two decades several numerical stability studies were performed in order to throw light on the question. At present the longest numerical integrations published are those of [@Ito2002], where six long-term numerical integrations of all nine planets, covering a time-span of several $10^9$ and $10^{11}$ years are discussed. Their fundamental conclusion is that the Solar System seems to be stable in terms of the Hill-criteria at least over a time-span of $\pm$4 Gyr. Moreover it turned out that during the integration period all the planetary orbital elements have been confined in a narrow region.
On the other hand according to Laskar’s semi-analytical secular perturbation theory [@Laskar88], the terrestrial planets’, especially Mercury’s and Mars’ eccentricities and inclinations show large and irregular variations on a time-scale of several $10^9$ year [@Laskar96].
Nowadays to study the stability of the Solar System, or its variants, as a representative of the different planetary systems has become part of the frontline research. Over the past few years the detection of planets outside the Solar System, the so called exoplanets, has greatly stimulated the stability studies of planetary systems. New exoplanets are being discovered on a regular basis; more than 150 (April, 2005) exoplanets are now known. There are 136 systems consisting of a central star and a gaseous giant planet, and 14 multiple systems with two, three and four planets. The so far discovered exoplanets have a minimum mass range ($m\cdot \sin(i_p)$) from 0.042 $m_J$ to 17.5 $m_J$ where $m_J$ is Jupiter’s mass and $i_p$ is the inclination of the orbital plane with respect to the plane of the sky. Since $i_p$ is unknown a precise mass cannot be determined, only a lower mass limit. Because of the technical limitations only planets of Neptune mass or above can be detected and then only if they are less than 5 AU or so from the star [^2]. Therefore more than 90 % of these planets are orbiting their host star well inside Jupiter’s orbit. There are major differences between the characteristics of the so far observed systems and those of the Solar System. Most of the planets have minimum masses substantially greater than that of Jupiter – up to six or even more times the mass of Jupiter. Dozens of planets are orbiting very close to their hosting star, with semimajor axes down to 0.04 AU. Finally, planetary orbits are found with large eccentricities, up to approximately 0.7, plus a few greater, significantly greater than the highest eccentricities observed for planets in our Solar System. These characteristics are depicted in Fig. \[fig1\], where the planets’ eccentricities are plotted against their semimajor axes, the locations of the Earth, Jupiter and Saturn are marked as diamonds. From Fig. \[fig1\] it is apparent that our planetary system may serve as model case for those planetary systems which have small eccentricities. Presumably these are also the ones where we may expect stable terrestrial planets moving in habitable zones [@Asghari2004].
![Eccentricities vs. the semimajor axes of the observed extrasolar planets. The $x$-axis is logarithmic. The positions of the Earth, Jupiter and Saturn are also indicated as diamonds with a plus sign in the middle.[]{data-label="fig1"}](./fig1.eps){width="1.0\linewidth"}
In previous papers [@Dvorak2002; @Dvorak2005] the dynamical evolution of a simplified Solar system was studied. The model consisted of the Sun, the three most massive terrestrial planets (Venus, Earth, Mars), Jupiter and Saturn. The masses of the inner planets were uniformly magnified by a mass factor $\kappa$. It turned out that the different systems remained stable up to 10 Myr for $\kappa \le 220$. Stable islands were found for $\kappa$ = 245 and 250, which is a well-known property in such regions which are close to the last stable orbit in the chaotic domain. We have shown that the dynamical coupling of Venus and Earth and that of Jupiter and Saturn remained unbroken for all studied $\kappa$. On the other hand the motion of Mars was not coupled to any other planet, what may be a reason for the fact that it was always Mars which caused the decay of the system after close approaches with Earth. However the remarkable stability of these model planetary systems suggests that exoplanetary systems with configuration like our Solar System may harbour moderately or even very massive terrestrial-like planets.
In the present work our aim was to study and analyze the dynamical evolution and the stability of the system with respect to the masses involved. Contrary to our previous study, in the present work only the mass of the Earth was increased via a mass factor $\kappa_E$. Furthermore, the examined systems may be considered as models for individual exoplanetary system, and the results can be applied to them. Section 2 explains our dynamical model, the applied methods and section 3 is devoted to a detailed description of the results in the $\kappa_E \in [1,600]$ region. Finally, we discuss the results and the implication for exoplanetary systems.
Description of the model and methods
====================================
The applied dynamical model consisted of the Sun, Venus, Earth, Mars, Jupiter and Saturn. We have chosen this model for two reasons. To speed up the numerical integrations, we have omitted Mercury, Uranus and Neptune. Because of the small mass of Mercury, it only slightly perturbs the motion of the terrestrial planets and it has a short orbital period which would require a reduction in the integration step-size resulting in increased CPU time. Even though Uranus and Neptune are massive planets, they are evolving around the Sun more than twice as far as Saturn, and so they do not influence the motion of the inner planets significantly. These simplifications are also justified by the fact, that the dynamics of the inner planets are dominated by Jupiter and Saturn. Furthermore the so far observed exosystems harbour at most three of four planets. The above modification of the Solar System gives grounds for the adjective “simplified”. Hereafter we will refer to our simplified Solar System as $S^3$. The other reason was to make the model more like an exoplanetary system. The mass of the Earth was magnified via a mass factor $\kappa_E$ (the masses of the other planets were unaltered), which resulted in such hypothetical planetary systems, whose characteristics agree very well with several exosystems (for example if we multiply the semimajor axes of the planets of 47 Ursae Majoris by 2.5, the resulting configuration is something like the Sun, Jupiter and Saturn system, if $\kappa_E$=1). According to these modifications, our models are parallel with those exosystems, which harbour planets with small eccentricities and inclinations. In Fig. \[fig2\] the masses of the planets are plotted as a function of $\kappa_E$. At $\kappa_E \approx 90$ Earth is as massive as Saturn and when $\kappa_E \approx 300$ Earth is as massive as Jupiter. The mass of the other planets are plotted by horizontal dashed lines. The mass distribution of the so far discovered exoplanetary systems is also depicted in Fig. \[fig2\] up to mass parameter = 2. The integrations were done in the $\kappa_E \in [1,600]$ region, which approximately corresponds to 0 $<$ mass parameter $\approx$ 2 interval, and therefore the mass distribution in Fig. \[fig2\] is displayed only in this interval.
We have considered the $S^3$ as a non-linear Hamiltonian system, governed only by classical Newtonian gravitational forces between the objects of the model. The planets have been taken as point masses, and the Earth’s Moon was not included in the models. The initial planetary orbital elements and the actual masses are listed in Table \[tab:2.1\].
{width="0.9\linewidth"}
To check whether the different $S^3$ setups belonging to different $\kappa_E$s could be stable over a long time interval we utilized the very precise numerical integration scheme, the Lie-integrator method. This method is based on the integration of differential equations with Lie-series and uses the property of recurrence formulae for the Lie-terms. The details of the method are described in [@Hanslmeier1984] and [@Lichtenegger84]. The scheme is particularly effective in the case of highly eccentric orbits. The accuracy of this integration technique is based on an automatic stepsize control, and it has been checked in several comparative test computations with other integrators. Although symplectic integrators are very effective when eccentricities remain small, but the Lie-integrator is a better choice in studies, where very large eccentricity orbits may occur.
The length of the integrations was fixed at 20 Myrs, in some cases at 100 Myrs, which was a trade-off between too long CPU time and the quality of the results. Since our interest focuses primarily on the inner three planets’ motions, for which the orbital time-scales are much shorter than those of the outer two giant planets, the 20 Myrs seems a justifiable choice, although it is known from earlier work (see for e.g. [@Jones2001; @Jones2002]) that exosystems can stay stable for hundreds of millions of years and then fall apart. In the case of weak chaos therefore the 100 and especially the 20 Myr timespan might be short and in some of the integrations where the terrestrial planets survived for 20 or 100 Myrs, they might not survive significantly longer. For the sake of a comparison all results were derived on the same computer. In some cases we have performed comparative integrations on different platforms (to obtain more information of the particular system).
The conservative definition of the point at which systems become unstable is when close encounter between two planets happens: two bodies approach one another within an area of the larger Hill radius. The consequence of such an event is the dramatic changes in the orbital elements of the two planets, and usually the escape of the planet with smaller mass. In this paper we define a system unstable, when an orbit crossing or a close encounter happens. This definition is somewhat more general and the instability can be directly connected to the eccentricity via the perihelia and aphelia distances. Henceforward we state that our model is dynamically unstable if orbit crossing or close encounters happen in the course of the integration. Using both criteria is clearly an extension of the conservative definition of instability. It is justifiable to incorporate the orbital crossing criteria in the definition since we know from experience that an orbital crossing in general leads to a close encounter in very short time. The main difference between the two definitions is the time-scale of instability. We note that orbital crossing does not lead to close encounter in all cases when certain resonances have adjusted the planetary motions in such a way that the planets avoid each other. For example this is the case in the Neptune-Pluto pair. In our models no such protection resonances are present, henceforward we use the above definition to distinguish between stable and unstable $S^3$ setups. The integrations were not stopped after one of the above criterion had been met, but were continued until the integration timespan was reached.
Venus Earth Mars Jupiter Saturn
---------- ------------ ------------- ------------- ------------ ------------
$a$ 0.723328 0.999999 1.523614 5.202627 9.545509
$e$ 0.006747 0.016716 0.093443 0.048370 0.052420
$i$ 3.394820 0.000545 1.850191 1.304638 2.485620
$\omega$ 54.847892 113.611521 286.492727 275.222227 338.025839
$\Omega$ 76.691772 349.288391 49.573832 100.470086 113.651098
$M$ 135.521541 78.172620 185.208769 233.733076 259.852365
$1/m_i $ 408 523.71 332 946.047 3 098 708.0 1047.348 3497.898
\[tab:2.1\]
For an indication of stability we used a straightforward check based on the eccentricity. This osculating orbital element shows the probability of orbital crossing and close encounter of two planets, and therefore its value provides information on the stability of the orbit. We examined the behavior of the eccentricities of the planets along the integration, and used the largest value as a stability indicator; in the following we call it the maximum eccentricity method (hereafter MEM). This is a reliable indicator of chaos, because the overlap of two or more resonances induce chaos and large excursions in the eccentricity. We know from experience, that instability comes from a chaotic growth of the eccentricity. This simple check has already been used in other stability studies, and was found to be quite a powerful indicator of the stability character of an orbit [@Dvorak2003; @Asghari2004].
The Laplace-Lagrange secular theory
-----------------------------------
In order to find a theoretical explanation for the decay of the system we have applied the Laplace-Lagrange first order secular theory. This linear theory yields accurate results under the following assumptions:
1. no mean motion commensurabilities,
2. no orbit-crossing, and
3. the eccentricities and inclinations are small enough.
Our models meet these criteria. However a complication is that the orbits of Jupiter and Saturn are close to a 5:2 commensurability. Since the appearance of the Laplace-Lagrange theory, as a first approximation it has been extensively used in the studies of motion of the planets and of other Solar System bodies. In several researches [@Knezevic86; @Laskar88] the results of the first and higher order secular theories were compared. According to these studies the secular frequencies calculated from the Laplace-Lagrange theory are sufficiently accurate for our present research goal. The main discrepancies are in the case of Jupiter and Saturn. As we are interested only in the motions of the inner planets, therefore we will apply the Laplace-Lagrange theory to our models. From the above assumptions it is apparent that the precision of the theory does not depend on the planetary masses, accordingly there is no theoretical limitation on the mass factor’s magnitude.
Since the eccentricities and inclinations may vanish at remote epochs, it is better to use the Lagrange orbital elements: $$\begin{aligned}
\left( \begin{array}{c}
h \\
k \\
\end{array} \right) =
e \cdot \begin{array}{c}
\sin \varpi \\
\cos \varpi \\
\end{array}
,\qquad
\left( \begin{array}{c}
p \\
q \\
\end{array} \right) =
i\cdot \begin{array}{c}
\sin \Omega \\
\cos \Omega \\
\end{array}
. \label{eq:2.1}\end{aligned}$$ Eq. (\[eq:2.1\]) associates the $h$, $k$, $p$ and $q$ Lagrangian-elements to the $e$, $i$, $\varpi$, and $\Omega$ Keplerian-elements, where $e$ denote the eccentricity, $i$ the inclination, $\varpi$ the longitude of perihelion and $\Omega$ the longitude of ascending node. Using these variables the general solution of the differential equations for the planets takes the following form: $$\begin{aligned}
\left( \begin{array}{c}
h_s \\
k_s \\
\end{array} \right) &=&
\sum_{j=1}^n M_s^{(j)} \begin{array}{c}
\sin \\
\cos \\
\end{array}
\left( g_jt+\beta_j\right), \label{eq:2.2} \\
\left( \begin{array}{c}
p_s \\
q_s \\
\end{array} \right) &=&
\sum_{j=1}^n L_s^{(j)} \begin{array}{c}
\sin \\
\cos \\
\end{array}
\left( f_jt+\gamma_j\right), \label{eq:2.3}\end{aligned}$$ where the $s$ index denotes the planet, the $j$ index denotes the mode, $N$ is the number of planets, $M_s^{(j)}$ and $L_s^{(j)}$ are the amplitudes, $g_j$ and $f_j$ denote the secular frequencies and $\beta_j$ and $\gamma_j$ are the angular phases.
The planet’s orbital elements are described by Eq. (\[eq:2.2\]) and Eq. (\[eq:2.3\]), which are the sum of harmonic oscillations. Using these formulae it can be calculated that the planet’s eccentricities and inclinations are varying between given limits with quasiperiodic oscillations. Due to the positive $g_j$ secular angular velocities the apsidal lines of the planets are rotating directly, whereas the nodes accordingly to the negative $f_j$ secular angular velocities are rotating indirectly. Upon these mean rotations quasiperiodic variations are superimposed. Both of the apsidal and nodal motions can be approximated by average angular velocities, which are to a first approximation equal with the frequencies of those harmonious terms which are multiplied by the largest amplitudes: $$\begin{aligned}
e_s \cdot \begin{array}{c}
\sin \varpi_s\\
\cos \varpi_s\\
\end{array}
\approx
M_s^{(J)} \begin{array}{c}
\sin \\
\cos \\
\end{array}
\left( g_Jt+\beta_J\right) \label{eq:2.4}\end{aligned}$$ where $M_s^{(J)} = \mathbf{max}_j |M_s^{(j)}|$ and the average angular velocity of the $s$’th planet is given by $g_J$.
In this manner on the basis of the amplitudes each secular frequency can be associate with each planet. This association is not mutually unambiguous. It may happen to associate a certain secular frequency to more then one planet.
The use of the linear theory seems to be in contradiction to the large eccentricities and inclinations that may be reached by a planet during the simulation. We emphasize that the forementioned theory was used at the beginning of the integration, when the inclinations and the eccentricities are small, and the above assumptions are therefore fullfilled.
The above described linear secular theory was implemented in the [MAPLE]{} computer algebra program. With the aid of this application the formulae of the theory can be evaluated in a few seconds and it gives the complete first order solution of the problem.
Results
=======
In this section we give a detailed description and overview of the results of our simulations. The model was integrated for more then 100 different values of $\kappa_E$. The interval of data output was 100 years, the total amount of data is approximately 6 GBs, and the total used CPU time is more than several thousands of days. All of the integrations were done on two Sun Fire 15000 supercomputers with 72 US-III+ 1200 MHz processor in each computer. In Table \[tab:3.1\] the first column lists the different $\kappa_E$ intervals, the second shows the stepsize ($\Delta \kappa_E$) in the mass factor. A small stepsize is taken for $4 \le \kappa_E \le 6$ in order to explore the interesting behaviour of the system leading – very surprisingly – to the escape of Mars. In the third column we list approximately the mass of Earth in Jupiter’s mass unit. The last column gives the time-span of numerical integrations.
--------------------- ------------------- ------------------ ------------ -- --
$\kappa_E$ interval $\Delta \kappa_E$ Earth mass in Total time
Jupiter mass \[Myr\]
1 – 25 1.0 \[1/300 – 1/12\] 20
4 – 6 0.1 \[1/75 – 1/50\] 100
30 – 200 5.0 \[1/10 – 2/3\] 20
210 – 300 10.0 \[2/3 – 1\] 20
330 – 600 30.0 \[1 – 2\] 20
--------------------- ------------------- ------------------ ------------ -- --
: Summary of the integrations.
\[tab:3.1\]
The $1 \leq \kappa_E \leq 25$ region
------------------------------------
As we expected, the orbital motions of the planets indicate long-term stability in most of our numerical experiments: no orbital crossings nor close encounters between any pair of planets took place in the course of the integrations. However, a suprising result was the discovery of the instability window for the $\kappa_E \in [4,6]$ interval: at several $\kappa_E$ values Mars escaped; the details can be found in the next subsection.
Fig. \[fig3\] depicts the results of the MEM. The maximum eccentricities (hereafter ME) of Jupiter and Saturn are actually constant, with a value of 0.06, and 0.088, respectively (they are not shown in Fig. \[fig3\]). As one can clearly see from Fig. \[fig3\] the MEs of the Earth and Venus are relatively small, and both curves show a similar behaviour as a consequence of the well-known coupling between them. We note, that the ME of Earth for $\kappa_E=1$ is 60% greater than for $\kappa_E=2$. After some oscillations of Venus’ and Earth’s MEs, they stay almost constant with a value of 0.041 and 0.035, respectively. In turn, the ME of Mars steadily increases with $\kappa_E$, and at $\kappa_E=5$ it suddenly reaches a very high value, $e_{Mars}=0.26091$ (the perihelion distance of Mars is $q_{Mars}
= 1.126$)! After this peak Mars’ ME drops down to its starting value 0.1231, and begins to increase slowly and gradually with $\kappa_E$.
![Results of the MEM in the $1 \le \kappa_E \le 25$ region ($\Delta \kappa_E = 1$). The $x$-axis is the $\kappa_E$, the $y$-axis is the maximum eccentricity (ME).[]{data-label="fig3"}](./fig3.eps){width="1.0\linewidth"}
{width="0.7\linewidth"} {width="0.7\linewidth"} {width="0.7\linewidth"}
The character of the variation of the planetary orbital elements does not change substantially over the course of the simulation, except for $\kappa_E$ around 5. The variations of the semimajor axes are in the order of $10^{-4}$–$10^{-3}$ AU for the inner planets, and for Jupiter and Saturn they are approximately $10^{-3}$, $10^{-2}$ AU, respectively. The variations in Jupiter’s and Saturn’s semimajor axes are one order of magnitude larger than those of the inner planets, which is a consequence of the 5:2 near commensurability between their motion. For all simulations in this mass factor region the behaviour of the semimajor axes did not change, and their excursions remained in the above range.
The eccentricities as a function of time of the inner planets for $\kappa_E$ = 1, 2, and 15 are shown in Fig. \[fig4\] (for the sake of a comparison the eccentricities for the actual masses are also depicted). As $\kappa_E$ goes from 1 to 2, there is a substantial change in $e$ for all three terrestrial planets. According to the MEM analysis (see Fig. \[fig3\]), the MEs of Earth and Venus decrease, which manifests themselves in greater perihelia and smaller aphelia distances (see Fig. \[fig4\]b). These changes decrease the probability of a close encounter between Venus and Earth, as can be inferred form Fig. \[fig4\]b. Beyond this $\kappa_E$ value the behaviour of Earth’s and Venus’ eccentricities do not change and in general their orbits turns into a more and more regular, quasiperiodic ones as $\kappa_E$ increases. Moreover their dynamical coupling strengthens. These features can be traced on the panels of Fig. \[fig4\].
In the case of Mars the character of the variation of the eccentricity is more complex:
- For the actual masses Mars’ eccentricity fluctuates with an amplitude of 0.1 and with a very long period of approximately 2.5 Myrs. To this fluctuation smaller amplitudes with shorter periods are added. It is clear from Fig. \[fig4\]a that Mars’ motion is not coupled to any of the other two terrestrial planets.
- At $\kappa_E$ = 2 Mars’ eccentricity shows a quite different behaviour: it oscillates with a very short period around a mean eccentricity of 0.1 with an amplitude of about 0.02. With increasing $\kappa_E$ the center of oscillation shifts to higher values, and reaches its maximum at 5, where the system is destabilized.
- For $\kappa_E > 5$ the motion of Mars gradually turns into a regular one. Moreover, dynamical coupling develops between all the three inner bodies, which is very well visible from Fig. \[fig4\]c.
The terrestrial planets seem to gradually form a subsystem as the mass factor increases. As the Earth becomes the dominant planet in this region and the motions of the two smaller neighboring planets are influenced primarily by the Earth (the major variation in the eccentricities of the Earth, Venus and Mars has the same period, see Fig. \[fig4\]c). With regard to this we may consider the inner planets as a collection of dynamically mutually dependent planets, namely a subsystem.
In this mass factor region the giant planets move on quiet orbits for the duration of all numerical integrations. Jupiter and Saturn also constitute a subsystem which is practically not affected from the terrestrial planets. An analysis of the eccentricities of all bodies for the subsequent mass factors shows that the model graduates into a system which consists of two separate and loosely dependent subsystems.
The instability window at $\kappa_E \approx 5$
----------------------------------------------
According to Fig. \[fig3\] the ME of Mars reached its maximum value at $\kappa_E$ = 5. To explore the dynamical evolution of the inner planets in more detail simulations were performed in the $\kappa_E$ = \[4,6\] region, with a stepsize of 0.1 for 100 Myrs. The results of the MEM are shown in Fig. \[fig5\], where several peaks ($e_M = 1.0$) in the curve of Mars’ ME are visible, which correspond to escape orbits. We note that there are two minima: at 4.4 and at 4.7. This feature is typical of chaotic systems, where small differences in initial conditions, round off errors or the applied computing architectures may result in different outcomes. To verify this, the two systems with $\kappa_E$ = 4.4 and 4.7 were integrated on different computers and both of the systems decayed. After $\kappa_E = 5.2$ the ME of Mars gradually drops down and at $\kappa_E = 6$ it is already less than 0.15.
![Results of the MEM in the $4 \le \kappa_E \le 6$ region. The dotted lines show the ME of the 20 Myr, whereas the solid lines show the ME of the 100 Myr integrations. The MEs of Jupiter and Saturn are not plotted. The instability window occupies the (4.2, 4.9) region. We note that the ordinate is logarithmic.[]{data-label="fig5"}](./fig5.eps){width="1.0\linewidth"}
Fig. \[fig6\] shows the evolution of the perihelia ($q$) and aphelia ($Q$) distances of the inner planets for $\kappa_E = 4.6$, for which case we measured the shortest dynamical lifetime. A striking feature is that Mars’ perihelion begins to decrease right at the start of the integration, as a consequence of the steep increase of its eccentricity. At 250 000 yr Mars’ $e$ reaches 0.16 then it decreases to 0.13, around this value it oscillates for 1 Myrs. Conversely, $q$ and $Q$ of the Earth and Venus stay in a well defined zone. At $\sim$ 1.5 Myr, the perihelion of Mars begins to oscillate with quite a large amplitude. The center of this oscillation grows secularly and after 10.5 Myr Mars becomes an Earth-crosser, moreover it crosses the orbit of Venus too. As a consequence of the several close encounters with Earth; Mars escapes from the system. This cascade mechanism is shown in Fig. \[fig6\]c, where the semimajor axes of the Earth and Venus are plotted.
![a.) the inner planets’ perihelia ($q$) and aphelia ($Q$) distances are shown as a function of time for the $\kappa_E=4.6$ simulation. b.) The eccentricity of Mars is depicted. Note that Mars’ orbit is excited to high eccentricity and it becomes Earth-crossing at $t \approx$ 10.5 Myr. c.) The semimajor axes of the Earth and Venus are shown together.[]{data-label="fig6"}](./fig6a.eps "fig:"){width="1.0\linewidth"} ![a.) the inner planets’ perihelia ($q$) and aphelia ($Q$) distances are shown as a function of time for the $\kappa_E=4.6$ simulation. b.) The eccentricity of Mars is depicted. Note that Mars’ orbit is excited to high eccentricity and it becomes Earth-crossing at $t \approx$ 10.5 Myr. c.) The semimajor axes of the Earth and Venus are shown together.[]{data-label="fig6"}](./fig6b.eps "fig:"){width="1.0\linewidth"} ![a.) the inner planets’ perihelia ($q$) and aphelia ($Q$) distances are shown as a function of time for the $\kappa_E=4.6$ simulation. b.) The eccentricity of Mars is depicted. Note that Mars’ orbit is excited to high eccentricity and it becomes Earth-crossing at $t \approx$ 10.5 Myr. c.) The semimajor axes of the Earth and Venus are shown together.[]{data-label="fig6"}](./fig6c.eps "fig:"){width="1.0\linewidth"}
Because of the chaotic nature of the system, different dynamical lifetimes for the different $\kappa_E$ values were found. Out of the 21 integrations for 100 Myr, Mars escaped four times and the system decayed [^3]. It must be stressed that the chosen length of the integration time is a problem, because it may be relatively short for such investigations. It is well-known that there exist orbits which are stable for a very long time interval and then, all of a sudden they evolve into chaotic orbits. These ’sticky orbits’ are common in nonlinear dynamical systems and therefore they exist also in planetary systems [@Dvorak1998; @Jones2002]. This class of orbits are embedded in those region of the phase space, where regular and irregular orbits are close to each other. We suspect that those systems which did not decay, may be ’sticky systems’, and will be unstable at a much longer time.
According to the chaos theory the extent of this instability window is a function of the numerical resolution (in our case the $\Delta \kappa_E$). Therefore it is not possible to find the exact size of this window, however with our resolution this instability window is in the interval 4.2 $< \kappa_E <$ 4.9
The surprising increase in the eccentricity of Mars may be a result of some secular resonances between the secular frequencies $g_j$ and $f_j$. Such phenomena were already found by Laskar in the Solar System, who reported that large and irregular variations can appear in the eccentricities and inclinations of the terrestrial planets, especially of Mercury and Mars on time scales of several Gyr [@Laskar96].
The secular variations of the orbital elements of the planets are calculated by means of the Laplace-Lagrange theory. Using our [MAPLE]{} application the $M_s^{(j)}$ and $L_s^{(j)}$ amplitudes, the $g_j$ and $f_j$ secular frequencies and $\beta_j$ and $\gamma_j$ were calculated in the $\kappa_E \in$ \[4,6\] interval on a very fine grid with a stepsize $\Delta\kappa_E =0.001$. In this region two frequencies, $f_2$ and $f_3$ have almost the same value (see Table \[tab:4.1\]), while the other ones are well separated. In Table \[tab:4.1\] the $f_5=0.0$ frequency is not included, and also the corresponding amplitude (2.8402 $\cdot 10 ^{-2}$) and the angular phase (106$^\circ$.17) were left out.
To visualize the dependence of $f_2$ and $f_3$ on the mass factor, they are shown in Fig. \[fig7\]a, while in Fig. \[fig7\]b their difference around $\kappa_E = 5$ is depicted. From Fig. \[fig7\]b it is obvious, that the two curves do not intersect each other: the difference between the two frequencies are several orders of magnitudes greater than the accuracy of the computation, which was set to $10^{-10}$. The $\Delta f=f_2 - f_3$ difference reaches its smallest value for $\kappa_E=5.03$, $\Delta f=0.1439$ ”/yr, which is at the border of the aforementioned instability window.
A study of Table \[tab:4.1\] shows that the largest amplitude are, in the solution for Mars, $L_3^{(2)}$ and $L_3^{(3)}$, for Jupiter $L_4^{(2)}$ and $L_4^{(3)}$ and for Saturn $L_5^{(2)}, L_5^{(3)}$. Accordingly the $f_2$ and $f_3$ frequencies, as was described in section 2.1, can be associated with Mars, Jupiter and Saturn. The orbital plane of Mars therefore rotates together with those of Jupiter and Saturn, giving rise to chaotic behaviour. The equality of two apsidal or nodal rates is referred to in Solar System as a secular resonance. In this case we have three secular resonances: $\dot \Omega_M \approx \dot\Omega_J$, $\dot\Omega_M \approx
\dot\Omega_S$ and $\dot\Omega_J \approx \dot\Omega_S$. We suspect that these secular resonances are the possible source of the observed chaos, and produce the instability window.
$f_j$ -47.381072 -7.157383
------------- -------------------------- -------------------------- -------------------------- --------------------------
$\gamma_j$ 76$^\circ$.92 306$^\circ$.01 128$^\circ$.89 297$^\circ$.61
$L_s^{(j)}$
Venus 5.2399 $\cdot 10 ^{-2}$ 7.5778 $\cdot 10 ^{-3}$ 1.0696 $\cdot 10 ^{-2}$ 2.5795 $\cdot 10 ^{-2}$
Earth -7.7184 $\cdot 10 ^{-3}$ 3.9381 $\cdot 10 ^{-3}$ 5.9598 $\cdot 10 ^{-3}$ 2.3939 $\cdot 10 ^{-2}$
Mars 2.4032 $\cdot 10 ^{-3}$ -4.2737 $\cdot 10 ^{-1}$ -4.3543 $\cdot 10 ^{-1}$ 1.3348 $\cdot 10 ^{-2}$
Jupiter 4.5852 $\cdot 10 ^{-6}$ 3.1885 $\cdot 10 ^{-3}$ -3.1088 $\cdot 10 ^{-3}$ -1.2056 $\cdot 10 ^{-4}$
Saturn -2.2595 $\cdot 10 ^{-6}$ -7.7642 $\cdot 10 ^{-3}$ 7.7424 $\cdot 10 ^{-3}$ -1.8405 $\cdot 10 ^{-4}$
\[tab:4.1\]
![On the a.) panel are the $f_2$ and $f_3$ secular frequencies, on the b.) panel their difference around 5.0 are plotted versus the $\kappa_E$ mass factor.[]{data-label="fig7"}](./fig7a.eps "fig:"){width="1.0\linewidth"} ![On the a.) panel are the $f_2$ and $f_3$ secular frequencies, on the b.) panel their difference around 5.0 are plotted versus the $\kappa_E$ mass factor.[]{data-label="fig7"}](./fig7b.eps "fig:"){width="1.0\linewidth"}
The $30 \leq \kappa_E \leq 600$ region
--------------------------------------
Fig. \[fig8\] summarizes the results of the MEM in the $30 \leq \kappa_E \leq 600$ region. As one can clearly see from Fig. \[fig8\] the MEs of the planets gradually increase. None of the curves show any peaks, consequently in this region no instability window was found. These numerical results are in line with the results of the linear theory. In this region each of the secular frequencies have quite different values. Beyond $\kappa_E \ge 540$ Mars escaped in all simulations. The MEs of the Earth, Jupiter and Saturn behave very similarly and we may say that from $\kappa_E = 100$ on the system is dominated by these planets and Venus and Mars may be considered as quasiasteroids.
![Results of the MEM in the $30 \leq \kappa_E \leq 600$ region. Note, that the $y$ axis is logarithmic.[]{data-label="fig8"}](./fig8.eps){width="1.0\linewidth"}
Generally the semimajor axes of the inner planets are confined in a narrow stripe in the order of $10^{-4}$–$10^{-3}$ AU. These stripes are defined by a very short quasiperiodic oscillation around the mean values of the planets’ semimajor axes. As the mass factor increases the amplitudes of the oscillations grow, the width of these stripes slightly broadens.
The typical evolution of the eccentricity of the inner planets are shown in Fig. \[fig9\]. The character of the curves are similar to Fig. \[fig4\]c, only the periods are shorter.
Since Venus, Earth and Mars have the same period in their eccentricities – just like the Jupiter and Saturn pair does – the system is separated into two subsystems: one consisting of Venus, Earth and Mars, and the other Jupiter and Saturn. This behaviour was already observed for smaller mass factor values. The strong coupling of the eccentricities of the two giant planets remains unbroken for all $\kappa_E$. The eccentricities of Venus and Mars are mainly determined by that of the Earth, which is the superposition of a long period ($T_l \sim 2.5\cdot10^{5}$ yr) variation with amplitude 0.035 and several shorter period ($\sim 10^{4}$ yr) and smaller amplitude variations (see Fig. \[fig9\]b). In the case of Venus, upon this long period variation several very short periods are superimposed (see Fig. \[fig9\]a). On the contrary the eccentricity of Mars flickers with a period $T_l$ around 0.1 (see Fig. \[fig9\]c).
![The eccentricities of the Venus, Earth and Mars (from top to bottom) for $\kappa_E = 30$ over a 1 million year time-span.[]{data-label="fig9"}](./fig9a.eps "fig:"){width="1.0\linewidth"} ![The eccentricities of the Venus, Earth and Mars (from top to bottom) for $\kappa_E = 30$ over a 1 million year time-span.[]{data-label="fig9"}](./fig9b.eps "fig:"){width="1.0\linewidth"} ![The eccentricities of the Venus, Earth and Mars (from top to bottom) for $\kappa_E = 30$ over a 1 million year time-span.[]{data-label="fig9"}](./fig9c.eps "fig:"){width="1.0\linewidth"}
As the mass factor increases the long period variations in the eccentricities of Jupiter and Saturn are more and more influenced by the Earth. In order to conserve the total angular momentum of the system the centers of oscillations of Jupiter’s and Saturn’s eccentricities are anticorrelated with the eccentricity of the Earth (see Fig. \[fig10\]b). In spite of the large perturbations from the Earth, the coupling between the motion of Jupiter and Saturn are still very well determined (see Fig. \[fig10\]a). This result is also supported by our previous study [@Dvorak2002], and therefore we may say that the dynamical coupling of Jupiter and Saturn essentially determines the dynamics of the Solar System’s bodies.
![a.) The time development of the eccentricities of Jupiter and Saturn. b.) The eccentricity of the Earth and the local maximums of Jupiter’s and Saturn’s eccentricities are shown for $\kappa_E = 250$. The coupling between the eccentricities can be well observed.[]{data-label="fig10"}](./fig10a.eps "fig:"){width="1.0\linewidth"} ![a.) The time development of the eccentricities of Jupiter and Saturn. b.) The eccentricity of the Earth and the local maximums of Jupiter’s and Saturn’s eccentricities are shown for $\kappa_E = 250$. The coupling between the eccentricities can be well observed.[]{data-label="fig10"}](./fig10b.eps "fig:"){width="1.0\linewidth"}
The time developments of the inclinations are very similar to those of the eccentricities:
- the variations of Jupiter’s and Saturn’s inclinations are coupled,
- the inclinations of Venus and Mars are primarily determined by that of the Earth,
- with growing $\kappa_E$ Jupiter’s and Saturn’s inclinations are increasingly influenced by the Earth.
When $\kappa_E = 540$ Mars’ eccentricity had grown secularly and at $T_e \approx$ 3.4 Myr, Mars passed near by Earth, which ejected the planet from the system (see Fig. \[fig11\]). In this case, we did not observe orbit-crossing. However the minimum distance between Earth and Mars was 0.164 AU, which is about twice the Hill radius of the massive “Earth” $R_H(\kappa_E = 540) = 0.081$ AU. A cataclysmic outcome is therefore to be expected [@Jones2005]. In the systems with $\kappa_E$ = 570 and 600 Mars was ejected after a sequence of orbit-crossing with Earth at $T_e \approx 4.35 \times 10^6$ and at $T_e \approx 5.51 \times 10^6$ years, respectively.
The stability of Venus is an intriguing property of the systems. In each of our simulations the orbital elements of Venus did not show any sign of chaos. The Earth and Venus are prevented from close encounters by a coupling of Venus’ eccentricity and $\Delta \varpi$; $\Delta \varpi$ is the difference between the longitude of perihelion of the Earth and Venus. Whenever the perihelion of Earth conjunctions with the aphelion of Venus ($\Delta \varpi = 180^\circ$), the eccentricity of Venus is around its minimum (see Fig. \[fig12\]), maximizing the distance between the two orbits. A similar protection mechanism also occurs with real Mars and the asteroid Pallas. This kind of coupling was also observed by [@Jones2002] for an ’Earth’ in the 47 UMa system.
![This plot shows the escape of Mars for $\kappa_E$ = 540.[]{data-label="fig11"}](./fig11.eps){width="1.0\linewidth"}
![This plot shows the relationship between Venus’ eccentricity and $\Delta \varpi$ during 20 Myrs for $\kappa_E$ = 540.[]{data-label="fig12"}](./fig12.eps){width="1.0\linewidth"}
Discussion
==========
We studied the dynamics of a simplified dynamical model of the Solar System where we included the three terrestrial planets Venus, Earth and Mars and the gas giants Jupiter and Saturn. This model was already studied in detail [@Dvorak2002] when the masses of all terrestrial planets were uniformly increased. It turned out that the different systems remained stable for 10 Myrs up to large mass factor values. As a continuation of this work the same model with a more massive Earth was studied and signs of chaotic behaviour were reported in [@Dvorak2005]. In this article a detailed exploration of the different $S^3$ setups is presented, and the existence of an instability window is revealed. As an important byproduct the plethora of the integrations can serve as a general model of exoplanetary systems with two massive planets close to the 5:2 mean motion resonance on low eccentric orbits for comparable mass ratios of the giant planets.
We investigated over 100 $S^3$ models where we increased the mass of the Earth by a mass factor $\kappa_E$ between $1 \le \kappa_E \le 600$. These new systems have been integrated using extensive numerical integrations for 20 million years (for selected systems up to 100 million years) to find out the effect of a very massive Earth on the stability of the whole system on one hand; on the other hand the model now can serve as example of exoplanetary systems of two or three massive planets.
It turned out, that even when the Earth had Jupiter’s mass ($\kappa_E \approx 300$) and beyond, the system was still stable, but when $\kappa_E \in (4.2\,,4.9)$, the motion of Mars become chaotic. In this instability window Mars’ orbital eccentricity finally reached values which led to close encounters of Mars with the Earth, and even with Venus. After a sequence of close encounters Mars escaped within some millions of years. Using the results of the Laplace-Lagrange secular theory we found secular resonances acting between the motions of the nodes of Mars, Jupiter and Saturn. These secular resonances give rise to strong chaos, which is the cause of the appearance of the instability window, and eventually the escape of Mars.
We also found an interesting coupling of Venus’ $e$ and $\Delta \varpi$, which protect Venus form a close encounter with Earth. This mechanism was observed in the Solar System and in 47 UMa with a hypothetical Earth [@Jones2005; @Jones2002].
According to these results, the stability of the Solar System depends on the masses of the planets, and small changes in these parameters may result in a different dynamical evolution of the planetary system. No other instability window in $\kappa_E$ were found; first results of additional computations where we increased the masses of Venus and separately also of Mars, showed signs of chaotic motions for some windows in $\kappa$ too, but a detailed study of these two other cases of a modified $S^3$ is in preparation. There we also intend to give a detailed comparison of the three systems, namely with a massive Venus, a massive Earth and a massive Mars.
Finally we note that all these models may be used as dynamical reference models for a better understanding of the stability of orbits in extrasolar planetary systems.
Acknowledgments {#acknowledgments .unnumbered}
===============
We would like to thank Professor B. Érdi for a critical reading of the original version of the paper and also the anonymous referee for several comments which greatly improved the clarity of the paper. We thank the Wissenschaftlich-Technisches Abkommen Österreich-Ungarn Projekt A12-2004, and the Hungarian Scientific Research Fund T043739. All numerical integrations were accomplished on the NIIDP (National Information Infrastructure Development Program) supercomputer in Hungary.
[99]{}
Asghari N. et al., 2004, A&A, 426, 353
Dvorak R., Süli Á. and Freistetter F., 2005, in eds. Z. Knezevic IAU Colloq. 197, Dynamics of populations of planetary systems, p. 63
Dvorak R., Pilat-Lohinger E., Funk B., Freistetter F., 2003, A&A, 398, L1
Dvorak R., Süli Á., 2002, Celest. Mech. & Dyn. Astron., 83, 77
Dvorak R., Contopoulos G., Efthymiopoulos Ch., Voglis N., 1998, Planetary and Space Science, 46, 1567
Hanslmeier A., Dvorak R., 1984, A&A, 132, 203
Ito T., Tanikawa K., 2002, MNRAS, 336, 483
Jones B. W., Sleep P. N. Chambers J. E., 2001, A&A, 366, 254
Jones B. W., Sleep P. N., 2002, A&A, 393, 1015
Jones B. W., Underwood D. R., Sleep P. N., 2005, ApJ, 622, 1091
Knezevic Z., 1986, Celest. Mech., 38, 123
Laskar J., 1988, A&A, 198, 341
Laskar J., 1996, Celest. Mech. & Dyn. Astron, 64, 115
Lecar M., Franklin F.A., Holman M.J., Murray N.W., 2001, Annu. Rev. A & A, 39, 581
Lichtenegger H., 1984, Celest. Mech., 34, 357
Murray N., Holman M., 1999, Sci. 283, 1877
Nesvorný D., Morbidelli A., 1999, Celest. Mech. & Dyn. Astron., 71, 243
\[lastpage\]
[^1]: E-mail: [email protected] (ÁS)
[^2]: With OGLE, which is based on optical gravitational lensing it is possible to detect exoplanets with only Earth-mass.
[^3]: By the decay of the system, we understand that the system has evolved into a different one, with a completely distinct configuration.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper, we study the problem of object counting with incomplete annotations. Based on the observation that in many object counting problems the target objects are normally repeated and highly similar to each other, we are particularly interested in the setting when only a few exemplar annotations are provided. Directly applying object detection with incomplete annotations will result in severe accuracy degradation due to its improper handling of unlabeled object instances. To address the problem, we propose a positiveness-focused object detector (PFOD) to progressively propagate the incomplete labels before applying the general object detection algorithm. The PFOD focuses on the positive samples and ignore the negative instances at most of the learning time. This strategy, though simple, dramatically boosts the object counting accuracy. On the CARPK dataset for parking lot car counting, we improved [email protected] from 4.58% to 72.44% using only 5 training images each with 5 bounding boxes. On the Drink35 dataset for shelf product counting, the [email protected] is improved from 14.16% to 53.73% using 10 training images each with 5 bounding boxes.'
author:
- Jianfeng Wang
- Rong Xiao
- Yandong Guo
- Lei Zhang
bibliography:
- 'main.bib'
title: |
Learning to Count Objects\
with Few Exemplar Annotations
---
Introduction
============
Object counting is to count the number of object instances in a single image or video sequence. It has many real-world applications such as traffic flow monitoring, crowdedness estimation, and product counting.
Existing approaches towards the counting problem can be roughly categorized as regression-based approach [@AnLV07; @ChanLV08; @ChenGXL13; @ChenLGX12; @KongGT06], density-based approach [@ZhangLWY15; @ArtetaLNZ14; @RodriguezLSA11] and detection-based approach [@HsiehLH17; @KamenetskyS15; @MoranduzzoM14]. The regression-based approach directly learns a mapping from the image to the number of instances. In contrast, the density-based approach first estimates a density map and then aggregates the density information to get the number of instances. Both categories of approaches provide little information of exact instance location, whereas the detection-based approach is capable of detecting the location of each individual instance and thus making the counting result more explainable. Due to this advantage, we mainly focus on the detection-based approach in this work.
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Different training image settings in detection-based object counting. The label names are shown in bottom left []{data-label="fig:isl"}](figs/CARPK_fsl.jpg "fig:"){width="\thirdsplitwidth\linewidth"} ![Different training image settings in detection-based object counting. The label names are shown in bottom left []{data-label="fig:isl"}](figs/CARPK_wsl.jpg "fig:"){width="\thirdsplitwidth\linewidth"} ![Different training image settings in detection-based object counting. The label names are shown in bottom left []{data-label="fig:isl"}](figs/CARPK_isl_4.jpg "fig:"){width="\thirdsplitwidth\linewidth"}
![Different training image settings in detection-based object counting. The label names are shown in bottom left []{data-label="fig:isl"}](figs/CocoBottle1024DrinkY_fsl.jpg "fig:"){width="\thirdsplitwidth\linewidth"} ![Different training image settings in detection-based object counting. The label names are shown in bottom left []{data-label="fig:isl"}](figs/CocoBottle1024DrinkY_wsl.jpg "fig:"){width="\thirdsplitwidth\linewidth"} ![Different training image settings in detection-based object counting. The label names are shown in bottom left []{data-label="fig:isl"}](figs/CocoBottle1024DrinkY_isl_9.jpg "fig:"){width="\thirdsplitwidth\linewidth"}
\(a) Fully-Supervised \(b) Weakly-Supervised \(c) Incompletely-Supervised
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
In recent years, the accuracy of object detection has been dramatically improved [@Girshick15; @RenHGS15; @LiuAESRFB16; @RedmonF17] thanks to the advance of deep convolutional neural network. However, the high accuracy is achieved normally with the large number of fully annotated bounding boxes. In the context of object counting, the number of instances even in a single image could be huge (e.g. from tens to hundreds), which not only presents a great challenge to object detection, but also requires tremendous annotation effort to build a high-accuracy object detector.
To save the annotation cost, weakly supervised approaches are generally adopted to train the model solely based on image-level annotations. However, such approaches still suffer from suboptimal accuracy due to the lack of instance-level (e.g. bounding box) annotations. To address this problem, we resort to a problem setting where each image is only partially annotated with a few bounding boxes and leaves other instances unlabeled. This setting is practically useful especially when the number of instances is large and it is tedious and costly to label all the instances. Another motivation is that in certain object counting problem, the large number of instances exhibits less variances, e.g. scale, color. Potentially, a few annotated bounding boxes might generalize well to achieve a high accuracy. We call the learning algorithm trained on such training data *incompletely-supervised learning*. Fig. \[fig:isl\] shows examples to illustrate the training image difference among the fully-supervised, weakly-supervised, and incompletely-supervised learning. For the fully-supervised setting, all cars and drinks are annotated. The total number of the car instances is as many as 58. In the weakly-supervised setting, we only have the information that the car or drinks exist in the image. Based on the observation that in many object counting problems the target objects are normally repeated and highly similar to each other, we are particularly interested in the setting when only a few exemplar annotations are provided, which is illustrated in Fig. \[fig:isl\](c)
In incompletely-supervised learning, if any unlabeled region is simply treated as background in object detection training, as the number of labeled instances is usually small, the trained model could over-fit the training data severely, and mistaken the unlabeled object as background. To address the issue, we propose a simple yet effective algorithm, named *positiveness-focused object detector* (PFOD), to progressively propagate the labels, which treats unlabeled regions as background first and then neglects them to mainly focus on the positive samples during the training. The intuition is to first learn a compact classifier which might over-fit the data, and then relax the learning by ignoring the unlabeled region to pull the unlabeled instances towards being positive. In this way, bounding box supervision could be automatically expanded during training before we apply a standard object detector.
Overall, our contributions can be summarized as follows.
1. We explicitly formulate the problem of incompletely-supervised learning, which focuses on the incomplete annotations for object counting.
2. We propose a progressive label propagation algorithm through positiveness-focused object detector to properly handle the incomplete labels.
3. We conduct extensive experiments demonstrates the effectiveness of the proposed PFOD to handle such training problems. On the CARPK dataset for parking lot car counting, we improved [email protected] from 4.58% to 72.44% using only 5 training images each with 5 bounding boxes. On the Drink35 dataset for shelf product counting, the [email protected] is improved from 14.16% to 53.73% using 10 training images each with 5 bounding boxes.
Related Work
============
Object Counting
---------------
The approaches towards object counting can be roughly categorized into three categories: regression-based approach, density-based approach, and detection-based approach.
### Regression-based Approach
predicts instance count directly based on global regressors with image features [@ChanLV08; @ChenGXL13; @ChenLGX12; @KongGT06]. For example, in [@ChanLV08], the video sequence is first segmented into different components of homogeneous motion, and a Gaussian process regression is learned for each segmented region to count the number of instances. Cumulative attribute representation [@ChenGXL13] was proposed and used to learn the regressor to handle the imbalanced training data. Inter-dependent features are used to mine the spatially importance among different region to learn the number of count [@ChenLGX12]. Feature normalization is taken into account to deal with perspective projections in [@KongGT06]. With the advantage of the deep learning, [@MundhenkKSB16] estimate the number of cars using CNNs. These regression-based approaches provide no clue of individual location of each object, which limits its potential applications.
### Density-based Approach
first maps the image to a density map, such that the integral over any sub region gives the count of objects within that region [@LempitskyZ10; @ArtetaLNZ14; @RodriguezLSA11; @ZhangLWY15]. In [@LempitskyZ10], the pixel-level density is learned by minimizing a regularized risk quadratic cost function. Based on the density map, an interactive counting system is introduced in [@ArtetaLNZ14] to incorporate the relevance feedback. Instead of hand-crafted feature, [@ZhangLWY15] focuses on the CNN-based density map and instance count estimation in cross-scene scenario. While the density map provides certain clue of the crowdedness, it still lacks the exact position of each instance.
### Detection-based Approach
gives the total count by localizing each object instance. Such approaches can be considered as the application of object detection and thus benefit a lot from the improvement of object detection [@Girshick15; @RenHGS15; @LiuAESRFB16; @RedmonF17]. However, directly applying object detection for object counting requires special focus on small-scale objects and extraordinary effort on massive instance labeling. In [@LinD10], a hierarchical part-template matching approach is proposed to detection humans, which requires careful feature and template design, while in [@HsiehLH17], the neural network learning is applied to detect and count car instances by incorporating layout information. With its large potential in applications demanding location information, we mainly work on the detection-based approach with special focus on incompletely-supervised learning.
Object Detection
----------------
Nowadays, mainstream object detection algorithms have changed to CNN-based implementation due to its powerful representation and high accuracy. Such algorithms can be roughly categorized into two-stage object detector [@GirshickDDM14; @Girshick15; @RenHGS15], and one-stage object detector [@LiuAESRFB16; @RedmonF17]. Two-stage object detectors such as Faster RCNN [@RenHGS15] first extract region proposals and then perform classification and bounding box regression, while single-stage objectors such as YOLO [@RedmonF17] directly output the bounding box locations and the classification results without generating the region proposal. All the object detection algorithms by default following the fully supervised setting - requiring large amount of training images with annotated bounding boxes to achieve a high accuracy.
To reduce the annotation effort, weakly supervised learning trains object detector only based on image-level labels. Most approaches [@BilenV16; @CinbisVS17; @KantorovOCL16; @TangWBL17] first generates multiple region proposals for each image and then leverage multi-instance learning algorithms to solve the problem. However, such approaches still suffer from suboptimal accuracy due to the lack of instance-level labels.
In both the fully and weakly supervised settings, each image is either labeled with full bounding box annotations or only image-level category information. In contrast, we are more interested in the incomplete supervised setting, where only a few exemplar bounding boxes are annotated and all others are unlabeled. This setting is practically more useful when the number of instances is large, for example to count the number of products on a retail store shelf, or the number of cars in a large parking lot.
Approach
========
Problem Definition
------------------
Although detection-based object counting resembles the general object detection problem, it still presents unique challenges when the total number of object instances in an image is large, e.g. from tens to hundreds, and the size of each instance is relatively small comparing with the image size. To save the tedious and costly labeling effort, we assume each image $I_n$ in the training set $\{I_n|n=1,\ldots,N\}$ is only labeled with a few instances (e.g. less than ten) though there might be tens or hundreds of instances in one image. For image $I_n$, denote by $\mathbf{x}_i^{n}$ its $i$-th annotated bounding box ($i = 1, \ldots, M_1(n)$). Since not all instances are annotated, the rest region could also contain the object instances. Without loss of generality, we assume the number of object categories is 1. For notation simplicity, we drop the $n$ from $\mathbf{x}_i^{n}$ and $M_1(n)$ when there is no ambiguity. Thus, the problem is how to build an effective object detector to count the object instances based on the incomplete annotations $\{\mathbf{x}_i^n\}$. We name the problem as *incompletely-supervised learning* for object counting.
### Comparison with other supervised learning problems.
The differences with other supervised learning settings in the context of object detection are also shown in Fig. \[fig:isl\]. Fully-supervised learning requires that all the bounding boxes are labeled. This represents the upper bound of detection-based object counting performance, but is costly to label every instance. Weakly-supervised learning assumes that only image-level category information is available. That means, we know there are some object instances in one image, but we have no information of their locations. Another problem setting is low-shot learning, where the number of training images is small, but each training image has full annotations. This is more like the fully-supervised learning but with only a few training images.
[@c@[ ]{}c@[ ]{}c@[ ]{}c@[ ]{}c@[ ]{}c@]{} ![Intuition of the proposed label propagation through a positiveness-focused object detector to manage the incompletely-supervised labels[]{data-label="fig:intuition"}](figs/ideal.eps "fig:"){width="\intuitionwidth\linewidth"} & ![Intuition of the proposed label propagation through a positiveness-focused object detector to manage the incompletely-supervised labels[]{data-label="fig:intuition"}](figs/baseline.eps "fig:"){width="\intuitionwidth\linewidth"} & ![Intuition of the proposed label propagation through a positiveness-focused object detector to manage the incompletely-supervised labels[]{data-label="fig:intuition"}](figs/phase1.eps "fig:"){width="\intuitionwidth\linewidth"} & ![Intuition of the proposed label propagation through a positiveness-focused object detector to manage the incompletely-supervised labels[]{data-label="fig:intuition"}](figs/phase2.eps "fig:"){width="\intuitionwidth\linewidth"} & ![Intuition of the proposed label propagation through a positiveness-focused object detector to manage the incompletely-supervised labels[]{data-label="fig:intuition"}](figs/phase3.eps "fig:"){width="\intuitionwidth\linewidth"} & ![Intuition of the proposed label propagation through a positiveness-focused object detector to manage the incompletely-supervised labels[]{data-label="fig:intuition"}](figs/phase4.eps "fig:"){width="\intuitionwidth\linewidth"}\
\
(a) Ideal & (b) Baseline & (c) Stage 1 & (d) Stage 2 & (d) Stage 3 & (e) Stage 4\
Label Propagation by Positiveness-Focused Object Detection
----------------------------------------------------------
As we have a few exemplar labels of the target objects, the most intuitive way is to treat the labels as seeds and carefully propagate them to unknown regions. We propose a simple yet effective *positiveness-focused object detector* (PFOD) to solve the propagation problem.
### Intuition.
Fig. \[fig:intuition\] illustrates the intuition of the label propagation by PFOD. The plus symbol denotes a positive sample while the minus symbol denotes negative. The sample in red is known (labeled) while the sample in blue is unknown. The line with the dotted line is the decision boundary. If all the positive and negative samples are known, we can easily figure out the decision boundary to separate true positive and true negative samples, as shown in Fig. \[fig:intuition\](a). When only one positive sample is known and all the rest are unknown, as shown in Fig. \[fig:intuition\](b), if we simply treat all the unknowns as negative sample, the decision boundary could mis-classify the unlabeled positive samples as being negative.
To compensate the lack of negative samples, we can introduce some images as extra negative data, as shown in green from Fig. \[fig:intuition\](c) to Fig. \[fig:intuition\](e). For object detection, it is indeed easy to find images without any target objects in the same domain. For example, we can use the PASCAL VOC [@EveringhamGWWZ10] 2007 data set as extra background images for the problem of car counting on the CARPK [@HsiehLH17] dataset. In Fig. \[fig:intuition\](c) we show how to propagate the labels. We first treat all the unlabeled samples (shown as both blue plus and minus symbols in Fig. \[fig:intuition\](c)) as negative samples to train the detector, and the learned decision boundary will be pushed close to the only positive sample (red plus symbol) as depicted with the red dotted line. Next, we ignore the unknown samples (note that we still have the extra negative samples shown as green minus symbols) and gradually update the learned classifier to classify the labeled positive sample and the extra negative samples. As the unknown positive samples (shown as blue plus symbols) are not taken as negative samples, they no longer push the decision boundary. As a result, the decision boundary (as shown as black dotted line) will be moved a bit further from the known positive sample and classify a few more unknown positive samples as positive samples.
If the propagation is carefully controlled, we can treat the newly classified positive samples as known labeled data for the next stage, as shown in Fig. \[fig:intuition\](d). We repeat the process above to iteratively learn the boundary and propagate the label set. Each iteration here is called a stage, and we use $L^{s}=\{\mathbf{x}_i|i=1,\ldots M_s\}$ to denote the label set in the $s$-th stage, where $M_s$ is the number of expanded bounding boxes in one image. Finally, we combine all the expanded positive samples and take all the others as background to learn the final decision boundary.
![Framework of the proposed label propagation by a positiveness-focused object detector (PFOD) []{data-label="fig:framework"}](figs/framework.pdf){width="0.99\linewidth"}
### Solution.
Fig. \[fig:framework\] shows the framework of the proposed strategy to learn the object detector. At the initial state of training, the bounding boxes we have are the labeled set, i.e. $L^1 = \{\mathbf{x}_i | i= 1, \ldots, M_1\}$ for each training image. Based on $L^s$, we train a positiveness-focused object detector (PFOD), which can be based on any object detector [@GirshickDDM14; @Girshick15; @RenHGS15; @LiuAESRFB16; @RedmonF17]. In this work, we choose YoloV2 [@RedmonF17] for its simplicity. The network first processes each training image in a batch manner by a fully convolutional neural network, and then outputs three components at each spatial position: bounding box coordinates, objectiveness to tell how confident the bounding box contains an object, and classification scores to tell which category the bounding box contains. Here we assume the number of categories is 1 and remove the classification module. The Euclidean loss is used for bounding box coordinate regression and objectiveness confidence regression. Specifically, for the objectiveness at spatial position $u$, the loss is defined as $$\begin{aligned}
l = (f_\theta(u) - \hat{f_u})^2\end{aligned}$$ where $\theta$ is the network parameter, learned iteratively through the mini-batch stochastic gradient descent (SGD) algorithm, $f_\theta(u)$ is the objectiveness score at position $u$, and $\hat{f_u}$ is 1 if it is identified as being positive for position $u$, and 0 otherwise based on the current label set. For the extra background images, the label is consistently set as 0 and the loss will be always enabled. For the training images in the target domain, we modify it as follows to implement PFOD,
$$\begin{aligned}
l = \begin{cases}
(f_\theta(u) - 1)^2 & \hat{f_u} = 1 \\
f_\theta^2(u) & \hat{f_u} = 0 \text{ \& } t \le T \\
0 & \hat{f_u} = 0 \text{ \& } t > T
\end{cases}\end{aligned}$$
where $t$ is the number of iterations in SGD, $T$ is a pre-defined parameter (200 in experiments) to determine how many iterations are needed to treat the unknown regions as background. After $T$ iterations, the detector training will only focus on the positive samples and the extra background data.
With the model trained by PFOD, we run the prediction over all the training images from target domain. The predicted bounding boxes with high probability scores (0.9 in experiments) will be merged into the original label set $L^s$ to form $L^{s + 1}$. We also discard any predicted bounding box if it has a high overlap (Intersection-over-Union $>$ 0.2 in experiments) with any of the original bounding boxes in $L^s$.
After $S$ stages, we feed the training images and the expanded bounding box set $L^{S+1}$ into a normal object detector training pipeline, in which the bounding boxes in $L^{S+1}$ are positive while all the rest are negative. We still apply the YoloV2 algorithm here for training and testing. Ideally, if all the unlabeled object instances could be propagated, the trained model should be able to achieve an accuracy on par with that trained from the full annotations.
Experiments
===========
Settings
--------
### Datasets.
We mainly evaluate the approaches on the CARPK [@HsiehLH17] dataset, which contains 989 training images and 459 testing images. The task is to detect and count the car instances in the image. Each training image has $42.7\pm 15.7$ annotated cars, while each testing image has $103.5 \pm 39.1$ cars. The images are collected by a drone on top of car parks. An example image is shown in the first row of Fig. \[fig:isl\]
Another interesting application is to count the number of drinks or products on retail store shelves. To demonstrate the effectiveness of the incompletely-supervised learning algorithm on other domains, we collect a small dataset *Drink35*, which contains 10 images as the training set and 25 images as the testing set. The task is to detect and count all the product instances. One example image is shown in the second row of Fig. \[fig:isl\].
To simulate the incompletely-supervised settings, we randomly select $C_i$ training images, and for each image we randomly select at most $C_a$ annotated cars. All the other unselected images are discarded during training. This training set is denoted as $\text{CARPK}\_C_i\_C_a$ or $\text{Drink35}\_C_i\_C_a$ for CARPK and Drink35 datasets, respectively. For example, CARPK\_5\_5 means the training set with 5 images and each with at most 5 annotated bounding boxes. Similarly, the suffix of $\_C_i\_\text{ALL}$ denotes the training set of $C_i$ images with all the annotated boxes, and $\_\text{ALL}\_C_a$ denotes all training images with at most $C_a$ annotated bounding boxes in each image. The test set is not altered for consistent evaluation.
We use the PASCAL VOC [@EveringhamGWWZ10] 2007 trainval set (5011 images) as the extra background images with all the original bounding box labels removed. Note that the images in PASCAL VOC 2007 contains the object of cars and drinks. We keep these images as negative samples because the cars and drinks in VOC 2007 are generally of difference appearances or views compared with the object instances in CARPK and Drink35.
### Criteria.
Since we focus on the detection-based approaches, we adopt the widely-used [@Girshick15; @RenHGS15; @LiuAESRFB16; @RedmonF17] [email protected] as one of the metrics, which measures the mean average precision (mAP) using 0.5 as the interaction-over-union (IoU) threshold.
Following [@LempitskyZ10; @HsiehLH17], we also use the Mean Absolute Error (MAE) and the Root Mean Squared Error (RMSE) to evaluate the accuracy of the counting results. MAE is defined as $\sum_{i}|n_i - n_i'|/N$, while RMSE as $\sum_{i}\sqrt{(n_i - n_i')^2}/N$, where $n_i$ is the number of objects predicted by the model for the $i$-th testing image, $n_i'$ is the ground-truth number of objects, and $N$ is the total number of testing images.
### Implementation details.
The data augmentation is of great importance for the network learning due to the small training set with incomplete labels. Motivated by the implementation of Yolo [@RedmonF17] and SSD [@LiuAESRFB16], we incorporate the random scaling, random aspect ratio distortion, and color jittering. Random rotation is also implemented for the car counting problem by multiple data samplers (motivated by SSD [@LiuAESRFB16]), so that non-rotated images are preferred than rotated images. That is, images with 0, 90, 180, and 270 degrees’ rotations are preferred than images with arbitrary rotations, and images with less than 10 degrees’ rotation are preferred than images with other arbitrary rotations. Specifically, for each training image, we have 25% chance to select non-rotated image; 25% chance to rotate images by $90$x degrees; 25% chance to rotate images by less than $\pm 10$ degrees plus a random $90$x rotation; and the last 25% chance to rotate images for any angle.
Another important parameter is that we use a large input resolution for both training and testing due to large number of objects and small object sizes in each image. During training, we resize the image so that the longer side randomly ranges from $832$ to $1664$, and then crop a subregion of $416\times 416$ as the network input. During testing, we resize the input image so that the whole area is close to $1248\times 1248$ while its aspect ratio is kept. Non-Maximum Compression (NMS) is used to filter the bounding box and the IoU threshold is 0.2.
The network backbone is Darknet19, which is the same as in YoloV2 [@RedmonF17]. We use 9 stages of PFOD to propagate the labels. In each stage of PFOD, we train the network with 10K iterations. The learning rate is 0.0001 for the first 100 iterations, 0.001 for the next 4900 iterations, 0.0001 for 4000 iterations and 0.00001 for the last 1000 iterations[^1]. The batch size is set to 64, and the weight decay is 0.0005. The last detector training shares the same parameters. The training takes 1.25 hours on 4 NVidia P100 GPUs to finish the 10K iterations. The implementation is based on Caffe [@jia2014caffe] under the environment of CUDA 8.0 and CUDNN 5.1.
We also report the accuracy without any label propagation, and directly train the object detector on the provided label set. This straightforward approach is denoted as OD as a naive baseline, and all the data augmentation and learning strategy parameters are the same with a single stage of PFOD.
In the incomplete annotation settings, we enable the extra background samples by replacing 16 images in each batch of 64 with the extra background images.
Method $C_i$ $C_a$ mAP(%) MAE RSME
----------------- ------- ------- ---------- --------------- ---------------
LPN[@HsiehLH17] 989 ALL NotAvail 23.80 36.79
989 ALL 96.67 2.94 (0.15) 3.94 (0.10)
$5$ ALL 91.97 4.38 (0.20) 5.58 (0.20)
$5$ $50$ 87.90 6.02 (0.05) 8.23 (0.05)
$5$ $25$ 61.26 49.28 (0.05) 54.45 (0.05)
$5$ $10$ 4.95 99.39 (0.05) 106.51 (0.05)
$5$ $5$ 4.58 101.93 (0.10) 108.99 (0.10)
$5$ $50$ 83.73 11.54 (0.05) 16.76 (0.05)
$5$ $25$ 80.27 16.52 (0.05) 21.45 (0.05)
$5$ $10$ 73.46 22.47 (0.05) 27.63 (0.05)
$5$ $5$ 72.44 23.63 (0.05) 26.36 (0.05)
: Accuracy on CARPK with different training images. $C_i$: number of training images from the whole training set; $C_a$: number of labeled bounding box annotations for each image; mAP: the higher, the better; MAE/RSME: the lower, the better. The value in parenthesis of MAE/RSME is the threshold to decide if the predicted bounding box is valid []{data-label="tbl:differentc1c2"}
Results on CARPK
----------------
The results are shown in Table \[tbl:differentc1c2\] for different numbers of training images ($C_i$) and different numbers of labeled bounding boxes ($C_a$) in each image. To count the number of objects, we need a threshold to determine if the predicted bounding box should be kept from the detector. Since different settings might favor different thresholds, we select the one with lowest MAE or RSME among $[0.05:0.05:0.7]$. We select the best threshold for MAE and RSME to examine the best performance under these two criteria. The threshold is in parentheses of Table \[tbl:differentc1c2\]. Note the criterion of mAP does not depend on the threshold. From the table, we have the following observations and discussions.
### What is the upper bound performance using all the training data?
In the fully supervised setting ($C_i = \text{ALL}$, $C_a = \text{ALL}$), our detector (OD) could achieve 96.67% [email protected], 2.94 MAE and 3.94 RMSE. This significantly outperforms the state-of-the-art method of LPN [@HsiehLH17], whose MAE is 23.80 and RMSE is 36.79 (mAP is not reported). Other baseline approaches are not shown here because the accuracy is lower than LPN. This demonstrates a strong object detector towards the counting problem.
### Do we need hundreds of training images for car counting?
We apply the OD on $C_i=5$ images with all the labeled annotations ($C_a = \text{ALL}$), and the detector can still achieves 91.97 [email protected], 3.00 MAE, and 5.58 RSME without extra background images. Compared with $C_i = \text{ALL}$, $C_a = \text{ALL}$ which uses 989 training images, this result is very encouraging. It clearly indicates that to develop a car counting algorithm, 5 or a few more images might be enough rather than several hundreds. The 5 image IDs are 20160524\_GF2\_00133, 20161030\_GF1\_00153, 20161030\_GF1\_00036, 20160331\_NTU\_00007, and\
20161030\_GF2\_00071 for reproducing the result.
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Straightforward object detection result on two training images. The model is trained on CARPK\_5\_5. The ground-truth bounding boxes are almost identical with the predicted boxes and thus are not shown for clarity. Confidence scores shown around each box are all close to 1 (0.8+). The threshold is 0.05, which means the confidence scores for unlabeled/undetected cars are less than 0.05. This is a clear indication of overfitting, which severely degrades the performance []{data-label="fig:overfitting_issue"}](figs/overfitting_20160524_GF2_00133.jpg "fig:"){width="0.48\linewidth"} ![Straightforward object detection result on two training images. The model is trained on CARPK\_5\_5. The ground-truth bounding boxes are almost identical with the predicted boxes and thus are not shown for clarity. Confidence scores shown around each box are all close to 1 (0.8+). The threshold is 0.05, which means the confidence scores for unlabeled/undetected cars are less than 0.05. This is a clear indication of overfitting, which severely degrades the performance []{data-label="fig:overfitting_issue"}](figs/overfitting_20161030_GF1_00153.jpg "fig:"){width="0.48\linewidth"}
(a) (b)
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
### What if the unlabeled regions are treated as negative samples?
When the number of bounding boxes is decreased to 5, the mAP significantly drops to 4.40. To identify the reason, we evaluate the trained model against the training images and show two examples in Fig \[fig:overfitting\_issue\]. The predicted bounding boxes are drawn on the image with confidence scores around each box. Only the boxes with confidence scores higher than 0.05 are displayed. The selected 5 labeled boxes are located at the same position with the predicted boxes and the probability is close to 1. Since the threshold here is 0.05, all the other regions including the unlabeled cars are classified as background with high confidence. This shows that the model overfits the training data severely, which degrades the accuracy.
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Label propagation process on training images of CARPK\_5\_5. The two numbers below the images are the number of expanded boxes and the number of correctly expanded. From the left to the right, the total number of true bounding boxes is 61, 58, 34, respectively. From the top to bottom, the images correspond to the training data for the third, fifth-th, seven-th and the last stage, respectively. The original sampled 5 labeled bounding boxes are shown in blue; the propagated boxes ares shown in yellow; the incorrectly (IoU $<$ 0.3) propagated is shown in red[]{data-label="fig:labelpropagation"}](figs/label_prop_20160524_GF2_00133_10_2.jpg "fig:"){width="\thirdsplitwidth\linewidth"} ![Label propagation process on training images of CARPK\_5\_5. The two numbers below the images are the number of expanded boxes and the number of correctly expanded. From the left to the right, the total number of true bounding boxes is 61, 58, 34, respectively. From the top to bottom, the images correspond to the training data for the third, fifth-th, seven-th and the last stage, respectively. The original sampled 5 labeled bounding boxes are shown in blue; the propagated boxes ares shown in yellow; the incorrectly (IoU $<$ 0.3) propagated is shown in red[]{data-label="fig:labelpropagation"}](figs/label_prop_20161030_GF1_00153_10_2.jpg "fig:"){width="\thirdsplitwidth\linewidth"} ![Label propagation process on training images of CARPK\_5\_5. The two numbers below the images are the number of expanded boxes and the number of correctly expanded. From the left to the right, the total number of true bounding boxes is 61, 58, 34, respectively. From the top to bottom, the images correspond to the training data for the third, fifth-th, seven-th and the last stage, respectively. The original sampled 5 labeled bounding boxes are shown in blue; the propagated boxes ares shown in yellow; the incorrectly (IoU $<$ 0.3) propagated is shown in red[]{data-label="fig:labelpropagation"}](figs/label_prop_20161030_GF2_00071_10_2.jpg "fig:"){width="\thirdsplitwidth\linewidth"}
\(a) expand=6, correct=6 \(b) expand=13, correct=13 \(c) expand=8, correct=8
![Label propagation process on training images of CARPK\_5\_5. The two numbers below the images are the number of expanded boxes and the number of correctly expanded. From the left to the right, the total number of true bounding boxes is 61, 58, 34, respectively. From the top to bottom, the images correspond to the training data for the third, fifth-th, seven-th and the last stage, respectively. The original sampled 5 labeled bounding boxes are shown in blue; the propagated boxes ares shown in yellow; the incorrectly (IoU $<$ 0.3) propagated is shown in red[]{data-label="fig:labelpropagation"}](figs/label_prop_20160524_GF2_00133_10_4.jpg "fig:"){width="\thirdsplitwidth\linewidth"} ![Label propagation process on training images of CARPK\_5\_5. The two numbers below the images are the number of expanded boxes and the number of correctly expanded. From the left to the right, the total number of true bounding boxes is 61, 58, 34, respectively. From the top to bottom, the images correspond to the training data for the third, fifth-th, seven-th and the last stage, respectively. The original sampled 5 labeled bounding boxes are shown in blue; the propagated boxes ares shown in yellow; the incorrectly (IoU $<$ 0.3) propagated is shown in red[]{data-label="fig:labelpropagation"}](figs/label_prop_20161030_GF1_00153_10_4.jpg "fig:"){width="\thirdsplitwidth\linewidth"} ![Label propagation process on training images of CARPK\_5\_5. The two numbers below the images are the number of expanded boxes and the number of correctly expanded. From the left to the right, the total number of true bounding boxes is 61, 58, 34, respectively. From the top to bottom, the images correspond to the training data for the third, fifth-th, seven-th and the last stage, respectively. The original sampled 5 labeled bounding boxes are shown in blue; the propagated boxes ares shown in yellow; the incorrectly (IoU $<$ 0.3) propagated is shown in red[]{data-label="fig:labelpropagation"}](figs/label_prop_20161030_GF2_00071_10_4.jpg "fig:"){width="\thirdsplitwidth\linewidth"}
\(d) expand=8, correct=8 \(e) expand=35, correct=35 \(f) expand=29, correct=29
![Label propagation process on training images of CARPK\_5\_5. The two numbers below the images are the number of expanded boxes and the number of correctly expanded. From the left to the right, the total number of true bounding boxes is 61, 58, 34, respectively. From the top to bottom, the images correspond to the training data for the third, fifth-th, seven-th and the last stage, respectively. The original sampled 5 labeled bounding boxes are shown in blue; the propagated boxes ares shown in yellow; the incorrectly (IoU $<$ 0.3) propagated is shown in red[]{data-label="fig:labelpropagation"}](figs/label_prop_20160524_GF2_00133_10_6.jpg "fig:"){width="\thirdsplitwidth\linewidth"} ![Label propagation process on training images of CARPK\_5\_5. The two numbers below the images are the number of expanded boxes and the number of correctly expanded. From the left to the right, the total number of true bounding boxes is 61, 58, 34, respectively. From the top to bottom, the images correspond to the training data for the third, fifth-th, seven-th and the last stage, respectively. The original sampled 5 labeled bounding boxes are shown in blue; the propagated boxes ares shown in yellow; the incorrectly (IoU $<$ 0.3) propagated is shown in red[]{data-label="fig:labelpropagation"}](figs/label_prop_20161030_GF1_00153_10_6.jpg "fig:"){width="\thirdsplitwidth\linewidth"} ![Label propagation process on training images of CARPK\_5\_5. The two numbers below the images are the number of expanded boxes and the number of correctly expanded. From the left to the right, the total number of true bounding boxes is 61, 58, 34, respectively. From the top to bottom, the images correspond to the training data for the third, fifth-th, seven-th and the last stage, respectively. The original sampled 5 labeled bounding boxes are shown in blue; the propagated boxes ares shown in yellow; the incorrectly (IoU $<$ 0.3) propagated is shown in red[]{data-label="fig:labelpropagation"}](figs/label_prop_20161030_GF2_00071_10_6.jpg "fig:"){width="\thirdsplitwidth\linewidth"}
\(g) expand=10, correct=10 \(h) expand=46, correct=46 \(i) expand=32, correct=32
![Label propagation process on training images of CARPK\_5\_5. The two numbers below the images are the number of expanded boxes and the number of correctly expanded. From the left to the right, the total number of true bounding boxes is 61, 58, 34, respectively. From the top to bottom, the images correspond to the training data for the third, fifth-th, seven-th and the last stage, respectively. The original sampled 5 labeled bounding boxes are shown in blue; the propagated boxes ares shown in yellow; the incorrectly (IoU $<$ 0.3) propagated is shown in red[]{data-label="fig:labelpropagation"}](figs/label_prop_20160524_GF2_00133_10_9.jpg "fig:"){width="\thirdsplitwidth\linewidth"} ![Label propagation process on training images of CARPK\_5\_5. The two numbers below the images are the number of expanded boxes and the number of correctly expanded. From the left to the right, the total number of true bounding boxes is 61, 58, 34, respectively. From the top to bottom, the images correspond to the training data for the third, fifth-th, seven-th and the last stage, respectively. The original sampled 5 labeled bounding boxes are shown in blue; the propagated boxes ares shown in yellow; the incorrectly (IoU $<$ 0.3) propagated is shown in red[]{data-label="fig:labelpropagation"}](figs/label_prop_20161030_GF1_00153_10_9.jpg "fig:"){width="\thirdsplitwidth\linewidth"} ![Label propagation process on training images of CARPK\_5\_5. The two numbers below the images are the number of expanded boxes and the number of correctly expanded. From the left to the right, the total number of true bounding boxes is 61, 58, 34, respectively. From the top to bottom, the images correspond to the training data for the third, fifth-th, seven-th and the last stage, respectively. The original sampled 5 labeled bounding boxes are shown in blue; the propagated boxes ares shown in yellow; the incorrectly (IoU $<$ 0.3) propagated is shown in red[]{data-label="fig:labelpropagation"}](figs/label_prop_20161030_GF2_00071_10_9.jpg "fig:"){width="\thirdsplitwidth\linewidth"}
\(j) expand=22, correct=22 \(k) expand=50, correct=49 \(l) expand=37, correct=33
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
### How effective is the proposed incompletely-supervised learning approach?
We apply the proposed PFOD on CARPK\_5\_5, and surprisingly the accuracy is boosted to 72.44 mAP, 23.63 MAE and 26.36 RSME. In terms of MAE and RSME, the accuracy has surpassed the LRN [@HsiehLH17] trained on the full training set of 989 images.
In Fig. \[fig:labelpropagation\], we illustrate the label propagation process by PFOD on CARPK\_5\_5. Each column corresponds to one training image. The two numbers below each images are the number of the expanded boxes (initially provided + propagated), and the number of correct boxes among those boxes. A box is correct if its IoU is larger than 0.3 with at least one bounding box in CARPK\_5\_ALL. From the figure, the correct bounding boxes used for training could be gradually populated. Taking the leftmost image as an example, the number of correct boxes is increased from the initial number 5 to 22. The rightmost one can have 33 correct boxes.
Meanwhile, we observe that the propagation is still not perfect - it introduces several false bounding boxes while missing a few cars. This is the reason why there is still a gap between this setting and CARPK\_5\_ALL, and will motivate us to continue investigating the problem.
### Will introducing more labels help?
By increasing the number of labeled boxes per training image from 5 to 50, the accuracy can be smoothly increased for both OD and PFOD. With less than 25 boxes in each image, the accuracy of PFOD is consistently higher than OD, while with 50 boxes, the accuracy is lower. The reason is that under the setting of 50 boxes, most of the true boxes are included, while the label propagation introduced some false boxes. That is, under the almost full annotations, it is enough to apply the OD instead of propagating the boxes.
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
OD on CARPK\_5\_5 OD on CARPK\_5\_ALL PFOD on CARPK\_5\_5
![Visualization of detection/counting results on 3 CARPK test images. Each row corresponds to one test image. Left: OD on CARPK\_5\_5; Middle: OD on CARPK\_5\_Full; Right: our PFOD on CARPK\_5\_5 []{data-label="fig:vis_car"}](figs/vis_CARPK_base_20161225_TPZ_00158.jpg "fig:"){width="\thirdsplitwidth\linewidth"} ![Visualization of detection/counting results on 3 CARPK test images. Each row corresponds to one test image. Left: OD on CARPK\_5\_5; Middle: OD on CARPK\_5\_Full; Right: our PFOD on CARPK\_5\_5 []{data-label="fig:vis_car"}](figs/vis_CARPK_full_20161225_TPZ_00158.jpg "fig:"){width="\thirdsplitwidth\linewidth"} ![Visualization of detection/counting results on 3 CARPK test images. Each row corresponds to one test image. Left: OD on CARPK\_5\_5; Middle: OD on CARPK\_5\_Full; Right: our PFOD on CARPK\_5\_5 []{data-label="fig:vis_car"}](figs/vis_CARPK_ours_20161225_TPZ_00158.jpg "fig:"){width="\thirdsplitwidth\linewidth"}
\(a) pred=9, gt=127 \(b) pred=130, gt=127 \(c) pred=122, 127
![Visualization of detection/counting results on 3 CARPK test images. Each row corresponds to one test image. Left: OD on CARPK\_5\_5; Middle: OD on CARPK\_5\_Full; Right: our PFOD on CARPK\_5\_5 []{data-label="fig:vis_car"}](figs/vis_CARPK_base_20161225_TPZ_00239.jpg "fig:"){width="\thirdsplitwidth\linewidth"} ![Visualization of detection/counting results on 3 CARPK test images. Each row corresponds to one test image. Left: OD on CARPK\_5\_5; Middle: OD on CARPK\_5\_Full; Right: our PFOD on CARPK\_5\_5 []{data-label="fig:vis_car"}](figs/vis_CARPK_full_20161225_TPZ_00239.jpg "fig:"){width="\thirdsplitwidth\linewidth"} ![Visualization of detection/counting results on 3 CARPK test images. Each row corresponds to one test image. Left: OD on CARPK\_5\_5; Middle: OD on CARPK\_5\_Full; Right: our PFOD on CARPK\_5\_5 []{data-label="fig:vis_car"}](figs/vis_CARPK_ours_20161225_TPZ_00239.jpg "fig:"){width="\thirdsplitwidth\linewidth"}
\(d) pred=13, gt=138 \(e) pred=152, gt=138 \(f) pred=93, gt=138
![Visualization of detection/counting results on 3 CARPK test images. Each row corresponds to one test image. Left: OD on CARPK\_5\_5; Middle: OD on CARPK\_5\_Full; Right: our PFOD on CARPK\_5\_5 []{data-label="fig:vis_car"}](figs/vis_CARPK_base_20161225_TPZ_00435.jpg "fig:"){width="\thirdsplitwidth\linewidth"} ![Visualization of detection/counting results on 3 CARPK test images. Each row corresponds to one test image. Left: OD on CARPK\_5\_5; Middle: OD on CARPK\_5\_Full; Right: our PFOD on CARPK\_5\_5 []{data-label="fig:vis_car"}](figs/vis_CARPK_full_20161225_TPZ_00435.jpg "fig:"){width="\thirdsplitwidth\linewidth"} ![Visualization of detection/counting results on 3 CARPK test images. Each row corresponds to one test image. Left: OD on CARPK\_5\_5; Middle: OD on CARPK\_5\_Full; Right: our PFOD on CARPK\_5\_5 []{data-label="fig:vis_car"}](figs/vis_CARPK_ours_20161225_TPZ_00435.jpg "fig:"){width="\thirdsplitwidth\linewidth"}
\(g) pred=0, gt=114 \(h) pred=115, gt=114 \(i) pred=107, gt=114
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
OD on Drink35\_ALL\_5 OD on full Drink35 PFOD on Drink35\_ALL\_5
![Visualization of the detection/counting results on Drink35 test images. Each row corresponds to one test image. Left: OD on Drink35\_ALL\_5. Middle: OD on the full Drink35; Right: our PFOD on Drink35\_ALL\_5 []{data-label="fig:vis_drink"}](figs/vis_CocoBottle1024DrinkYW2_base_OldPicture_a_324_JPG.jpg "fig:"){width="\thirdsplitwidth\linewidth"} ![Visualization of the detection/counting results on Drink35 test images. Each row corresponds to one test image. Left: OD on Drink35\_ALL\_5. Middle: OD on the full Drink35; Right: our PFOD on Drink35\_ALL\_5 []{data-label="fig:vis_drink"}](figs/vis_CocoBottle1024DrinkYW2_full_OldPicture_a_324_JPG.jpg "fig:"){width="\thirdsplitwidth\linewidth"} ![Visualization of the detection/counting results on Drink35 test images. Each row corresponds to one test image. Left: OD on Drink35\_ALL\_5. Middle: OD on the full Drink35; Right: our PFOD on Drink35\_ALL\_5 []{data-label="fig:vis_drink"}](figs/vis_CocoBottle1024DrinkYW2_ours_OldPicture_a_324_JPG.jpg "fig:"){width="\thirdsplitwidth\linewidth"}
\(a) pred=0, gt=58 \(b) pred=49, gt=58 \(c) pred=36, gt=58
![Visualization of the detection/counting results on Drink35 test images. Each row corresponds to one test image. Left: OD on Drink35\_ALL\_5. Middle: OD on the full Drink35; Right: our PFOD on Drink35\_ALL\_5 []{data-label="fig:vis_drink"}](figs/vis_CocoBottle1024DrinkYW2_base_OldPicture_d_7_jpg.jpg "fig:"){width="\thirdsplitwidth\linewidth"} ![Visualization of the detection/counting results on Drink35 test images. Each row corresponds to one test image. Left: OD on Drink35\_ALL\_5. Middle: OD on the full Drink35; Right: our PFOD on Drink35\_ALL\_5 []{data-label="fig:vis_drink"}](figs/vis_CocoBottle1024DrinkYW2_full_OldPicture_d_7_jpg.jpg "fig:"){width="\thirdsplitwidth\linewidth"} ![Visualization of the detection/counting results on Drink35 test images. Each row corresponds to one test image. Left: OD on Drink35\_ALL\_5. Middle: OD on the full Drink35; Right: our PFOD on Drink35\_ALL\_5 []{data-label="fig:vis_drink"}](figs/vis_CocoBottle1024DrinkYW2_ours_OldPicture_d_7_jpg.jpg "fig:"){width="\thirdsplitwidth\linewidth"}
\(d) pred=14, gt=51 \(e) pred=46, gt=51 \(f) pred=45, gt=51
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Fig. \[fig:vis\_car\] visualizes the detection and counting results on three testing images based on the model with OD trained on CARPK\_5\_5, OD trained on CARPK\_5\_ALL, and the model with PFOD on CARPK\_5\_5. The yellow boxes are the correct bounding boxes while the red one is the incorrect one. A predicted box is correct if it has an IoU larger than 0.3 with one of the ground truth bounding boxes. The two numbers under the image are the number of instances predicted by the model and the ground-truth number of instances, respectively.
Results on Drink35
------------------
With the full annotations, OD can achieve 79.53% [email protected], 7.92 MAE and 11.50 RSME. This can be treated as the upper bound performance as we have used all the labels. If each training image is provided with only 5 labeled instance, the OD’s accuracy degrades to 14.16% mAP, 29.52 MAE and 38.62 RSME. In contrast, using PFOD, the accuracy could be jumped to 53.73%, 11.16 MAE and 13.10 RSME. Fig. \[fig:vis\_drink\] shows two example images detected/counted by the three approaches.
Conclusion
==========
We have studied the problem of object counting when there are only a few exemplar annotations available. The problem is more practical especially when the number of object instances is large. We formulate the problem as incompletely-supervised learning in the context of object detection. Since not all the bounding boxes are provided, we cannot simply treat other regions as background which will lead to severe overfitting and performance degradation. To address the problem, we have proposed a positiveness-focused object detector to progressively propagate the incomplete labels to more object instances. Our experimental results over two applications have demonstrated that this simple yet effective approach significantly boosts the accuracy with only a few manually annotations.
[^1]: The total number iterations could be greatly reduced as the number of training images is small. However, as the training time is not a concern in this work, we leave the training optimization as a future work.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: '[^1] [^2] This paper studies energy efficiency (EE) and average throughput maximization for cognitive radio systems in the presence of unslotted primary users. It is assumed that primary user activity follows an ON-OFF alternating renewal process. Secondary users first sense the channel possibly with errors in the form of miss detections and false alarms, and then start the data transmission only if no primary user activity is detected. The secondary user transmission is subject to constraints on collision duration ratio, which is defined as the ratio of average collision duration to transmission duration. In this setting, the optimal power control policy which maximizes the EE of the secondary users or maximizes the average throughput while satisfying a minimum required EE under average/peak transmit power and average interference power constraints are derived. Subsequently, low-complexity algorithms for jointly determining the optimal power level and frame duration are proposed. The impact of probabilities of detection and false alarm, transmit and interference power constraints on the EE, average throughput of the secondary users, optimal transmission power, and the collisions with primary user transmissions are evaluated. In addition, some important properties of the collision duration ratio are investigated. The tradeoff between the EE and average throughput under imperfect sensing decisions and different primary user traffic are further analyzed.'
author:
- '[^3]'
title: Spectral and Energy Efficiency in Cognitive Radio Systems with Unslotted Primary Users and Sensing Uncertainty
---
Cognitive radio, collision constraints, energy efficiency, interference power constraint, optimal frame duration, optimal power control, probability of detection, probability of false alarm, renewal processes, throughput, unslotted transmission.
Introduction
============
Cognitive radio is a promising innovative technology, leading to more efficient spectrum management and utilization. In cognitive radio systems, unlicensed users (i.e., cognitive or secondary users) are allowed to either continuously share the spectrum licensed with legacy users (i.e., primary users) without causing any significant interference, or periodically monitor the primary user activity via spectrum sensing and then perform transmissions according to sensing decisions.
Motivation
----------
Increasing global energy demand, consequent environmental concerns in terms of high levels of greenhouse gas emissions, high energy prices, and operating costs currently have triggered extensive research efforts in energy efficient communication systems. Optimal and efficient use of energy resources is paramount importance for cognitive radio systems in order to effectively utilize limited transmission power of battery-powered cognitive radios and support additional signal processing requirements such as spectrum sensing. Many existing works have focused on the design of optimal resource allocation schemes to maximize the spectrum efficiency (SE) and energy efficiency (EE) of the cognitive users. In particular, the authors in [@stotas] determined the optimal power control and sensing duration to maximize the ergodic capacity of cognitive radio systems operating in multiple narrowband channels under two different transmission schemes, namely sensing-based spectrum sharing and opportunistic spectrum access. The authors in [@akin] characterized the effective capacity of secondary users and the corresponding optimal power control policy in the presence of sensing errors.
The work in [@hu] mainly focused on the design of the optimal sensing duration and sensing decision threshold to maximize the weighted sum of the EE and SE. The authors in [@zhang] analyzed the optimal sensing duration that maximizes the EE of secondary user subject to a constraint on the detection probability. In [@xiong], the optimal subcarrier assignment and power allocation were proposed to maximize the worst case EE, (i.e., by considering the secondary user with the lowest EE) or to maximize the average EE of secondary users in an OFDM-based cognitive radio network. The work in [@park] studied the optimal power control scheme that maximizes the sum of EEs of the cognitive femto users for 5G communications. The authors in [@ramamonjison] developed energy-efficient power control algorithms for secondary users in a two-tier cellular network. In these works, it is assumed that primary users transmit in a time-slotted fashion, i.e., the activity of the primary users (e.g., active or inactive) remains the same during the entire frame duration.
In practice, primary and secondary user transmissions may not necessarily be synchronized. For instance, the primary user traffic can be bursty and may change its status during the transmission phase of the secondary users. In such cases, the assumption of time-slotted primary user transmission adopted in most studies (as also seen in the above-mentioned works) does no longer hold. In unslotted scenarios, it is assumed that ON-OFF periods of the primary user transmissions are random variables, following certain specific distributions. Exponential distribution is a commonly used model (see e.g., [@pei] – [@zarrini]). In particular, the authors in [@pei] determined the optimal frame duration that maximizes the throughput of the secondary users with perfect sensing decisions under collision constraints, assuming that the primary user activity changes only once within each frame. By adopting the same assumptions for the primary user activity as in the previous work, the authors in [@wang] mainly focused on the throughput of secondary users operating in the presence of multiple primary users with imperfect channel sensing results. In the same setting, the work in [@zarrini] mainly analyzed the optimal frame duration that maximizes the secondary user throughput. In [@guerrini], the exact secondary user throughput was determined and joint optimization of the sensing duration and frame period in the presence of sensing errors was performed by assuming that the primary user changes its status multiple times. The authors in [@macdonald] – [@dhakal] studied the impact of primary user activity on the sensing performance. The works in [@tang], [@pradhan] analyzed the sensing-throughput tradeoff for a secondary user in the presence of random arrivals and departures of the primary user and multiple transitions of primary user activity during sensing duration.
Main Contributions
------------------
The recent work in [@zhang2] analyzed general EE-SE relation for overlay, underlay and interweave cognitive radio systems. In particular, the authors introduced a general EE optimization problem and derived closed-form EE expressions for these systems. However, the authors did not consider collision constraints in the optimization problem, the transmission power was not instantaneously adapted according to channel conditions and the perfect sensing was assumed for interweave cognitive radio systems. In practice, cognitive radio systems, which employ spectrum sensing mechanisms to learn the channel occupancy by primary users, generally operate under sensing uncertainty arising due to multipath fading, shadowing and hidden node problem. Such kind of events can be incorporated into sensing uncertainty [@sharma], [@kaushik]. Therefore, motivated mainly by the fact that the optimal power control policies that maximize the EE or maximize the SE under constraints on EE (and hence address the tradeoff between EE and SE) have not been derived in the presence of unslotted primary users and imperfect sensing results, we have the following key contributions in this paper:
- We derive, in closed-form, the optimal power control policy that maximizes the EE of the secondary users operating with unslotted primary users subject to peak/average transmit power, average interference power and collision constraints in the presence of sensing errors. Hence, the power level has been adapted instantaneously according to the channel power gains of both the transmission link between the secondary transmitter and the secondary receiver and the interference link between the secondary transmitter and the primary receiver. We do not impose any limitations on the number of transitions of the primary user activity unlike the studies in [@pei] – [@zarrini] where the primary user activity changes only once. We assume that the primary user can change its status between ON and OFF states multiple times.
- In order to consider the EE and SE requirements of the secondary users jointly, we obtain the optimal power control scheme that maximizes the average througput of the secondary users while satisfying the minimum required EE in the presence of unslotted primary users.
- We propose low-complexity algorithms for jointly finding the optimal power control policy and frame duration.
- We analyze several important properties of the collision duration ratio and relations among sensing performance, secondary user throughput, EE, optimal frame duration and the resulting collisions with the primary user.
The rest of the paper is organized as follows. Section \[sec:system\_model\] introduces the primary user activity model, opportunistic spectrum access scheme and collision constraints. In Sections \[sec:EE\_power\] and \[sec:SE\_power\], optimal power control schemes that maximize the EE of the secondary users and the average throughput under a minimum EE constraint are derived, respectively. The algorithms for jointly determining the optimal power control and frame duration are also developed. Numerical results are provided and discussed in Section \[sec:num\_results\] before giving the main concluding remarks in Section \[sec:conclusion\].
System Model {#sec:system_model}
============
In this paper, we consider a cognitive radio system consisting of a pair of primary transmitter and receiver, and a pair of secondary transmitter and receiver. Secondary users opportunistically access the channel licensed to the primary users. In the following subsections, we describe the primary user activity model, opportunistic spectrum access policy of the secondary users, and the formulation of the collision constraint imposed for the protection of the primary users.
Primary User Activity Model
---------------------------
Differing from the majority of the studies (which assume that the primary users adopt a time-slotted transmission scheme), we consider a continuous, i.e., unslotted transmission structure as shown in Fig. \[fig:frame\_PU\_SU\] at the top of next page.
{width="80.00000%"}
We assume that the primary user activity follows a semi-Markov process with ON and OFF states. We have adopted exponential distribution model for the primary user traffic due to its popularity and existence in real systems. In particular, the recent works [@riihijarvi] – [@yin] have confirmed the exponential distribution for time-domain utilization of certain licensed channels through experimental simulation results. Also, the recent measurement study [@willkomm] has shown the exponential distribution of call arrival times in CDMA-based systems. Also, the exponentially distributed traffic model for the primary user is common assumption for cognitive radio systems in the literature. The parameters of the exponentially distributed primary user traffic can be found using blind and non-blind algorithms based on maximum likelihood estimation and adaptive sampling techniques as proposed in [@gabran]. In this model, the ON state indicates that the primary user is transmitting while the OFF state represents that the channel is not occupied by the primary user. Such a process is also known as an alternating renewal process. The durations of ON and OFF periods are independent of each other and are exponentially distributed with means $\lambda_0$ and $\lambda_1$, respectively, and therefore have probability density functions $$\begin{aligned}
f_{\text{ON}}(t)=\frac{1}{\lambda_0} {{\mathrm{e}}}^{-\frac{t}{\lambda_0}}, \text{ and }
f_{\text{OFF}}(t)=\frac{1}{\lambda_1} {{\mathrm{e}}}^{-\frac{t}{\lambda_1}}.\end{aligned}$$ Hence, the prior probabilities of channel being vacant or occupied by the primary user can be expressed, respectively, as $$\begin{aligned}
\label{eq:prior_probs}
\Pr\{{\mathcal{H}}_0\}=\frac{\lambda_0}{\lambda_0+\lambda_1}, \hspace{0.5cm} \Pr\{{\mathcal{H}}_1\}&=\frac{\lambda_1}{\lambda_0+\lambda_1}.\end{aligned}$$
Opportunistic Spectrum Access by the Secondary Users
----------------------------------------------------
Secondary users employ frames of duration $T_{{\rm{f}}}$. In the initial duration of $\tau$ seconds, secondary users perform channel sensing and monitor the primary user activity. Subsequently, data transmission starts in the remaining frame duration of $T_{{\rm{f}}}-\tau$ seconds only if the primary user activity is not detected, the event of which is denoted by ${\hat{\mathcal{H}}}_0$. In our analysis, we consider that sensing duration is much shorter than the mean duration of ON period of the primary user traffic, therefore it is reasonable to assume that the primary user activity is constant during the sensing duration. Spectrum sensing is modeled as a simple binary hypothesis testing problem with two hypotheses ${\mathcal{H}}_0$ and ${\mathcal{H}}_1$ corresponding to the absence and presence of the primary user signal, respectively. Many spectrum sensing methods have been proposed [@axell], and the corresponding sensing performance is characterized by two parameters, namely the probabilities of detection and false alarm, which are defined as $$\begin{aligned}
P_{{\rm{d}}}&= \Pr\{{\hat{\mathcal{H}}}_1|{\mathcal{H}}_1\}, \hspace{0.5cm} P_{{\rm{f}}}=\Pr\{{\hat{\mathcal{H}}}_1|{\mathcal{H}}_0\},\end{aligned}$$ where ${\hat{\mathcal{H}}}_1$ denotes the event that the primary user activity is detected. We note that any sensing method can be employed in the rest of the analysis since the results depend on the sensing performance only through the probabilities of detection and false alarm, and the sensing duration.
Collision Constraints
---------------------
We first describe the secondary users’ collisions with the primary users, which can lead to considerable performance degradation in the primary user communication. Subsequently, we impose a constraint on the ratio of the average collision duration to the transmission duration in order to protect the primary users. Depending on the true nature of the primary user activity at the beginning of the frame, collisions between the primary and secondary users can occur in the following two cases:
- *Case $1$*: The channel is not occupied by the primary user and is correctly detected as idle at the beginning of the frame. Even if the primary user is not actually transmitting initially, it is possible for the primary user to start data transmission at any time during the current frame, which results in a collision event. By conditioning on the correct detection of the initial absence of the primary user, the ratio of the average collision duration to data transmission duration, which is called the collision duration ratio, can be expressed as $$\begin{aligned}
\mathscr{P}_{c,0}=\frac{{\mathbb{E}}\{T_{c | {\mathcal{H}}_0, {\hat{\mathcal{H}}}_0}\}}{T_{{\rm{f}}}-\tau},\end{aligned}$$ where ${\mathbb{E}}\{\cdot\}$ denotes the expectation, and $T_{c | {\mathcal{H}}_0,{\hat{\mathcal{H}}}_0}$ is a random variable representing the collision duration between the secondary and primary users given that the primary user is inactive initially at the beginning of the frame (event ${\mathcal{H}}_0$) and the sensing decision is idle (event ${\hat{\mathcal{H}}}_0$). It is assumed that the primary user is in the OFF state at first and taking into account the possible multiple transitions between ON and OFF states. In this setting, the recursive expression of $T_{c | {\mathcal{H}}_0,{\hat{\mathcal{H}}}_0}$ is written as $$\begin{aligned}
\small
\begin{split}\label{eq:recursive_collision}
\hspace{-0.4cm}T_{c | {\mathcal{H}}_0,{\hat{\mathcal{H}}}_0} (t|X,\!Y) \!=\!\!\begin{cases}0 &t \leq X \\
T_{{\rm{f}}}-\tau-X \hspace{1cm}&X \! \leq \! t \! \leq \! X\!+\!Y\\
Y\!+\!T_{c | {\mathcal{H}}_0,{\hat{\mathcal{H}}}_0} (t\!-\!X\!-\!Y) &X+Y \leq t,
\end{cases}
\end{split}
\normalsize\end{aligned}$$ where $X$ denotes the first OFF state, which is exponentially distributed with mean $\lambda_1$ and and $Y$ represents the first ON state, which is exponentially distributed with mean $\lambda_0$. Using (\[eq:recursive\_collision\]), ${\mathbb{E}}\{T_{c | {\mathcal{H}}_0, {\hat{\mathcal{H}}}_0}\}$ is calculated in the following: $$\begin{aligned}
{\mathbb{E}}\{T_{c | {\mathcal{H}}_0, {\hat{\mathcal{H}}}_0}\} = \int \int_{xy} T_{c | {\mathcal{H}}_0,{\hat{\mathcal{H}}}_0} (t|X,Y) f_{XY}(x,y)dxdy.\end{aligned}$$ Then, the closed form expression for ${\mathbb{E}}\{T_{c |{\mathcal{H}}_0,{\hat{\mathcal{H}}}_0}\}$ can be found by using Laplace transform and following the same steps as in [@jiang Theorem 2]. Hence, $P_{c,0}$ is given by $$\begin{aligned}
\mathscr{P}_{c,0}=\Pr\{{\mathcal{H}}_1\}-\frac{\lambda_0 \Pr\{{\mathcal{H}}_1\}^2}{T_{{\rm{f}}}-\tau}\bigg(1-{{\mathrm{e}}}^{-\frac{T_{{\rm{f}}}-\tau}{\lambda_0 \Pr\{{\mathcal{H}}_1\}}}\bigg).\end{aligned}$$
- *Case $2$*: The primary user is actually present in the channel at the beginning of the frame, however the secondary user miss-detects the primary user activity, resulting in a collision right away due to sensing error. Multiple collisions can also occur if the primary user turns OFF and then back ON in a single frame once or multiple times. Similar to the first case, by conditioning on the miss detection event, the collision duration ratio can be found as $$\begin{aligned}
\mathscr{P}_{c,1}&=\frac{{\mathbb{E}}\{T_{c |{\mathcal{H}}_1,{\hat{\mathcal{H}}}_0}\}}{T_{{\rm{f}}}-\tau}
\\
&=\Pr\{{\mathcal{H}}_1\}\!+\!\frac{\lambda_1 \Pr\{{\mathcal{H}}_0\}^2}{T_{{\rm{f}}}-\tau}\bigg(1\!-\!{{\mathrm{e}}}^{-\frac{T_{{\rm{f}}}-\tau}{\lambda_0 \Pr\{{\mathcal{H}}_1\}}}\bigg)\end{aligned}$$ where $T_{c | {\mathcal{H}}_1, {\hat{\mathcal{H}}}_0}$ is a random variable describing the collision duration between the secondary and primary users given that the primary user is active at the beginning of the frame but sensing decision is incorrectly an idle channel.
Based on the above two cases, the collision duration ratio averaged over the true nature of the primary user activity given the idle sensing decision ${\hat{\mathcal{H}}}_0$ can be expressed as $$\begin{aligned}
\mathscr{P}_c= \Pr\{{\mathcal{H}}_0|{\hat{\mathcal{H}}}_0\}\mathscr{P}_{c,0}+\Pr\{{\mathcal{H}}_1|{\hat{\mathcal{H}}}_0\}\mathscr{P}_{c,1}\end{aligned}$$ where $\Pr\{{\mathcal{H}}_0|{\hat{\mathcal{H}}}_0\}$ and $\Pr\{{\mathcal{H}}_1|{\hat{\mathcal{H}}}_0\}$ denote the conditional probabilities of the primary user being active or inactive given the idle sensing decision, respectively, which can be written in terms of $P_{{\rm{d}}}$ and $P_{{\rm{f}}}$ as $$\begin{aligned}
\label{eq:PH0_hH0_exp}
\Pr\{{\mathcal{H}}_0|{\hat{\mathcal{H}}}_0\}&=\frac{\Pr\{{\mathcal{H}}_0\}(1-P_{{\rm{f}}})}{\Pr\{{\mathcal{H}}_0\}(1-P_{{\rm{f}}})+\Pr\{{\mathcal{H}}_1\}(1-P_{{\rm{d}}})},\end{aligned}$$ $$\begin{aligned}
\label{eq:PH1_hH0_exp}
\Pr\{{\mathcal{H}}_1|{\hat{\mathcal{H}}}_0\}&=\frac{\Pr\{{\mathcal{H}}_1\}(1-P_{{\rm{d}}})}{\Pr\{{\mathcal{H}}_0\}(1-P_{{\rm{f}}})+\Pr\{{\mathcal{H}}_1\}(1-P_{{\rm{d}}})}.\end{aligned}$$ In the following, we provide two key properties of $\mathscr{P}_c$.
\[prop:1\] The average collision duration ratio $\mathscr{P}_c$ under idle sensing decision has the following properties:
- It is an increasing function of the frame duration $T_{{\rm{f}}}$ for $P_{{\rm{f}}}<P_{{\rm{d}}}$ and a decreasing function for $P_{{\rm{f}}}>P_{{\rm{d}}}$.
- It takes values between $\Pr\{{\mathcal{H}}_1|{\hat{\mathcal{H}}}_0\}$ and $\Pr\{{\mathcal{H}}_1\}$.
*Proof:* See Appendix \[appendix1\].
![Average collision duration vs. frame duration $T_{{\rm{f}}}$ in the cases of imperfect sensing and perfect sensing.[]{data-label="fig:Pc_Tf_illustrate"}](Pc_T.eps){width="50.00000%"}
In Fig. \[fig:Pc\_Tf\_illustrate\], we illustrate $\mathscr{P}_c$ as a function of the frame duration $T_{{\rm{f}}}$ when $P_{{\rm{f}}}<P_{{\rm{d}}}$, i.e., correct detection probability is greater than the false alarm probability. Note that this is generally the desired case in practice in which the probability of detection is expected to be greater than $0.5$ and the probability of false alarm be less than $0.5$ for reliable sensing performance. In the figure, both imperfect sensing and perfect sensing are considered. For the case of imperfect sensing, $\mathscr{P}_c$ takes values between $\Pr\{{\mathcal{H}}_1|{\hat{\mathcal{H}}}_0\}$ and $\Pr\{{\mathcal{H}}_1\}$. For perfect sensing, $\mathscr{P}_c$ is first $0$ since $\Pr\{{\mathcal{H}}_1|{\hat{\mathcal{H}}}_0\}=0$, which corresponds to no collision event initially, as expected, and then $\mathscr{P}_c$ starts to increase with increasing $T_{{\rm{f}}}$ as it becomes more likely that the primary user initiates a transmission and secondary users collide with the primary users.
Energy-Efficient Optimal Power Control and Frame Duration {#sec:EE_power}
=========================================================
Average Transmit Power and Average Interference Power Constraints
-----------------------------------------------------------------
In this subsection, we determine the optimal power control policy and frame duration that maximize the EE of the secondary users in the presence of sensing uncertainty and unslotted primary users. We consider average transmit power and average interference power constraints. The latter constraint is imposed by the secondary transmitter to maintain a long-term power budget and hence long battery life by limiting the average transmit power by $P_{{\text{avg}}}$, which is the maximum average transmit power limit. The former constraint is imposed to satisfy the long-term QoS requirements of the primary users by limiting the average interference power by $Q_{{\text{avg}}}$, which represents the maximum average received interference power limit at the primary receiver. Regulatory bodies (e.g., Federal Communication Commissions (FCC)) sets an interference temperature limit, $Q_{{\text{avg}}}$ which provides the maximum amount of tolerable interference at the primary receiver for a given frequency band at a particular location. In this setting, the optimization problem can be formulated as $$\begin{aligned}
\label{eq:EE_max_Pavg_Qavg}
&\hspace{-2cm}\max_{
\substack{T_{{\rm{f}}}, P(g,h)}} \eta_{EE}=\frac{R_{{\text{avg}}}}{\big(\frac{T_{{\rm{f}}}-\tau}{T_{{\rm{f}}}}\big)P({\hat{\mathcal{H}}}_0){\mathbb{E}}_{g,h}\{P(g,h)\}+P_{c_r}} \\ \label{eq:coll_constraint}
\text{subject to}\hspace{0.4cm}&\mathscr{P}_c \leq \mathscr{P}_{c,max} \\ \label{eq:avg_transmit_power}
&\Big(\frac{T_{{\rm{f}}}-\tau}{T_{{\rm{f}}}}\Big)\Pr\{{\hat{\mathcal{H}}}_0\}{\mathbb{E}}_{g,h}\big\{P(g,h)\big\} \le P_{{\text{avg}}} \end{aligned}$$ $$\begin{aligned}
\label{eq:avg_inter_power}
&\Big(\frac{T_{{\rm{f}}}-\tau}{T_{{\rm{f}}}}\Big)\mathscr{P}_c\Pr\{{\hat{\mathcal{H}}}_0\}{\mathbb{E}}_{g,h}\big\{P(g,h)|g|^2\big\} \le Q_{{\text{avg}}} \\
& P(g,h) \geq 0 \\
& T_{{\rm{f}}} \geq \tau,\end{aligned}$$ where the EE in the objective function is defined as the ratio of average throughput of the secondary users to the total power consumption, including average transmission power and circuit power, denoted by $P_{c_r}$. Above, $P(g,h)$ denotes the instantaneous transmission power as a function of the channel fading coefficient $g$ of the interference link between the secondary transmitter and the primary receiver, and the channel fading coefficient $h$ of the transmission link between the secondary transmitter and the secondary receiver. The average transmission rate expression, $R_{{\text{avg}}}$ is given in (\[eq:Ravg\]),
$$\small
\begin{split} \label{eq:Ravg}
R_{{\text{avg}}}=\Big(\frac{T_{{\rm{f}}}-\tau}{T_{{\rm{f}}}}\Big){\mathbb{E}}_{g,h}\Big\{&\Pr\{{\mathcal{H}}_0\}(1-P_{{\rm{f}}})\Big[\log_2\Big(1+\frac{P(g,h) |h|^2}{N_0}\Big)(1-\mathscr{P}_{c,0})+\log_2\Big(1+\frac{P(g,h) |h|^2}{N_0+\sigma_s^2}\Big)\mathscr{P}_{c,0}\Big]\\ +&\Pr\{{\mathcal{H}}_1\}(1-P_{{\rm{d}}})\Big[\log_2\Big(1+\frac{P(g,h) |h|^2}{N_0}\Big)(1-\mathscr{P}_{c,1})+\log_2\Big(1+\frac{P(g,h) |h|^2}{N_0+\sigma_s^2}\Big)\mathscr{P}_{c,1}\Big]\!\Big\},
\end{split}
\normalsize$$
------------------------------------------------------------------------
where $N_0$ and $\sigma_s^2$ represent the variances of the additive Gaussian noise and primary user’s received faded signal, respectively. It is assumed that the secondary transmitter has perfect channel side information (CSI), i.e., perfectly knows the values of $g$ and $h$. While the assumption of perfect CSI is idealistic, channel knowledge can be obtained rather accurately if the mobility in the environment and channel variations are relatively slow. More specifically, secondary transmitter can acquire channel knowledge once the secondary receiver learns the channel and sends this information via an error-free feedback link. Also, the knowledge of the interference link, $g$, can be obtained through direct feedback from the primary receiver [@alouini], indirect feedback from a third party such as a band manager [@peha] or by periodically sensing the pilot symbols sent by the primary receiver under the assumption of channel reciprocity [@zhao].
In (\[eq:coll\_constraint\]), $\mathscr{P}_{c,max}$ denotes the maximum tolerable collision duration ratio, which needs to be greater than $P({\mathcal{H}}_1 | {\hat{\mathcal{H}}}_0)$ based on Proposition \[prop:1\] because, otherwise, the constraint cannot be satisfied. Since $\mathscr{P}_c$ is an increasing function of $T_{{\rm{f}}}$ when $P_{{\rm{f}}}<P_{{\rm{d}}}$, the collision constraint in (\[eq:coll\_constraint\]) provides an upper bound on the frame duration $T_{{\rm{f}}}$ as follows: $$\begin{aligned}
T_{{\rm{f}}} \leq \mathscr{P}_c^{-1}(\mathscr{P}_{c,max}).\end{aligned}$$ Above, $\mathscr{P}_c^{-1}(.)$ is the inverse function of $\mathscr{P}_c$.
As the frame duration increases, the secondary users have more time for data transmission, which leads to higher throughput, consequently higher EE. On the other hand, the primary user is more likely to become active with increasing transmission duration. In this case, the secondary users may collide with the primary transmission more frequently, which reduces the throughput, and hence EE. Therefore, there indeed exists an optimal frame duration that achieves the best tradeoff between the EE of the secondary users and collisions with the primary users. It can be easily verified that the EE is not a concave function of the frame duration $T_{{\rm{f}}}$ since the second derivative of the EE with respect $T_{{\rm{f}}}$ is less than, greater than or equal to zero depending on the values of the sensing parameters and prior probabilities of primary user being active and idle. However, the optimal frame duration which maximizes the EE can easily be obtained using a one-dimensional exhaustive search within the interval $(\tau, \mathscr{P}_c^{-1}(\mathscr{P}_{c,max})]$. For a given frame duration, we derive the optimal power control policy in the following result.
\[teo:1\] The optimal power control that maximizes the EE of the secondary users operating subject to the average transmit power constraint in (\[eq:avg\_transmit\_power\]) and average interference power constraint in (\[eq:avg\_inter\_power\]) in the presence of sensing errors and unslotted primary users is given by $$\begin{aligned}
\label{eq:opt_power}
&\hspace{1.8cm}P_{\text{opt}}(g,h) =\bigg[\frac{A_0+\sqrt{\Delta_0}}{2}\bigg]^{+} \\
\text{where } &A_0=\frac{\log_2(e)}{(\alpha+\lambda)+\nu\mathscr{P}_c|g|^2}-\frac{2N_0+\sigma_s^2}{|h|^2} \\
&\hspace{-1cm}\Delta_0\!=\!A_0^2\!-\!\frac{4}{|h|^2}\bigg(\!\frac{N_0(N_0\!+\!\sigma_s^2)}{|h|^2}\!-\frac{\log_2(e)(N_0+(1-\mathscr{P}_c) \sigma_s^2)}{(\alpha+\lambda)+\nu\mathscr{P}_c|g|^2}\!\bigg).\end{aligned}$$ Above, $(x)^+=\max\{0,x\}$ and $\alpha$ is a nonnegative parameter. Morever, $\lambda$ and $\nu$ are the Lagrange multipliers which can be jointly obtained by inserting the above optimal power control into the constraints in (\[eq:avg\_transmit\_power\]) and (\[eq:avg\_inter\_power\]), respectively.
*Proof:* See Appendix \[appendix2\].
Initialize $P_{{\rm{d}}}=\mathscr{P}_{{\rm{d}},init}$, $P_{{\rm{f}}}=\mathscr{P}_{{\rm{f}},init}$, $\epsilon > 0$, $\delta > 0$, $t > 0$, $\alpha^{(0)}=\alpha_{\text{init}}$, $\lambda^{(0)}=\lambda_{\text{init}}$, $\nu^{(0)}=\nu_{\text{init}}, \mathscr{P}_{c,max}=\mathscr{P}_{c,max,\text{init}}$ $T_{{\rm{f}},\text{opt}}=0, P_{\text{opt}}(g,h)=0$ $k \leftarrow 0$ $n \leftarrow 0$ calculate $P_{\text{opt}}(g,h)$ using (\[eq:opt\_power\]); update $\lambda$ and $\nu$ using subgradient method as follows: $n \leftarrow n+1$ $\alpha^{(k+1)}=\frac{R_{{\text{avg}}}}{\big(\frac{T_{{\rm{f}}}-\tau}{T_{{\rm{f}}}}\big)\Pr\{{\hat{\mathcal{H}}}_0\}{\mathbb{E}}_{g,h}\{P_{\text{opt}}(g,h)\}+P_{c_r}}$ $k \leftarrow k+1$ $\eta_{EE}= \alpha^{(k)}$ $T_{{\rm{f}},opt}={\operatornamewithlimits{arg\,max}}\eta_{EE}$ by bisection search $P^*_{\text{opt}}(g,h)=[P_{\text{opt}}(g,h)]_{T_{{\rm{f}}}=T_{{\rm{f}},opt}}$
The values of $\lambda$ and $\nu$ can be obtained numerically via the projected subgradient method. In this method, $\lambda$ and $\nu$ are updated iteratively in the direction of a negative subgradient of the Lagrangian function $\mathcal{L}(P(g,h),\lambda,\nu,\alpha)$ (given in (\[eq:lagrangian\_function\]) in Appendix \[appendix2\]) until convergence as follows:
$$\begin{aligned}
\small
\begin{split} \label{eq:subgradient}
&\lambda^{(n+1)}=\Big(\lambda^{(n)}-t\Big(P_{{\text{avg}}}-\Big(\frac{T_{{\rm{f}}}-\tau}{T_{{\rm{f}}}}\Big)\Pr\{{\hat{\mathcal{H}}}_0\}{\mathbb{E}}_{g,h}\big\{P(g,h)\big\}\Big)\Big)^+
\end{split}
\normalsize\end{aligned}$$
$$\begin{aligned}
\small
\begin{split} \label{eq:subgradient_Qavg}
\nu^{(n+1)}\!=\!\!\Big(\!\nu^{(n)}\!-\!t\Big(\!Q_{{\text{avg}}}\!-\!\!\Big(\frac{T_{{\rm{f}}}\!-\!\tau}{T_{{\rm{f}}}}\Big)\mathscr{P}_c\Pr\{{\hat{\mathcal{H}}}_0\}{\mathbb{E}}_{g,h}\big\{\!P(g,h)|g|^2\!\big\} \!\Big)\!\Big)^{\!+},
\end{split}
\normalsize\end{aligned}$$
where $n$ is the iteration index and $t$ is the step size. For a constant $t$, $\lambda$ and $\nu $ are shown to converge to the optimal values within a small range [@boyd2].
In Table \[table:algorithm1\], we provide our low-complexity algorithm for jointly finding the optimal power control policy and frame duration, which maximize the EE of the secondary users in the presence of unslotted primary users and imperfect sensing decisions. In the table, for a given value of $\alpha$ and frame duration $T_{{\rm{f}}}$, the optimal power control is obtained when $F(\alpha) \leq \epsilon$ is satisfied, where $F(\alpha)$ is defined in (\[eq:F\_alpha\]) in Appendix \[appendix2\] and $\alpha$ is a nonnegative parameter. The solution is optimal if $F(\alpha)=0$, otherwise $\epsilon$-optimal solution is obtained.
The proposed power control algorithm consists of two nested loops. In the outer loop, Dinkelbach’s method iteratively solves the energy efficiency maximization problem by solving a sequence of parameterized concave problems. It is shown that Dinkelbach’s method has a super-linear convergence rate [@schaible], and hence the sequence converges to an optimal solution in a small number of iterations. In the inner loop, Lagrange multipliers are updated using the subgradient method, which involves the computation of subgradient and simple projection operations. The subgradient method is widely used to find Lagrange multipliers due to its simplicity, easy implementation, the speed for computing a direction, and the global convergence property [@zhao]. In addition, the optimal frame duration is obtained by bisection search, which is the simplest root finding method. In particular, the bisection method halves the search interval at each iteration and its time complexity is logarithmic. Hence, the proposed algorithm is computationally efficient.
The optimal power control policy in (\[eq:opt\_power\]) is a decreasing function of average collision duration ratio, $\mathscr{P}_c$. In particular, when the secondary users have higher $\mathscr{P}_c$, less power is allocated in order to limit the interference inflicted on the primary user transmission. Also, the proposed power control policy depends on sensing performance through $\mathscr{P}_c$, which is a function of detection and false alarm probabilities, $P_{{\rm{d}}}$ and $P_{{\rm{f}}}$, respectively.
By setting $\alpha=0$ in (\[eq:EE\_concave\]) in Appendix \[appendix2\], the optimization problem becomes $$\begin{aligned}
\max_{
\substack{P(g,h) \geq 0}} R_{{\text{avg}}}\end{aligned}$$ which corresponds to the throughput maximization problem. Therefore, solving the above optimization problem or equivalently inserting $\alpha=0$ into the proposed scheme in (\[eq:opt\_power\]), we can readily obtain the optimal power control strategy that maximizes the average throughput of secondary users in the presence of unslotted primary users.
By inserting $\alpha=0$, $\mathscr{P}_{c,0}=0$ and $\mathscr{P}_{c,1}=1$ into (\[eq:opt\_power\]), we can see that the optimal power control scheme has a similar structure to the scheme that maximizes the throughput of secondary users operating over a single frequency band given in [@stotas eq. (36)], where it is assumed that the primary users do not change their activity during the entire frame duration of the secondary users, i.e., a time-slotted transmission scheme. Hence, our results can be specialized to the time-slotted case by setting $\mathscr{P}_{c,0}$ and $\mathscr{P}_{c,1}$ equal to $0$ and $1$, respectively.
Peak Transmit Power and Average Interference Power Constraints
--------------------------------------------------------------
In this subsection, we consider that the secondary user transmission is subject to peak transmit power and average interference power constraints. Under these assumptions, the optimization problem can be expressed as $$\begin{aligned}
\label{eq:EE_max_Ppk_Qavg}
&\hspace{-1.6cm}\max_{
\substack{T_{{\rm{f}}}, P(g,h)}} \eta_{EE}=\frac{R_{{\text{avg}}}}{\big(\frac{T_{{\rm{f}}}-\tau}{T_{{\rm{f}}}}\big)\Pr\{{\hat{\mathcal{H}}}_0\}{\mathbb{E}}_{g,h}\{P(g,h)\}+P_{c_r}} \\ \label{eq:coll_constraint2}
\text{subject to }&\mathscr{P}_c \leq \mathscr{P}_{c,max} \\ \label{eq:peak_transmit_power}
&P(g,h) \le P_{{\rm{pk}}} \\ \label{eq:avg_inter_power1}
&\Big(\frac{T_{{\rm{f}}}-\tau}{T_{{\rm{f}}}}\Big)\mathscr{P}_c\Pr\{{\hat{\mathcal{H}}}_0\}{\mathbb{E}}_{g,h}\big\{P(g,h)|g|^2\big\} \le Q_{{\text{avg}}} \\
& P(g,h) \geq 0 \\ \label{eq:T_lower}
& T_{{\rm{f}}} \geq \tau,\end{aligned}$$ where $P_{{\rm{pk}}}$ represents the peak transmit power limit at the secondary transmitter. Subsequently, the optimal power control policy is determined in the following result.
\[teo:2\] For a given frame duration $T_{{\rm{f}}}$, the optimal power control scheme subject to the constraints in (\[eq:peak\_transmit\_power\]) – (\[eq:T\_lower\]) is obtained as $$\begin{aligned}
\label{eq:opt_power_Ppk}
&\hspace{1cm}P_{\text{opt}}(g,h) =\min\bigg\{\bigg[\frac{A_1+\sqrt{\Delta_1}}{2}\bigg]^{+},P_{{\rm{pk}}} \bigg\} \\
\text{where } &A_1=\frac{\log_2(e)}{\alpha+\mu\mathscr{P}_c|g|^2}-\frac{2N_0+\sigma_s^2}{|h|^2} \\
&\hspace{-1cm}\Delta_1\!=\!A_1^2\!-\!\frac{4}{|h|^2}\bigg(\!\frac{N_0(N_0\!+\!\sigma_s^2)}{|h|^2}\!-\frac{\log_2(e)(N_0+(1\!-\!\mathscr{P}_c) \sigma_s^2)}{\alpha+\mu\mathscr{P}_c|g|^2}\!\bigg).\end{aligned}$$ Above, $\mu$ is the Lagrange multiplier associated with the average interference power constraint in (\[eq:avg\_inter\_power1\]).
Since we follow similar steps as in the proof of Theorem \[teo:1\], the proof is omitted for the sake of brevity.
Different from the optimal power control strategy in Theorem \[teo:1\], the instantaneous transmission power level in (\[eq:opt\_power\_Ppk\]) is limited by $P_{{\rm{pk}}}$ due to the peak transmit power constraint, which imposes stricter limitations than the average transmit power constraint.
Setting $\alpha=0$ in (\[eq:opt\_power\_Ppk\]), we obtain the optimal power control strategy which maximizes the throughput of secondary users with unslotted primary users, which is in agreement with the result derived in [@ozcan].
Algorithm 1 can be easily modified to maximize the EE of the secondary users under peak transmit power and average interference constraints by calculating the power level, $P_{\text{opt}}(g,h)$ through the expression in (\[eq:opt\_power\_Ppk\]) and updating the Lagrange multiplier $\mu$ similarly as in (\[eq:subgradient\_Qavg\]).
Spectrally-Efficient Optimal Power Control and Frame Duration with a Minimum EE Constraint {#sec:SE_power}
==========================================================================================
Average Transmit Power and Average Interference Power Constraints
-----------------------------------------------------------------
In this subsection, we analyze the EE-SE tradeoff by formulating the optimal power control problem to maximize the average throughput of the secondary users subject to a minimum EE constraint, and average transmit power, average interference power and collision constraints. The optimization problem is formulated as follows: $$\begin{aligned}
\label{eq:throughput_max_Pavg_Qavg}
&\hspace{0.8cm}\max_{
\substack{T_{{\rm{f}}} \geq \tau,P(g,h) \geq 0}} R_{{\text{avg}}}\\ \label{eq:coll_constraint3}
\text{subject to } &\mathscr{P}_c \leq \mathscr{P}_{c,max} \\
\label{eq:EEmin_Pavg_Qavg} &\frac{R_{{\text{avg}}}}{\big(\frac{T_{{\rm{f}}}-\tau}{T_{{\rm{f}}}}\big)\Pr\{{\hat{\mathcal{H}}}_0\}{\mathbb{E}}_{g,h}\{P(g,h)\}+P_{c_r}} \geq \text{EE}_{\text{min}}\\ \label{eq:avg_inter_power3}
&\Big(\frac{T_{{\rm{f}}}-\tau}{T_{{\rm{f}}}}\Big)\mathscr{P}_c\Pr\{{\hat{\mathcal{H}}}_0\}{\mathbb{E}}_{g,h}\big\{P(g,h)|g|^2\big\} \le Q_{{\text{avg}}} \\
\label{eq:avg_transmit_power3}
&\Big(\frac{T_{{\rm{f}}}-\tau}{T_{{\rm{f}}}}\Big)\Pr\{{\hat{\mathcal{H}}}_0\}{\mathbb{E}}_{g,h}\big\{P(g,h)\big\} \le P_{{\text{avg}}},\end{aligned}$$ where $\text{EE}_{\text{min}}$ denotes the minimum required EE. The optimal power control is determined in two steps. In the first step, we determine the average power level at which the required minimum EE is achieved. In the second step, we optimally allocate the transmission power in order to maximize the average throughput of the secondary users by combining the power level obtained in the first step under average transmit power and average interference power constraints. In this regard, we first provide the following result.
\[prop:2\] For a given frame duration $T_{{\rm{f}}}$, the average power level that satisfies the minimum required EE can be obtained as $$\begin{aligned}
\label{eq:Pavg_EEmin}
P_{{\text{avg}}}^*=\Big(\frac{T_{{\rm{f}}}-\tau}{T_{{\rm{f}}}}\Big)\Pr\{{\hat{\mathcal{H}}}_0\}{\mathbb{E}}_{g,h}\big\{P^*(g,h)\big\}\end{aligned}$$ and $P^*(g,h)$ is given by $$\begin{aligned}
\label{eq:P0_EEmin}
&\hspace{2.4cm}P^*(g,h) =\bigg[\frac{A_3+\sqrt{\Delta_3}}{2}\bigg]^{+},\\ \label{eq:A3}
\text{where } &A_3=\frac{(1+\eta)\log_2(e)}{\eta \text{EE}_{\text{min}}}-\frac{2N_0+\sigma_s^2}{|h|^2} \\ \nonumber
&\hspace{-0.6cm}\Delta_3\!=\!A_3^2\!-\!\frac{4}{|h|^2}\bigg(\!\frac{N_0(N_0\!+\!\sigma_s^2)}{|h|^2}\! \\&\label{eq:delta3}-\frac{(1\!+\!\eta)\log_2(e)(N_0\!+\!(1\!-\!\Pr\{{\hat{\mathcal{H}}}_0\}\mathscr{P}_c) \sigma_s^2)}{\eta \text{EE}_{\text{min}}}\!\bigg).\end{aligned}$$ The optimal value of $\eta$ can be found by solving the equation below: $$\begin{aligned}
\label{eq:eta_EEmin}
R_{{\text{avg}}}+ \eta \bigg(\Big(\frac{T_{{\rm{f}}}-\tau}{T_{{\rm{f}}}}\Big)\Pr\{{\hat{\mathcal{H}}}_0\}{\mathbb{E}}\big\{P^*(g,h)\big\}\bigg)=0.\end{aligned}$$
*Proof:* See Appendix \[appendix3\].
Using the results in Proposition \[prop:2\], the throughput optimization problem subject to the minimum EE constraint is equivalent to the throughput maximization under an average power constraint with the power limit, $P^*_{{\text{avg}}}$, which achieves the minimum required EE. By combining this power limit with the average transmit power constraint in (\[eq:avg\_transmit\_power3\]), we define the operating average transmission power as follows: $$\begin{aligned}
\label{eq:Pavg_optimal_minEE}
P_{op}=\begin{cases} P^*_{{\text{avg}}} &\text{if } P_{{\text{avg}}} \geq P^*_{{\text{avg}}} \\ P_{{\text{avg}}} &\text{if } P_{{\text{avg}}} < P^*_{{\text{avg}}} \\& \text{and} \left. \eta_{EE} \right |_{\text{s.t.} (\ref{eq:avg_transmit_power}) \text{ and } (\ref{eq:avg_inter_power})} \geq \text{EE}_{\text{min}} \\
0 &\text{if } P_{{\text{avg}}} < P^*_{{\text{avg}}} \\& \text{and} \left. \eta_{EE} \right |_{\text{s.t.} (\ref{eq:avg_transmit_power}) \text{ and } (\ref{eq:avg_inter_power})} < \text{EE}_{\text{min}}
\end{cases}\end{aligned}$$
{width="100.00000%"}
The operating average transmission power is determined according to three cases as illustrated in Fig. \[fig:Op\_power\].\
[***Case (i)***]{}: When $P_{{\text{avg}}}$ is larger than $P^*_{{\text{avg}}}$, average transmit power constraint $P_{{\text{avg}}}$ is loose. Since operating at average transmission power level greater than $P^*_{{\text{avg}}}$ violates the minimum required EE constraint, we set $P_{op}=P^*_{{\text{avg}}}$ and the optimal transmission power control policy is obtained by satisfying $P^*_{{\text{avg}}}$ with equality. This case is illustrated in Fig. \[fig:Op\_power\].(i).\
[***Case (ii)***]{}: As shown in Fig. \[fig:Op\_power\].(ii), when $P_{{\text{avg}}}$ is less than $P^*_{{\text{avg}}}$ and the EE achieved at $P_{{\text{avg}}}$ is greater than $\text{EE}_{\text{min}}$, average transmit power constraint $P_{{\text{avg}}}$ is dominant. Since average transmission power is limited by $P_{{\text{avg}}}$, we set $P_{op}=P_{{\text{avg}}}$ and the optimal transmission power control policy is found when $P_{{\text{avg}}}$ is satisfied with equality.\
[***Case (iii)***]{}: As demonstrated in Fig. \[fig:Op\_power\].(iii), when $P_{{\text{avg}}} < P^*_{{\text{avg}}}$ and the EE achieved at $P_{{\text{avg}}}$ is less than $\text{EE}_{\text{min}}$, there is no feasible solution, and hence we set $P_{op}=0$
In the following result, we identify the optimal power control strategy.
\[teo:3\] For a given frame duration $T_{{\rm{f}}}$, if $P_{{\text{avg}}} < P^*_{{\text{avg}}}$ and the maximum EE subject to the constraints in (\[eq:avg\_transmit\_power\]) and (\[eq:avg\_inter\_power\]) is less than $ \text{EE}_{\text{min}}$, the power level is set to zero, i.e., $P^*_0(g,h) =0$, otherwise we allocate the power according to $$\begin{aligned}
\label{eq:opt_power4}
&\hspace{2cm}P_{\text{opt}}(g,h) =\bigg[\frac{A_4+\sqrt{\Delta_4}}{2}\bigg]^{+}\end{aligned}$$
$$\begin{aligned}
\label{eq:A4}
\text{where } &A_4=\frac{\log_2(e)}{\vartheta+\varphi\mathscr{P}_c|g|^2}-\frac{2N_0+\sigma_s^2}{|h|^2} \\
\label{eq:delta4}
&\hspace{-0.9cm}\Delta_4\!=\!A_4^2\!-\!\frac{4}{|h|^2}\bigg(\!\frac{N_0(N_0\!+\!\sigma_s^2)}{|h|^2}\!-\frac{\log_2(e)(N_0\!+\!(1\!-\!\mathscr{P}_c) \sigma_s^2)}{\vartheta+\varphi\mathscr{P}_c|g|^2}\!\bigg).\end{aligned}$$
Above, $\vartheta$ and $\varphi$ are the Lagrange multipliers associated with the average transmit power constraint, $\min(P_{{\text{avg}}},P^*_{{\text{avg}}})$ and interference power constraint in (\[eq:avg\_inter\_power3\]), respectively.
*Proof:* See Appendix \[appendix4\].
In Table \[table:algorithm2\], we provide the details of an algorithm for jointly finding the optimal power control policy and frame duration that maximize the average throughput of the secondary users subject to constraints on collision duration ratio, the minimum required EE, average transmit power and interference power in the presence of unslotted primary users.
For a given $P_{{\rm{d}}}$, $P_{{\rm{f}}}$, $\mathscr{P}_{c,max}$, $\text{EE}_{\text{min}}$, initialize $\eta^{(0)}=\eta_{\text{init}}$, $\vartheta^{(0)}=\vartheta_{\text{init}}$, $\varphi^{(0)}=\varphi_{\text{init}}$ $T_{{\rm{f}},\text{opt}}=0, P_{\text{opt}}(g,h)=0$ Find the optimal value of $\eta$ that solves (\[eq:eta\_EEmin\]) by using a root-finding function. Calculate $P_{{\text{avg}}}^*=\Big(\frac{T_{{\rm{f}}}-\tau}{T_{{\rm{f}}}}\Big)P({\hat{\mathcal{H}}}_0){\mathbb{E}}_{g,h}\big\{P^*(g,h)\big\}$ where $P^*(g,h)$ is given in (\[eq:P0\_EEmin\]). $P_{\text{opt}}(g,h)=0$ $P_{op}=\min(P_{{\text{avg}}},P^*_{{\text{avg}}})$ and calculate $P_{\text{opt}}(g,h)$ using (\[eq:opt\_power4\]) Update $\vartheta$ and $\varphi$ using subgradient method Calculate $R_{{\text{avg}}}$ using (\[eq:Ravg\]) $T_{{\rm{f}},opt}={\operatornamewithlimits{arg\,max}}R_{{\text{avg}}}$ by bisection search $P^*_{\text{opt}}(g,h)=[P_{\text{opt}}(g,h)]_{T_{{\rm{f}}}=T_{{\rm{f}},opt}}$
Peak Transmit Power and Average Interference Power Constraints
--------------------------------------------------------------
In this subsection, we consider that the objective function in (\[eq:throughput\_max\_Pavg\_Qavg\]) is subject to the constraints in (\[eq:coll\_constraint3\])- (\[eq:avg\_inter\_power3\]) and the peak transmit power constraint $P(g,h) < P_{pk}$ instead of the average transmit power constraint. In this case, we derive the optimal power control as follows:
\[teo:4\]The average power level at which the minimum required EE is achieved can be determined by inserting the power control given below in (\[eq:P0\_EEmin\_Ppk\]) into (\[eq:Pavg\_EEmin\]): $$\begin{aligned}
\label{eq:P0_EEmin_Ppk}
P^*(g,h) =\bigg\{\bigg[\frac{A_3+\sqrt{\Delta_3}}{2}\bigg]^{+},P_{{\rm{pk}}}\bigg\},\end{aligned}$$ where $A_3$ and $\Delta_3$ are given in (\[eq:A3\]) and (\[eq:delta3\]), respectively. If the maximum EE at $P_{{\rm{pk}}}$ is less than $ \text{EE}_{\text{min}}$, the power level is set to zero, i.e., $P_{\text{opt}}(g,h) =0$, otherwise the optimal power control can be found as $$\begin{aligned}
\label{eq:opt_power5}
P_{\text{opt}}(g,h) =\min \bigg\{ \bigg[\frac{A_4+\sqrt{\Delta_4}}{2}\bigg]^{+},P_{{\rm{pk}}} \bigg\}\end{aligned}$$ Above, $A_4$ and $\Delta_4$ are given in (\[eq:A4\]) and (\[eq:delta4\]), respectively.
*Proof:* We follow similar steps as in the proof of Proposition \[prop:2\] and Theorem \[teo:3\] with peak transmit power constraint in consideration. Therefore, the power levels are limited by $P_{{\rm{pk}}}$ in this case. $\square$
Numerical Results {#sec:num_results}
=================
In this section, we present and discuss the numerical results for the optimal power control and frame duration, which maximize the EE or throughput of the secondary users, and analyze the resulting collisions with the unslotted primary users. Unless mentioned explicitly, the noise variance is $N_0=0.01$ and the variance of primary user’s received signal is $\sigma_s^2=0.1$. Also, the mean values of the durations of ON and OFF periods, denoted by $\lambda_0$ and $\lambda_1$, are set to $650$ ms and $352$ ms, respectively so that $\Pr\{{\mathcal{H}}_0\} \approx 0.65$, corresponding to the setting in the voice over Internet protocol (VoIP) traffic. The step size $t$ and tolerance $\epsilon$ are chosen as $0.1$ and $10^{-5}$, respectively. The circuit power $P_{c_r}$ is set to $1$. We consider a Rayleigh fading environment, and hence the channel power gains of the transmission link and interference link are exponentially distributed with unit mean.
It is assumed that the secondary users employ energy detection scheme for spectrum sensing, and hence the probabilities of detection and false alarm are expressed, respectively as [@tandra], [@liang] $$\begin{aligned}
P_{{\rm{d}}}&=\mathcal{Q}\bigg(\Big(\frac{\varepsilon}{N_0}-\frac{\sigma_s^2}{N_0}-1\Big)\sqrt{\frac{\tau f_s}{2\frac{\sigma_s^2}{N_0}+1}}\bigg)\\
P_{{\rm{f}}}&=\mathcal{Q}\bigg(\Big(\frac{\varepsilon}{N_0}-1\Big)\sqrt{\tau f_s}\bigg),\end{aligned}$$ where $\mathcal{Q}(x)=\int_{x}^{\infty}\frac{1}{\sqrt{2\pi}}{{\mathrm{e}}}^{-t^2/2}dt$ is the Gaussian $Q$-function, $\varepsilon$ represents the decision threshold and $f_s$ denotes the sampling frequency. The decision threshold $\varepsilon$ can be chosen to satisfy the target detection and false alarm probabilities, denoted by $\bar{P}_{{\rm{d}}}$ and $\bar{P}_{{\rm{f}}}$, respectively and the resulting sensing duration $\tau$ is expressed as $$\begin{aligned}
\tau=\frac{1}{f_s}\bigg(\frac{\mathcal{Q}^{-1}(\bar{P}_{{\rm{f}}})-\sqrt{2\sigma_s^2+1}\mathcal{Q}^{-1}(\bar{P}_{{\rm{d}}})}{\sigma_s^2}\bigg)^2.\end{aligned}$$ In the numerical computations, $f_s$ is set to $100$ kHz.
[0.32]{} {width="\textwidth"}
[0.32]{} {width="\textwidth"}
[0.32]{} {width="\textwidth"}
[0.32]{} {width="\textwidth"}
[0.32]{} {width="\textwidth"}
[0.32]{} {width="\textwidth"}
In Fig. \[fig:R\_Tf\], we plot the average throughput of the secondary users, $R_{{\text{avg}}}$, as a function of the frame duration $T_{{\rm{f}}}$ for $P_{{\rm{pk}}}=10$ dB and different average power constraints, namely $Q_{{\text{avg}}}=-15$ dB, $Q_{{\text{avg}}}=-10$ dB and $Q_{{\text{avg}}}=0$ dB. We consider target detection probability $\bar{P}_{{\rm{d}}}=0.9$ and false alarm probability $\bar{P}_{{\rm{f}}}=0.1$, and hence $\tau$ becomes $7.21$ ms. Transmission power level is chosen according to $\min\Big\{P_{{\rm{pk}}},\Big(\frac{T_{{\rm{f}}}}{T_{{\rm{f}}}-\tau}\Big)\frac{Q_{{\text{avg}}}}{\Pr\{{\mathcal{H}}_0,{\hat{\mathcal{H}}}_0\}\mathscr{P}_{c,0}+\Pr\{{\mathcal{H}}_1,{\hat{\mathcal{H}}}_0\}\mathscr{P}_{c,1}}\Big\}$. In this setting, average throughput formulation in (\[eq:Ravg\]) is also verified through Monte Carlo simulations with $100000$ runs. It is seen that $R_{{\text{avg}}}$ initially increases with increasing transmission duration. After reaching a peak value, $R_{{\text{avg}}}$ begins to diminish as the secondary user starts colliding with primary user transmissions more frequently, degrading the performance. It is also observed that as the interference power constraint gets looser, i.e., as $Q_{{\text{avg}}}$ changes from $-15$ to $0$ dB, higher throughput is achieved since secondary user transmits at higher power levels. As illustrated in the figure, $R_{{\text{avg}}}$ is not a concave function of $T_{{\rm{f}}}$. However, $R_{{\text{avg}}}$ curves are seen to exhibit a quasiconcave property and there exists an optimal frame duration that maximizes the throughput.
In Fig. \[fig:EE\_Topt\_Pd\], we display the maximum EE $\eta_{EE}$, average collision duration ratio $\mathscr{P}_c$, and the optimal frame duration $T_{{\rm{f}},opt}$ as functions of the probability of detection $P_{{\rm{d}}}$. We set the maximum collision limit as $\mathscr{P}_{c,max}=0.2$. It is assumed that the average transmit power constraint is $P_{{\text{avg}}}=10$ dB and average interference power constraint is $Q_{{\text{avg}}}=-20$ dB. We consider both the transmission with the optimal power control policy and constant-power transmission. For the constant power case, power is not adaptively varied with respect to the channel power gains of the transmission link and interference link. On the other hand, optimal power control derived in (\[eq:opt\_power\]) is a function of both $h$ and $g$. As $\bar{P}_{{\rm{d}}}$ increases while keeping $\bar{P}_{{\rm{f}}}$ fixed at $0.1$ and hence sensing performance improves, secondary user has a higher EE. In addition, collision duration ratio decreases with increasing detection probability in both cases of optimal power control and constant power. For $P_{{\rm{d}}}$ values less than $0.585$, collision constraint is not satisfied for any value of the frame duration $T_{{\rm{f}}}$, and therefore the secondary user throughput is $0$. When $P_{{\rm{d}}}$ takes values between $0.585$ and $0.6$, maximum EE is achieved at the maximum collision limit, i.e, when $\mathscr{P}_c=0.2$. It is also observed that the optimal power control leads transmissons with a larger frame duration while satisfying the maximum allowed collision limit and achieving a higher EE compared to constant-power transmissions.
In Fig. \[fig:EE\_Topt\_Pf\], we plot the maximum EE $\eta_{EE}$, average collision duration ratio $\mathscr{P}_c$, and the optimal frame duration $T_{{\rm{f}},opt}$ as functions of the probability of false alarm $P_{{\rm{f}}}$. We consider the same setting as in the previous figure. It is seen that as $P_{{\rm{f}}}$ increases while keeping $P_{{\rm{d}}}$ fixed at $0.9$, sensing performance degrades and secondary users experience more false alarm events, which leads to more collisions with the primary user transmission. Therefore, secondary user has a lower EE in both cases of optimal power control and constant power. We also notice in Fig. \[fig:EE\_Topt\_Pf\](a) that the optimal power control outperforms constant-power transmissions.
The maximum EE, $\eta_{EE}$, as a function of peak/average transmit power constraints is illustrated in Fig. \[fig:EE\_Pavg\_Ppk\]. Regarding the average interference constraint, we consider two scenarios: $Q_{{\text{avg}}}=-10$ dB and $Q_{{\text{avg}}}=-20$ dB. Target probabilities of detection and false alarm are set to $0.8$ and $0.1$, respectively, for which the corresponding sensing duration is $4.85$ ms. In addition, the frame duration is selected to maximize the EE. It can be seen from the figure that for low values of $P_{{\text{avg}}}$ and $P_{{\rm{pk}}}$, average interference power constraints are loose, and hence the power is determined by either the average or peak transmit power constraint, which results in the same EE regardless of whether $Q_{{\text{avg}}}=-10$ dB or $Q_{{\text{avg}}}=-20$ dB. The EE of the secondary users increases with increasing peak/average transmit power levels. As expected, peak transmit power constraint yields lower EE compared to that achieved under the average transmit power constraint since the instantaneous transmission power is limited by $P_{{\rm{pk}}}$ under the peak transmit power constraint, which imposes stricter limitations than the average transmit power constraint. As the constraints become less stringent and the peak and average transmit power levels are further increased, the maximum EE levels off and becomes the same under peak/average transmit power constraints since the power starts being allocated according to only the average interference constraint, $Q_{{\text{avg}}}$, due to this constraint being the dominant one.
In Fig. \[fig:throughput\_EEmin\_gain\], we display the maximum average throughput as a function of the EE gain in percentage for different levels of primary traffic. More specifically, we consider a normal traffic load, i.e., VOIP traffic with $\lambda_ 0 = 650$ ms and $\lambda_1 = 352$ ms as assumed before, and also heavy traffic load with $\lambda_0=350$ ms and $\lambda_1=650$ ms so that $\Pr\{{\mathcal{H}}_0\} \approx 0.37$. It is assumed that $\mathscr{P}_{c,max}=0.3$, average transmit power constraint is $P_{{\text{avg}}}=0$ dB and average interference power constraint is $Q_{{\text{avg}}}=10$ dB, and $\bar{P}_{{\rm{d}}}=0.8$, $\bar{P}_{{\rm{f}}}=0.1$, and hence $\tau=4.85$ ms. The frame duration for normal traffic and heavy traffic are chosen optimally as $T_{{\rm{f}}}=125$ ms and $T_{{\rm{f}}}=36$ ms, respectively, in order to maximize the EE in each traffic model. The EE gain is calculated as the ratio of the minimum required EE, $\text{EE}_{\text{min}}$, to the maximum EE achieved with the proposed power control in (\[eq:opt\_power\]). It is seen that a tradeoff between the EE and SE indeed exists, i.e., as the EE gain increases, the maximum average throughput of the secondary users decreases. We also note that the primary user with a heavy traffic load occupies the channel more often, and hence the secondary users have less opportunity to access the channel. In this heavy-load scenario, secondary users experience more frequent collisions with the primary user transmission. As a result, secondary users have lower throughput in the presence of heavy primary-user traffic compared to the case with a normal primary-user traffic.
In Fig. \[fig:throughput\_EEmin\_Qavg\], we display the maximum average throughput as a function of the average interference power constraint, $Q_{{\text{avg}}}$, under a minimum EE constraint, namely $\text{EE}_{\min}=1$ bit/joule in the presence of primary users with normal and heavy traffic loads. The frame duration is selected to maximize the system performance for each case. We assume imperfect spectrum sensing with $P_{{\rm{d}}}=0.8$ and set $P_{{\rm{f}}}=0.1$ and $P_{{\text{avg}}}=P_{{\rm{pk}}}=4$ dB, $\mathscr{P}_{c,max}=0.3$. As $Q_{{\text{avg}}}$ increases, the secondary users transmit with higher power levels, resulting in higher throughput. However, increasing $Q_{{\text{avg}}}$ further than a certain threshold does not provide performance improvements since the power starts being limited by either $P_{{\text{avg}}}$ or $P_{{\rm{pk}}}$. In addition, secondary users have higher throughput with longer transmission duration when the primary user has a normal traffic load rather than a heavy one.
Conclusion {#sec:conclusion}
==========
In this paper, we have derived the optimal power control policies that maximize the EE or maximize the average throughput of the secondary users while satisfying a minimum required EE level, in the presence of unslotted primary users, imperfect sensing, and average/peak transmit power, average interference power and collision constraints. We have also provided low-complexity algorithms to jointly optimize the transmission power and frame duration. Numerical results reveal important relations and tradeoffs between the EE and throughput performance of the secondary users. We have addressed how secondary user’s EE, collisions with the primary user transmissions, and the optimal frame duration vary as a function of the probabilities of detection and false alarm. It is also shown that optimal power control policy significantly enhances the system performance compared to the constant power scheme. The impact of the primary traffic on the system performance is analyzed as well. In particular, we have observed that secondary users achieve smaller throughput when the primary user has a heavy traffic load.
Proof of Proposition \[prop:1\] {#appendix1}
-------------------------------
*Proof:* The first derivative of $\mathscr{P}_c$ with respect to frame duration $T_{{\rm{f}}}$ is given (\[eq:Pc-derivative\]) at the top of next page.
$$\begin{aligned}
\small
\begin{split}
\frac{\partial \mathscr{P}_c}{\partial T_{{\rm{f}}}}&=\Big(\Pr\{{\mathcal{H}}_0|{\hat{\mathcal{H}}}_0\}\lambda_0\Pr\{{\mathcal{H}}_1\}^2-\Pr\{{\mathcal{H}}_1|{\hat{\mathcal{H}}}_0\}\lambda_1\Pr\{{\mathcal{H}}_0\}^2\Big) \bigg(\frac{1-{{\mathrm{e}}}^{-\frac{T_{{\rm{f}}}-\tau}{\lambda_0 \Pr\{{\mathcal{H}}_1\}}}}{(T_{{\rm{f}}}-\tau)^2}-\frac{1}{\lambda_0\Pr\{{\mathcal{H}}_1\}(T_{{\rm{f}}}-\tau)} {{\mathrm{e}}}^{-\frac{T_{{\rm{f}}}-\tau}{\lambda_0 \Pr\{{\mathcal{H}}_1\}}}\bigg).
\end{split}\label{eq:Pc-derivative}
\normalsize\end{aligned}$$
------------------------------------------------------------------------
The expression inside the first parenthesis can easily be seen to be greater than zero if $P_{{\rm{f}}}<P_{{\rm{d}}}$ and less than zero if $P_{{\rm{f}}}>P_{{\rm{d}}}$ by using the formulations in (\[eq:prior\_probs\]), (\[eq:PH0\_hH0\_exp\]) and (\[eq:PH1\_hH0\_exp\]). In order to show that the expression inside the second parenthesis is always nonnegative, we compare it with zero as follows: $$\begin{aligned}
\label{eq:second_paranthesis}
\frac{1-{{\mathrm{e}}}^{-\frac{T_{{\rm{f}}}-\tau}{\lambda_0 \Pr\{{\mathcal{H}}_1\}}}}{(T_{{\rm{f}}}-\tau)^2}-\frac{1}{\lambda_0\Pr\{{\mathcal{H}}_1\}(T_{{\rm{f}}}-\tau)} {{\mathrm{e}}}^{-\frac{T_{{\rm{f}}}-\tau}{\lambda_0 \Pr\{{\mathcal{H}}_1\}}} > 0.\end{aligned}$$ Above inequality can be rewritten as $$\begin{aligned}
\label{eq:comparision}
\bigg(1+\frac{T_{{\rm{f}}}-\tau}{\lambda_0\Pr\{{\mathcal{H}}_1\}}\bigg){{\mathrm{e}}}^{-\frac{T_{{\rm{f}}}-\tau}{\lambda_0 \Pr\{{\mathcal{H}}_1\}}}<1.\end{aligned}$$ Left-hand side of (\[eq:comparision\]) is a decreasing function since its first derivative with respect to frame duration $T_{{\rm{f}}}$ is $-\frac{T_{{\rm{f}}}-\tau}{(\lambda_0 \Pr\{{\mathcal{H}}_1\})^2}{{\mathrm{e}}}^{-\frac{T_{{\rm{f}}}-\tau}{\lambda_0 \Pr\{{\mathcal{H}}_1\}}} \le 0$. Since it is a decreasing function and it takes values between $(0,1)$ for $T_{{\rm{f}}}> \tau$, the inequality in (\[eq:comparision\]) and hence the inequality in (\[eq:second\_paranthesis\]) hold. With this, we have shown that the expression inside the second parenthesis in (\[eq:Pc-derivative\]) is nonnegative, and therefore the first derivative of $\mathscr{P}_c$ is greater than zero if $P_{{\rm{f}}}<P_{{\rm{d}}}$ and less than zero if $P_{{\rm{f}}}>P_{{\rm{d}}}$, proving the property that $\mathscr{P}_c$ is increasing with $T_{{\rm{f}}}$ if $P_{{\rm{f}}}<P_{{\rm{d}}}$ and decreasing with $T_{{\rm{f}}}$ if $P_{{\rm{f}}}>P_{{\rm{d}}}$.
Also, it can be easily verified that $\mathscr{P}_c$ takes values between $\Pr\{{\mathcal{H}}_1|{\hat{\mathcal{H}}}_0\}$ and $\Pr\{{\mathcal{H}}_1\}$. In particular, we examine the limit of $\mathscr{P}_c$ as $T_{{\rm{f}}}$ approaches $\tau$ and $\infty$ as follows: $$\begin{aligned}
\nonumber
\lim_{T_{{\rm{f}}} \rightarrow \tau} \mathscr{P}_c & = \Pr\{{\mathcal{H}}_0|{\hat{\mathcal{H}}}_0\}\bigg(\Pr\{{\mathcal{H}}_1\}-\frac{\lambda_0 \Pr\{{\mathcal{H}}_1\}^2}{\lambda_0\Pr\{{\mathcal{H}}_1\}}\bigg) \\ \nonumber & \hspace{0.8cm} + \Pr\{{\mathcal{H}}_1|{\hat{\mathcal{H}}}_0\}\bigg(\Pr\{{\mathcal{H}}_1\}+\frac{\lambda_1 \Pr\{{\mathcal{H}}_0\}^2}{\lambda_0\Pr\{{\mathcal{H}}_1\}}\bigg) \\ & = \Pr\{{\mathcal{H}}_1|{\hat{\mathcal{H}}}_0\} \\ \nonumber
\lim_{T_{{\rm{f}}} \rightarrow \infty} \mathscr{P}_c & = \Pr\{{\mathcal{H}}_0|{\hat{\mathcal{H}}}_0\}\Pr\{{\mathcal{H}}_1\}+ \Pr\{{\mathcal{H}}_1|{\hat{\mathcal{H}}}_0\}\Pr\{{\mathcal{H}}_1\} \\& =\Pr\{{\mathcal{H}}_1\}\end{aligned}$$
Proof of Theorem \[teo:1\] {#appendix2}
--------------------------
*Proof:* The objective function in (\[eq:EE\_max\_Pavg\_Qavg\]) is quasiconcave since the average throughput in the numerator is composed of positive weighted sum of logarithms which are strictly concave and the power consumption in the denominator is both affine and positive. Therefore, the optimal power value can be found iterativaly by using Dinkelbach’s method [@dinkelbach]. The optimization problem is first transformed into the equivalent parameterized concave problem as follows: $$\begin{aligned}
\label{eq:EE_concave}
&\hspace{-2cm}\max_{
\substack{P(g,h) \geq 0}} \bigg\{R_{{\text{avg}}}-\alpha \bigg(\Big(\frac{T_{{\rm{f}}}-\tau}{T_{{\rm{f}}}}\Big)\Pr\{{\hat{\mathcal{H}}}_0\}{\mathbb{E}}\{P(g,h)\}+P_{c_r}\bigg)\bigg\} \\ \label{eq:avg_transmit_power_eq}
\text{subject to } &\Big(\frac{T_{{\rm{f}}}-\tau}{T_{{\rm{f}}}}\Big)\Pr\{{\hat{\mathcal{H}}}_0\}{\mathbb{E}}\big\{P(g,h)\big\} \le P_{{\text{avg}}} \\ \label{eq:avg_inter_power_eq}
&\Big(\frac{T_{{\rm{f}}}-\tau}{T_{{\rm{f}}}}\Big)\mathscr{P}_c\Pr\{{\hat{\mathcal{H}}}_0\}{\mathbb{E}}\big\{P(g,h)|g|^2\big\} \le Q_{{\text{avg}}},\end{aligned}$$ where $\alpha$ is a nonnegative parameter. At the optimal value of $\alpha^*$, the following condition is satisfied $$\begin{aligned}
\label{eq:F_alpha}
F(\alpha^*)\!=\!R_{{\text{avg}}}\!-\!\alpha^*\bigg(\!\Big(\frac{T_{{\rm{f}}}-\tau}{T_{{\rm{f}}}}\Big)\Pr\{{\hat{\mathcal{H}}}_0\}{\mathbb{E}}\{P(g,h)\}\!+\!P_{c_r}\!\bigg)\!=\!0.\end{aligned}$$ Explicitly, the solution of $F(\alpha^*)$ is equivalent to the solution of the EE maximization problem in (\[eq:EE\_max\_Pavg\_Qavg\]). It is shown that Dinkelbach’s method converges to the optimal solution at a superlinear convergence rate. The detailed proof of convergence and further details can be found in [@schaible]. Since the parameterized problem in (\[eq:EE\_concave\]) is concave for a given $\alpha$, the optimal power levels can be obtained by using the Lagrangian optimization approach as follows: $$\begin{aligned}
\small
\begin{split} \label{eq:lagrangian_function}
&\mathcal{L}(P(g,h),\lambda,\nu,\alpha)\!=\!R_{{\text{avg}}}-\alpha \bigg(\!\Big(\frac{T_{{\rm{f}}}\!-\!\tau}{T_{{\rm{f}}}}\Big)\Pr\{{\hat{\mathcal{H}}}_0\}{\mathbb{E}}\{P(g,h)\}\!+\!P_{c_r}\!\!\bigg) \\ &\hspace{1.8cm}-\lambda\bigg(\Big(\frac{T_{{\rm{f}}}-\tau}{T_{{\rm{f}}}}\Big)\Pr\{{\mathcal{H}}_0\}{\mathbb{E}}\{P(g,h)\}-P_{{\text{avg}}}\bigg) \\
&\hspace{1.8cm}-\nu\bigg(\Big(\frac{T_{{\rm{f}}}-\tau}{T_{{\rm{f}}}}\Big)\mathscr{P}_c\Pr\{{\hat{\mathcal{H}}}_0\}{\mathbb{E}}\big\{P(g,h)|g|^2\big\}-Q_{{\text{avg}}} \bigg),
\end{split}
\normalsize\end{aligned}$$ where $\lambda$ and $\nu$ are the nonnegative Lagrange multipliers. The Lagrange dual problem is defined as $$\begin{aligned}
\min_{\lambda, \nu \ge 0} \max_{P(g,h) \geq 0} \mathcal{L}(P(g,h),\lambda,\nu,\alpha).\end{aligned}$$ For fixed $\lambda$ and $\nu$ values, and each fading state, we express the subproblem using the Lagrange dual decomposition method [@palomar]. According to the Karush-Kuhn-Tucker (KKT) conditions, the optimal power control $P_{\text{opt}}(g,h)$ must satisfy the set of equations and inequalities below: $$\begin{aligned}
\nonumber
& \frac{\Pr\{{\hat{\mathcal{H}}}_0\}}{\log_e(2)}\Big(\frac{T_{{\rm{f}}}-\tau}{T_{{\rm{f}}}}\Big) \bigg[\bigg(\frac{(1-\mathscr{P}_c)|h|^2}{N_0+P_{\text{opt}}(g,h)|h|^2}\bigg)\\ \nonumber &\!+\! \bigg(\frac{\mathscr{P}_c|h|^2}{N_0\!+\!\sigma_s^2+P_{\text{opt}}(g,h)|h|^2}\bigg)\bigg]\! -\! (\alpha\!+\!\lambda)\Big(\frac{T_{{\rm{f}}}-\tau}{T_{{\rm{f}}}}\Big) \Pr\{{\hat{\mathcal{H}}}_0\} \\ \label{eq:KKT_condition1} &\hspace{2.8cm}-\nu\Big(\frac{T_{{\rm{f}}}-\tau}{T_{{\rm{f}}}}\Big) \Pr\{{\hat{\mathcal{H}}}_0\}\mathscr{P}_c|g|^2=0\\ \label{eq:average_transmit_cond3} &\lambda\bigg(\Big(\frac{T_{{\rm{f}}}-\tau}{T_{{\rm{f}}}}\Big)\Pr\{{\hat{\mathcal{H}}}_0\}{\mathbb{E}}\{P_{\text{opt}}(g,h)\}-P_{{\text{avg}}}\bigg)=0,
\\ \label{eq:average_inter_cond3}&\nu\bigg(\Big(\frac{T_{{\rm{f}}}-\tau}{T_{{\rm{f}}}}\Big)\mathscr{P}_c\Pr\{{\hat{\mathcal{H}}}_0\}{\mathbb{E}}\big\{P_{\text{opt}}(g,h)|g|^2\big\}-Q_{{\text{avg}}} \bigg) = 0, \\ \label{eq:lagrange_multipliers}
&\lambda \ge 0, \nu \ge 0.\end{aligned}$$
Solving (\[eq:KKT\_condition1\]) and incorporating the nonnegativity of the transmit power yield the desired result in (\[eq:opt\_power\]). $\square$
Proof of Proposition \[prop:2\] {#appendix3}
-------------------------------
In order to find the operating power level, which satisfies the minimum required EE, we consider that the objective function in (\[eq:throughput\_max\_Pavg\_Qavg\]) is subject to only a minimum EE constraint in (\[eq:EEmin\_Pavg\_Qavg\]). Since $R_{{\text{avg}}}$ is a concave function of the transmission power and the feasible set defined by the minimum EE constraint is a convex set, KKT conditions are both sufficient and necessary for the optimal solution. The constraint in (\[eq:EEmin\_Pavg\_Qavg\]) can be rewritten as follows $$\begin{aligned}
\hspace{-0.2cm}R_{{\text{avg}}}- \text{EE}_{\text{min}} \Big(\big(\frac{T_{{\rm{f}}}-\tau}{T_{{\rm{f}}}}\big)\Pr\{{\hat{\mathcal{H}}}_0\}{\mathbb{E}}\{P(g,h)\}+P_{c_r}\Big) \geq 0.\end{aligned}$$ By defining $\eta$ as the Lagrange multiplier associated with the above constraint, the Lagrangian function is expressed as $$\begin{aligned}
\nonumber
&\mathcal{L}(P(g,h),\eta)=(1+\eta)R_{{\text{avg}}} \\&\hspace{1cm}-\eta \text{EE}_{\text{min}} \Big(\big(\frac{T_{{\rm{f}}}-\tau}{T_{{\rm{f}}}}\big)\Pr\{{\hat{\mathcal{H}}}_0\}{\mathbb{E}}\{P(g,h)\}+P_{c_r}\Big).\end{aligned}$$ By setting the derivative of the above function with respect to $P(g,h)$ equal to zero at the optimal power level, we obtain the equation in (\[eq:Lagrangian\_EEmin\]) given at the top of next page.
$$\begin{aligned}
\begin{split} \label{eq:Lagrangian_EEmin}
\hspace{-0.4cm}&\left. \frac{\partial \mathcal{L}(P(g,h),\eta)}{\partial P(g,h)} \right |_{P(g,h)=P^*(g,h)} \!\!\!\!\!=(1+\eta)\frac{\Pr\{{\hat{\mathcal{H}}}_0\}}{\log_e(2)}\Big(\frac{T_{{\rm{f}}}\!-\!\tau}{T_{{\rm{f}}}}\Big)\bigg[\bigg(\frac{(1-\mathscr{P}_c)|h|^2}{N_0+P^*(g,h)|h|^2}\bigg)\!+\!\bigg(\frac{\mathscr{P}_c|h|^2}{N_0+\sigma_s^2+P^*(g,h)|h|^2}\bigg)\bigg]\\&\hspace{13cm}-\eta \text{EE}_{\text{min}}\Big(\frac{T_{{\rm{f}}}-\tau}{T_{{\rm{f}}}}\Big) \Pr\{{\hat{\mathcal{H}}}_0\}\!=\!0.
\end{split}\end{aligned}$$
------------------------------------------------------------------------
Solving the equation in (\[eq:Lagrangian\_EEmin\]) leads to the desired characterization in (\[eq:P0\_EEmin\]) and the Lagrange multiplier, $\eta$ can be determined by satisfying the minimum EE constraint with equality or solving (\[eq:eta\_EEmin\]). Consequently, the average transmission power is obtained by inserting (\[eq:P0\_EEmin\]) into (\[eq:Pavg\_EEmin\]). $\square$
Proof of Theorem \[teo:3\] {#appendix4}
--------------------------
The Lagrangian function is expressed as $$\small
\begin{split}
&\mathcal{L}(P(g,h),\vartheta,\varphi)=R_{{\text{avg}}} - \vartheta \Big(\Big(\frac{T_{{\rm{f}}}-\tau}{T_{{\rm{f}}}}\Big)P({\hat{\mathcal{H}}}_0){\mathbb{E}}\{P(g,h)\} \\&\!-\!\min(P_{{\text{avg}}},\!P^*_{{\text{avg}}})\!\Big)\!-\!\varphi \Big(\!\Big(\frac{T_{{\rm{f}}}\!-\!\tau}{T_{{\rm{f}}}}\Big)\mathscr{P}_c\Pr\{{\hat{\mathcal{H}}}_0\}{\mathbb{E}}\{\!P(g,\!h)|g|^2\!\}\!-\!Q_{{\text{avg}}})\!\Big).
\end{split}
\normalsize$$ Setting the derivative of the above function with respect to transmission power, $P(g,h)$, to zero and arranging the terms give the desired optimal power control in (\[eq:opt\_power4\]).$\square$
[99]{}
S. Stotas and A. Nallanathan, “Optimal sensing time and power allocation in multiband cognitive radio networks," *IEEE Trans. Wireless Commun.*, vol. 59, no. 1, pp. 226–235, Jan. 2011.
S. Akin and M. C. Gursoy, “Effective capacity analysis of cognitive radio channels for quality of service provisioning," *IEEE Systems Journal*, vol. 7, no. 1, pp. 68–76, Nov. 2010.
H. Hu, H. Zhang, and Y.-C. Liang, “On the spectrum- and energy-efficiency tradeoff in cognitive radio networks," *IEEE Trans. Commun.*, vol. 64, no. 2, pp. 490–501, Feb. 2016.
J. Zhang, F..-C. Zheng, X.-Q. Gao, H.-B. Zhu, “Sensing-energy efficiency tradeoff for cognitive radio networks," *IET Commun.*, vol. 8, no. 18, pp. 3414–3423, Dec. 2014.
C. Xiong, L. Lu, and G. Y. Li, “Energy-efficient spectrum access in cognitive radios," *IEEE J. Sel. Areas Commun.*, vol. 32, no. 3, pp. 550–562, Mar. 2014.
H. Park and T. Hwang, “Energy-efficient power control of cognitive femto users for 5G communications," *IEEE J. Sel. Areas Commun.*, vol. 34, no. 4, Apr. 2016.
R. Ramamonjison and V. K. Bhargava, “Energy efficiency maximization framework in cognitive downlink two-tier networks," *IEEE Trans. Wireless Commun.*, vol. 14, no. 3, pp. 1468–1479, Mar. 2015.
Y. Pei, A. T. Hoang, and Y. C. Liang, “Sensing-throughput tradeoff in cognitive radio networks: how frequently should spectrum sensing be carried out," in *Proc. of IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC)*, pp. 1–5, Sept. 2007.
W. Tang, M. Z. Shakir, M. A. Imran, R. Tafazolli, and M. S. Alouini, “Throughput analysis for cognitive radio networks with multiple primary users and imperfect spectrum sensing," *Communications, IET*, vol. 6, no. 17, pp. 2787–2795, 2012.
S. Zarrini and T. J. Lim, “Throughput-sensing tradeoff of cognitive radio networks based on quickest sensing," in *Proc. of IEEE International Conference on Communications (ICC)*, pp. 1–5, June 2011.
M. Guerrini, L. Rugini, and P. Banelli, “Sensing-throughput tradeoff for cognitive radios," in *Proc. of IEEE Workshop on Signal Processing Advances in Wireless Communications (SPAWC)*, pp. 115–119, June 2013.
S. L. MacDonald and D. C. Popescu, “Impact of primary user activity on the performance of energy-based spectrum sensing in cognitive radio systems," in *Proc. of the IEEE Global Commun. Conf. (Globecom)*, pp. 3224-3228, Dec. 2013.
T. Wang, Y. Chen, E. Hines, and B. Zhao, “Analysis of effect of primary user traffic on spectrum sensing performance," in *Proc. of IEEE International Conf. on Commun. (ICC)*, pp. 1-5, Aug. 2009.
J. Y. Wu, P. H. Huang, T. Y. Wang, and V. W. S. Wong, “Energy detection based spectrum sensing with random arrival and departure of primary user’s signal," in *Proc. of the IEEE Global Commun. Conf. (Globecom)*, pp. 380 - 384, Dec. 2013
P. Dhakal, S. K. Sharma. S. Chatzinotas, B. Ottersten, and C. Riviello, “Effect of primary user traffic on largest eigenvalue based spectrum sensing technique," in *Proc. of the International Conference on Cognitive Radio Oriented Wireless Networks (Crowncom)*, pp. 67-78, May 2016.
L. Tang, Y. Chen, E. L. Hines, and M. S. Alouini, “Effect of primary user traffic on sensing-throughput tradeoff for cognitive radios," *IEEE Trans. Wireless Commun.*, vol. 10, no. 4, pp. 1063–1068, April 2011.
H. Pradhan, A. S. Kalamkar, and A. Banerjee, “Sensing-throughput tradeoff in cognitive radio with random arrivals and departures of multiple primary users," *IEEE Commun. Letters*, vol. 19, no. 3, pp. 415–418, March 2015.
W. Zhang, C. X. Wang, D. Chen, and H. Xiong, “Energy-spectral efficiency tradeoff in cognitive radio networks,” *IEEE Trans. on Veh. Tech.* vol. 65, no. 4, pp. 2208-2218, April 2016.
S. K. Sharma, T. E. Bogale, S. Chatzinotas, B. Ottersten, L. B. Le and X. Wang, “Cognitive radio techniques under practical imperfections: A survey,” *IEEE Commun. Surveys & Tutorials*, vol. 17, no. 4, pp. 1858-1884, Fourthquarter 2015.
A. Kaushik, S. K. Sharma, S. Chatzinotas, B. Ottersten and F. K. Jondral, “Sensing-throughput tradeoff for interweave cognitive radio system: A deployment-centric viewpoint,” *IEEE Trans. Wireless Commun.*, vol. 15, no. 5, pp. 3690-3702, May 2016.
J. Riihijärvi, J. Nasreddine, and P. Mähönen, “Impact of primary user activity patterns on spatial spectrum reuse opportunities," in *Proc. of IEEE Wireless Conference (EW)*, pp. 962–968, April 2010.
M. Wellens, J. Riihijärvi, and P. Mähönen, “Empirical time and frequency domain models for spectrum use," *Phys. Commun.*, vol. 2, no. 1–2, pp. 10-32, 2009.
S. Yin, D. Chen, Q. Zhang, M. Liu, and S. Li, “Mining spectrum usage data: A large-scale spectrum measurement study,," *IEEE Trans. Mobile Comput.*, vol. 11, no. 6, pp. 1033–1046, Jun. 2012.
D. Willkomm, S. Machiraju, J. Bolot, and A. Wolisz, “Primary users in cellular networks: A large-scale measurement study," in *Proc. of IEEE Inter. Symp. on New Frontiers in Dynamic Spectrum Access Networks (DySPAN)*, Oct. 2008.
W. Gabran, C. H. Liu, P. Pawelczak and D. Cabric, “Primary User Traffic Estimation for Dynamic Spectrum Access," in *IEEE Journal on Selected Areas in Communications*, vol. 31, no. 3, pp. 544-558, March 2013.
E. Axell, G. Leus, E. G. Larsson, and H. V. Poor, “Spectrum sensing for cognitive radio: State-of-the-art and recent advances," *IEEE Signal Process. Mag.*, vol. 29, no. 3, pp. 101–116, May 2012.
C. Jiang, Y. Chen, K. J. R. Liu, and Y. Ren, “Renewal-theoretical dynamic spectrum access in cognitive radio network with unknown primary behavior," *IEEE J. Sel. Areas in Commun.*, vol. 31, no. 3, pp. 406–416, March 2013.
M. -S. Alouini and A. J. Goldsmith, “Capacity of Rayleigh fading channels under different adaptive transmission and diversity-combining techniques," *IEEE Trans. Veh. Tech.*, vol. 48, no. 4, pp. 1165–1181, July 1999.
J. M. Peha, “Approaches to spectrum sharing," *IEEE Commun. Mag.*, vol. 43, no. 2, pp. 10–12, Feb. 2005.
Q. Zhao, S. Geirhofer, L. Tong, and B. M. Sadler, “Opportunistic spectrum access via periodic channel sensing," *IEEE Trans. Signal Process.*, vol. 56, no. 2, pp. 785–796, Feb. 2008.
D. Palomar and M. Chiang, “A tutorial on decomposition methods for network utility maximization,” *IEEE J. on Sel. Areas in Commun.*, vol. 24, no. 8, pp. 1439–1451, Aug. 2006.
S. Boyd, L. Xiao, and A. Mutapcic, “Subgradient methods," Lecture Notes of EE392o, Standford University, Autumn Quarter 2003-2004.
S. Schaible, “Fractional programming. [slowromancap2@]{}, On Dinkelbach’s algorithm," *Management Science*, vol. 22, no. 8, pp. 868–873, 1976.
X. Zhao, P. B. Luh, and J. Wang, “Surrogate gradient algorithm for Lagrangian relaxation,” J. Optimization Theory and Applications," *J. Optimization Theory and Application*, vol. 100, no. 3, pp. 699–712, Mar. 1999.
W. Dinkelbach, “On nonlinear fractional programming," *Management Science*, vol. 13, pp. 492–498, Mar. 1967.
G. Ozcan, M. C. Gursoy, and J. Tang, “Power control for cognitive radio systems with unslotted primary users under sensing uncertainty," in *Proc. of the IEEE Inter. Conf. on Commun. (ICC)*, pp. 1428–1433, London, June 2015.
R. Tandra and A. Sahai, “SNR walls for signal detection," *IEEE J. on Sel. Topics Signal Process.*, vol. 2, no. 1, pp. 4–17, Feb. 2008.
R. Tandra and A. Sahai, “SNR walls for signal detection," *IEEE J. on Sel. Topics Signal Process.*, vol. 2, no. 1, pp. 4–17, Feb. 2008.
Y. C. Liang, Y. Zeng, E. C. Y. Peh and A. T. Hoang, “Sensing-Throughput Tradeoff for Cognitive Radio Networks," *IEEE Trans. on Wireless Commun.*, vol. 7, no. 4, pp. 1326-1337, April 2008.
[^1]: The material in this paper was presented in part at the IEEE International Conference on Communications (ICC), London, UK, June 2015.
[^2]: This research is supported by NSF grants CNS-1443966 and ECCS-1443994.
[^3]: Gozde Ozcan, M. Cenk Gursoy and Jian Tang are with the Department of Electrical Engineering and Computer Science, Syracuse University, Syracuse, NY, 13244 (e-mail: [email protected], [email protected], [email protected]).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We give a surgery formula for the asymptotic behavior of the sequence given by the logarithm of the higher dimensional Reidemeister torsion. Applying the resulting formula to Seifert fibered spaces, we show that the growth of the sequences has the same order as the indices and we give the explicit values for the limits of the leading coefficients. There are finitely many possibilities as the limits of the leading coefficients for a Seifert fibered space. We also show that the maximum is given by $-\chi \log 2$ where $\chi$ is the Euler characteristic of the base orbifold for a Seifert fibered space. These limits of the leading coefficients give a locally constant function on a character variety. This function takes the maximum $-\chi \log 2$ only on the top-dimensional components of the ${\mathrm{SU}(2)}$-character varieties for Seifert fibered homology spheres.'
address: 'Department of Mathematics, Akita University, 1-1 Tegata-Gakuenmachi, Akita, 010-8502, Japan'
author:
- Yoshikazu Yamaguchi
bibliography:
- 'asymptoticsTorsionSeifert.bib'
title: A surgery formula for the asymptotics of the higher dimensional Reidemeister torsion and Seifert fibered spaces
---
Introduction
============
This paper is devoted to the study of the asymptotics of the higher dimensional Reidemeister torsion for [*Seifert fibered spaces*]{}. The higher dimensional Reidemeister torsion is defined for a $3$-manifold and a sequence of homomorphisms from the fundamental group into special linear groups. This sequence is given by the composition of an [*${\mathrm{SL}_{2}({\mathbb{C}})}$-representation*]{} of a fundamental group with $n$-dimensional irreducible representations of ${\mathrm{SL}_{2}({\mathbb{C}})}$. It is of interest to observe the asymptotic behavior of the higher dimensional Reidemeister torsion on the index $n$.
This study is motivated by the works of W. M[ü]{}ller [@Muller:AsymptoticsAnalyticTorsion] and P. Menal-Ferrer and J. Porti [@FerrerPorti:HigherDimReidemeister], which revealed the relationship between the asymptotic behaviors of the Ray–Singer and the Reidemeister torsions for a hyperbolic $3$-manifold and its hyperbolic volume. Their invariants are defined by a sequence of ${\mathrm{SL}_{n}({\mathbb{C}})}$-representations induced from the holonomy representation corresponding to the complete hyperbolic structure. The sequences of the Ray–Singer and the Reidemeister torsions have exponential growth and the logarithms have the order of $n^2$. They showed that the leading coefficient of the logarithm converges to the product of the hyperbolic volume and $-1/(4\pi)$.
We can also consider the similar sequences given by the Reidemeister torsions for non–hyperbolic $3$-manifolds, especially for Seifert fibered spaces. From the results of M[ü]{}ller and Menal-Ferrer and Porti, it is expected that the growth of the logarithm of the higher dimensional Reidemeister torsion for a Seifert fibered space has the order of smaller than $n^2$ and the leading coefficient converges to some geometric quantity of the Seifert fibered space. Actually we will see that the order of growth is the same as $n$ and the limits of the leading coefficients form a finite set, in which the maximum is given by $-\chi \log 2$ where $\chi$ is the Euler characteristic of the base orbifold of a Seifert fibered space.
To observe the asymptotic behaviors for Seifert fibered spaces, we establish a surgery formula for gluing solid tori to a compact orientable $3$-manifold with torus boundary, since every Seifert fibered space admits the canonical decomposition into the trivial $S^1$-bundle over a compact surface and several solid tori (for details, see Subsection \[subsec:Seifert\]).
There is a problem which we have to consider on the choice of homomorphisms from fundamental groups into ${\mathrm{SL}_{2}({\mathbb{C}})}$ since Seifert fibered spaces do not admit complete hyperbolic structures. Concerning this problem, we focus on the [*acyclicity*]{} properties for the induced twisted chain complexes defined by ${\mathrm{SL}_{n}({\mathbb{C}})}$-representations for hyperbolic $3$-manifolds. It was shown in [@Raghunathan65:CohomologyVanishing; @FerrerPorti:TwistedCohomology] that for a hyperbolic $3$-manifold, the induced ${\mathrm{SL}_{2N}({\mathbb{C}})}$-representations from the holonomy representation define [*acyclic*]{} chain complexes, [i.e.,]{}all of those homology groups vanish. We will restrict our attention to ${\mathrm{SL}_{2}({\mathbb{C}})}$-representations such that all the induced ${\mathrm{SL}_{2N}({\mathbb{C}})}$-representations define the acyclic twisted chain complexes.
The following is a surgery formula (Theorem \[thm:surgery\_formula\]) for the limits of sequences given by the logarithms of the higher dimensional Reidemeister torsions under our acyclicity conditions (see Definition \[def:acyclicity\_conditions\] for the acyclicity conditions).
Let ${M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}}$ be a closed orientable $3$-manifold obtained by gluing solid tori $S_1, \ldots, S_m$ with the slopes $\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}$ to a compact orientable manifold $M$ with torus boundary.
We denote by $\rho$ an ${\mathrm{SL}_{2}({\mathbb{C}})}$-representation of $\pi_1({M})$ which can be extended to a homomorphism of $\pi_1({M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}})$ and by $\rho_{2N}$ the $2N$-dimensional representation induced from $\rho$. Then the asymptotics of $\log|{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}};\rho_{2N})}|$ is expressed as follows:
1. $\displaystyle{
\lim_{N \to \infty}
\frac{
\log|{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}};\rho_{2N})}|
}{
(2N)^2
}
=
\lim_{N \to \infty}
\frac{
\log|{\mathop{\mathrm{Tor}}\nolimits ({M};\rho_{2N})}|
}{
(2N)^2
},
}$
2. $\displaystyle{
\lim_{N \to \infty}
\frac{
\log|{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}};\rho_{2N})}|
}{
2N
}
=
\lim_{N \to \infty}
\frac{
\log|{\mathop{\mathrm{Tor}}\nolimits ({M};\rho_{2N})}|
}{
2N
}
-\log 2 \sum_{j=1}^m \frac{1}{\lambda_j}
}$
where $2\lambda_j$ is the order of $\rho(\ell_j)$ and $\ell_j$ is the homotopy class of the core in $S_j$.
Applying the above surgery formula to the decomposition of Seifert fibered spaces, we can show the asymptotic behaviors for Seifert fibered spaces.
Let ${M\hbox{$\big(\frac{1}{-b}, \frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}}$ denote a Seifert fibered space with the Seifert index: $$\{b, (o, g); (\alpha_1, \beta_1), \ldots, (\alpha_m, \beta_m)\}.$$ Then we can express the asymptotic behavior of $\log |{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{1}{-b}, \frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}};\rho_{2N})}|$ as follows:
1. $\displaystyle{
\lim_{N \to \infty}
\frac{
\log |{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{1}{-b}, \frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}};\rho_{2N})}|
}{
(2N)^2
}
=0
}$,
2. $\displaystyle{
\lim_{N \to \infty}
\frac{
\log |{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{1}{-b}, \frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}};\rho_{2N})}|
}{
2N
}
= - \Big(2-2g - \sum_{j=1}^m \frac{\lambda_j - 1}{\lambda_j}\Big) \log 2
}$
where $2\lambda_j$ is the order of $\rho(\ell_j)$ for the exceptional fiber $\ell_j$. In particular, the second equality can be written as $$\lim_{N \to \infty}
\frac{
\log |{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{1}{-b}, \frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}};\rho_{2N})}|
}{
2N
}
= -\Big(2-2g - \sum_{j=1}^m \frac{\alpha_j -1}{\alpha_j}\Big) \log 2
- \Big(\sum_{j=1}^m\Big(\frac{1}{\lambda_j} - \frac{1}{\alpha_j}\Big) \Big)\log 2.$$
The first term in the right hand side is equal to $-\chi \log 2$ where $\chi$ is the Euler characteristic of the base orbifold. Each $\lambda_j$ turns out to be a divisor of $\alpha_j$. Hence we can see that the maximum of the above limit is given by $-\chi \log 2$. Note that we only require that the original ${\mathrm{SL}_{2}({\mathbb{C}})}$-representation sends a regular fiber to $-{I}$ for a Seifert fibered space.
Moreover the limit of the leading coefficient is determined by each component in the representation spaces. By the invariance of Reidemeister torsion under the conjugation of representations, we can assign each component of the character variety to the limit of the leading coefficient, namely, we can define a locally constant function on the character variety for a Seifert fibered space. We will discuss on which components our locally constant function takes the maximum and minimum for the ${\mathrm{SU}(2)}$-character varieties of Seifert fibered homology $3$-spheres in detail. We can find the components which give the maximum in top dimensional components of the ${\mathrm{SU}(2)}$-character variety. In particular, for Seifert fibered homology $3$-spheres given by sequences of prime integers, Theorem \[thm:max\_min\_leading\_coeff\] shows that all top-dimensional components give the maximum and some $0$-dimensional components give the minimum.
Organization {#organization .unnumbered}
------------
We review the definition of the Reidemeister torsion and the construction of the higher dimensional ones in Section \[sec:preliminaries\]. Section \[sec:surgery\_formula\] is devoted to establish our surgery formula under the acyclicity conditions which are deduced from the observation in Subsection \[subsec:acyc\_condition\]. The examples of the surgery formula for integral surgeries along torus knots are exhibited in Subsection \[subsec:example\_torusknots\]. We discuss the asymptotic behaviors of the sequences given by the logarithm of the higher dimensional Reidemeister torsion for Seifert fibered spaces in Section \[sec:asymptotics\_Seifert\]. We review on Seifert fibered spaces and prepare notations in Subsection \[subsec:Seifert\]. Subsection \[subsec:asymptotics\_Seifert\] gives a general formula of the asymptotic behavior for a Seifert fibered space. Furthermore we observe the relation between the limits of the leading coefficients and components in the ${\mathrm{SU}(2)}$-character varieties for Seifert fibered homology $3$-spheres in Subsection \[subsec:leadingCoeff\_Seifert\]. The last Subsection \[subsec:examples\] gives the explicit examples of limits of the leading coefficients and the ${\mathrm{SU}(2)}$-character varieties.
Preliminaries {#sec:preliminaries}
=============
Although one can find the similar preliminaries in [@yamaguchi:RtorTorusKnots], we give a review on the Reidemeister torsion, needed in our observation, to make this article self-contained.
Reidemeister torsion {#section:Rtorsion}
--------------------
### Torsion for acyclic chain complexes {#torsion-for-acyclic-chain-complexes .unnumbered}
[*Torsion*]{} is an invariant defined for based chain complexes. We denote by $(C_*, {\mathbf{c}^{*}} = \cup_i {\mathbf{c}^{i}})$ a [*based*]{} chain complex: $$C_*: 0 \to C_n \xrightarrow{{\partial_{n}}}
C_{n-1} \xrightarrow{{\partial_{n-1}}}
\cdots \xrightarrow{{\partial_{2}}}
C_1 \xrightarrow{{\partial_{1}}}
C_0 \to 0$$ where each chain module $C_i$ is a vector space over a field ${\mathbb{F}}$ and equipped with a basis ${\mathbf{c}^{i}}$. We are mainly interested in an [*acyclic*]{} chain complex $C_*$ which has the trivial homology group, [i.e.,]{}$H_*(C_*) = {\mbox{\boldmath{$0$}}}$. The chain complex $C_*$ also has a basis determined by the boundary operators ${\partial_{i}}$, which arises from the following decomposition of chain modules.
We suppose that a based chain complex $(C_*, {\mathbf{c}^{*}})$ is acyclic. For each boundary operator ${\partial_{i}}$, we denote $\ker {\partial_{i}} \subset C_i$ by $Z_i$ and the image of ${\partial_{i}}$ by $B_i \subset C_{i-1}$. The chain module $C_i$ is expressed as the direct sum of $Z_i$ and the lift of $B_i$, denoted by ${\tilde{B}}_i$. Moreover we can rewrite the kernel $Z_i$ as the image of boundary operator ${\partial_{i+1}}$: $$\begin{aligned}
C_i
&= Z_i \oplus {\tilde{B}}_i \\
&= {\partial_{i+1}}{\tilde{B}}_{i+1} \oplus {\tilde{B}}_i \end{aligned}$$ where $Z_i = B_{i+1}$ is written as ${\partial_{i+1}}{\tilde{B}}_{i+1}$.
We denote by ${\tilde{\mathbf{b}}^{i}}$ a basis of ${\tilde{B}}_{i}$. Then the set ${\partial_{i+1}}({\tilde{\mathbf{b}}^{i+1}}) \cup {\tilde{\mathbf{b}}^{i}}$ forms a new basis of the vector space $C_i$. We define the [*torsion*]{} of $(C_*, {\mathbf{c}^{*}})$ as the following alternating product of determinants of base change matrices: $$\label{eqn:def_torsion_complex}
\mathrm{Tor}(C_*, {\mathbf{c}^{*}})
= \prod_{i \geq 0}
\left[
{\partial_{i+1}}({\tilde{\mathbf{b}}^{i+1}}) \cup {\tilde{\mathbf{b}}^{i}}\,/\, {\mathbf{c}^{i}}\right]^{(-1)^{i+1}}
\in {\mathbb{F}}^* = {\mathbb{F}}\setminus \{0\}$$ where $[ {\partial_{i+1}}({\tilde{\mathbf{b}}^{i+1}}) \cup {\tilde{\mathbf{b}}^{i}}\,/\, {\mathbf{c}^{i}}]$ denotes the determinant of the base change matrix from the given basis ${\mathbf{c}^{i}}$ to the new one ${\partial_{i+1}}({\tilde{\mathbf{b}}^{i+1}}) \cup {\tilde{\mathbf{b}}^{i}}$.
Note that the right hand side is independent of the choice of bases ${\tilde{\mathbf{b}}^{i}}$. The alternating product in is determined by the based chain complex $(C_*, {\mathbf{c}^{*}})$.
### Reidemeister torsion for CW–complexes {#reidemeister-torsion-for-cwcomplexes .unnumbered}
We apply the torsion of a based chain complex to the [*twisted chain complex*]{} given by a CW–complex and a homomorphism from its fundamental group to some linear group. Let $W$ denote a finite CW–complex and $(V, \rho)$ a representation of $\pi_1(W)$, which means $V$ is a vector space over ${\mathbb{F}}$ and $\rho$ is a homomorphism from $\pi_1(W)$ into ${\mathrm{GL}}(V)$. We will call $\rho$ a ${\mathrm{GL}}(V)$-representation of $\pi_1(W)$ simply.
\[def:twistedcomplex\] We define the twisted chain complex $C_*(W; V_\rho)$ which consists of the twisted chain module as: $$C_i (W; V_\rho) :=V \otimes_{{\mathbb{Z}}[\pi_1(W)]} C_i ({{\widetilde W}};{\mathbb{Z}})$$ where ${{\widetilde W}}$ is the universal cover of $W$ and $C_i({{\widetilde W}};{\mathbb{Z}})$ is a left ${\mathbb{Z}}[\pi_1(W)]$-module in which the action of $\pi_1(W)$ is given by the covering transformation. In taking the tensor product, we regard $V$ as a right ${\mathbb{Z}}[\pi_1(W)]$-module under the homomorphism $\rho^{-1}$. We identify a chain ${\mbox{\boldmath{$v$}}} \otimes \gamma c$ with $\rho(\gamma)^{-1}({\mbox{\boldmath{$v$}}}) \otimes c$ in $C_i (W; V_\rho)$.
We call $C_*(W;V_\rho)$ the twisted chain complex with the coefficient $V_\rho$. Choosing a basis of the vector space $V$, we give a basis of the twisted chain complex $C_*(W; V_\rho)$. To be more precise, let $\{e^{i}_1, \ldots, e^{i}_{m_i}\}$ be the set of $i$-dimensional cells of $W$ and $\{{\mbox{\boldmath{$v$}}}_1, \ldots, {\mbox{\boldmath{$v$}}}_{d}\}$ a basis of $V$ where $d = \dim_{{\mathbb{F}}} V$. Choosing a lift ${\tilde{e}}^{i}_j$ of each cell and taking tensor product with the basis of $V$, we have the following basis of $C_i(W;V_\rho)$: $${\mathbf{c}^{i}}(W;V)=
\{
{\mbox{\boldmath{$v$}}}_1 \otimes {\tilde{e}}^{i}_1, \ldots, {\mbox{\boldmath{$v$}}}_d \otimes {\tilde{e}}^{i}_1,
\ldots,
{\mbox{\boldmath{$v$}}}_1 \otimes {\tilde{e}}^{i}_{m_i}, \ldots, {\mbox{\boldmath{$v$}}}_d \otimes {\tilde{e}}^{i}_{m_i}
\}.$$ We denote by $H_*(W;V_\rho)$ the homology group and call it [*the twisted homology group*]{} and say that $\rho$ is acyclic if the twisted homology group vanishes. Regarding $C_*(W; V_\rho)$ as a based chain complex, we define the Reidemeister torsion for $W$ and an acyclic representation $(V, \rho)$ as the torsion of $C_*(W;V_\rho)$, [i.e.,]{}$$\label{eqn:def_RtorsionCW}
\mathrm{Tor}(W; V_\rho) = \mathrm{Tor}(C_*(W;V_\rho), {\mathbf{c}^{*}}(W;V))
\in {\mathbb{F}}^*$$ up to a factor in $\{\pm \det(\rho(\gamma)) \,|\, \gamma \in \pi_1(W)\}$ since we have many choices of lifts ${\tilde{e}}^{i}_j$ and orders and orientations of cells $e^{i}_j$. We call $\mathrm{Tor}(W; V_\rho)$ the Reidemeister torsion of $W$ and a ${\mathrm{GL}}(V)$-representation $\rho$.
We mention some well–definedness of the torsion :
- The acyclicity of $C_*(W;V_\rho)$ implies that the Euler characteristic of $W$ is zero. Then the torsion of $C_*(W;V_\rho)$ is independent of the choice of a basis in $V$.
- If we choose an $\mathrm{SL}(V)$-representation $\rho$ with an even dimensional $V$, then the Reidemeister torsion $\mathrm{Tor}(W; V_\rho)$ has no indeterminacy.
- The Reidemeister torsion has an invariance under the conjugation of representations.
The following lemma for torsion will be needed to derive our surgery formula:
\[lemma:MultLemma\] Let $0 \to (C'_*, \bar{\mathbf{c}}^{*}) \to (C_*, {\mathbf{c}^{*}}) \to (C''_*, \bar{\bar{\mathbf{c}}}^{*}) \to 0$ be the short exact sequence of based chain complexes such that $[\bar{\mathbf{c}}^{i} \cup \bar{\bar{\mathbf{c}}}^{i} / {\mathbf{c}^{i}}] =1$ for all $i$. Suppose that any two of the complexes are acyclic. Then the third one is also acyclic and the torsion of the three complexes are well-defined. Furthermore we have the next equality: $$\mathrm{Tor}(C_*, {\mathbf{c}^{*}})
= (-1)^{\sum_{i \geq 0} \beta'_{i-1}\beta''_{i}}
\mathrm{Tor}(C'_*, \bar{\mathbf{c}}^{*}) \mathrm{Tor}(C''_*, \bar{\bar{\mathbf{c}}}^{*})$$ where $\beta'_i= \dim_{\mathbb{F}}\partial C'_{i+1}$ and $\beta''_i = \dim_{\mathbb{F}}\partial C''_{i+1}$.
We refer to Milnor’s survey [@Milnor:1966] and Turaev’s book [@Turaev:2000] for more details on Reidemeister torsion.
Higher dimensional Reidemeister torsion for ${\mathrm{SL}_{2}({\mathbb{C}})}$-representations
---------------------------------------------------------------------------------------------
We will consider a sequence of the Reidemeister torsion of a finite CW-complex $W$. This sequence corresponds to the sequence of the ${\mathrm{SL}_{n}({\mathbb{C}})}$-representations of $\pi_1(W)$, induced by an ${\mathrm{SL}_{2}({\mathbb{C}})}$-representation. Let $\rho$ be a homomorphism from $\pi_1(W)$ to ${\mathrm{SL}_{2}({\mathbb{C}})}$. Then the pair $({\mathbb{C}}^2, \rho)$ is an ${\mathrm{SL}_{2}({\mathbb{C}})}$-representation of $\pi_1(W)$ by the standard action of ${\mathrm{SL}_{2}({\mathbb{C}})}$ to ${\mathbb{C}}^2$. It is known that the pair of the symmetric product ${\mathop{\mathrm{Sym}^{n-1}({\mathbb{C}}^2)}}$ and the induced action by ${\mathrm{SL}_{2}({\mathbb{C}})}$ gives an $n$-dimensional irreducible representation of ${\mathrm{SL}_{2}({\mathbb{C}})}$. The symmetric product ${\mathop{\mathrm{Sym}^{n-1}({\mathbb{C}}^2)}}$ can be identified with the vector space $V_n$ of homogeneous polynomials on ${\mathbb{C}}^2$ with degree $n-1$, [i.e.,]{}$$V_n =
\mathrm{span}_{{\mathbb{C}}}\langle
z_1^{n-1}, z_1^{n-2}z_2, \ldots, z_1^{n-k-1} z_2^k, \ldots,z_1 z_2^{n-2}, z_2^{n-1}
\rangle$$ and the action of $A \in {\mathrm{SL}_{2}({\mathbb{C}})}$ is expressed as $$\label{eqn:action_SL2}
A \cdot p(z_1, z_2) = p (A^{-1} \left(\begin{smallmatrix} z_1 \\ z_2\end{smallmatrix} \right))
\quad \text{for} \quad p(z_1, z_2) \in V_n.$$ We write $(V_n, \sigma_n)$ for the representation given by the action of ${\mathrm{SL}_{2}({\mathbb{C}})}$ where $\sigma_n$ denotes the homomorphism from ${\mathrm{SL}_{2}({\mathbb{C}})}$ into ${\mathrm{GL}}(V_n)$. In fact the image of $\sigma_n$ is contained in ${\mathrm{SL}_{n}({\mathbb{C}})}$. We will see this after the definition of the higher dimensional Reidemeister torsion.
We write $\rho_n$ for the composition $\sigma_n \circ \rho$. Then we can take $V_n$ as a coefficient of a twisted chain complex for $W$ since the vector space $V_n$ is a right ${\mathbb{Z}}[\pi_1(W)]$-module of $\pi_1(W)$. We denote by $C_*(W; V_n)$ this twisted chain complex of $W$ defined by $(V_n, \rho_n)$. We will drop the subscript $\rho_n$ in the coefficient for simplicity when no confusion can arise.
When the twisted chain complex $C_*(W; V_n)$ is acyclic, we define the higher dimensional Reidemeister torsion for $W$ and $\rho_n$ as ${\mathop{\mathrm{Tor}}\nolimits (W;V_n)}$ and denote by ${\mathop{\mathrm{Tor}}\nolimits (W;\rho_n)}$ since the coefficient $V_n$ of $C_*(W; V_n)$ is determined by $\rho$ and $n$.
Increasing $n$ to infinity, we obtain the sequence of the Reidemeister torsion ${\mathop{\mathrm{Tor}}\nolimits (W;V_n)}$ when $C_*(W; V_n)$ is always acyclic. We will observe the asymptotic behaviors of these sequences for Seifert fibered spaces in the subsequent sections.
We also review eigenvalues of the image $\sigma_n (A)$ for $A \in {\mathrm{SL}_{2}({\mathbb{C}})}$. Let $a^{\pm 1}$ be the eigenvalues of $A$. By direct calculation, we can see that the eigenvalues of $\sigma_n(A) \in {\mathrm{SL}_{n}({\mathbb{C}})}$ are given by $a^{-n+1}, a^{-n+3}, \ldots, a^{n-1}$, [i.e.,]{} the weight space of $\sigma_n$ is $\{-n+1, -n+3, \ldots, n-1\}$ and the multiplicity of each weight is $1$.
\[remark:weight\] If $n > 1$ is even (resp. odd), the eigenvalues of $\sigma_n(A)$ are the odd (resp. even) powers, by odd (resp. even) integers from $1$ (resp. $0$) to $n-1$, of the eigenvalues of $A$. This implies that $\det \sigma_n(A) = 1$ for any $A \in {\mathrm{SL}_{2}({\mathbb{C}})}$ and $n \geq 1$.
Surgery formula for the asymptotics of higher dimensional Reidemeister torsion {#sec:surgery_formula}
==============================================================================
We will give a surgery formula to observe the asymptotic behavior of the sequence given by the higher dimensional Reidemeister torsions. Our surgery formula is based on Lemma \[lemma:MultLemma\] (Multiplicativity Lemma) in Section \[section:Rtorsion\]. We will consider a connected compact orientable $3$-manifold ${M}$ with torus boundary $\partial {M}= T^2_1 \cup \ldots \cup T^2_m$. We denote a pair of meridian and longitude on $T^2_j$ by $(q_j, h_j)$, [i.e.,]{}$\pi_1(T^2_j) = \langle q_j, h_j \,|\, [q_j, h_j] = 1\rangle$. By Dehn filling with slopes $\alpha_1 / \beta_1, \ldots, \alpha_m / \beta_m$, we obtain a closed $3$-manifold ${M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}}$. Here a slope $\alpha_j / \beta_j$ is the unoriented isotopy class of the essential simple loop $\alpha_j q_j + \beta_j h_j$ on the $j$-th boundary component $T^2_j$, [i.e.,]{}$${M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}}= {M}\cup (\cup_{j=1}^m D^2_j \times S^1_j)
\quad \text{where}\quad \partial D^2_j \times \{*\} \sim \alpha_j q_j + \beta_j h_j \quad (\forall j).$$
Our purpose is to express the higher dimensional Reidemeister torsion of resulting manifolds ${M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}}$ by those of ${M}$ and solid tori $S_j = D^2_j \times S^1_j$ $(j=1, \ldots, m)$. We start with a homomorphism $\rho$ from $\pi_1({M})$ to ${\mathrm{SL}_{2}({\mathbb{C}})}$. When $\rho$ satisfies that the equations $\rho(q_j)^{\alpha_j} \rho(h_j)^{\beta_j} = {I}$, it extends to a homomorphism of the fundamental group of the resulting manifold and also defines the higher dimensional representations $\rho_n$ of $\pi_1({M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}})$. Then we can consider the short exact sequence with the coefficient $V_n$: $$\label{eqn:MayerVietorisV_n}
0 \to \oplus_{j=1}^m C_*(T^2_j;V_n)
\to C_*({M};V_n) \oplus (\oplus_{j=1}^m C_*(S_j;V_n))
\to C_*({M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}};V_n )
\to 0.$$
If the left and middle parts in the short exact sequence are acyclic, then the higher dimensional Reidemeister torsion of ${M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}}$ is expressed as $$\label{eqn:HighRTorSurgeredMfd}
{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}};\rho_n)}
=
\pm
{\mathop{\mathrm{Tor}}\nolimits ({M};\rho_n)} \cdot
\prod_{j=1}^m {\mathop{\mathrm{Tor}}\nolimits (S_j;\rho_n)} \cdot
\prod_{j=1}^m {\mathop{\mathrm{Tor}}\nolimits (T^2_j;\rho_n)}^{-1}.$$ by Lemma \[lemma:MultLemma\] (Multiplicativity Lemma).
We have seen that if $n$ is odd, then the image $\sigma_n(A)$ always has the eigenvalue $1$ for any $A \in {\mathrm{SL}_{2}({\mathbb{C}})}$. This implies that the twisted chain complex $C_*(S_j;V_n)$ is never acyclic when $n=2N-1$ (this will be seen in the following Subsection \[subsec:acyc\_condition\]). Hence we will focus on even dimensional representations $\rho_{2N}$ to apply Multiplicativity Lemma for acyclic Reidemeister torsions.
In Subsection \[subsec:acyc\_condition\], we will give equivalent conditions for the twisted chain complexes for $T^2_j$ and $S_j$ to be acyclic for all $2N$. At least, we have to work on our surgery formula under the resulting conditions. We will derive from Eq. a surgery formula for the asymptotic behaviors of the sequences obtained by the higher dimensional Reidemeister torsion of ${M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}}$ and ${M}$ in Subsection \[subsec:surgery\_formula\]. The last Subsection \[subsec:example\_torusknots\] gives the example of our surgery formula in the case that ${M}$ is a torus knot exterior.
Acyclicity conditions for the boundary and solid tori {#subsec:acyc_condition}
-----------------------------------------------------
First, we review the twisted chain complexes of $T^2$ and $D^2 \times S^1$. The torsion of $D^2 \times S^1$ coincides with that of core $\{0\} \times S^1$ since they are simple homotopy equivalent. We consider the twisted chain complexes of $S^1$ instead of $D^2 \times S^1$. Under the cell decomposition $$S^1 = e^0 \cup e^1 \quad \text{and} \quad T^2 = e^0 \cup e^1_1 \cup e^1_2 \cup e^2,$$ the twisted chain complexes with the coefficient $V_n$ of $T^2$ and $S^1$ are described as follows: $$\begin{aligned}
C_*(S^1; V_n) &\co
0 \to C_1(S^1 ;V_n)=V_n \xrightarrow{L - {I}} C_0(S^1;V_n)=V_n \to 0, \\
C_*(T^2; V_n) &\co
0 \to C_2(T^2;V_n) = V_n \xrightarrow{{\partial_{2}}} C_1(T^2 ;V_n)=V_n^{\oplus 2} \xrightarrow{{\partial_{1}}} C_0(T^2 ;V_n)=V_n \to 0, \\
& \qquad
{\partial_{2}}=
\begin{pmatrix}
-H + {I}\\
Q - {I}\end{pmatrix}, \quad
{\partial_{1}} = (Q -{I}, H-{I})\end{aligned}$$ where $L$, $Q$ and $H$ denote ${\mathrm{SL}_{n}({\mathbb{C}})}$-matrices corresponding the simple closed loops $\ell = e^0 \cup e^1$ in $S^1$, $q = e^0 \cup e^1_1$ and $h = e^0 \cup e^1_2$ in $T^2$.
The twisted homology group $H_1(S^1;V_n)$ is the eigenspace of $L$ for the eigenvalue $1$. As mentioned in Remark \[remark:weight\], if $n$ is odd, then the ${\mathrm{SL}_{n}({\mathbb{C}})}$-matrix $L$ always has the eigenvalue $1$. Hence $C_*(S^1;V_{2N-1})$ can not be acyclic. Here and subsequently, we focus only on the twisted chain complexes given by the even dimensional vector spaces $V_{2N}$.
It is known that every abelian subgroup in ${\mathrm{SL}_{2}({\mathbb{C}})}$ is moved by conjugation into either the maximal abelian subgroups ${\mathrm{Hyp}}$ or ${\mathrm{Para}}$: $${\mathrm{Hyp}}:= \left\{\left.
\begin{pmatrix}
z & 0 \\
0 & z^{-1}
\end{pmatrix}
\,\right|\,
z \in {\mathbb{C}}\setminus\{0\}
\right\},\,
{\mathrm{Para}}:=
\left\{\left.
\begin{pmatrix}
\pm 1 & w \\
0 & \pm 1
\end{pmatrix}
\,\right|\,
w \in {\mathbb{C}}\right\}.$$
Since the conjugation of representations induces an isomorphism between twisted homology groups, we can assume that ${\mathrm{SL}_{2}({\mathbb{C}})}$-representations send $\pi_1(S^1)$ and $\pi_1(T^2)$ into ${\mathrm{Hyp}}$ or ${\mathrm{Para}}$.
We describe the acyclicity conditions for the twisted chain complexes $C_*(S^1;V_{2N})$ and $C_*(T^2;V_{2N})$ by the terminologies of the ${\mathrm{SL}_{2}({\mathbb{C}})}$-matrices corresponding to generators of $\pi_1(S^1)$ and $\pi_1(T^2)$.
\[prop:acyclicity\_circle\] Let $\rho$ be an ${\mathrm{SL}_{2}({\mathbb{C}})}$-representation of $\pi_1(S^1) = \langle \ell \rangle$. The composition $\rho_{2N} = \sigma_{2N} \circ \rho$ is acyclic for all $N \geq 1$ if and only if $\rho(\ell)$ is neither of odd order nor parabolic with the trace $2$.
The dimensions of $H_0(S^1;V_{2N})$ and $H_1(S^1;V_{2N})$ are same since the Euler characteristic of $S^1$ is zero. The homology group $H_1(S^1;V_{2N})$ is the eigenspace of $L=\rho_{2N}(\ell)$ for the eigenvalue $1$. The acyclicity of $\rho_{2N}$ is equivalent for the ${\mathrm{SL}_{2N}({\mathbb{C}})}$-matrix $L$ not to have the eigenvalue $1$.
If $\rho(\ell) \in {\mathrm{Hyp}}$ is not of odd order, then the eigenvalues of $L$ forms $\{e_{\ell}^{\pm (2k-1)} \,|\, k=1, \ldots, N\}$ where $e_\ell^{\pm 1}$ are the eigenvalues of $\rho(\ell)$. Since $e_\ell$ is not of odd order, the ${\mathrm{SL}_{2N}({\mathbb{C}})}$-element $L$ does not have the eigenvalue $1$ for all $N$. Thus $\rho_{2N}$ is acyclic for all $N$. If $\rho(\ell) \in {\mathrm{Para}}$ has the trace $-2$, then the eigenvalue of $L$ is just $-1$ for all $N$. The ${\mathrm{SL}_{2N}({\mathbb{C}})}$-representation $\rho_{2N}$ is also acyclic for all $N$.
Conversely suppose that $\rho(\ell)$ has the order $2 k_\ell -1$. Then the set of eigenvalues of $L$ contain $1$ when $N \geq k_\ell$. The twisted homology group $H_1(S^1;V_{2N})$ is not trivial for $N \geq k_\ell$. Suppose that $\rho(\ell) \in {\mathrm{Para}}$ has the trace $2$. Then the eigenvalue of $L$ is just $1$ for all $N$. The twisted homology group $H_1(S^1;V_{2N})$ is not trivial for all $N$.
\[prop:acyclicity\_torus\] Let $\rho$ be an ${\mathrm{SL}_{2}({\mathbb{C}})}$-representation of $\pi_1(T^2) = \langle q, h \,|\, [q, h]=1\rangle$. The composition $\rho_{2N} = \sigma_{2N} \circ \rho$ is acyclic if and only if either $\rho(q)$ or $\rho(h)$ is neither of odd order nor parabolic with the trace $2$.
The homology group $H_2(T^2;V_{2N})$ is generated by the common eigenvectors of $Q=\rho_{2N}(q)$ and $H=\rho_{2N}(h)$ for the eigenvalue $1$. Since the Euler characteristic of $T^2$ is zero, by Poincaré duality, the twisted homology group $H_*(T^2;V_{2N})$ vanishes if and only if $H_2(T^2;V_{2N}) = {\mbox{\boldmath{$0$}}}$. The acyclicity of $\rho_{2N}$ is equivalent to exist no common eigenvectors of $Q$ and $H$ for the eigenvalue $1$.
If $\rho(q)$ is neither of odd order nor parabolic with the trace $2$, then $Q$ does not have the eigenvalue $1$ for all $N$ as in the proof of Proposition \[prop:acyclicity\_circle\]. We have no common eigenvectors of $Q$ and $H$ for the eigenvalue $1$. Hence $\rho_{2N}$ is acyclic for all $N$.
Conversely suppose that $\rho(q)$ and $\rho(h)$ have the orders $2 k_q -1$ and $2k_h-1$. Then at the weight $(2k_q -1)(2k_h-1)$, we have a common eigenvector of $Q$ and $H$ for the eigenvalue $1$ when $N$ is sufficiently large. Thus $\rho_{2N}$ is not acyclic for sufficiently large $N$. Suppose that $\rho(q)$ and $\rho(h)$ are parabolic with the trace $2$. Then $Q$ and $H$ are always upper triangular matrix whose all diagonal entries are $1$. We have a common eigenvector of $Q$ and $H$ for the eigenvalue $1$. Hence $\rho_{2N}$ is not acyclic for all $N$.
One can show that the ${\mathrm{SL}_{2N-1}({\mathbb{C}})}$-representation $\rho_{2N-1}$ of $\pi_1(T^2)$ is not acyclic for all $N \geq 1$ by the similar argument in Proposition \[prop:acyclicity\_torus\].
Surgery formula for the asymptotic behaviors {#subsec:surgery_formula}
--------------------------------------------
We show a surgery formula for the asymptotic behaviors of the higher dimensional Reidemeister torsions of a closed $3$-manifold ${M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}}= {M}\cup (\cup_{j=1}^m) D^2_j \times S^1_j$. To apply Lemma \[lemma:MultLemma\] (Multiplicativity Lemma) for acyclic chain complexes, we assume that the following acyclicity conditions for the twisted chain complexes $\partial {M}= \cup_{j=1}^m T^2_j$ and solid tori $S_j = D^2_j \times S^2_j$ with the presentations of fundamental groups: $$\pi_1(T^2_j) = \langle q_j, h_j \,|\, [q_j, h_j] = 1 \rangle, \quad
\pi_1(S_j) = \langle \ell_j \rangle.$$
\[def:acyclicity\_conditions\] Let $\rho$ be a homomorphism from $\pi_1({M})$ to ${\mathrm{SL}_{2}({\mathbb{C}})}$ such that $\rho(q_j^{\alpha_j} h_j^{\beta_j}) = {I}$ for all $j=1, \ldots m$. We use the same symbol $\rho$ for the induced homomorphism of $\pi_1({M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}})$ and assume that for all $j = 1, \ldots, m$
1. \[item:acyclicity\_boundary\] either $\rho(q_j)$ or $\rho(h_j)$ is of even order and;
2. \[item:acyclicity\_solid\_torus\] the order of $\rho(\ell_j)$ is also even.
We will call the above conditions & .
The acyclicity conditions guarantee that all twisted chain complexes of $T^2_j$ and $S_j$ are acyclic. Our acyclicity conditions are more restricted as compared with the conditions in Propositions \[prop:acyclicity\_torus\] & \[prop:acyclicity\_circle\]. However in the case that the resulting manifold ${M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}}$ is a Seifert fibered space, it is reasonable to assume our conditions as seen in Section \[sec:asymptotics\_Seifert\].
Under the acyclic conditions, if an ${\mathrm{SL}_{2N}({\mathbb{C}})}$-representation $\rho_{2N}$ of $\pi_1({M})$ is acyclic, then we can express the higher dimensional Reidemeister torsion of ${M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}}$ as $$\label{eqn:torsion_explicit_product}
{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}};\rho_{2N})}
= {\mathop{\mathrm{Tor}}\nolimits ({M};\rho_{2N})} \cdot \prod_{j=1}^m {\mathop{\mathrm{Tor}}\nolimits (S_j;\rho_{2N})} \cdot \prod_{j=1}^m {\mathop{\mathrm{Tor}}\nolimits (T^2_j;\rho_{2N})}^{-1}.$$ by applying Lemma \[lemma:MultLemma\] (Multiplicativity Lemma). Note that every integer $\beta'_i$ in Lemma \[lemma:MultLemma\] is even from the acyclicity of $C_*(T^2_j;V_{2N})$.
Then the asymptotics of $\log|{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}};\rho_{2N})}|$ for $2N$ is determined by that of $\log|{\mathop{\mathrm{Tor}}\nolimits ({M};\rho_{2N})}|$ as follows.
\[thm:surgery\_formula\] Let $\rho$ be an ${\mathrm{SL}_{2}({\mathbb{C}})}$-representation of $\pi_1({M})$ satisfying $\rho(q_j^{\alpha_j} h_j^{\beta_j}) = {I}$ and the acyclicity conditions in Definition \[def:acyclicity\_conditions\]. Suppose that $\rho_{2N}$ of $\pi_1({M})$ is acyclic for all $N$. Then the asymptotics of $\log|{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}};\rho_{2N})}|$ is expressed as follows:
1. $\displaystyle{
\lim_{N \to \infty}
\frac{
\log|{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}};\rho_{2N})}|
}{
(2N)^2
}
=
\lim_{N \to \infty}
\frac{
\log|{\mathop{\mathrm{Tor}}\nolimits ({M};\rho_{2N})}|
}{
(2N)^2
},
}$
2. $\displaystyle{
\lim_{N \to \infty}
\frac{
\log|{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}};\rho_{2N})}|
}{
2N
}
=
\lim_{N \to \infty}
\frac{
\log|{\mathop{\mathrm{Tor}}\nolimits ({M};\rho_{2N})}|
}{
2N
}
-\log 2 \sum_{j=1}^m \frac{1}{\lambda_j}
}$
where $2\lambda_j$ is the order of $\rho(\ell_j)$ and $\ell_j$ is the homotopy class of $\{0\} \times S^1_j \subset S_j$.
By Eq. , the logarithm $\log |{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}};\rho_{2N})}|$ is expressed as $$\begin{aligned}
&\log |{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}};\rho_{2N})}| \\
& \quad= \log|{\mathop{\mathrm{Tor}}\nolimits ({M};\rho_{2N})}|
+ \sum_{J=1}^m \log|{\mathop{\mathrm{Tor}}\nolimits (S_j;\rho_{2N})}|
- \sum_{j=1}^m \log|{\mathop{\mathrm{Tor}}\nolimits (T^2_j;\rho_{2N})}|.
\end{aligned}$$ Applying the following Propositions \[prop:torsion\_torus\] & \[prop:torsion\_circle\], we obtain Theorem \[thm:surgery\_formula\].
\[prop:torsion\_torus\] Let $\rho$ be an ${\mathrm{SL}_{2}({\mathbb{C}})}$-representation of $\pi_1(T^2)=\langle q, h \,|\, [q, h] =1\rangle$ such that either $\rho(q)$ or $\rho(h)$ is of even order. Then ${\mathop{\mathrm{Tor}}\nolimits (T^2;\rho_{2N})} = 1$ for all $N \geq 1$.
This follows from the direct calculation.
\[prop:torsion\_circle\] Let $\rho$ be an ${\mathrm{SL}_{2}({\mathbb{C}})}$-representation of $\pi_1(S^1)=\langle \ell \rangle$ such that $\rho(\ell)$ has an even order. Then we have the following limits of $\log|{\mathop{\mathrm{Tor}}\nolimits (S^1;\rho_{2N})}|$:
1. \[item:torsion\_circle\_square\_order\] $\displaystyle{\lim_{N \to \infty}
\frac{\log|{\mathop{\mathrm{Tor}}\nolimits (S^1;\rho_{2N})}|}{(2N)^2} = 0}$,
2. \[item:leading\_coeff\_circle\] $\displaystyle{\lim_{N \to \infty}
\frac{\log|{\mathop{\mathrm{Tor}}\nolimits (S^1;\rho_{2N})}|}{2N} = \frac{-\log 2}{\lambda}}$
where $2\lambda$ is the order of $\rho(\ell)$.
We begin with computing the Reidemeister torsion ${\mathop{\mathrm{Tor}}\nolimits (S^1;\rho_{2N})}$. We can express the Reidemeister torsion as $
{\mathop{\mathrm{Tor}}\nolimits (S^1;\rho_{2N})} = \det (\rho_{2N}(\ell) - {I})^{-1}.
$ The set of the eigenvalues of $\rho_{2N}(\ell)$ is obtained from the eigenvalues $e^{\pm \pi \eta \sqrt{-1}/\lambda}$ of $\rho(\ell)$, where $\eta$ is odd and $\eta$ and $\lambda$ are coprime. It turns into $\{e^{\pm \pi (2k-1) \eta \sqrt{-1}/\lambda} \,|\, k = 1, \ldots, N\}$. Hence the Reidemeister torsion ${\mathop{\mathrm{Tor}}\nolimits (S^1;\rho_{2N})}$ turns out $$\begin{aligned}
{\mathop{\mathrm{Tor}}\nolimits (S^1;\rho_{2N})}
&=
\prod_{k=1}^m \{(e^{\pi (2k-1) \eta \sqrt{-1} / \lambda} -1)((e^{-\pi (2k-1) \eta \sqrt{-1} / \lambda} -1))\}^{-1} \\
&=
\prod_{k=1}^N \left(2 \sin \frac{\pi (2k-1)\eta}{2\lambda} \right)^{-2}.
\end{aligned}$$ The logarithm $\log |{\mathop{\mathrm{Tor}}\nolimits (S^1;\rho_{2N})}|$ is expressed as $$\label{eqn:log_torsion_circle}
\log |{\mathop{\mathrm{Tor}}\nolimits (S^1;\rho_{2N})}|
= 2N \log 2^{-1} + 2 \sum_{k=1}^N \log \left|\sin \frac{\pi (2k-1)\eta}{2\lambda} \right|^{-1}.$$ We can now proceed to compute the limits & .
From the following inequality $$\left|\sin \frac{\pi}{2 \lambda} \right|
\leq \left| \sin \frac{\pi (2k-1)\eta}{2 \lambda} \right|
\leq 1$$ it follows that $$\label{eqn:ineq_log_circle}
N \log \left|\sin \frac{\pi}{2 \lambda} \right|^{-1}
\geq
\sum_{k=1}^N \log \left| \sin \frac{\pi (2k-1)\eta}{2 \lambda} \right|^{-1}
\geq 0.$$ By the inequality and explicit form of $\log |{\mathop{\mathrm{Tor}}\nolimits (S^1;\rho_{2N})}|$, we can assert $$\lim_{N \to \infty}
\frac{\log|{\mathop{\mathrm{Tor}}\nolimits (S^1;\rho_{2N})}|}{(2N)^2} = 0.$$
We can express the limit as $$\label{eqn:leading_term_circle}
\lim_{N \to \infty}
\frac{\log|{\mathop{\mathrm{Tor}}\nolimits (S^1;\rho_{2N})}|}{2N}
= \log 2^{-1} +
\lim_{N \to \infty}
\frac{1}{N} \sum_{k=1}^{N} \log \left| \sin \frac{\pi (2k-1)\eta}{2\lambda} \right|^{-1}.$$ The second term in the right hand side of can be rewritten as $$\lim_{N \to \infty}
\frac{1}{N} \sum_{k=1}^{N} \log \left| \sin \frac{\pi (2k-1)\eta}{2\lambda} \right|^{-1}
=
\lim_{N \to \infty}
\frac{1}{N} \sum_{k=1}^{N}
\log \left| \sin \left( \frac{\pi \eta}{2\lambda} + \frac{\pi (k-1)\eta}{\lambda} \right)\right|^{-1}$$ The sequence $\{\log \left| \sin \left(\pi \eta / (2\lambda) + \pi (k-1)\eta /\lambda \right)\right|^{-1}\}_{k=1, 2, \ldots}$ has the minimum period $\lambda$ since $\eta$ and $\lambda$ are coprime. By Lemma \[lemma:limit\_average\], we can rewrite as $$\lim_{N \to \infty}
\frac{1}{N} \sum_{k=1}^{N}
\log \left| \sin \left( \frac{\pi \eta}{2\lambda} + \frac{\pi (k-1)\eta}{\lambda} \right)\right|^{-1}
=
\frac{1}{\lambda} \sum_{k=1}^\lambda
\log \left| \sin \left( \frac{\pi \eta}{2\lambda} + \frac{\pi (k-1)\eta}{\lambda} \right)\right|^{-1}.$$ The right hand side of turns into $$\begin{aligned}
\log 2^{-1} +
\frac{1}{\lambda} \sum_{k=1}^{\lambda}
\log \left| \sin \left( \frac{\pi \eta}{2\lambda} + \frac{\pi (k-1)\eta}{\lambda} \right)\right|^{-1}
&=
\frac{1}{\lambda}
\log \prod_{k=1}^{\lambda}
\left| 2\sin \left( \frac{\pi \eta}{2\lambda} + \frac{\pi (k-1)\eta}{\lambda} \right)\right|^{-1} \\
&=
\frac{1}{\lambda}
\log \left| 2\sin \left( \frac{\pi \eta}{2} \right)\right|^{-1}
\end{aligned}$$ by $|2\sin (n\theta)| = \prod_{k=0}^{n-1} |2 \sin (\theta + k\pi/n)|$. Therefore we obtain the limit $$\lim_{N \to \infty}
\frac{\log|{\mathop{\mathrm{Tor}}\nolimits (S^1;\rho_{2N})}|}{2N}
= \frac{-\log 2}{\lambda}$$ since $\eta$ is odd.
We obtain $|2\sin (n\theta)| = \prod_{k=0}^{n-1} |2 \sin (\theta + k\pi/n)|$ from substituting $z = e^{-2\theta \sqrt{-1}}$ to $|z^n - 1| = \prod_{k=0}^{n-1} | z - e^{2\pi k\sqrt{-1} / n}|$.
\[lemma:limit\_average\] Let $\{a_k\,|\, a_k \in {\mathbb{R}}\}_{k \geq 1}$ be a sequence such that $a_k \geq 0$ and $a_{k+N_0} = a_k$. Then we have the following limit: $$\lim_{N \to \infty}
\frac{a_1 + \cdots + a_N}{N}
= \frac{
a_1 + \cdots + a_{N_0}}{N_0}.$$
It follows that $
\left[\frac{N}{N_0}\right] \sum_{k=1}^{N_0} a_k
\leq
\sum_{k=1}^{N} a_k
\leq
\left(\left[\frac{N}{N_0}\right] + 1\right)
\sum_{k=1}^{N_0} a_k
$ where $[\,x\,]$ denotes the maximal integer less than or equal to $x$. Note that $\frac{N}{N_0} - 1 < \left[\frac{N}{N_0}\right] \leq \frac{N}{N_0}$.
In Lemma \[lemma:limit\_average\], it is not required that the period $N_0$ is minimum. However we have the same average for any period $N_0$.
Example for Dehn fillings of torus knot exteriors {#subsec:example_torusknots}
-------------------------------------------------
We give examples of Theorem \[thm:surgery\_formula\] for integral surgeries along torus knots in $S^3$. Let ${M}$ be the $(p, q)$-torus knot exterior which is obtained by removing an open tubular neighbourhood of the knot from $S^3$. After gluing a solid torus along the slope $1/n$ $(n \in {\mathbb{Z}})$ on $\partial {M}$, we have an integral homology $3$-sphere ${M}(\frac{1}{n})$. Since we consider the $(p, q)$-torus knot exterior, the resulting manifold ${M}(\frac{1}{n})$ is a Brieskorn homology $3$-sphere of index $(p, q, pqn \pm 1)$. Here the sign in $pqn \pm 1$ depends on the orientation of the preferred longitude on $\partial {M}$.
The $(p, q)$-torus knot group admits the following presentation: $$\pi_1({M}) =
\langle x, y \,|\, x^p=y^q \rangle.$$ In this presentation, we can express a pair of meridian $m$ and longitude $\ell$ as $$m = x^{-u} y^v, \quad \ell = m^{pq}x^{-p}$$ where $u$ and $v$ are integers satisfying that $pv-qu=1$. Then the Brieskorn homology sphere ${M}(\frac{1}{n})$ has the index $(p, q, pqn+1)$. We consider irreducible ${\mathrm{SL}_{2}({\mathbb{C}})}$-representations $\rho$ of $\pi_1({M})$ such that $\rho(m \ell^{n}) = {I}$, [i.e.,]{}they extend to irreducible ${\mathrm{SL}_{2}({\mathbb{C}})}$-representations of $\pi_1({M}(\frac{1}{n})) = \langle x, y \,|\, x^p=y^q, m\ell^n=1 \rangle$. Here irreducible means that there are no common non–trivial eigenvectors among all elements in $\rho(\pi_1({M}))$. Under the assumption of irreducibility for $\rho$, the central element $x^p (=y^q)$ must be sent to $\pm {I}$. The requirement that $\rho(m \ell^n) = {I}$ turns into $\rho(m)^{npq+1} = \pm {I}$. Hence we have the constrains on the order of $\rho(x)$, $\rho(y)$ and $\rho(m)$ for every irreducible ${\mathrm{SL}_{2}({\mathbb{C}})}$-representation of $\pi_1({M}(\frac{1}{n}))$.
The conjugacy classes of irreducible ${\mathrm{SL}_{2}({\mathbb{C}})}$-representations of $\pi_1({M}(\frac{1}{n}))$ form a finite set. Each member of the finite set corresponds to a triple of integers. This was shown by D. Johnson [@Johnson:unpublished] and he also gave the explicit form of Reidemeister torsion for acyclic ${\mathrm{SL}_{2}({\mathbb{C}})}$-representations as follows.
The conjugacy classes of irreducible ${\mathrm{SL}_{2}({\mathbb{C}})}$-representations $\rho$ of $\pi_1({M}(\frac{1}{n}))$ are given by triples $(a, b, c)$ such that
1. $0 < a < p$, $0 < b < q$, $a \equiv b \quad \mathrm{mod}\, 2$,
2. $0 < c < r =|pqn+1|$, $c \equiv na \quad \mathrm{mod}\, 2$,
3. ${{\rm tr}\,}\rho(x) = 2 \cos \pi a / p$,
4. ${{\rm tr}\,}\rho(y) = 2 \cos \pi b / q$,
5. ${{\rm tr}\,}\rho(m) = 2 \cos \pi c / r$.
The Reidemeister torsion is given by $${\mathop{\mathrm{Tor}}\nolimits ({M}(\hbox{$\frac{1}{n}$});\rho)} =
\begin{cases}
2^{-4} \sin^{-2} \frac{\pi a}{2p} \sin^{-2} \frac{\pi b}{2q} \sin^{-2} \frac{\pi(cpq - r)}{2r}
& a \equiv b \equiv 1, c \equiv n \quad \mathrm{mod}\,2\\
\text{non-acyclic}
& a \equiv b \equiv 0\, or\, c \not \equiv n \quad \mathrm{mod}\,2
\end{cases}$$ for $\rho \in (a, b, c)$.
In the remainder of this subsection, we denote by $(a, b, c)$ the corresponding conjugacy class of irreducible ${\mathrm{SL}_{2}({\mathbb{C}})}$-representations. We also refer to [@Freed:Brieskorn] for the Reidemeister torsion of Brieskorn homology $3$-spheres.
The parameters $a$ and $b$ determine the image of the central element $x^p (=y^q)$ by $(-{I})^a (=(-I)^b)$.
To apply Theorem \[thm:surgery\_formula\], we need to find
- a condition on $(a, b, c)$ for all $\rho_{2N}|_{\pi_1({M})}$ to be acyclic and
- the orders of ${\mathrm{SL}_{2}({\mathbb{C}})}$-elements in the acyclicity conditions (Definition \[def:acyclicity\_conditions\]).
The author has shown in [@yamaguchi:RtorTorusKnots Proposition 3.1] that $\rho|_{\pi_1({M})}$ induces an acyclic ${\mathrm{SL}_{2N}({\mathbb{C}})}$-representation for all $N$ if and only if the parameters $a$ and $b$ of $\rho|_{\pi_1({M})}$ satisfy that $a \equiv b \equiv 1$ mod $2$.
Since the surgery slope for ${M}(\frac{1}{n}) = {M}\cup D^2 \times S^1$ is $1/n$, the homotopy class of the core $\{0\} \times S^1$ is given by $\ell^{\pm 1}$ in $\pi_1({M}(\frac{1}{n}))$. To check the acyclicity conditions, we only need to find the order of $\rho(\ell)$. Let $\rho$ be in the conjugacy class $(a, b, c)$ such that $a \equiv b \equiv 1$ mod $2$. Then it follows from $c \equiv na$ mod 2 that $$\rho(\ell)^r = (-{I})^{pqc - r} = -{I}.$$ Hence the eigenvalues of $\rho(\ell)$ are given by $e^{\pm \pi \eta \sqrt{-1}/r}$ for some odd integer $\eta$, which shows that $\rho(\ell)$ has an even order.
Let us apply Theorem \[thm:surgery\_formula\] to the Brieskorn homology $3$-sphere ${M}(\frac{1}{n})$. We obtain $$\begin{aligned}
\lim_{N \to \infty}
\frac{
\log| {\mathop{\mathrm{Tor}}\nolimits ({M}(\frac{1}{n});\rho_{2N})} |
}{
(2N)^2
}
&=
\lim_{N \to \infty}
\frac{
\log| {\mathop{\mathrm{Tor}}\nolimits ({M};\rho_{2N})} |
}{
(2N)^2
} \label{eqn:torus_knot_squared}\\
\lim_{N \to \infty}
\frac{
\log | {\mathop{\mathrm{Tor}}\nolimits ({M}(\frac{1}{n});\rho_{2N})} |
}{
2N
}
&=
\lim_{N \to \infty}
\frac{
\log | {\mathop{\mathrm{Tor}}\nolimits ({M};\rho_{2N})} |
}{
2N
}
- \frac{\log 2}{r'}
\label{eqn:torus_knot_leading}\end{aligned}$$ where $2r'$ is the order of $\rho(\ell)$. Note that it is seen from the g.c.d. $(p, r) = (q, r) = 1$ that $r' = r / (c, r)$.
It has shown in [@yamaguchi:RtorTorusKnots Theorem 4.2] that the right hand side in vanishes. The higher Reidemeister torsion of the torus knot exterior ${M}$ is expressed as, by [@yamaguchi:RtorTorusKnots Proposition 4.1], $${\mathop{\mathrm{Tor}}\nolimits ({M};\rho_{2N})}
=
\frac{2^{2N}}{
\prod_{k=1}^N 4^2 \sin^2 \frac{\pi (2k-1)a}{2p} \sin^2 \frac{\pi (2k-1)b}{2q}}.$$ By a similar argument to the proof of Proposition \[prop:torsion\_circle\], we can see that the limit in the right hand of turns out $$\lim_{N \to \infty}
\frac{
\log | {\mathop{\mathrm{Tor}}\nolimits ({M};\rho_{2N})} |
}{
2N
}
= \left( 1 - \frac{1}{p'} - \frac{1}{q'} \right)\log 2$$ where $p' = p / (p,a)$ and $q' = q / (q,b)$.
\[thm:asymptotics\_Brieskorn\] The growth of $\log |{\mathop{\mathrm{Tor}}\nolimits ({M}(\frac{1}{n});\rho_{2N})}|$ has the same order as $2N$ for any acyclic irreducible representation $\rho$ of $\pi_1({M}(\frac{1}{n}))$. Moreover if $\rho$ is contained in the conjugacy class $(a, b, c)$, then the leading coefficient in $2N$ converges as $$\lim_{N \to \infty}
\frac{
\log | {\mathop{\mathrm{Tor}}\nolimits ({M}(\frac{1}{n});\rho_{2N})} |
}{
2N
}
=
\left(
1 - \frac{1}{p'} - \frac{1}{q'} - \frac{1}{r'}
\right)\log 2$$ where $p' = p / (a, p)$, $q' = q / (b, q)$ and $r' = r / (c, r)$.
In the case that $(a, p) = (b, q) = (c, r) = 1$, the leading coefficient converges to the maximum $(1 - 1/p - 1/q - 1/r) \log2$
For more details on the limits of $\log | {\mathop{\mathrm{Tor}}\nolimits ({M};\rho_{2N})} | / (2N)$, we refer to [@Yamaguchi13:kokyuroku].
Asymptotics of higher dimensional Reidemeister torsion for Seifert fibered spaces {#sec:asymptotics_Seifert}
=================================================================================
We will apply Theorem \[thm:surgery\_formula\] to Seifert fibered spaces and study the asymptotic behaviors of their higher dimensional Reidemeister torsions. We will see the growth of the logarithm of the higher dimensional Reidemeister torsion has the same order as the dimension of representation and we will also give the explicit limit of the leading coefficient.
The limits of the leading coefficients are determined by each component in the ${\mathrm{SL}_{2}({\mathbb{C}})}$-representation space of the fundamental group of a Seifert fibered space, that is to say, we obtain a locally constant function on the ${\mathrm{SL}_{2}({\mathbb{C}})}$-representation space. From the invariance of Reidemeister torsion under the conjugation of representations, we also obtain a locally constant function on the character variety of the fundamental group of a Seifert fibered space.
We will focus on ${\mathrm{SU}(2)}$-character varieties for Seifert fibered homology spheres and describe explicit values of the locally constant functions. Our calculation shows that these locally constant functions take the maximum values on the top dimensional components and the explicit maximums are given by $-\chi \log 2$ where $\chi$ is the Euler characteristic of the base orbifold of a Seifert fibered homology sphere.
We start with a brief review on Seifert fibered spaces in Subsection \[subsec:Seifert\]. Subsection \[subsec:asymptotics\_Seifert\] shows the application of Theorem \[thm:surgery\_formula\] to Seifert fibered spaces. We will observe the relation between limits of the leading coefficients and components in ${\mathrm{SU}(2)}$-character varieties in Subsection \[subsec:leadingCoeff\_Seifert\].
Seifert fibered spaces {#subsec:Seifert}
----------------------
A Seifert fibered space is referred as an $S^1$-fibration over a closed $2$-orbifold. We consider the orientable Seifert fibered space given by the following Seifert index: $$\{b, (o, g); (\alpha_1, \beta_1), \ldots, (\alpha_m, \beta_m)\}.$$ where $\alpha_j \geq 2$ $(j=1,\ldots, m)$ and each pair of $\alpha_j$ and $\beta_j$ is coprime. For a Seifert index, we refer to [@NeumannJankins:Seifert; @Orlik:SeifertManifold].
We can regard a Seifert fibered space as an $S^1$-bundle over a closed orientable surface $\Sigma$ with $m+1$ exceptional fibers, where the genus of $\Sigma$ is $g$. From this viewpoint, we can decompose a Seifert fibered space into tubular neighbourhoods of exceptional fibers and their complement. Set $\Sigma_* = \Sigma \setminus \mathrm{int}(D^2_0 \cup \ldots \cup D^2_m)$ where $D^2_0, \ldots, D^2_m$ are disjoint disks in $\Sigma$. Let ${M}$ be the trivial $S^1$-bundle $\Sigma_* \times S^1$. We have a canonical decomposition of the Seifert fibered space as the following union of ${M}$ and solid tori: $$\begin{aligned}
&{M}\cup (S_0 \cup S_1 \cup \cdots \cup S_m)\\
&= {M\hbox{$\big(\frac{1}{-b}, \frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}}.\end{aligned}$$ The solid torus $S_0$ corresponds to the triviality obstruction $b$ and the others $S_j$ $(1 \leq j \leq m)$ correspond to the exceptional fibers with the index $(\alpha_j, \beta_j)$. Then the fundamental group of ${M\hbox{$\big(\frac{1}{-b}, \frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}}$ admits the following presentation: $$\begin{gathered}
\pi_1({M\hbox{$\big(\frac{1}{-b}, \frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}}) \\
= \langle
a_1, b_1, \ldots, a_g, b_g, q_1, \ldots. q_m, h \,|\,
[a_i, h] = [b_i, h] = [q_j, h] = 1, \\
q_j^{\alpha_j} h^{\beta_j} = 1,
q_1 \cdots q_m [a_1, b_1] \cdots [a_g, b_g] = h^b
\rangle\end{gathered}$$ where $a_i$ and $b_j$ correspond to generators of $\pi_1(\Sigma)$ and $q_j$ is the corresponding to the circle $\partial D^2_j \subset \Sigma$ and $h$ is the homotopy class of a regular fiber in ${M}$. Note that the presentation of $\pi_1({M})$ is given by $\langle a_1, b_1, \ldots, a_g, b_g, q_0, q_1, \ldots. q_m, h \,|\,
[a_i, h] = [b_i, h] = [q_j, h] = 1,
q_1 \cdots q_m [a_1, b_1] \cdots [a_g, b_g] = q_0
\rangle$.
We review the acyclicity of ${\mathrm{SL}_{n}({\mathbb{C}})}$-representations of the fundamental group of ${M}$. We are supposed to consider the sequences of Reidemeister torsions for acyclic chain complexes $C_*({M};V_{2N})$ ($N \geq 1$) derived from an ${\mathrm{SL}_{2}({\mathbb{C}})}$-representation $\rho$ of $\pi_1({M})$. It was shown in T. Kitano [@Kitano:RtorsionSeifertSL2] that $C_*({M};{\mathbb{C}}^n)$ is acyclic if and only if $\rho_n(h) = -{I}$ where $({\mathbb{C}}^n, \rho_n)$ is an ${\mathrm{SL}_{n}({\mathbb{C}})}$-representation of $\pi_1({M})$.
We also touch the first homology groups of Seifert fibered spaces since we will consider Seifert fibered homology spheres in Subsection \[subsec:leadingCoeff\_Seifert\]. It is known that the first homology group of ${M\hbox{$\big(\frac{1}{-b}, \frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}}$ is expressed as $H_1({M\hbox{$\big(\frac{1}{-b}, \frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}};{\mathbb{Z}}) \simeq {\mathbb{Z}}^{2g}\oplus T$ where $T$ is a finite abelian group with the order $\alpha_1 \cdots \alpha_m |b + \sum_{j=1}^m \beta_j / \alpha_j|$ if $b + \sum_{j=1}^m \beta_j / \alpha_j$ is not zero. In the case that $b + \sum_{j=1}^m \beta_j / \alpha_j = 0$, the homology group $H_1({M\hbox{$\big(\frac{1}{-b}, \frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}};{\mathbb{Z}})$ has the free rank $2g+1$. Hence, for any Seifert fibered homology sphere ${M\hbox{$\big(\frac{1}{-b}, \frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}}$, the genus of the base orbifold is zero and we have the equation: $$\alpha_1 \cdots \alpha_m \Big(b + \sum_{j=1}^{m} \frac{\beta_j}{\alpha_j}\Big) = 1,$$ in particular, which implies that $\alpha_j$ are pairwise coprime.
The asymptotic behavior of the higher dimensional Reidemeister torsions for Seifert fibered spaces {#subsec:asymptotics_Seifert}
--------------------------------------------------------------------------------------------------
We apply Theorem \[thm:surgery\_formula\] to a Seifert fibered space ${M\hbox{$\big(\frac{1}{-b}, \frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}}$ for ${\mathrm{SL}_{2}({\mathbb{C}})}$-representations $\rho$ which satisfy $\rho(h) = -{I}$, that is to say, the central element $h$ is sent to the non-trivial central element $-{I}$ of ${\mathrm{SL}_{2}({\mathbb{C}})}$. We begin with observing the acyclicity for ${M}$, $\partial {M}$ and the solid tori $S_j$ and confirm that the condition that $\rho(h) = -{I}$ is only needed in our situation.
It is shown from [@Kitano:RtorsionSeifertSLn the proof of Proposition 3.1] that $C_*({M};V_{2N})$ is acyclic if and only if $\rho_{2N}(h)=-{I}$. It holds for all $N$ that $\rho_{2N}(h) = -{I}$ under the assumption that $\rho(h) = -{I}$ since every weight of $\sigma_{2N}$ for the eigenvalues of $\rho(h)$ is odd. Also we have shown that such an ${\mathrm{SL}_{2}({\mathbb{C}})}$-representation $\rho$ satisfies the acyclicity condition for all boundary components $T^2_j$ since $\pi_1(T^2_j)$ is presented as $$\pi_1(T^2_j) =
\langle q_j, h \,|\, [q_j, h] = 1 \rangle$$ and $\rho(h)$ has the order of $2$. Furthermore we can see that all conditions in our surgery formula (Theorem \[thm:surgery\_formula\]) are satisfied under the assumption that $\rho(h) = -{I}$.
Suppose that an ${\mathrm{SL}_{2}({\mathbb{C}})}$-representation $\rho$ of $\pi_1({M\hbox{$\big(\frac{1}{-b}, \frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}})$ sends $h$ to $-{I}$. Then $\rho$ satisfies the acyclicity conditions in Definition \[def:acyclicity\_conditions\] and the restriction of $\rho$ gives an acyclic twisted chain complex $C_*({M};V_{2N})$ for any $N$.
It remains to prove that the acyclicity for the twisted chain complexes of $S_j$ (the condition in Definition \[def:acyclicity\_conditions\]). This follows from the following Lemma \[lemma:order\_ell\].
\[lemma:order\_ell\] Let $\rho$ be an ${\mathrm{SL}_{2}({\mathbb{C}})}$-representations of $\pi_1({M\hbox{$\big(\frac{1}{-b}, \frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}})$ such that $\rho(h)=-{I}$ and $\ell_j$ denote the homotopy class of the core of $S_j$ for $j=0, \ldots, m$. Then $\rho(\ell_j)$ is of even order for all $j$.
Set $\alpha_0 = 1$ and $\beta_0=-b$, which are the corresponding slope to the solid torus $S_0$. We can express each $\ell_j$ $(j=0, 1, \ldots, m)$ as $
\ell_j = q_j^{\mu_j} h^{\nu_j}
$ where integers $\mu_j$ and $\nu_j$ satisfy that $\alpha_j \nu_j - \beta_j \mu_j = -1$ and $0 < \mu_j < \alpha_j$. We will show that every $\rho(\ell_j)^{\alpha_j}$ turns into $-{I}$. For $j=0$, the matrix $\rho(\ell_0)^{\alpha_0} (= \rho(\ell_0))$ turns out $$\begin{aligned}
\rho(\ell_0)^{\alpha_0}
&= \rho(q_1 \cdots q_m [a_1, b_1]\cdots[a_g, b_g])^{\alpha_0 \mu_0} \rho(h)^{\alpha_0 \nu_0} \\
&= \rho(h)^{b\mu_0+\nu_0} \\
&= -{I}.
\end{aligned}$$ Similarly $\rho(\ell_j)^{\alpha_j}$ turns into $$\rho(q_j)^{\alpha_j \mu_j} \rho(h)^{\alpha_j \nu_j} = \rho(h)^{\alpha_j \nu_j - \beta_j \mu_j} = -{I}.$$ Hence, for all $j$, the eigenvalues of $\rho(\ell_j)$ is given by $e^{\pm \pi \eta_j \sqrt{-1}/\alpha_j}$ where some odd integer $\eta_j$, which implies that the order of $\rho(\ell)$ is even.
For every $j=0, 1, \ldots, m$, it holds that $\rho(\ell_j)^{\alpha_j} = -{I}$ and $\rho(\ell_j)^{2\alpha_j} = {I}$. The order of $\rho(\ell_j)$ must be less than or equal to $2 \alpha_j$. However the order of $\rho(\ell_0)$ is always $2$.
We turn to the higher dimensional Reidemeister torsion of $M = \Sigma_* \times S^1$ for $\rho_{2N}$. We have the following explicit values under the assumption that $\rho(h)=-{I}$.
\[prop:log\_tor\_mfd\] The Reidemeister torsion ${\mathop{\mathrm{Tor}}\nolimits ({M};\rho_{2N})}$ is given by $2^{-2N(1-2g-m)}$, [i.e.,]{} $\log|{\mathop{\mathrm{Tor}}\nolimits ({M};\rho_{2N})}| = -2N(1-2g-m) \log2$.
Now we are in position to apply our surgery formula (Theorem \[thm:surgery\_formula\]) to a Seifert fibered space ${M\hbox{$\big(\frac{1}{-b}, \frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}}$. From Proposition \[prop:log\_tor\_mfd\] and Theorem \[thm:surgery\_formula\], we can derive the asymptotic behavior of the higher dimensional Reidemeister torsion for ${M\hbox{$\big(\frac{1}{-b}, \frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}}$.
\[thm:asymptotics\_Seifert\] Let $\rho$ be an ${\mathrm{SL}_{2}({\mathbb{C}})}$-representation of ${M\hbox{$\big(\frac{1}{-b}, \frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}}$ such that $\rho(h) = -{I}$. Then we can express the asymptotics of $\log |{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{1}{-b}, \frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}};\rho_{2N})}|$ as follows:
1. $\displaystyle{
\lim_{N \to \infty}
\frac{
\log |{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{1}{-b}, \frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}};\rho_{2N})}|
}{
(2N)^2
}
=0
}$,
2. \[item:leading\_term\_Seifert\] $\displaystyle{
\lim_{N \to \infty}
\frac{
\log |{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{1}{-b}, \frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}};\rho_{2N})}|
}{
2N
}
= - \Big(2-2g - \sum_{j=1}^m \frac{\lambda_j - 1}{\lambda_j}\Big) \log 2
}$
where $2\lambda_j$ is the order of $\rho(\ell_j)$.
In particular, if $\lambda_j$ is equal to $\alpha_j$ for all $j$, then we have $$\begin{aligned}
\lim_{N \to \infty}
\frac{
\log |{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{1}{-b}, \frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}};\rho_{2N})}|
}{
2N
}
&= -\Big(2-2g - \sum_{j=1}^m \frac{\alpha_j -1}{\alpha_j}\Big) \log 2\\
&= -\chi \log 2
\end{aligned}$$ where $\chi$ is the Euler characteristic of the base orbifold.
Applying Theorem \[thm:surgery\_formula\], we obtain $$\lim_{N \to \infty}
\frac{
\log |{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{1}{-b}, \frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}};\rho_{2N})}|
}{
(2N)^2
}
=
\lim_{N \to \infty}
\frac{
\log |{\mathop{\mathrm{Tor}}\nolimits ({M};\rho_{2N})}|
}{
(2N)^2
}
= 0$$ by Proposition \[prop:log\_tor\_mfd\]. Also it follows that $$\begin{aligned}
\lim_{N \to \infty}
\frac{
\log |{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{1}{-b}, \frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}};\rho_{2N})}|
}{
2N
}
&=
\lim_{N \to \infty}
\frac{
\log |{\mathop{\mathrm{Tor}}\nolimits ({M};\rho_{2N})}|
}{
2N
}
- \Big(\sum_{j=0}^m \frac{1}{\lambda_j}\Big) \log 2\\
&=
-(1-2g-m) - \Big( 1 + \sum_{j=1}^{m} \frac{1}{\lambda_j}\Big) \log 2\\
&=
- \Big(2-2g - \sum_{j=1}^{m} \frac{\lambda_j - 1}{\lambda_j} \Big) \log 2.
\end{aligned}$$
\[remark:special\_case\_lambda\] It follows from the proof of Lemma \[lemma:order\_ell\] that $\rho(\ell_j)^{2\alpha_j} = {I}$ for all $j$. Each $\lambda_j$ in Theorem \[thm:asymptotics\_Seifert\] is a divisor of the corresponding $\alpha_j$.
\[cor:maximal\_leading\_term\] The value $- \chi \log 2$ is the maximum in the limits of Theorem \[thm:asymptotics\_Seifert\] for all ${\mathrm{SL}_{2}({\mathbb{C}})}$-representations sending $h$ to $-{I}$.
We can rewrite in Theorem \[thm:asymptotics\_Seifert\] as $$\begin{aligned}
\lim_{N \to \infty}
\frac{
\log | {\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}};\rho_{2N})} |
}{
2N
}
&= - \Big(2-\sum_{j=1}^m \frac{\lambda_j -1}{\lambda_j}\Big) \log 2\notag\\
&= - \chi \log 2 - \log 2\sum_{j=1}^m \Big(\frac{1}{\lambda_j} - \frac{1}{\alpha_j}\Big)
\label{eqn:another_form_leading_term}.
\end{aligned}$$ Our claim follows from that each $\lambda_j$ is a divisor of $\alpha_j$ for $j=1, \ldots, m$.
We also give the explicit form of the higher dimensional Reidemeister torsion for a Seifert fibered space. The following is the direct application of Lemma \[lemma:MultLemma\] (Multiplicativity Lemma).
Let $\rho$ be an ${\mathrm{SL}_{2}({\mathbb{C}})}$-representation of $\pi_1({M\hbox{$\big(\frac{1}{-b}, \frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}})$ such that $\rho(h)=-{I}$. Then we can express ${\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{1}{-b}, \frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}};\rho_{2N})}$ as $${\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{1}{-b}, \frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}};\rho_{2N})}
=
2^{-2N(2-2g-m)}
\cdot \prod_{j=1}^m \prod_{k=1}^N \Big(2 \sin \frac{\pi(2k-1)\eta_j}{2\alpha_j}\Big)^{-2}$$ where $e^{\pm \pi \eta_j \sqrt{-1} / \alpha_j}$ are the eigenvalues of $\rho(\ell_j)$.
For the Reidemeister torsion of Seifert fibered spaces ($g>1$) with more general irreducible ${\mathrm{SL}_{n}({\mathbb{C}})}$-representations, we refer to [@Kitano:RtorsionSeifertSLn].
We do not require the irreducibility of $\rho_{2N} = \sigma_{2N} \circ \rho$. However our assumption that $\rho(h)=-{I}$ guarantees the acyclicity of $\rho_{2N}$ for all $N$.
The leading coefficients and the ${\mathrm{SU}(2)}$-character varieties for Seifert fibered homology spheres {#subsec:leadingCoeff_Seifert}
------------------------------------------------------------------------------------------------------------
We have shown the explicit limits of the leading coefficients in the higher dimensional Reidemeister torsions for Seifert fibered spaces ${M\hbox{$\big(\frac{1}{-b}, \frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}}$. The limit of the leading coefficient depends only on the order of $\rho(\ell_j)$ for an ${\mathrm{SL}_{2}({\mathbb{C}})}$-representation $\rho$ of $\pi_1({M\hbox{$\big(\frac{1}{-b}, \frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}})$. In $\pi_1({M\hbox{$\big(\frac{1}{-b}, \frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}})$, we have the relations that $\ell_j = q_j^{\mu_j} h^{\nu_j}$ and $q_j^{\alpha_j} h^{\beta_j} = 1$ where $\alpha_j \nu_j - \beta_j \nu_j = -1$. Under the assumption that $\rho(h) = -{I}$, the order of $\rho(\ell_j)$ is determined by the order $\rho(q_j)$, [i.e.,]{}the eigenvalues of $\rho(q_j)$.
Here and subsequently, following the previous studies [@FintushelStern:InstantonSeifert; @KirkKlassen:RepresentationSeifert; @BauerOkonek], we assume that $\rho$ is irreducible. Irreducible representations with the same eigenvalues for the generators $q_1, \ldots, q_m$ form a set with a structure of variety. When we also consider their conjugacy classes, it is known that the set of conjugacy classes also has a structure of variety.
We focus on Seifert fibered homology spheres and ${\mathrm{SU}(2)}$-representations of their fundamental groups. It has shown by [@FintushelStern:InstantonSeifert; @KirkKlassen:RepresentationSeifert; @BauerOkonek] that the set of conjugacy classes of irreducible ${\mathrm{SU}(2)}$-representations for a Seifert fibered homology sphere can be regarded as the set of smooth manifolds with even dimensions. In the remain of paper, we deal with Seifert fibered homology spheres ${M\hbox{$\big(\frac{1}{-b}, \frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}}$ with $b=0$. We will denote it briefly by ${M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}}$ and write the set of conjugacy classes of irreducible ${\mathrm{SU}(2)}$-representations of $\pi_1({M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}})$ as $${\mathcal{R}({M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}})}
= \mathrm{Hom}^{\mathrm{irr}}(\pi_1({M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}}), {\mathrm{SU}(2)}) \,\big/\, \mathrm{conj.}$$ which is called [*the ${\mathrm{SU}(2)}$-character variety*]{}.
Each component in ${\mathcal{R}({M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}})}$ determined by the set of eigenvalues for ${\mathrm{SU}(2)}$-elements corresponding to $q_1, \ldots, q_m$. By the relation that $q_j^{\alpha_j} h^{\beta_j}=1$, the eigenvalues of $\rho(q_j)$ for an irreducible ${\mathrm{SU}(2)}$-representation $\rho$ are given by $
e^{\pm \pi \xi_j \sqrt{-1} / \alpha_j}
$ ($0 \leq \xi_j \leq \alpha_j$). We will use the $m$-tuple $(\xi_1, \ldots, \xi_m)$ to denote the corresponding component in the ${\mathrm{SU}(2)}$-character variety ${\mathcal{R}({M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}})}$ (for details, we refer to [@FintushelStern:InstantonSeifert; @KirkKlassen:RepresentationSeifert; @BauerOkonek]).
\[prop:Fact\_CharVar\] Let $\rho$ be an irreducible ${\mathrm{SU}(2)}$-representation of ${M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}}$. Suppose that the conjugacy class of $\rho$ is contained in a component $(\xi_1, \ldots, \xi_m)$. Then the dimension of $(\xi_1, \ldots, \xi_m)$ is equal to $2(n-3)$ where $n$ is the number of $\xi_j$ such that $\xi_j \not = 0$, $\alpha_j$, [i.e.,]{}$\rho(q_j) \not = \pm {I}$, in $j=1, \ldots, m$.
We will find components of ${\mathcal{R}({M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}})}$ containing the conjugacy class of an irreducible ${\mathrm{SU}(2)}$-representation $\rho$ which makes the leading coefficient in the logarithm of $|{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}};\rho_{2N})}|$ converge to $-\chi \log 2$.
\[prop:leading\_term\_char\_var\] Let $\rho$ be an irreducible ${\mathrm{SU}(2)}$-representation such that $\rho(h) = -{I}$. The leading coefficient of $\log|{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}};\rho_{2N})}|$ converges to $-\chi \log 2$ if and only if the conjugacy class of $\rho$ is contained in a $2(m-3)$-dimensional component $(\xi_1, \ldots, \xi_m)$ of ${\mathcal{R}({M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}})}$ such that $\alpha_j$ and $\xi_j$ are coprime for all $j$.
Before proving Proposition \[prop:leading\_term\_char\_var\], let us observe a relation between the order of $\rho(\ell_j)$ and $\xi_j$.
\[lemma:xi\_alpha\] Suppose that an irreducible ${\mathrm{SU}(2)}$-representation $\rho$ satisfies that $\rho(h) = -{I}$ and its conjugacy class is contained in a component $(\xi_1, \ldots, \xi_m)$. Then the order of $\rho(\ell_j)$ is equal to 2$\alpha_j$, [i.e.,]{}$\lambda_j = \alpha_j$, if and only if $\xi_j$ and $\alpha_j$ are coprime.
We can express $\ell_j$ as $\ell_j = q_j^{\mu_j} h^{\nu_j}$ where $\alpha_j \nu_j - \beta_j \mu_j = -1$. Since the eigenvalues of $\rho(q_j)$ are given by $e^{\pm \pi \xi_j \sqrt{-1}/\alpha_j}$, we can diagonalize $\rho(\ell_j)$ as $$\rho(\ell_j) \sim
\begin{pmatrix}
e^{ \pi \eta_j \sqrt{-1}/\alpha_j} & 0 \\
0 & e^{ -\pi \eta_j \sqrt{-1}/\alpha_j}
\end{pmatrix}
\quad (\eta_j = \mu_j \xi_j - \alpha_j \nu_j).$$ Suppose that $\xi_j$ and $\alpha_j$ are coprime. It follows from $(\alpha_j, \mu_j) = 1$ that the g.c.d. $(\alpha_j, \eta_j)$ coincides with $(\alpha, \xi_j) = 1$. Since the order of $\rho(\ell_j)$ is $2\lambda_j$, we can see that $\rho(\ell_j)^{\lambda_j} = -{I}$. Hence $\alpha_j$ divides $\lambda_j$, which implies that $\lambda_j = \alpha_j$ from that $\lambda_j$ is a divisor of $\alpha_j$. Similarly, when the order of $\rho(\ell_j)$ is $2\alpha_j$, the g.c.d. $(\alpha_j, \xi_j)$ must be $1$.
Under the assumption that $\rho(h)=-{I}$, if $\xi_j$ is equal to $0$ or $\alpha_j$, then $\rho(\ell_j) = -{I}$. Hence $\lambda_j = 1$.
According to Eq. , every $\lambda_j$ coincides with $\alpha_j$ for all $j$ if and only if the leading coefficient of $\log|{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}};\rho_{2N})}|$ converges to $-\chi \log 2$. By Lemma \[lemma:xi\_alpha\], we can rephrase $\lambda_j = \alpha_j$ for all $j$ as $[\rho] \in (\xi_1, \ldots, \xi_m)$ with the g.c.d. $(\alpha_j, \xi_j)=1$ for all $j$. In particular, it is seen from Proposition \[prop:Fact\_CharVar\] that the dimension of $(\xi_1, \ldots, \xi_m)$ is equal to $2(m-3)$.
In special cases that every $\alpha_j$ is prime in the Seifert index of ${M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}}$, we obtain a simple correspondence between the limits of $\log|{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}};\rho_{2N})}| / (2N)$ and components of ${\mathcal{R}({M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}})}$.
\[thm:max\_min\_leading\_coeff\] Let $\rho$ be an irreducible ${\mathrm{SU}(2)}$-representation of $\pi_1({M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}})$ such that $\rho(h)= -{I}$. Suppose that every $\alpha_j$ is prime and $\alpha_1 < \cdots < \alpha_m$. If the conjugacy class of $\rho$ is contained in a component $(\xi_1, \ldots, \xi_m)$, then the leading coefficient of $\log|{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}};\rho_{2N})}|$ converges to $$\label{eqn:simple_form_leading_coeff}
\lim_{N \to \infty}
\frac{
\log|{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}};\rho_{2N})}|
}{
2N
}
= - \Big(2-\sum_{\xi_j \not = 0, \alpha_j} \frac{\alpha_j -1}{\alpha_j}\Big)\log 2.$$ If the set $\{ [\rho] \in {\mathcal{R}({M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}})} \,|\, \rho(h) = -{I}\}$ has $0$ and $2(m-3)$-dimensional components, then the limit takes the maximum $-\chi \log 2$ on only all top-dimensional components and takes the minimum on some $0$-dimensional components.
As seen in Remark \[remark:special\_case\_lambda\], we have $\lambda_j =1$ if $\xi_j = 0$ or $\alpha_j$. Substituting $\lambda_j = 1$ into the equality above Eq. for the corresponding index $j$, we have the limit . From Proposition \[prop:leading\_term\_char\_var\] and our assumption, it follows that the limit takes the maximum $-\chi \log 2$ on all top-dimensional components.
It remains to prove that the limit takes the minimum on a $0$-dimensional component. Since the limit is expressed as Eq. , we consider the minimum of $\sum_{\xi_j \not = 0, \alpha_j} (\alpha_j -1)/\alpha_j$. Each $0$-dimensional component is given by $(\xi_1, \ldots, \xi_m)$ for all $\xi_j = 0, \alpha_j$ except three $\xi_{j_1}$, $\xi_{j_2}$ and $\xi_{j_3}$. We need to consider two cases: (i) $\alpha_1=2$ and (ii) $\alpha_1 \geq 3$. In the case that $\alpha_1 = 2$, $\xi_1$ must be $1$ for all components $(\xi_1, \ldots, \xi_m)$ since we have $\rho(q_1)^{\alpha_1} = -{I}$ from the assumption that $\rho(h)=-{I}$. The sum $(\alpha_{j_1} -1) / \alpha_{j_1} + (\alpha_{j_2} -1) / \alpha_{j_2} + (\alpha_{j_3} -1) / \alpha_{j_3}$ turns into $$\frac{1}{2} +
\frac{\alpha_{j_2} -1}{\alpha_{j_2}} +
\frac{\alpha_{j_3} -1}{\alpha_{j_3}}
< \frac{5}{2}.$$ On the other hand, it is easily seen that for higher dimensional components, $$\frac{1}{2} +
\frac{\alpha_{i_2} -1}{\alpha_{i_2}} +
\frac{\alpha_{i_3} -1}{\alpha_{i_3}} +
\frac{\alpha_{i_4} -1}{\alpha_{i_4}} +\cdots
\geq \frac{1}{2} + \frac{3 - 1}{3} + \frac{5-1}{5} + \frac{7-1}{7}
> \frac{5}{2}.$$ Hence the minimum lies in a $0$-dimensional component.
In the other case that $\alpha_1 \geq 3$, it is clear that $$\frac{\alpha_{j_1} -1}{\alpha_{j_1}} +
\frac{\alpha_{j_2} -1}{\alpha_{j_2}} +
\frac{\alpha_{j_3} -1}{\alpha_{j_3}}
< 3.$$ On the other hand, we can see that $$\frac{\alpha_{i_1} -1}{\alpha_{i_1}} +
\frac{\alpha_{i_2} -1}{\alpha_{i_2}} +
\frac{\alpha_{i_3} -1}{\alpha_{i_3}} +
\frac{\alpha_{i_4} -1}{\alpha_{i_4}} +\cdots
\geq
\frac{3-1}{3} + \frac{5-1}{5} + \frac{7-1}{7} + \frac{11-1}{11}
> 3.$$ The minimum of the limits lies on $0$-dimensional components.
Examples for Seifert fibered homology spheres {#subsec:examples}
---------------------------------------------
We will see two examples of Theorem \[thm:max\_min\_leading\_coeff\] and an example which shows that $\log |{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{\alpha_1}{\beta_1}, \ldots, \frac{\alpha_m}{\beta_m}\big)$}};\rho_{2N})}| / (2N)$ does not converges to the maximum $-\chi \log 2$ on all top–dimensional components if some $\alpha_j$ is not prime and $(\alpha_j, \xi_j) \not = 1$.
### $M\hbox{$\big(\frac{2}{\beta_1}, \frac{3}{\beta_2}, \frac{7}{\beta_3}\big)$}$
We can choose that $\beta_1=1$ and $\beta_2 = \beta_3 = -1$ by the requirement that $2 \cdot 3 \cdot 7 (\beta_1 / 2 + \beta_2 / 3 + \beta_3 / 7) =1$. The Brieskorn homology $3$-sphere ${M\hbox{$\big(\frac{2}{1}, \frac{3}{-1}, \frac{7}{-1}\big)$}}$ also corresponds to the surgery along $(2, 3)$-torus knot with slope $1$ in Subsection \[subsec:example\_torusknots\]. From the presentation: $$\pi_1({M\hbox{$\big(\frac{2}{1}, \frac{3}{-1}, \frac{7}{-1}\big)$}})
=\langle
q_1, q_2, q_3, h \,|\,
[q_j, h]=1, q_j^{\alpha_j} h^{\beta_j} = 1, q_1 q_2 q_3 =1
\rangle,$$ every irreducible ${\mathrm{SU}(2)}$-representation sends $h$ to $-{I}$ and the ${\mathrm{SU}(2)}$-character variety of ${M\hbox{$\big(\frac{2}{1}, \frac{3}{-1}, \frac{7}{-1}\big)$}}$ consists of $(\xi_1, \xi_2, \xi_3) = (1, 1, 3)$ and $(1, 1, 5)$. For details about the computation of ${\mathrm{SU}(2)}$-character varieties, we refer to [@FintushelStern:InstantonSeifert; @KirkKlassen:RepresentationSeifert] and [@Saveliev99:LectureTopology3mfd Lecture $14$].
Let $\rho$ be an irreducible ${\mathrm{SU}(2)}$-representation of ${M\hbox{$\big(\frac{2}{1}, \frac{3}{-1}, \frac{7}{-1}\big)$}}$. By the relations that $\ell_j = q_j^{\mu_j} h^{\nu_j}$ and $\alpha_j \nu_j - \beta_j \mu_j = -1$ ($0 < \mu_j < \alpha_j$), we obtain that $$\rho(\ell_1) = \rho(q_1),\,
\rho(\ell_2)= - \rho(q_2)^2,\,
\rho(\ell_3) = - \rho(q_3)^6.$$ From Eqs. $\rho(q_1)^2 = -{I}$, $\rho(q_2)^3 = -{I}$ and $\rho(q_3)^7 = -{I}$, the orders $2\lambda_j$ of $\rho(\ell_j)$ are given by $$2\lambda_1 = 4,\,
2\lambda_2 = 6,\,
2\lambda_3 = 14.$$ By Theorem \[thm:asymptotics\_Seifert\], for the both cases of $[\rho] \in (1, 1, 3)$ and $[\rho] \in (1, 1, 5)$, we can see that $$\lim_{N \to \infty}
\frac{
\log |{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{2}{1}, \frac{3}{-1}, \frac{7}{-1}\big)$}};\rho_{2N})}|}{2N}
= \Big(1 - \frac{1}{2} - \frac{1}{3} -\frac{1}{7}\Big) \log2.$$ The limit of the leading coefficient takes the maximum $-\chi \log 2$ on all top–dimensional components (see also Theorem \[thm:asymptotics\_Brieskorn\]).
### $M\hbox{$\big(\frac{2}{\beta_1}, \frac{3}{\beta_2}, \frac{5}{\beta_3}, \frac{7}{\beta_4}\big)$}$
Let us choose $\beta_1=1$, $\beta_2 = \beta_3 = -2$ and $\beta_4=4$. The subvariety $\{[\rho] \in {\mathcal{R}({M\hbox{$\big(\frac{2}{1}, \frac{3}{-2}, \frac{5}{-2}, \frac{7}{4}\big)$}})} \,|\, \rho(h) = -{I}\}$ consists of eight points and six $2$-dimensional spheres. (For details, see [@Saveliev99:LectureTopology3mfd Lecture $14$])
Each $0$-dimensional component corresponds to the parameter $(\xi_1, \xi_2, \xi_3, \xi_4)$ in: $$\label{eqn:2dim_compo}
\left\{
\begin{array}{c}
(1, 0, 2, 2),\, (1, 0, 2, 4),\, (1, 0, 2, 6),\, (1, 0, 4, 4), \\
(1, 2, 0, 2),\, (1, 2, 0, 4),\, (1, 2, 2, 0),\, (1, 2, 4, 0)
\end{array}
\right\}$$ and each $2$-dimensional components are given by $(\xi_1, \xi_2, \xi_3, \xi_4)$ in $$\left\{
\begin{array}{cc}
(1, 2, 2, 2),\, (1, 2, 2, 4),\, (1, 2, 2, 6), \\
(1, 2, 4, 2),\, (1, 2, 4, 4),\, (1, 2, 4, 6)
\end{array}
\right\}.$$ We can express $\ell_j$ ($1 \leq j \leq 4$) as $$\ell_1 = q_1,\,
\ell_2 = q_2 h^{-1},\,
\ell_3 = q_3^2 h^{-1},\,
\ell_4 = q_4 h.$$
Let $\rho$ be an irreducible ${\mathrm{SU}(2)}$-representation of $\pi_1({M\hbox{$\big(\frac{2}{1}, \frac{3}{-2}, \frac{5}{-2}, \frac{7}{4}\big)$}})$ such that $\rho(h) = -{I}$. We have the following table between the $0$-dimensional components and the orders of $\rho(\ell_j)$ for $j=1, 2, 3, 4$: $$
$(\xi_1, \xi_2, \xi_3, \xi_4)$ $\lambda_j$: the half of the order of $\rho(\ell_j)$
-------------------------------- ---------------------------------------------------------------------
$(\xi_1, 0, \xi_3, \xi_4)$ $\lambda_1 = 2$, $\lambda_2 = 1$, $\lambda_3 = 5$, $\lambda_4 = 7$
$(\xi_1, \xi_2, 0, \xi_4)$ $\lambda_1 = 2$, $\lambda_2 = 3$, $\lambda_3 = 1$, $\lambda_4 = 7$
$(\xi_1, \xi_2, \xi_3, 0)$ $\lambda_1 = 2$, $\lambda_2 = 3$, $\lambda_3 = 5$, $\lambda_4 = 1$.
$$ Hence for $[\rho] \in (\xi_1, \xi_2, \xi_3, \xi_4)$ $(\xi_2 = 0)$, by Theorem \[thm:asymptotics\_Seifert\], we obtain $$\begin{aligned}
\lim_{N \to \infty}
\frac{
\log |{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{2}{1}, \frac{3}{-2}, \frac{5}{-2}, \frac{7}{4}\big)$}};\rho_{2N})}|}{2N}
&=
-\Big(2 - \sum_{j \not = 2} \frac{\lambda_j -1}{\lambda_j}\Big)\log 2\\
&= -\chi \log 2 - \frac{2}{3} \log 2,\end{aligned}$$ for $[\rho] \in (\xi_1, \xi_2, \xi_3, \xi_4)$ ($\xi_3 = 0$), $$\begin{aligned}
\lim_{N \to \infty}
\frac{
\log |{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{2}{1}, \frac{3}{-2}, \frac{5}{-2}, \frac{7}{4}\big)$}};\rho_{2N})}|}{2N}
&=
-\Big(2 - \sum_{j \not = 3} \frac{\lambda_j -1}{\lambda_j}\Big)\log 2\\
&= -\chi \log 2 - \frac{4}{5} \log 2\end{aligned}$$ and for $[\rho] \in (\xi_1, \xi_2, \xi_3, \xi_4)$ ($\xi_4 = 0$), $$\begin{aligned}
\lim_{N \to \infty}
\frac{
\log |{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{2}{1}, \frac{3}{-2}, \frac{5}{-2}, \frac{7}{4}\big)$}};\rho_{2N})}|}{2N}
&=
-\Big(2 - \sum_{j \not = 4} \frac{\lambda_j -1}{\lambda_j}\Big)\log 2\\
&= -\chi \log 2 - \frac{6}{7} \log 2.\end{aligned}$$
When the conjugacy class $[\rho]$ is contained in the $2$-dimensional components in , then it is seen that $\lambda_1 = 2$, $\lambda_2 = 3$, $\lambda_3 = 5$ and $\lambda_4 = 7$. Hence we obtain the maximum of the limits: $$\lim_{N \to \infty}
\frac{
\log |{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{2}{1}, \frac{3}{-2}, \frac{5}{-2}, \frac{7}{4}\big)$}};\rho_{2N})}|}{2N}
= -\chi \log 2.$$ The limit of $\log |{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{2}{1}, \frac{3}{-2}, \frac{5}{-2}, \frac{7}{4}\big)$}};\rho_{2N})}| / (2N)$ takes the minimum at the components $(1, 2, 2, 0)$ and $(1, 2, 4, 0)$.
### $M\hbox{$\big(\frac{5}{\beta_1}, \frac{6}{\beta_2}, \frac{7}{\beta_3}\big)$}$
Let us choose $\beta_1 = 3$, $\beta_2 = -1$ and $\beta_3 = -3$. The subvariety $\{ [\rho] \in {\mathcal{R}({M\hbox{$\big(\frac{5}{3}, \frac{6}{-1}, \frac{7}{-3}\big)$}})} \,|\, \rho(h) = -{I}\}$ consists of eight points which are given by the following set of $(\xi_1, \xi_2, \xi_3)$: $$\left\{
\begin{array}{c}
(1, 1, 1),\, (1, 3, 3),\, (1, 5, 5), \\
(3, 1, 5),\, (3, 3, 1),\, (3, 3, 3),\, (3, 3, 5),\, (3, 5, 3)
\end{array}
\right\}$$
We separate the above subvariety into two subset $X_1$ and $X_2$ as follows: $$\begin{aligned}
X_1 &:= \{ (1, 1, 1), (1, 5, 5), (3, 1, 5), (3, 5, 3) \}\\
X_2 &:= \{ (1, 3, 3), (3, 3, 1), (3, 3, 3), (3, 3, 5) \}\end{aligned}$$ where every component is $0$-dimensional. Every components in $X_1$ satisfies that $(\alpha_j, \xi_j)=1$ for all $j$. On the other hand, each component in $X_2$ satisfies that $(\alpha_1, \xi_1)=(\alpha_3, \xi_3)= 1$ and $(\alpha_2, \xi_2) = 3$.
In the case that the conjugacy class $[\rho]$ is contained in $X_1$, we can also compute similarly $\lambda_1 = 5$, $\lambda_2 = 6$ and $\lambda_3 = 7$. Therefore we obtain the following limit, by Theorem \[thm:asymptotics\_Seifert\] $$\lim_{N \to \infty}
\frac{
\log |{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{5}{3}, \frac{6}{-1}, \frac{7}{-3}\big)$}};\rho_{2N})}|
}{
2N
}
= -\chi \log 2.$$
In the case of the conjugacy class $[\rho]$ is contained in $X_2$, the order of $\rho(\ell_2)$ is $4$ since $\ell_2 = q_2^{5} h^{-1}$ and $\rho(q_2)^2 = -{I}$. Hence $\lambda_2$ equals to $2$. Similar computations yield $\lambda_1 = 5$ and $\lambda_3 = 7$. Furthermore Theorem \[thm:asymptotics\_Seifert\] gives the following limit: $$\begin{aligned}
\lim_{N \to \infty}
\frac{
\log |{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{5}{3}, \frac{6}{-1}, \frac{7}{-3}\big)$}};\rho_{2N})}|
}{
2N
}
&= -\Big(2 - \sum_{j=1}^3 \frac{\lambda_j - 1}{\lambda_j}\Big) \log 2\\
&= -\chi \log 2 - \frac{1}{3} \log 2.\end{aligned}$$
Therefore we have the top-dimensional components where the limit of the leading coefficient in $\log |{\mathop{\mathrm{Tor}}\nolimits ({M\hbox{$\big(\frac{5}{3}, \frac{6}{-1}, \frac{7}{-3}\big)$}};\rho_{2N})}|$ does not take the maximum.
Acknowledgment {#acknowledgment .unnumbered}
==============
The author wishes to express his thanks to Joan Porti, Takayuki Morifuji and Takahiro Kitayama for helpful suggestions to start the computation in this manuscript. This research was supported by Research Fellowships of the Japan Society for the Promotion of Science for Young Scientists.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Understanding to what extent stem cell potential is a cell-intrinsic property, or an emergent behavior coming from global tissue dynamics and geometry, is a key outstanding question of stem cell biology. Here, we propose a theory of stem cell dynamics as a stochastic competition for access to a spatially-localized niche, giving rise to a “stochastic conveyor-belt” model. Cell divisions produce a steady cellular stream which advects cells away from the niche, while random rearrangements enable cells away from the niche to be favourably repositioned. Importantly, even when assuming that all cells in a tissue molecularly equivalent, the model predicts a common (“universal”) functional dependence of the long-term clonal survival probability on the position within the niche, as well as the emergence of a well-defined number of “functional” stem cells, dependent only on the rate of random movements vs. mitosis-driven advection. We test the predictions of this theory on datasets on pubertal mammary gland tips, embryonic kidney tips as well homeostatic intestinal crypt, and find good quantitative agreement for the number of functional stem cells in each organ, as well as the predicted functional dependence of the competition.'
author:
- 'Bernat Corominas-Murtra$^1$, Colinda L.G.J. Scheele$^2$, Kasumi Kishi$^1$, Saskia I.J. Ellenbroek$^2$, Benjamin D. Simons$^{3,4,5}$, Jacco van Rheenen$^2$, Edouard Hannezo$^1$'
title: Stem cell lineage survival as a noisy competition for niche access
---
[^1]
Many biological tissues are renewed via small numbers of stem cells, which divide to produce a steady stream of differentiated cells and balance homeostatic cell loss. Although novel experimental approaches in the past decade have produced key insights into the number, identity, and (often stochastic) dynamics of stem cells in multiple organs, an outstanding question remains as to whether stem cell potential is a cell-intrinsic, “inherited” property, or rather an extrinsic, context-dependent state emerging from the collective dynamics of a tissue and cues from local “niches”, or microenvironments [@Wang:2010; @Simons:2011; @Blanpain:2013; @Watt:2013; @Blanpain:2014; @Plaks:2015; @Moris:2016; @Yang:2017]. Although recent experiments have provided evidence for the latter in settings such as the growing mammary gland [@Scheele:2017], adult interfollicular epidermis [@Rompolas:2013; @Rompolas:2016], spermatogenesis [@Kitadate:2019] or the intestinal epithelium [@Ritsma:2014], a more global theoretical framework allowing to quantitatively interpret these findings is still lacking.
The case of the intestinal crypt serves as a paradigmatic example of the dynamics of tissue renewal, and is one of the fastest in mammals [@Ritsma:2014]. The intestinal crypt consists of a small invagination in the intestine where the epithelial cells populating the intestinal walls are constantly produced. The very bottom of the crypt hosts a small number of proliferative, Lgr5+ stem cells, [@Barker:2007] that divide and push the cells located above them to the transit amplification (TA) region, where cells lose self-renewal potential. Cells are eventually shed in the villus a few days later, constituting a permanent “conveyor-belt” dynamics. Lineage tracing approaches, which irreversibly label a cell and its progeny [@Blanpain:2013], have been used to ask which cell type will give rise to lineages that renew the whole tissue and have revealed that all Lgr5+ cells can stochastically compete in an equipotent manner on the long term [@Snippert:2010; @Lopez-Garcia:2010; @Snippert:2011; @Klein:2011], but still display positional-dependent short term biases for survival [@Ritsma:2014]. Interestingly, similar conclusions have been achieved in pubertal mammary gland development [@Scheele:2017], where branching morphogenesis is performed through the proliferation of the cells in the terminal end buds of the ducts, the region where the mammary stem cells (MaSCs) reside [@Scheele:2017; @Visvader:2014]. In both cases, intravital imaging revealed random cellular motions enabling a cell to move against the cellular flow/drift defined by the conveyor belt dynamics. Moreover, in the intestine, tissue damage, or genetic ablation of all Lgr5+ stem cells, caused Lgr5- cells to recolonize the crypts and re-express Lgr5+ to function as stem cells [@Ritsma:2014], arguing for extensive reversibility and flexibility in the system [@vanEs:2012]. In addition, Lgr5- and Lgr5+ cells were also shown to nearly equally contribute to intestinal morphogenesis [@Guiu:2019]. Altogether, this supports proposals that the definition of stem cell potential should evolve to emphasize, instead of molecular markers, the functional ability of replace lost cells via mitosis [@Krieger:2015; @Post:2019] and survive long-term in tissues. However, this new definition raises a number of outstanding conceptual problems: What then defines the number of functional stem cells in a tissue? How can short-term biases be reconciled with long-term equipotency? Is there a sharp distinction between stem and non-stem cells, or is there instead a continuum of stem cell potential together with flexible transition between states? Qualitatively, it is clear that fluctuations and positional exchange are needed to prevent a single cell in the most favourable position to be the unique “functional” stem cell (defined as cells whose lineage colonizes a tissue compartment on the long-term). Incorporating these features in a dynamical model of stem cell growth and replacement, able to make predictions e.g., over the probability of lineage perpetuation, would represent an important step towards the understanding of how stem cells operate in the process of tissue development and renewal.
In this paper we develop a reaction-diffusion formalism for stem cell renewal in the presence of noise and local niches, taking into account local tissue geometry as well as cell division and random cell movements –see Fig. (\[fig:Schema\]a–c). Importantly, within this purely extrinsic and dynamical approach, which does not need to posit any intrinsic “stem cell identity”, a well-defined number of functional stem cell emerges, which only depends on the geometry and a balance between the noisiness of cell movements and division rates advecting cells away from niche regions. This model also predicts that “stem cell potential” should decay continuously as a function of distance from the niche, with a “universal” Gaussian functional dependence. We test this prediction against published live-imaging datasets for the homeostatic intestinal crypt [@Ritsma:2014] and during the branching of embryonic kidney explants [@Riccio:2016], and find a good quantitative agreement for the full survival probability of cells depending on their initial position relative to the niche. Furthermore, we use our theoretical results to extract the amplitude of the random positional fluctuations in the developing mammary gland using lineage tracking experiments [@Scheele:2017]. This enables us to predict the number of functional stem cells for this system, finding values fully consistent with estimates previously reported.
Dynamics of tissue renewal and development
==========================================
To develop the model, we first consider the simplest situation of a one-dimensional column of cells, with a rigid boundary condition at the base (mimicking, for instance, the bottom of the crypt), so that each cell division produces a pushing force upwards transmitted to the cells above (or in the case of growing mammary gland or kidney, driving ductal elongation). This model is motivated by its simplicity, as it is able to qualitatively derive the essential traits of the complex dynamics studied here. As we shall see, further refinements, aimed at making predictions for real systems, consider much more realistic geometries. From this simple dynamics, we define the number of functional stem cells as the typical number of cells that have a non-negligible probability to produce long-term progenies (without “losing” the competition against other cells). If the dynamics was fully devoid of noise (a simple conveyor belt) and all cell divisions were symmetric, then one of the bottom-most cells would always win the competition. Thus, there would be a single row of functional stem cells, which is the limiting case of described in Ref. [@Lopez-Garcia:2010] of symmetric and stochastic 1D competition along a ring of equipotent cells. However, as mentioned above, an extensive amount of noise in cellular movements and rearrangement is observed, to different levels, in multiple settings via live-imaging, for instance in the mammary gland [@Scheele:2017], kidney morphogenesis [@Riccio:2016; @Packard:2013], or intestinal crypts [@Ritsma:2014]. Intuitively, such rearrangements are expected to increase the number of “functional” stem cells, as re-arrangements allow cells away from the niche to relocate in favourable positions, and would thus provide a biophysical mechanism for setting the number of stem cells assumed in models such as Ref. [@Lopez-Garcia:2010].
Lineage dynamics
----------------
To further develop our intuition, and provide a quantitative criterion for how re-arrangements affect the number of cells which participate effectively in the competition, we defined a minimal model of such “stochastic-conveyor belt” dynamics. We start with the equation that describes the stochastic movement of a single cell: First, one must account for the push-up force due to the proliferation of cells located lower in the system at rate $k_d$. Furthermore, it must incorporate a probability for cells to randomly switch position, accounting for the random positional fluctuations, with amplitude $k_r$ (although the exact nature and implementation of this noise does not impact the results, as shown and discussed in Appendix, sections A1 and C). With these ingredients, the stochastic dynamics is governed by the equation: $$dz=k_d zdt+\sqrt{k_r}dW\quad,\nonumber
%\label{eq:Lang}$$ where $dW$ is the standard Brownian differential and $z$ the distance of the cell from the origin. This dynamics represents an [*Ornstein-Uhlenbeck process*]{} [@Gardiner:1983; @VanKampen:2007], with the peculiarity that the drift term is positive. To analyze the evolution of the lineage, we have to introduce a proliferative term in the dynamics. Let $\rho_n(z,t)$ be the expected density of the lineage $n$ at time $t$ at position $z$, where $n$ is defined such that the mother cell occupies position $z=n$ at $t=0$. At the next time step, $k_d\rho_n(z,t)$ new cells of this lineage will appear at position $z$ due to cell proliferation. The lineage density evolution dynamics is thus described by a reaction-diffusion equation [@Britton:1986; @Grindrod:1996], in which the reaction term describes the exponential growth of the lineage density and the diffusion term follows an Ornstein-Uhlenbeck-like dynamics (see Fig. (\[fig:Schema\]a,b) and the Appendix, sections A1,2): $$\frac{\partial \rho_n}{\partial t}=-k_d\frac{\partial }{\partial z}(z\rho_n)+k_r\frac{\partial^2 \rho_n}{\partial z^2}+k_d\rho_n\quad.
\label{eq:FisherKPP}$$ Note that here we have assumed that all cells are of the same type and that properties such as cell divisions and rearrangements do not change spatially; that is, $k_r$ and $k_d$ are constant throughout the system, two assumptions that we lift in subsequent sections. This particular stochastic process has no stationary solutions. However, considering initial conditions $t_0=0$, $z_0=\delta(z-n)$ and natural boundary conditions, one can still compute the time-dependent solutions, which are well approximated by a normal distribution whose mean and variance define a front wave that runs and widens exponentially fast in time outwards (see Appendix, section A2 for details): $$\rho_n(z,t)\approx \sqrt{\frac{k_d}{2k_r}}{\rm exp}\left\{-\frac{k_d}{2 k_r}\left(\frac{z-ne^{k_d t}}{e^{k_d t}}\right)^2\right\}\quad.
\label{eq:rhoexplicitMain}$$
Equation (\[eq:rhoexplicitMain\]) tracks the density of a single lineage through time and space. The whole system dynamics, however, is composed of different lineages competing to reach monoclonality. In addition, a given position can only contain a single cell, making a rigorous analytic treatment intractable. However, since the competition for a given position is neutral, one can make the approximation that the above derived densities reflect the likelihood that a given lineage occupies the considered position. In that context, the relative value between two different densities would therefore reflect the rate at which a given lineage would outperform the other. Consistently, if we consider all lineages competing, we conclude that the probability of lineage survival after a large time period can be derived directly from the normalization of the asymptotic densities, i.e.: $$p(c_n)\approx\frac{\rho_n(\infty)}{\sum_j\rho_j(\infty)}\quad,$$ where, interestingly, $\rho_n(\infty)\equiv \lim_{t\to \infty}\rho_n(z,t)$ is a constant independent of the position $z$. Importantly, this leads us to a well-defined, stationary probability of lineage survival: $$p(c_n)\propto {\rm exp}\left\{-\frac{k_d}{2 k_r}n^2\right\}\quad.
\label{eq:p(c_k)}$$ The above equation, which is a central result of the paper, predicts the probability that a cell starting at position $n$ will “win the competition” and colonize the whole one-dimensional system (see Appendix, section A4 for details).
Functional stem cell numbers and dynamics in the stochastic conveyor belt
=========================================================================
The prediction for the probability of long-term lineage survival is surprisingly simple, decaying as a Gaussian distribution as a function of position away the niche, with a length scale that is simply the amplitude of the stochastic fluctuations divided by the proliferation rate, $\sim\sqrt{k_r/k_d}$ (see equation (\[eq:p(c\_k)\])). Intuitively, cells close to the origin have the highest chance to win and survive, whereas this probability drops abruptly for cells starting the competition further away, i.e. around $N_s$ cell diameters away from the base, with: $$N_s^{1D} = 1+\sqrt{\frac{k_r}{k_d}}\quad.
\label{eq:NsTrivial}$$ Note that the first term satisfies the boundary condition this at, in the case $k_r=0$, the system has a single functional stem cell (located at the base) in 1D. Equation (\[eq:NsTrivial\]) thus implies that multiple rows of cells possess long-term self-renewal potential (as assessed for example, in a lineage tracing assay), emerging through their collective dynamics, and with a number that depends only on the ratio of the division to rearrangement rates (resp. $k_d$ and $k_r$). Although equation (\[eq:NsTrivial\]) is a 1D criterion, we show (Appendix Text, section B) that it holds and can be generalized in more complex geometries. In particular, in a cylindrical 2D geometry, we show the functional stem cell number would simply be the same number $N_s^{1D}$ of cell rows (arising from the stochastic conveyor belt dynamics), multiplied by the number of cells per row (fixed by the geometry of the tissue). Therefore, this generalizes the results of Ref. [@Lopez-Garcia:2010], as we do not fix the stem cell number $N_s$ explicitly, which rather emerges from an interplay between geometry and stochastic conveyor belt dynamics, together with the competitive dynamics being qualitatively different in the flow direction (Appendix Text, section A).
In spite of the approximations outlined above, stochastic numerical simulations of the model system show excellent agreement with equation (\[eq:p(c\_k)\])(see Fig. (\[fig:Schema\]d,e)). We also note that although we model cells with identical intrinsic properties, the model can be extended be more complex cases, for instance with non-neutral cells (see below) or cells away from the niche going through irreversible differentiation. Moreover, although we have assumed here that positional rearrangements occur between two cells, more complex sources of noise can be considered, and lead to the same qualitative results. These include, for instance, post-mitotic dispersal, as seen during the branching morphogenesis of the kidney uteric bud [@Packard:2013] and where daughter cells can travel long distances outside the epithelium post-division, or correlated “tectonic” movements of the epithelium, where cells could collectively reposition relative to the niche, as proposed during mammary or gut morphogenesis [@Scheele:2017; @Guiu:2019] (see Appendix section C for details).
We now turn to experimental data to test whether the proposed dynamics can help predict the number of functional stem cells in several organs, as well as the evolution of the survival probability with starting position of a clone. Although the division rate $k_d$ is well-known in most systems considered, the stochastic movement rate $k_r$ is harder to estimate, and can potentially vary widely, from rather small in intestinal crypts [@Ritsma:2014], to large in mammary and kidney tips, with extensive clonal fragmentation and random cell movements [@Scheele:2017; @Riccio:2016].
Predictions on clonal dynamics and survival
-------------------------------------------
Intra-vital live-imaging provides an ideal platform to test the model, as it provides both knowledge of the starting position of a given cell as well as its clonal time evolution (whereas classical lineage tracing relies on clonal ensembles obtained from fixed samples). In small intestinal crypts, different Lgr5+ cells have been predicted to have very different lineage survival potential on the short-term, depending on their position within the stem cell niche, resulting in an effective number of stem cells smaller than the number of Lgr5+ cells [@Ritsma:2014; @Kozar:2013]. We thus reanalyzed quantitatively this dataset by plotting the survival probability of a clone as a function of its starting position $n$ (see Fig. (\[fig:RealSurvival\]) after a given time period, $p(c_n,\tau)$. We then compared this to a 2D stochastic simulation of the model (see Appendix section D for details). Importantly, we found a good qualitative and quantitative agreement between model and data, with the survival probability decaying smoothly with the starting position (see Fig. (\[fig:RealSurvival\]b)). The only parameter here was $k_r/k_d = 1$, which fits well with short-term live imaging experiments and the idea of cell division promoting cell rearrangement [@Ritsma:2014].
To back these simulations with an analytical prediction on stem cell numbers, the details of tissue geometry must be taken into account (with the number of cells per row $i$ needing to be estimated, while the number of rows participating in the competition arising as an emergent property from the 1D model). A good approximation is based on that fact that the crypt can be abstracted as a hemispherical monolayer with radius $R$ (measured in units of cell diameter) coupled to a cylindrical region (see Appendix, section B for details), so that one can get the number of stem cells, $N^{2D}_s$, as: $$N^{2D}_s\approx 2\pi R^2\left[1-\cos\left\{ \frac{1}{R}\left(1+\sqrt{\frac{\pi}{2}\frac{k_r}{k_d}}\right)\right\}\right]\quad.
\label{eq:Ns}$$ The above prediction is similar to the 1D prediction in equation (\[eq:NsTrivial\]), although it takes into account the effect of tissue geometry. With $k_r/k_d \approx 1$ as above, and estimating $R \approx 2$ for the radius, our simple theory then predicts that the number of functional stem cells should be $N^{2D} \approx 8$, which agrees well with measurements of [@Ritsma:2014], as well as inferred numbers from continuous clonal labelling experiments [@Kozar:2013], which is expected, as our model reduces to the 1D ring model of Ref. [@Lopez-Garcia:2010] for low $k_r/k_d$.
We then sought to test the model further using a published dataset on embryonic kidney branching in explants [@Riccio:2016], which has been recently noted to be a highly stochastic process, with neighbouring cells at the start of the tracing ending up either surviving long-term in tips or being expelled to ducts. Moreover, Ref. [@Riccio:2016] observed extensive random cell intercalations, in addition to the previously described mitotic dispersal [@Packard:2013], where cells extrude from the epithelium post-division and reinsert at a distance of $d_c$ cell diameters away. Importantly, these processes can still be captured as an effective diffusion coefficient $k_r$ in our framework (see Appendix, section D for details). Specifically, knowing that the fluctuations may occur at each duplication, and that they imply a displacement up to $d_c \approx 2-4$ cell lengths, we can estimate that $k_r/k_d \approx d^2_c\quad$ at the minimum (i.e. discounting other fluctuations). Note that the conveyor belt dynamics applies exactly as in the crypt: The only difference is that the reference frame from which the dynamics is observed now changes and, instead of taking the bottom of the organ as the rest reference, we take the newly created ductal cells produced by the tips. In that reference frame, the dynamics predicts a continuous elongation of the tips of the branches, as observed in reality (see Fig. (\[fig:RealSurvival\_Kidney\]a) and Appendix, section A for details).
The above observation argues again that noise will play a key role in kidney tip cell dynamics. Strikingly, extracting from Ref. [@Riccio:2016] the probability of survival as a function of distance from the edge of a tip, we found that the 2D simulations of our model provided again an excellent prediction for the full probability distribution (see Fig. (\[fig:RealSurvival\_Kidney\]b)), with cells much further away (compared to the intestinal crypt) having a non-negligible probability to go back and contribute. Again, the only fit parameter was the ratio $k_r/k_d = 16$, which agrees well with our estimate of the noise arising from mitotic dispersal. Taking into account the full 2D geometry as above, and estimating in this case a tip radius of $R=3-5$ cells, this predicts $N_s \approx 90\pm 10$, which could be tested in clonal lineage tracing experiments.
These two examples show that the same model and master curve for the survival probability of clones can be used in different organs to understand their stem cell dynamics, and shows that ratios of relocation to advection $k_r/k_d$ can be widely different even in systems with similar division rates $k_d$.
Number of functional stem cells in the developing mammary gland {#sec:Mam}
---------------------------------------------------------------
Next, we sought to test the model in the setting of mammary gland morphogenesis, where extensive cell movements have been reported within tips via intravital live-imaging [@Scheele:2017], with rapid rearrangements occurring on time scales of a few hours (see Fig. \[fig:geometry\] and Fig. S7 of the SI). In this case, however, tips cannot be followed for long-enough for survival probabilities to be inferred as in Figs. (\[fig:RealSurvival\]) and (\[fig:RealSurvival\_Kidney\]) for intestine and kidney, respectively. However, extensive clonal dispersion has been observed in quantitative clonal lineage tracing experiments during pubertal growth [@Davis:2016; @Scheele:2017], and we therefore sought to develop a statistical method to infer the value of noise from these experiments.
Turning back to published lineage-tracing datasets, where single mammary stem cells are labelled at the beginning of puberty (3 weeks of age) and traced until either 5 weeks or 8 weeks of age, clones in tips displayed extensive fragmentation, which is expected to be directly related to the ratio $k_r/k_d$ (see Fig. (\[fig:geometry\]c,d)). We thus ran as above two-dimensional simulations of our stochastic conveyor belt model (see Appendix, section D, for details), using measured values of the tip width and length to set the geometry. As a metric for clonal dispersion, we then computationally measured for each labelled cell the distance to its closest clonal neighbour: for a fully cohesive clone, all cells should be touching and the distance to the closest neighbour should be always one cell diameter. Increasing the value of $k_r/k_d$ robustly increased the closest neighbour distance. We then performed the same measurements in the experimental data set, both for the 5 weeks and 8 weeks time points (see Fig. (\[fig:geometry\]c,d)), but also for luminal and basal cell types separately, given the demonstrated unipotency of these cell populations in pubertal development [@Davis:2016; @Scheele:2017]. We found highly consistent results in all four cases (average closest distance of around $1.85$ cell diameter) which allowed us to infer a ratio of (see Appendix, section D): $$k_r/k_d \approx 2-5\quad,
\label{eq:mammary}$$ in mammary gland, emphasizing the importance of considering stochasticity in the conveyor belt picture. Indeed, we found that, with this fitting parameter, the model reproduced well the probability distribution of closest distances, both at the 5 weeks and 8 weeks time points (see Fig. (\[fig:geometry\]c,d)).
In addition to this value, we must again pay attention to the geometry of the mammary tip, with basal cells forming a 2D monolayer (similar to the previous cases) while luminal cells form multiple layers in 3D within the tip. Assuming that the intercalation between cells occurs mainly at the same layer, the system of luminal cells in the tip of the mammary gland can be abstracted as $R-1$ successive hemispherical 2D layers. Let us emphasize the dependence of $N^{2D}$, as defined by equation (\[eq:Ns\]), on $R$, writing $N^{2D}_s\equiv N^{2D}_s(R)$. In that case, the amount of luminal stem cells can be inferred as: $$N_s^{3D}=\sum_{R'<R}N_s^{2D}(R')\quad.
\label{eq:Ns3D}$$
Taking the fitted range of $k_r/k_d\in (2,5)$, together with an estimation of the radius of $R=5\pm 2$, equation \[eq:Ns3D\] then predicts that a number of luminal stem cell per tip of $N_s^{3D}=162\pm 114$, in good quantitative agreement with experimental estimates from lineage tracing of $N_s^{exp}=172 \pm 102$ (mean$\pm$s.d.) [@Scheele:2017].
For basal cells, using the same parameters for a 2D monolayer, equation \[eq:Ns\] predicts that $N_s^{2D}=33\pm 14$, against empirical observations reporting an amount of basal stem cells of at least 15 [@Davis:2016], and $N_s^{exp}=93 \pm 76$ (mean$\pm$s.d.) [@Scheele:2017]. Although the prediction thus falls in the correct range, the under-estimation of basal stem cell number may be due to the highly anisotropic geometry of basal stem cells.
Discussion
==========
The main objective of this study is to provide new insights to the question of whether stem cell function is a cell-intrinsic, inherited, property, or rather an extrinsic, context-dependent notion emerging from the collective dynamics of a tissue [@Simons:2011; @Blanpain:2013; @Moris:2016; @Yang:2017]. To that end, we took a complementary standpoint to the one based on the classification of molecular markers and their potential functional role, adopting a purely dynamical/geometrical approach. In that sense, we tested to what extent the observed phenomena can be attributed to extrinsic, context dependent properties arising from the dynamics and geometry of the system. We analyzed in detail the dynamical properties of the process of tissue growth and self-renewal, taking into account the presence of stochastic cell rearrangements, cell duplication rates and tissue geometry. The combination of these ingredients gives rise to a complex reaction-diffusion process that can be abstracted as a stochastic-conveyor belt; that is, cells follow an exponential growth and are dragged along the tissue by an Ornstein-Uhlenbeck process with positive drift. Within this framework, we can predict several critical quantities, which emerge solely from the dynamics. Among others, the hierarchy of lineage survival, which has been shown to follow a Gaussian distribution that matches previously reported data, from kidney development and intestinal crypt renewal. From it, one can infer the effective size of the stem cell niche or functional number of stem cells, showing very good agreement with recent quantitative observations both in the crypt and in the development of the kidney and mammary gland. In the latter, we identified the traces of random rearrangements from the fragmentation of clones in lineage tracing data, obtaining an interesting steady ratio around $k_r/k_d \approx 3$, which is enough to allow realistic predictions of functional stem cell numbers in each mammary tip. Furthermore, the theory also makes non-trivial predictions on the survival probability of cells as a function of distance from niches, which we could validate both in intestinal crypt homeostasis and kidney morphogenesis, although the ratio of noise to advection $k_r/k_d$ differed by more than an order of magnitude in both systems [@Ritsma:2014; @Riccio:2016]. We finally evaluated the impact of the presence of mutant, fast replicating cells on the lineage survival dynamics, which were revealed to be strongly dependent on the number of functional stem cells in wild-type. This provides simple predictions that could guide further experimental work. It is worth emphasizing that these predictions rely only on the geometry of the tissue, and the ratio between the stochastic rearrangements and cell duplication rates.
We also emphasize that, although we have sketched here the simplest source of noise in cellular movements (random exchange of position in cell neighbors), our approach and results are in fact highly robust to different types of microscopic assumptions, and should thus be seen as representative of a general class of models for stem cell dynamics with advection and noise, rather than a specific microscopic mechanism. In mammary gland and kidney morphogenesis, direct cell-cell rearrangements are observed [@Scheele:2017; @Riccio:2016], while kidney also displays mitotic dispersal [@Packard:2013], where noise arises from the randomness of cell re-insertion in the layer after division. Furthermore, on short-time scales, directed cellular movements have been observed in kidney tip morphogenesis, with Ret and Etv4 mutant clones being statistically overtaken by wild-type cells, leading to the proposal that Ret/Etv4 were involved in directional movement towards tips [@Riccio:2016]. However, tips maintain heterogeneity in Ret expression through branching, arguing that cells must shuttle between high-Ret and low-Ret states [@Riccio:2016]. In our model, short-term directional movements based on Ret signalling, which is itself variable would still result in long-term diffusion motion, effectively give an additional contribution to $k_r$ on long time scales. Finally, we also envisioned a source of noise $k_r$ which does not result in clonal dispersion. We show indeed that “tectonic” and collective movements of the monolayer away from the niche (see Fig. S5 of the SI), would again give rise to the exact same results on the shape of the survival probability as described here (again resulting in effectively higher $k_r$). Such tectonic movements are particularly relevant in developmental settings, such as gut morphogenesis, where the shape of the intestine changes, for instance via villi bending, which displaces collectively cells from villi regions to niche/crypt regions [@Guiu:2019]. This is also relevant for branching morphogenesis, as proposed in Ref. [@Scheele:2017], because branching collectively repositions cells into tip regions due to global shape changes of the epithelium, and could be an additional source of noise explaining the high value of $k_r$ we inferred in kidney morphogenesis, for instance.
What are the consequences of this intrinsic stochasticity, beyond the emergence of a structured stem cell region? We made the observation that the size of the stem cell niche, understood as the fraction of cell-lineages with non-negligible probability of colonizing the system, may differ from the dividing region. This result is relevant because it complements and enriches the concept of a stem cell niche. Moreover, it is worth pointing out that a positionally-defined stem cell potential, jointly with stochastic fluctuations, can serve as a protection mechanism against death or malfunctioning of the tissue. Indeed, the existence of fluctuations can give rise to the possibility of spontaneous replacement ensuring the healthy functioning of the tissue. The exploration of the consequences of such stochasticity and trade-offs is an active field of research in stem cell modelling, both in the context of cancer [@Michor:2004; @Sottoriva:2015; @Altrock:2015; @Williams:2016; @Tkadlec:2018], but also for instance in ageing [@Hormoz:2013].
The proposed framework is in fact completely general and can be, in principle, applied to any tissue dynamics in which the replication occurs from a given spatial point, such that the growth can be projected in a given, well defined single coordinate. Nevertheless, this strong geometrical constraint is not always playing a predominant role, for instance in the context of an “open niche” such as spermatogenesis [@Jorg:2019] or skin homeostasis [@Rompolas:2016] where most cells attached to the basement membrane form a 2D layer of equipotent progenitors. In spermatogenesis, competition instead occurs at the level of diffusible mitogen/fate determinants consumption [@Kitadate:2019; @Jorg:2019], which shares some conceptual similarities with our model. Outside of epithelial tissues, macrophages and fibroblasts have been shown to form a stable circuit of interactions, giving robustness to the density regulation of each population [@Zhou:2018]. Therefore, our approach must be taken as part of a more general enterprise, namely, the role of the complex, global dynamics of the system in defining the functional stem cells. Indeed, understanding the interaction between the dynamics we propose here and the role of the molecular markers/functional differentiation dynamics *in vivo* would be a natural next step. For example, the prediction of the number of stem cells in the crypt, even though accurate with respect to live-imaging functional data, does not explain why Lgr5 levels display a clear on/off pattern inside vs. outside the crypt, and reversal to a stem cell fate is not always a fast process dependent solely on niche cues. Therefore, a natural follow up should explore the relation between the emergent properties of the dynamics and the molecular markers, jointly to their functional differentiation [@Itzkovitz:2012; @Moor:2018]. In addition, more complex geometries, and with that, other tissues, could be explored within our framework.
Acknowledgements
================
We thank all members of the Hannezo, Simons and Van Rheenen groups for stimulating discussions.
[10]{}
Wang J, Xu L, Wang E, Huang S (2010) The potential landscape of genetic circuits imposes the arrow of time in stem cell differentiation. 99(1):29–39.
Simons, B. D. ans Clevers H (2011) Strategies for homeostatic stem cell self-renewal in adult tissues. 145:851–862.
Blanpain C, Simons BD (2013) Unravelling stem cell dynamics by lineage tracing. 14(8):489.
Watt FM, Huck WT (2013) Role of the extracellular matrix in regulating stem cell fate. 14(8):467.
Blanpain C, Fuchs E (2014) Plasticity of epithelial stem cells in tissue regeneration. 344(6189):1242281.
Plaks V, Kong N, Werb Z (2015) The cancer stem cell niche: how essential is the niche in regulating stemness of tumor cells? 16(3):225–238.
Moris N, Pina C, Arias AM (2016) Transition states and cell fate decisions in epigenetic landscapes. 17(11):693.
Yang H, Adam RC, Ge Y, Hua ZL, Fuchs E (2017) Epithelial-mesenchymal micro-niches govern stem cell lineage choices. 169(3):483–496.
Scheele C, et al. (2017) Identity and dynamics of mammary stem cells during branching morphogenesis. 542(7641):313–317.
Rompolas P, Mesa KR, Greco V (2013) Spatial organization within a niche as a determinant of stem-cell fate. 502(7472):513.
Rompolas P, et al. (2016) Spatiotemporal coordination of stem cell commitment during epidermal homeostasis. 352(6292):1471–1474.
Kitadate Y, et al. (2019) Competition for mitogens regulates spermatogenic stem cell homeostasis in an open niche. 24(1):79–92.
Ritsma, L. ea (2014) Intestinal crypt homeostasis revealed at single-stem-cell level by in vivo live imaging. 507:362–365.
Barker Nea (2007) Identification of stem cells in small intestine and colon by marker gene lgr5. 449:1003–1007.
Snippert, H. J. ea (2010) Intestinal crypt homeostasis results from neutral competition between symmetrically dividing lgr5 stem cells. 143:134–144.
López-García C, Klein AM, Simons BD, Winton DJ (2010) Intestinal stem cell replacement follows a pattern of neutral drift. 330:822–825.
Snippert HJ, Clevers H (2011) Tracking adult stem cells. 12:113–122.
Klein AM, Simons BD (2011) Universal patterns of stem cell fate in cycling adult tissues. 138(15):3103–3111.
Visvader JE, Stingl J (2014) Mammary stem cells and the differentiation hierarchy: current status and perspectives. 28:1143–1158.
van Es, J. H. ea (2012) Dll1+ secretory progenitor cells revert to stem cells upon crypt damage. 14:1099–1104.
Guiu J, et al. (2019) Tracing the origin of adult intestinal stem cells. 570(7759):107.
Krieger T, Simons BD (2015) Dynamic stem cell heterogeneity. 142(8):1396–1406.
Post Y, Clevers H (2019) Defining adult stem cell function at its simplest: The ability to replace lost cells through mitosis. 25(2):174–183.
Riccio P. ea (2016) Ret and etv4 promote directed movements of progenitor cells during renal branching morphogenesis. 14(2):e1002382.
Packard A, et al. (2013) Luminal mitosis drives epithelial cell dispersal within the branching ureteric bud. 27(3):319–330.
Gardiner CW (1983) [*Handbook of Stochastic Methods for Physics, Chemistry and the Natural Sciences*]{}. (Springer-Verlag:Berlin).
Van Kampen NG (2007) [*Stochastic Processes in Physics and Chemistry*]{}. (North Holland), 3 edition.
Britton N (1986) [*Reaction-Diffusion Equations and Their Applications to Biology*]{}. (Academic Press).
Grindord P (1996) [*The Theory and Applications of Reaction-Diffusion Equations*]{}. (Oxford University Press).
Kozar S, et al. (2013) Continuous clonal labeling reveals small numbers of functional stem cells in intestinal crypts and adenomas. 13(5):626–633.
Davis FM, et al. (2016) Single-cell lineage tracing in the mammary gland reveals stochastic clonal dispersion of stem/progenitor cell progeny. 7:13053.
Visvader JE (2011) Cells of origin in cancer. 469(7330):314.
Snippert HJ, Schepers AG, van Es JH, Simons BD, Clevers H (2014) Biased competition between lgr5 intestinal stem cells driven by oncogenic mutation induces clonal expansion. 15(1):62–69.
S[á]{}nchez-Dan[é]{}s A, et al. (2016) Defining the clonal dynamics leading to mouse skin tumour initiation. 536(7616):298.
Williams MJ, Werner B, Barnes CP, Graham TA, Sottoriva A (2016) Identification of neutral tumor evolution across cancer types. 48(3):238.
Altrock PM, Liu LL, Michor F (2015) The mathematics of cancer: integrating quantitative models. 15(12):730.
Michor F, Iwasa Y, Nowak MA (2004) Dynamics of cancer progression. 4(3):197.
Sottoriva A, et al. (2015) A big bang model of human colorectal tumor growth. 47(3):209.
Tkadlec J, Pavlogiannis A, Chatterjee K, Nowak MA (2018) Fixation probability and fixation time in structured populations. .
Hormoz S (2013) Stem cell population asymmetry can reduce rate of replicative aging. 331:19–27.
Jörg DJ, Kitadate Y, Yoshida S, Simons BD (2019) Competition for stem cell fate determinants as a mechanism for tissue homeostasis. .
Zhou X, et al. (2018) Circuit design features of a stable two-cell system. 172(4):744–757.
Itzkovitz S, et al. (2012) Single-molecule transcript counting of stem-cell markers in the mouse intestine. 14(1):106.
Moor AE, et al. (2018) Spatial reconstruction of single enterocytes uncovers broad zonation along the intestinal villus axis. 175(4):1156–1167.
I. Karatzas, and S. E. Shreve (1988) [*Brownian Motion and Stochastic Calculus*]{} Springer-Verlag:Berlin.
C. Guillot,and T. Lecuit (2013) Mechanics of epithelial tissue homeostasis and morphogenesis [*Science*]{} 340, 1185–1189
W. Klingenberg (1978) [*A course on differential geometry*]{} Translated by D. Hoffman. Springer-Verlag:Berlin.
W. K" uhnel (2006) [*Differential Geometry: Curves–Surfaces–Manifolds*]{} 2nd Edition. (Student Mathematical Library series, 77) American Mathematical Society:New york.
M. Alieva et al. (2014) Imaging windows for long-term intravital imaging. IntraVital: 3, 2
Basic dynamics
==============
The simplest abstraction of our system is composed by a column of cells arranged in a finite segment $[0,N]$, in which the length unit is scaled to the average length of a cell. Each cell divides at constant rate $k_d$. We consider that the system can only grow towards the positive axis and that the density of cells remains constant. This division creates a positive push up force because newly born cells occupy upper adjacent positions in the array, pushing the cells that were there to higher positions. As long as a cell reaches the position $N$ and is pushed, it disappears from the system. Both intestinal crypts and elongating mammary or kidney tips can be described by this dynamics, up to a change of referential frame: intestinal crypts expel differentiated cells with the crypt base staying stationary, while differentiated cells stay in place while the tip moves forward in the case of mammary gland or kidney development. The coordinate $x$ or $z$ through which the dynamics take place will depict only the distance to the origin of the organ, without assumptions on the global movement of it: In the crypt that will represent the distance to the bottom, in the case of the kidney or the mammary gland, the distance to the tip –see figure (\[fig:Frame\]). In addition, the position of the cells can fluctuate stochastically at rate $k_r$ (either via local cell-cell rearrangements, or more global movements of cells relative to the niche, see sections below for more details). Importantly, the results that we derive, as described below, are highly generic to different types of assumptions on the microscopic dynamics, as long as the basis features of advection and rearrangements exist in a system (sketched in Fig. 1 of the main text).
Before proceeding with the details of the derivations and numerical computations, we present an overview of the strategy followed to derive the result. The aim is to convey a bird-eye view over the whole strategy, and to highlight the assumptions underlying the approach. The modelling and conceptual steps are the following:
1. The first step is to define the conveyor-belt dynamics at the single cell level. The probability that a cell that started at position $k$ is at a given position of the organ at time $t$. This dynamics obey a Ohrnstein-Uhlenbeck process with positive drift. The evolution of the probabilities is governed by the following Fokker-Planck equation: $$\frac{1}{k_d}\frac{\partial p_k}{\partial t}=-\frac{\partial }{\partial x}(xp_k)+\frac{k_r}{2k_d}\frac{\partial^2 p_k}{\partial x^2}\quad.$$
2. We then add the [*reaction*]{} term, that is, the cell, aside the random fluctuations provided above, also divides at a certain rate. The reaction-diffusion equation governing the density of lineage $c_k$ is: $$\frac{1}{k_d}\frac{\partial \rho_k}{\partial t}=-\frac{\partial }{\partial x}(x\rho_k)+\frac{k_r}{k_d}\frac{\partial^2 \rho_k}{\partial x^2}+\rho_k\quad.$$
3. In spite the above system has no stationary solutions, the density of the lineage $c_k$ at a given position of the organ goes to a constant over the whole organ as $t\to \infty$. We call this density $\rho_k(\infty)$.
4. We [*assume*]{} that $\rho_k(\infty)$ is proportional to the probability that the lineage $c_k$ will colonize the whole organ. Therefore: $$p(c_k)\approx\frac{\rho_k(\infty)}{\sum_j\rho_j(\infty)}\quad.$$
5. This probability turns out to be a gaussian-like distribution depending only on the ratio $k_r/k_d$: $$p(c_k)\sim {\rm exp}\left\{-\frac{k_d}{2k_r}k^2\right\}\quad.$$
6. Once the 1D problem is solved, we go to more general geometries. As long as the push up force is projected towards a single coordinate of the system, the general reaction-diffusion reads: $$\frac{1}{k_d}\frac{\partial \rho_k}{\partial t}=-\frac{\partial }{\partial z}\left[\left(\frac{dS_k}{dz}\right)^{-1}S_k \rho_k\right]+\frac{k^z_r}{2k_d}\frac{\partial^2 \rho_k}{\partial z^2}+\rho_k\quad,$$ where: $$S_k=\int . . .\int_0^{z_k}\sqrt{g}dx_1. . . d{z}\quad,$$ $g$ is the determinant of the metric tensor, $z$ is the coordinate over which the push-up force due to the fluctuations is projected and $k^z_r$ the explicit value of these fluctuations projected over the relevant coordinate $z$.
7. After some approximations, we reach a differential equation that can be integrated, for the case of geometries corresponding to cylinders coupled to hemispheric regions, as is the case of the organs studied: $$\frac{1}{k_d}\frac{\partial \rho_k}{\partial t}\approx-\frac{2}{\pi} \frac{\partial }{\partial z}(z\rho_k)+\frac{k_r}{4k_d}\frac{\partial^2 \rho_k}{\partial z^2}+\rho_k\quad.
\label{eq:general_Fig}$$
8. We finally study the cases where different $k_r/k_d$ rates are present in the same system.
Single cell dynamics
--------------------
This process is a mixture of Brownian motion with a given amplitude $k_r$ and a drift parameter that depends on the position. In the continuous limit, the position of the cell, $x$, can be described as a random variable satisfying the following stochastic equation: $$dx=k_dxdt+\sqrt{k_r}dW\quad,%\nonumber
\label{eq:LangevinA}$$ being $dW$ the differential of the standard Wiener process with mean $0$ and variance $1$. In other words, we are describing a kind of [*Ornstein-Uhlenbeck*]{} process, with positive drift $k_d$ [@Gardiner:1983; @VanKampen:2007]. The above described stochastic process has no stationary solutions, which is in agreement with the conveyor belt-like dynamics of the systems under study: all cells will sooner or later be pushed out from the system (in the case of the crypt, or will be left behind, as in the case of the mammary gland development, note that for finite systems, this will consist of a single remaining winner lineage colonizing the whole tissue). One can still compute the time-dependent solutions, considering initial conditions $t_0=0$, $x_0=\delta(x-k)$ and natural boundary conditions[^2] as follows: First, we observe that: $$d\left(e^{-k_d t} X(t)\right)=dX(t)e^{-k_d t}-k_de^{-k_d t}X(t)dt\quad.$$ Then, multiplying both sides of equation (\[eq:LangevinA\]) by $e^{-k_d t}$, and after some algebra, one finds that: $$d\left(e^{-k_d t} X(t)\right)=\sqrt{k_r}e^{-k_dt}dW\quad,$$ leading to: $$X(t)=x_0e^{k_dt}+\int_0^t \sqrt{k_r}e^{k_d (t-s)} dW\quad.
\label{eq:Solution}$$ The integral is a standard stochastic integral with respect to a Wiener process. According to [*Ito’s isometry*]{} [@Karatzas:1988] one has that the law governing the random variable described by the integral is a normal distribution $\mathbf{N}(0,\sigma^2(t))$. In our case this reads: $$\int_0^t e^{k_d (t-s)}\sqrt{k_r} dW \sim{\mathbf N}\left(0,\int_0^t\left|\sqrt{k_r}e^{k_d (t-s)}\right|^2ds\right)\quad,$$ which means that he explicit form of $\sigma^2(t)$, is thus given by: $$\begin{aligned}
\sigma^2(t)=\int_0^t\left|\sqrt{k_r}e^{k_d (t-s)}\right|^2ds=\frac{k_r}{2k_d}\left(e^{2k_dt}-1\right)\quad.
\label{eq:sigmaT}\end{aligned}$$ Finally, from equation (\[eq:Solution\]), we conclude that the time dependent mean, $\mu(t)$, is: $$\mu(t, x_0)=x_0e^{k_d t}\quad.$$ By setting $$\tau\equiv k_dt\quad,
\label{eq:tautau}$$ and $x_0=k$ as the starting position, one gets the time evolution of the probability distribution of for the position of the random walker which started at $k$: $$p_k(x,\tau)=\sqrt{\frac{k_d}{2\pi k_r\left(e^{2\tau}-1\right)}}{\rm exp}\left\{{-\frac{k_d}{2k_r}\frac{\left(x-ke^{\tau}\right)^2}{\left(e^{2\tau}-1\right)}}\right\}\;.
\label{eq:solutionP}$$ In words, the solution is given by random variable following a normal distribution whose mean and variance run exponentially fast in time through the positive axis. In Fig. (\[fig:Dynamics\]a) of this Appendix, we plotted some snapshots of this time dependent probability.
Lineage dynamics
----------------
The Fokker-Planck equation accounting for the probability that a given random walker starting at $k$ will be at $x$ at a given time $\tau$ –as defined in equation (\[eq:tautau\])–, for the stochastic process described by equation (\[eq:LangevinA\]), $p_k(x,\tau)$, is given by: $$\frac{\partial p_k}{\partial \tau}=-\frac{\partial }{\partial x}(xp_k)+\frac{k_r}{2k_d}\frac{\partial^2 p_k}{\partial x^2}\quad.
\label{eq:FP}$$ If we want to study the [*density*]{} of cells of the lineage started by cell $k$ in a given position, $\rho_k(x,\tau)$, we have to take into account that, in a given time step, $\rho_k(x,\tau)$ new cells of the lineage will emerge in such a position. Therefore, to describe the whole process, we need a diffusive part, given by the differential operator $-\frac{\partial }{\partial x}x+\frac{k_r}{k_d}\frac{\partial^2 }{\partial x^2}$ of equation (\[eq:FP\]) and a reactive part, given by $\rho_k(x,\tau)$. To derive the reactive part, we take into account the following reasoning: If the density of a given lineage at position $x$ at time $\tau$ is $\rho_k(x,\tau)$, one expects that, in average, new $\rho_k(x,\tau)$ cells will emerge within this interval in a time unit –as it is standard in the exponential growth. So, one has that the equation accounting for the time evolution –in units of $\tau=k_d t$– of such density is: $$\frac{\partial \rho_k}{\partial \tau}=-\frac{\partial }{\partial x}(x\rho_k)+\frac{k_r}{k_d}\frac{\partial^2 \rho_k}{\partial x^2}+\rho_k\quad.
\label{eq:FisherKPP}$$ The above equation is a reaction diffusion equation whose diffusive part observes a Ornstein-Uhlenbeck processes with positive drift and whose reaction part is given by an standard exponential growth. Knowing that the solution of equation (\[eq:FP\]) is given by $p_k(x,\tau)$ as defined in equation (\[eq:solutionP\]), the solution of equation (\[eq:FisherKPP\]), with initial conditions $\rho_k(x,0)=\delta(k-x)$, is given by: $$\rho_k(x, \tau)=A\cdot e^\tau p_k(x,\tau)\quad.
\label{eq:rho}$$ We observe that, according to equation (\[eq:rho\]), the term $e^\tau p_k(x,\tau)$ can be safely approximated by $$e^\tau p_k(x,\tau)\approx \sqrt{\frac{k_d}{2\pi k_r}}{\rm exp}\left\{-\frac{k_d}{2k_r}\left(\frac{x-ke^{\tau}}{e^{\tau}}\right)^2\right\}\quad.
\label{eq:Safeapprox}$$
Determination of the integration constant
-----------------------------------------
We observe that, even though random fluctuations in cell positions are allowed, epithelial tissues remain confluent [@Guillot:2013], and maintain a constant density at homeostasis. This imposes that, in the limit of a continuous array of cells and for $\tau\to\infty$, the overall density of cells of the site $x$ must stabilize to $1$, and must not depend on the position $x$. We will call that condition the [*confluent tissue condition*]{}. Consistent with the above reasoning, we observe that: $$\int_0^N e^\tau p_x(x',\tau) dx\to c\in \mathbb{R}^+\quad.$$ To determine the constant $A$ of equation (\[eq:rho\]), we impose the [*confluent tissue condition*]{}. According to the above equation, this is satisfied by: $$A\equiv \frac{1}{c}\quad.$$ Let us remark that we consider $x\in [0,N]$. Then, knowing that, in this case: $$\lim_{\tau\to \infty}\int_0^N e^\tau p_x(x',\tau) dx=\int_0^N \lim_{\tau\to \infty} \{e^\tau p_x(x',\tau)\} dx\quad,$$ we compute the limit, taking into account equation (\[eq:Safeapprox\]): $$\lim_{\tau\to \infty} e^\tau p_x(x',\tau)=\sqrt{\frac{k_d}{2\pi k_r}}{\rm exp}\left\{-\frac{k_d}{2k_r}x^2\right\}\quad,$$ which is a gaussian with zero mean and $\sigma^2=k_r/k_d$. so that: $$\begin{aligned}
c&=&\int_0^N\sqrt{\frac{k_d}{2\pi k_r}}{\rm exp}\left\{-\frac{k_d}{2k_r}x^2\right\}dx\nonumber\\
&=&\frac{1}{2}{\mathbf{erf}}\left(N\sqrt{\frac{k_d}{2k_r}}\right)\nonumber\\
&\approx&\frac{1}{2}\quad,\nonumber\end{aligned}$$ where ${\bf erf}(x)$ is the [*error function*]{}. The approximation $c\approx 1/2$ holds as soon as $N\gg 1$. Therefore, in that approximation, $A=2$.
By setting the integration constant $A=2$, the process fills up space in a correct manner that is: The density of cells is always conserved as $1$ cell$\times$length unit, as expected for a confluent homeostasic tissue. The final expression for $\rho(k,\tau)$ will thus be: $$\rho_k(x,\tau)\approx \sqrt{\frac{2k_d}{\pi k_r}}{\rm exp}\left\{-\frac{k_d}{2k_r}\left(\frac{x-ke^{\tau}}{e^{\tau}}\right)^2\right\}\quad.
\label{eq:rhoexplicit}$$ In Fig. (\[fig:Dynamics\]b) of this Appendix we plotted some snapshots of this time dependent density.
Lineage survival probability
----------------------------
The first task is to demonstrate that the above dynamics leads, in the long term, to a system composed by descendants of the same cell, i.e., that only a single lineage is present in the system for $\tau \gg 1$. In the case of a 1D system defined over the interval $[0,N]$ with $k_r=0$, this result is trivial, since the lineage of the cell located at the bottom of the system will occupy the whole system whenever it has divided enough times to cover all the positions from $0$ to $N$. In the case $k_r >0$, the strategy to see that it runs towards monoclonality is the following: Let $\Omega_N$ be all the potential configurations the system can have in terms of lineage configuration whenever the system has $2,3, . . .,N$ lineages alive. That is $\sigma\in \Omega_N$ will be a sequence of $N$ numbers each labelling the lineage the cell in a given position belongs to: $$\sigma=\sigma^1, . . .,\sigma^N\;;\quad \sigma^k\in\{0,1, . . ., N\}\quad.$$ In units of $\tau$ and given a configuration $\sigma$ of lineages, for $\Delta \tau >\frac{1}{k_d}\log_2 N$ all the cells will potentially have produced more than $N$ new cells. Therefore, the probability that a lineage has been expelled by the system will be larger than zero. Let us call this probability $p(\downarrow,\Delta \tau)$. Let us define $\tilde{p}_\Omega$ as: $$\tilde{p}_\Omega=\min_{\sigma\in\Omega}\{p(\downarrow,\Delta \tau)\}\quad,$$ that is, the configuration for which the probability of losing a lineage after $\Delta\tau$ steps is minimum. By construction, we know that from whatever configuration containing more than a single lineage has a non-zero probability of losing at least one of them after $\Delta \tau$, so: $$\tilde{p}_\Omega>0\quad.$$ The probability of [*not*]{} losing any lineage whatever the configuration of the system after $n$ steps of duration $\Delta \tau$, $p(=,n\Delta \tau)$ will be bounded as: $$0\leq p(=,n\Delta \tau) \leq \left(1-\tilde{p}_\Omega\right)^{n}\quad,$$ and the probability of losing [*at least one*]{} lineage will be bounded as well as: $$1\geq p(\downarrow,\Delta \tau)\geq 1- \left(1-\tilde{p}_\Omega\right)^{n}\quad.$$ In consequence, $$\begin{aligned}
&&p(=,n\Delta \tau)\to 0\nonumber\\
&&p(\downarrow,\Delta \tau)\to 1\nonumber\quad.\end{aligned}$$ Since the loss of a lineage is a completely irreversible process, the above equation tells us that it is expected that the system will lose lineages until only one survives. Note that for a 2D geometry but $k_r=0$, the system reduces to the stochastic voter model along a 1D ring (the $N_0$ cells at position 0) which was proposed and tested experimentally in Ref. [@Lopez-Garcia:2010], and where the process of monoclonal conversion is diffusive, occurring on time scales of $N_0^2 /k_d$.
Now that we know that the system will reach monoclonality, the next question is to ask which lineage will win the competition. In particular, we are interested in the probability that a given lineage colonizes the whole system as a function of the position of the cell that defined the lineage at $t=0$. To that end, we first compute the asymptotic lineage density, $\rho_k(\infty)$, that reads: $$\rho_k(\infty)\equiv\lim_{\tau\to \infty}\rho_k(x,\tau)=\sqrt{\frac{2k_d}{\pi k_r}}e^{-\frac{k_d}{2k_r}k^2}\quad.
\label{eq:rhoinf=}$$ We observe that this density is independent of $x$ –see Fig. (\[fig:Dynamics\]b) of this SI. Now we make the following assumption: The competition between cells at different levels term is absorbed in a mean field approach by the drift push-up force. In this context, we conclude that the probability of lineage survival $p(c_k)$ –that is, the probability that the whole tissue from $0$ to $N$ will be occupied by cells of the lineage $k$– can be derived directly from the normalization of the asymptotic densities $\rho_k(\infty)$: $$p(c_k)\propto\frac{\rho_k(\infty)}{\sum_j\rho_j(\infty)}\quad,\nonumber$$ which leads to: $$p(c_k)= \frac{1}{Z_N}{\rm exp}\left\{-\frac{k_d}{2k_r}k^2\right\}\quad,
\label{eq:p(c_k)=}$$ being $Z_N$ the normalization constant, namely, $Z_N\equiv \sum_{j\leq N}{\rm exp}\left\{-\frac{k_d}{2k_r}j^2\right\}$.
Dynamics in more general geometries {#sec:General}
===================================
In general we will assume that there is a coordinate $z$ over which the displacement induced by the division takes place. All the dynamics will be, in consequence, studied from its projection over this coordinate –see Fig. (\[fig:CryptGeom\]) of this Appendix for the special cases of hemispheric and spheric geometries. In the case of a $1$D-system, as the one described above, this coordinate is the length, $x$. In the case of a hemisphere, assuming that the push-up force is exerted from the bottom pole, this coordinate is the arc length defined from the position of the cell to the bottom pole itself –see Fig. (\[fig:CryptGeom\]a) of this SI. To gain intuition, consider the surface of the hemisphere with radius $R$: The cells at the bottom pole divide and push the ones on top of them up through the surface. The cell under consideration is located at whatever position defining an arc from the bottom pole equal to $z=R\varphi_k=z_k$, where $\varphi_k$ is the polar angle, meaning that there is an arc of $k$ cells from the given cell to the pole of the hemisphere –see Fig. (\[fig:CryptGeom\]a) of this SI. The successive divisions of cells located at $z_i<z_k$ will result into a net displacement along the angular coordinate $\varphi$ of the cell located initially at $z_k=R\varphi_k$, going from $z_k=R\varphi_k$ to $z_{k'}=R\varphi_k'$, with $\varphi_k'>\varphi_k$. The linear displacement along the surface will be $\Delta z=R(\varphi_k'-\varphi_k)$. Displacements along the other coordinate will have no effect in the push up force.
Drift term
----------
The push-up force or drift term will be described by the function $h(z)$, and will be defined as: $$h(z)=\frac{dz}{dt}\quad.
\label{eq:Defh}$$ In the case of a $1$D-system, as the one described by equation (\[eq:FisherKPP\]), one has that $h(z)=k_d z$. To properly study this dynamics over mor general geometries, let us consider a Riemannian manifold equipped with a metric tensor $\mathbf{g}$, with components $g_{ij}$ [@Klingenberg:1978]. Crucial to our purposes is the property of [*local flatness*]{} [@Klingenberg:1978; @Kuhnel:2006]. Roughly speaking, this implies that, for small enough regions of the manifold, the geometry has euclidean properties. Let us consider that the push up force due to duplications has an origin and is exerted along the direction of a single coordinate $z$ as well. As we did above, the surface/volume units are given such that an average cell has a surface/volume of $1$ in the corresponding units. Consider the starting position of our cell to be $z_k$ along the coordinate $z$ along which the displacement due to the push up force takes place. If the other coordinates are given by $x_1, . . .,x_{n-1}$, the surface/volume encapsulated below this position is given by: $$S_k=\int . . .\int_0^{z_k}\sqrt{g}dx_1. . . d{z}\quad,$$ where $g$ is the determiner of the metric tensor, i.e.: $$g=\left|
\begin{array}{lllll}
g_{11} & g_{12} &. & .&. \\
g_{21} & g_{22} & & & \\
.& & . & & \\
.& & &. &\\
.& & & &.\\
\end{array}
\right|$$ Cells are assumed to divide at rate $k_d$. That implies that $k_d S_k$ new cells will be produced [*below*]{} the cell located at $z_k$. This will create an extra surface/volume of: $$\frac{dS_k}{dt}=k_d S_k\quad,$$ that will project into the coordinate $z$. Using that: $$\frac{dS_k}{dt}=\frac{dS_k}{dz}\frac{dz}{dt}\quad,$$ and, then, equation (\[eq:Defh\]), one can find the general expression for this projection, which reads: $$h(z)=k_d\left(\frac{dS_k}{dz}\right)^{-1}S_k\quad.
\label{eq:h}$$
Projection of the fluctuations
------------------------------
We are only interested on the projection of the dynamics over the coordinate $z$ along which the system grows, as in the other coordinates the competition is neutral and has no net effect in the lineage survival statistics. If the reported fluctuations are $k_r$, we will refer to the projection to the coordinate $z$ as $k_r^z$. In general, $k_r^z$ will be a function $k_r^z=f(k_r,z)$. According to the above results, we will have that the general equation for the evolution of cell lineage densities along the coordinate $z$ will read: $$\frac{\partial \rho_k}{\partial t}=-\frac{\partial }{\partial z}(h(z)\rho_k)+\frac{k^z_r}{2}\frac{\partial^2 \rho_k}{\partial z^2}+\rho_k\quad.
\label{eq:FisherKPPGen}$$ In the case $h(z)$ can be approached as a linear function, i.e., $h(z)\sim a k_d z$ and $k^z_r$ as $k^z_r=bf(k_r)$, one can reproduce the reasoning provided to derive the lineage survival probability, equation (\[eq:p(c\_k)=\]), and obtain: $$p(c_k)\propto {\rm exp}\left\{-\frac{a}{2b}\frac{k_d}{f(k_r)}k^2\right\}\quad.$$
To gain intuition, imagine that we report experimental fluctuations of amplitude $k_r$ –see Fig. (\[fig:Fluct\]) of this SI. That is, in a time unit, the cells move randomly over the manifold $k_r$ steps. We are in a $2$D isotropic, locally flat surface with generic orthogonal coordinates $y,z$ –for example, $R\times$ the azimuthal angle $\theta$ and $R\times$ the polar angle $\varphi$ over a sphere surface. The amplitude of the fluctuations after time $t$ is known to be $\sim \sqrt{k_rt}$, a distance defined over the surface. In the case we consider the projection over the coordinate $z$, thanks to the local flatness [@Kuhnel:2006], assuming that $\sqrt{k_r}\ll R$ and using only symmetry reasonings, one has that since the displacement is given by $(\Delta y,\Delta z)=(\sqrt{k_r^y},\sqrt{k_r^z})$: $$\sqrt{(\sqrt{k_r^y})^2+(\sqrt{k_r^z})^2}=\sqrt{k_r}\quad,$$ and the fluctuations are isotropic, then: $$k_r^z=k_r^y\quad,$$ and the only solution to the above problem is that: $$k^z_r=\frac{k_r}{2}\quad.
\label{eq:ukz}$$ In the case we are dealing with a spherical surface, we are projecting the fluctuations over the polar angle $z=R\varphi$ –see Fig. (\[fig:Fluct\]) of this SI. The stochastic differential equation that will describe the movement of a single cell in this manifold will be, for $z\leq \frac{\pi}{2}R$: $$dz=h(z)dt+\sqrt{\frac{k_r}{2}}dW\quad,$$ where $dW$ is the differential of the standard Brownian motion with average $0$ and variance $1$ in one dimension.
Evolution of densities: Hemispherical approach
----------------------------------------------
Let us now consider a detailed version of the geometry of the crypt. This consists in a half sphere, $H$, whose arc length from the bottom pole to the end is is $z _R=\frac{\pi}{2}R$ coupled to a cylinder $C$ of length $L$ and radius $R$. The cells populate both the surface of the hemisphere and the cylinder. The push-up force is directed towards the top of the cylinder –see Fig. (\[fig:CryptGeom\]a) of this SI. In the arc that goes from the bottom pole to the end of the hemisphere there are $z_R$ cells. Again, the units are given considering the average size of the cell as the length/surface/volume unit. Therefore, the cells will be labelled in terms of the geodesic distance over the hemisphere to the bottom pole. The density of the lineages will be given by: $$(\rho_0,\rho_1,\rho_2, . . .,\rho_{z_{R-1}},\rho_{z_R} )\quad.$$ Since the coordinate $R$ is constant, the only dynamically relevant information will come from the angle $\varphi$. Each position $k$ in the arc $(0,1,2, . . .,z_k,. . .,z_R-1,z_R)$ describing the initial point of a cell lineage can be rewritten as: $$z_k=R\varphi_k\;,\quad \varphi_k\in\left(0,\frac{\pi}{2}\right)\quad.$$ i.e., $\varphi_k=\frac{z_k}{R}$. The metric tensor for this hemispheric surface is [@Kuhnel:2006]: $${\mathbf g}=\left(
\begin{array}{cc}
R^2 & 0\\
0 &R^2\sin^2(\varphi)
\end{array}
\right)\quad.$$ Computing the determiner of ${\mathbf g}$, $g$: $$g=\left|
\begin{array}{cc}
R^2 & 0\\
0 &R^2\sin^2(\varphi)
\end{array}
\right|=R^4\sin^2\varphi\quad.$$ one can compute the surface element as [@Kuhnel:2006]: $$dS=\sqrt{g}d\theta d\varphi\quad.$$ In consequence, the area under the position of the cell $k$ in the in the hemisphere $H$, located at the arc position $z_k$, will be: $$S^H_k=\int_0^{2\pi}\int_0^{\frac{z_k}{R}}\sqrt{g}d\theta d\varphi=2\pi R^2\left(1-\cos\left( \frac{z_k}{R}\right)\right)\quad.$$ By direct application of equation (\[eq:h\]), we have that the push-up force inside the hemisphere $H$ is given by: $$h^H(z)=k_d R\left[\frac{1-\cos\left(\frac{z_k}{R}\right)}{\sin \left(\frac{z_k}{R}\right)}\right]\quad.
\label{eq:h^H}$$ Finally, from equation (\[eq:ukz\]) we know that –see also Fig. (\[fig:Fluct\]) of this SI: $$k_r^z=\frac{k_r}{2}\quad,$$ leading, according to equation (\[eq:FisherKPPGen\]) to the general dynamical equation for $0\leq z\leq \frac{\pi}{2}R$ in a hemispherical surface to be: $$\frac{\partial \rho_k}{\partial t}=-k_d R\frac{\partial }{\partial z}\left( \frac{1-\cos\left(\frac{z_k}{R}\right)}{\sin \left(\frac{z_k}{R}\right)}\rho_k\right)+\frac{k_r}{4}\frac{\partial^2 \rho_k}{\partial z^2}+\rho_k\quad.
\label{eq:FisherKPPSphere}$$ The above equation is difficult to deal with. However, we observe that in the region of interest, $z\in\left[0,\frac{\pi}{2}R\right]$, equation (\[eq:h\^H\]) can be approximated as: $$\tilde{h}(z)\sim \frac{2k_d}{\pi}z\quad,$$ leading to an error bounded as: $$\max_{z\in\left[0,\frac{\pi}{2} R\right]}\left |\left | h(z)-\frac{2k_d}{\pi}z\right |\right |<0.09 k_dR\quad,$$ according to numerical tests. With this approximation, we have that equation (\[eq:FisherKPPSphere\]) can be rewritten approximately as: $$\frac{\partial \rho_k}{\partial t}\approx-\frac{2k_d}{\pi} \frac{\partial }{\partial z}(z\rho_k)+\frac{k_r}{4}\frac{\partial^2 \rho_k}{\partial z^2}+\rho_k\quad,
\label{eq:FisherKPPapprox}$$ which is the general kind of Ornstein-Uhlenbeck equations we have been working so far.
Coupling to a cylinder
----------------------
In the case the cell is at the position $k$ in the cylindric region $C$, the area under it will be given by $S^H_{\frac{\pi}{2}R}$, the area of the whole hemisphere, and the remaining surface due to the cell is the cylinder. Knowing that for the cylindric coordinates $\sqrt{g}=R$, then: $$S^C_k=S^H_{\frac{\pi}{2}R}+\int_0^{2\pi} d\theta \int^{z_k}_{\frac{\pi}{2}R}\sqrt{g}dz=S^H_{\frac{\pi}{2}R}+2\pi R\left (z_k-\frac{\pi}{2}R\right)\quad.$$ Completing the picture, the push force felt by a cell in the cylindric region $C$ is given by: $$h^C(z)=k_d\left( z_k+\left(1-\frac{\pi}{2}\right)R\right)\quad.
\label{eq:h^C}$$ It is easy to check that: $$\begin{aligned}
\lim_{z\to \frac{\pi}{2}R^+}h^H(z)&=&\lim_{z\to \frac{\pi}{2}R^-}h^C(z)\nonumber\\
\lim_{z\to \frac{\pi}{2}R^+}\frac{d}{dz}h^H(z)&=&\lim_{z\to \frac{\pi}{2}R^-}\frac{d}{dz}h^C(z)\quad.\nonumber\end{aligned}$$ Therefore, one can define a function, $h(z)$ as: $$\begin{aligned}
h^{H,C}(z)=\left\{
\begin{array}{ll}
h^H(z)\;{\rm if}\;z\leq \frac{\pi}{2}R\\
h^C(z)\;{\rm if}\;z>\frac{\pi}{2}R\quad,
\end{array}
\right.
\label{eq:h(HC)}\end{aligned}$$ which is always well defined. In addition, the projection of the fluctuations will be the same in both regions, since the only relevant property is the local dimension, which in both cases is $2$, leading to a $k_r^z=k_r/2$. Consequently, equation (\[eq:FisherKPPGen\]) can be rewritten consistently for all the hemisphere/cylinder system as: $$\frac{\partial \rho_k}{\partial t}=-\frac{\partial }{\partial z}(h^{H,C}(z)\rho_k)+\frac{k_r}{4}\frac{\partial^2 \rho_k}{\partial z^2}+\rho_k\quad.
\label{eq:FisherKPPHC}$$
Lineage survival: Detailed geometry approach
--------------------------------------------
To guess the probability of lineage survival, we restrict ourselves to the hemispheric region $H$ of the crypt. As discussed in section \[sec:General\] of this Appendix, the existence of linear functions approximating the drift and fluctuation parameters of the general reaction-diffusion equation (\[eq:FisherKPPGen\]) leads to gaussian-like lineage survival probabilities. According to the approximations leading to equation (\[eq:FisherKPPapprox\]), we have that, considering the hemispheric region of the crypt: $$p(c_k)\propto {\rm exp}\left\{-\frac{2}{\pi}\frac{k_d}{k_r}k^2\right\}\quad.$$
Number of stem cells in the system
----------------------------------
To finish with, we compute the size of the emerging stem-cell region. The fluctuations will project over the surface from the bottom pole to the cell located at: $$z_s\approx 1+\sqrt{\frac{\pi}{2}\frac{k_r}{k_d}}\quad.$$ The position of this cell will define an angle of $$\varphi_s\approx\frac{1}{R}\left(1+\sqrt{\frac{\pi}{2}\frac{k_r}{k_d}}\right)\quad.$$ Therefore, the amount of cells below this angle, the stem cell number will be: $$N_s^{2D}\approx 2\pi R^2\left[1-\cos\left\{\frac{1}{R}\left(1+\sqrt{\frac{\pi}{2}\frac{k_r}{k_d}}\right)\right\}\right]\quad.
\label{eq:N_sCrypt}$$
Higher dimensionality with radial symmetry
------------------------------------------
Imagine that our tissue can be abstracted as something isomporphic to a disk or sphere. In consequence, we will only consider the $r$ coordinate –see Fig. (\[fig:CryptGeom\]b) of this SI. Let the units of volume be in number of cells. That is, we assume that, on average, the volume of cells is around $1$. The fluctuations can occur in all directions with equal probability. Therefore, if we know that the rate at which fluctuations occur is $k_r$, following the same reasoning we used in section \[sec:General\], we can approximate the effective rate projected towards the $r$ axis by: $$k_r^r\sim\frac{k_r}{{D}}\quad,$$ where $D$ is the dimension of the volume. The diffusion term in the reaction diffusion equation will be given by: $$h(r)\sim \frac{k_r}{{D}}\frac{\partial^2 \rho_k}{\partial r^2}\quad,$$ Now we compute the drift term. For that, we apply directly equation (\[eq:h\]). In the case of a $3$D sphere growing from the centre, we have, if $V(r)$ is the volume encapsulated by the surface of radius $r$, the displacement of a cell at this position will be determined by: $$\frac{d V(r)}{dt}=k_d V(r)\quad,$$ If $V(r)=\frac{4}{3}\pi r^3$, then, according to the above equation: $$\frac{d V(r)}{dt}\propto r^2\frac{dr}{dt}\quad.$$ Thus, the push up force will, accordingly, result in a radial displacement of: $$\frac{dr}{dt}\propto k_d r\;,\quad h(r)\propto k_d r\quad.$$ The above speed will define the drift term $h(r)$. Again, this will lead to a gaussian distribution of lineage cell survival and, therefore, to a well defined stem-cell region.
Noise in the stochastic conveyor-belt from “tectonic” epithelial movements
==========================================================================
In this section, we explore an alternative source for noise in determining the number of functional stem cells, i.e. the possibility of global rearrangements of the epithelium relative to the optimal position (bottom of the crypt/edge of the tip –see Fig. (\[fig:Tectonic\]a) of this Appendix for a sketch. This is motivated by experiments in intestinal morphogenesis or mammary morphogenesis, where global three-dimensional bending of the epithelial modifies the location of the niche (see main text).
Importantly, performing full stochastic simulations of this process in one-dimensions revealed a strikingly similar paradigm compared to the version of the model introduced in the main text, with survival probabilities decaying as normal distributions away from the central, optimal position for survival – see Fig. (\[fig:CryptGeom\]b) of this SI. Moreover, the variance of these probabilities, which define the number of functional stem cells, also scale as $k_r/k_d$ as expected in the model from the main text –Fig. (\[fig:CryptGeom\]c) of this SI.
This confirms that such tectonic movements can also be described in our coarse-grained model, simply renormalizing in long-term dynamics the intensity of the noise term $k_r$ in the system (although one would expect tectonic movements to significantly change the short-term dynamics). Interestingly, this allows for the system to be “noisy”, i.e. many functional stem cells to contribute to the long-term dynamics, without any clonal dispersion, showing that one must be careful in equating the two directly. This would in particular be relevant for the dynamics of intestinal crypts, where cells away from starting position 0 have been shown experimentally to still contribute long-term (see Fig. 2 of the main text), but where little clonal fragmentation was observed (raising the possibility that such tectonic collective movements could occur to reposition cells towards/away from the best location).
Numerical simulations and parameter estimation
==============================================
[**1D simulation.**]{} A one-dimensional array of 20 cells was initialised, where every cell was given an index $x=\{0,1,...,19\}$ to identify their starting position in the crypt, corresponding to their lineage. 0 is the most advantageous position at the bottom of the crypt, and 19 is at the top where it will be removed from the crypt by any single division event below. The simulation parameters were $k_d$, the probability that a cell divides, and $k_r$, the probability that a cell switches positions with its neighbour in either direction. During the simulation, cells were chosen at random with equal probability to decide whether to divide or change neighbors. The simulation was terminated when the array was colonized by a single cell lineage –we denote this as a “win” by that lineage.
The probability of survival of a given lineage, $p(c_k)$ was calculated as the number of wins divided by the number of simulation runs. The simulation was repeated 2000 times. For every $\frac{k_r}{k_d}$, $p(c_k)$ was plotted as a function of the starting position $k$, and the data was fitted with the following function: $$p(c_k) = \frac{1}{a \cdot \sigma \sqrt{2 \pi}} {\rm exp} \Big\{ -\frac{k^2}{2\sigma ^2} \Big\}$$ where $a$ is a normalization constant. Consistently with the theoretical predictions given in equation (\[eq:p(c\_k)=\]), we find that: $$\sigma\sim \sqrt{\frac{k_r}{k_d}}\quad.$$
$\\$ [**2D simulation.**]{} A two-dimensional array of size 8x20 cells was initialised, where every cell was given an index $x=\{0,1,...,19\}$ to identify their starting row in the crypt, corresponding to their lineage. 0 is the most advantageous row and 20 is the upper boundary after which cells are removed (Note that this boundary condition is largely irrelevant for the results because cells at these rows have vanishing chance to contribute to a winning lineage). Periodic boundary conditions were applied to every row. This simulation uses the same parameters $k_d$ and $k_r$, but now with a maximum of 4 possible neighbours for intercalation. The simulation was repeated 2000 times, and plotting the probability of survival of a given lineage shows that it fits to: $$p_{2D}(k) = \frac{1}{a \cdot \sigma \sqrt{2 \pi}} {\rm exp}\Big\{ -\frac{k^2}{2\sigma ^2} \Big\} %\cdot$$ where $a$ is a normalization constant. In agreement with the theoretical predictions given in equation (\[eq:p(c\_k)=\]), we find that: $$\sigma\sim \sqrt{\frac{2k_r}{k_d}}\quad.$$ Both fits for intestinal homeostatic crypts and developing kidney shown in the main text were performed using this 2D simulation. In both cases, in order to build confidence intervals, in a non-parametric way, for our predictions of clonal survival as a function of position, we simulated 2000 times the number $N$ of labelling events as in the data ($N=45$ for intestine, $N=24$ for kidney), and calculated the mean survival probability, as well as the $68\%$ confidence interval around this prediction (i.e. 1 standard deviation around the mean). In both cases, we found that all of the experimental datapoint was contained within this interval. Finally, although on the long-term, survival probabilities converge towards a steady-state universal gaussian distribution, the live-imaging datasets were experimentally acquired on finite timescales, requiring simulations to examine the dynamics of clonal conversion. We thus state below for each organ the duration of the simulations $T$ (rescaled by the cell division rate $k_d$).
For the intestinal crypts, clones were initialized only at positions 0,1,2,3 and 4, to match with the experimental set-up where clones were only traced from Lgr5+ stem cells. Moreover, we defined a clone as “lost” in the system as soon as it didn’t have any cells in this compartment (positions 0-4), again to match the way that the experimental data was recorded in Ref. [@Ritsma:2014]. One should note that this assumption is expected to be largely irrelevant for our findings, given the results of Section 5: $k_r/k_d$ in this system is small enough that $N_s < 5$, so that the probability for clones to come back in the Lgr5+ region after having left it during the early phase of clonal competition is vanishingly small. Plots in the main figure for intestinal crypts used the following parameters: $k_r/k_d=1$ and a runtime of $T k_d = 0.5$ (we note that the latter is relatively small, corresponding experimentally to a bit less than one full cell division in 3 days in Ref. [@Ritsma:2014], which could be linked to the method of intravital imaging (typical timescales reported in intestine are 1-2 days).
For the kidney tips, clones were initialized evenly in positions 0-10, which was also the definition for the compartment of clonal survival. Plots in the main figure used the following parameters: $k_r/k_d=16$ (see main text for details on this parameter estimation) and a runtime of $T k_d = 2$ (which is the typical average number of divisions seen in the experimental dataset during the time course, see Ref. [@Riccio:2016]). The experimental data of Ref. [@Riccio:2016] assigns to cells a starting position on a 10x10 grid, and notes whether a clone still remains in the tip at the end of the observation period. As position $(10,10)$ was the edge of the tip, we calculated the euclidian distance of all coordinates $(i,j)$ from position $(10,10)$, which is the starting distance reported in Fig. 3 of the main text. One should note that because of the longer runtime of experiments, predictions in kidney are much-closer to their steady-state universal form.
Finally, we also used these 2D simulations to fit the clonal fragmentation seen in mammary gland (where no long-term live-imaging was possible to follow clonal survival). To match experiments where cells were labelled in mouse of 3 weeks and collected at 5 weeks, and where the typical cell division rate is $16$ hours, we used $T k_d=20$, although the predictions are largely insensitive to this timing. We then measured for each labelled cell the distance to the closest labelled cell, and built probability distributions for these nearest cell-cell distance. As expected, for low values of $k_r/k_d$, cells are always close neighbours, whereas the average cell-cell distance increased monotonously with $k_r/k_d$. In figure (\[fig:Fragmentation\]) of this Appendix we describe the process of inference of the $k_r/k_d$ parameter. In Fig. 4 of the main text we report the explicit probability distributions for the nearest cell-cell distance inferred from clonal dispersion observed during the development of the mammary gland.
Experimental procedures
=======================
Mice
----
All mice were females from a mixed background, housed under standard laboratory conditions, and received food and water ad libitum. All experiments were performed in accordance with the Animal Welfare Committee of the Royal Netherlands Academy of Arts and Sciences, The Netherlands.
Intravital microscopy
---------------------
R26-CreERT2;R26-mTmG female mice were IP injected with 0.2mg/25gTamoxifen diluted in sunflower oil (Sigma) at 3 weeks of age to induce sporadic recombination in the developing mammary gland. At 5 weeks of age, a mammary imaging window was implanted near the 4th and 5th mammary gland –for details, see [@Alieva:2014]. Mice were anesthetized using isoflurane (1.5% isoflurane/medical air mixture) and placed in a facemask with a custom designed imaging box. Imaging was performed on an inverted Leica SP8 multiphoton microscope with a chameleon Vision-S (Coherent Inc., Santa Clare, CA, www.coherent.com), equipped with four HyD detectors: HyD1 (<455nm), HyD2 (455–490nm), HyD3 (500–550nm) and HyD4 (560–650nm). Collagen I (second harmonic generation) was excited with a wavelength of 860nm and detected with HyD1, GFP and Tomato were excited with a wavelength of 960nm and detected with HyD3 and HyD4. Mammary gland tips were imaged at an interval of 20-30 minutes using a Z-step size of 3$\mu$m over a minimum period of 8 hours. All images were acquired with a 25$\times$ (HCX IRAPO N.A. 0.95 WD 2.5mm) water objective.
Quantitative image analysis
---------------------------
Clonal dispersion in the developing mammary tips was measured in lineage traced whole mount glands from R26-CreERt2;R26-Confetti mice as previously described [@Scheele:2017]. In brief, R26-CreERt2;R26-Confetti female mice were injected at 3 weeks of age with 0.2mg/25g Tamoxifen to achieve clonal density labelling (<1 cell per tip). Lineage traced mice were sacrificed at mid-puberty (5 weeks of age) or at the end of puberty (8 weeks). Mammary glands were dissected, fixed in periodate-lysing-paraformaldehyde (PLP) buffer (1% paraformaldehyde (PFA, Electron Microscopy Science), 0.01M sodium periodate, 0.075M L-lysine and 0.0375M P-buffer (0.081M Na2HPO4 and 0.019M NaH2PO4) (pH 7.4)) for 2 hours at room temperature (RT), and incubated for 2 hours in blocking buffer containing 1% bovine serum albumin (Roche Diagnostics), 5% normal goat serum (Monosan) and 0.8% Triton X-100 (Sigma-Aldrich) in PBS. Subsequently, glands were incubated with primary antibodies anti-K14 (rabbit, Covance, PRB155P, 1:700) or anti-E-cadherin (rat, eBioscience, 14-3249-82, 1:700), and secondary antibodies goat anti-rabbit or goat anti-rat, both conjugated to Alexa-647 (Life Technologies, A21245 and A21247 respectively, 1:400). Mammary glands were mounted on a microscopy slide with Vectashield hard set (H-1400, Vector Laboratories), and imaging was performed using a Leica TCS SP5 confocal microscope, equipped with a 405nm laser, an argon laser, a DPSS 561nm laser and a HeNe 633nm laser. All images were acquired with a 20x (HCX IRAPO N.A. 0.70 WD 0.5mm) dry objective using a Z-step size of 5$\mu$m (total Z-stack around 200$\mu$m). 3D tile scan images of whole-mount mammary glands were used to manually reconstruct the tips. Labelled confetti cells were annotated in the schematic outline of the tips including information on the confetti colour for the mammary glands (GFP=green, YFP=yellow, RPF=red and CFP=cyan). The length and the width of the tips were measured, and the coordinates of each labelled confetti cell in the tip were determined. The coordinates were used to calculate the position of each labelled cell within the ductal tip, as well as the minimal distance to the nearest neighboring cells with the same confetti color (as a measure of the clonal dispersion within the stem cell zone).
[^1]: To whom correspondence should be addressed. E-mail: [email protected], [email protected], [email protected]\
$^*$Author contributions: B.C.-M., B.D.S, J.v.R and E. H. designed research; B.C.-M., C.L.G.J.S, K.K, S. I.J. E., and EH performed research and analyzed data; B.C-M and E.H. wrote the paper with input from C.L.G.J.S, B.D.S and J.v.R.\
$^*$B.C.-M. and C.L.G.L.S. contributed equally to this work.
[^2]: A rigorous approach to this problem would require a reflecting boundary condition at $x=0$. Imposing such boundary condition would make the whole problem much more difficult and, eventually intractable. The reason by which we adopted natural boundary conditions is due to the fact that the dynamics in this system is extremely imbalanced and runs essentially in only one direction. If one takes equation (\[eq:solutionP\]) at $x=0$ we observe that the probability of being at $x=0$ decays as $\sim e^{-\tau}$, for any starting point $k>0$, as it is the case in our system. This tells us that the probability of visiting regions $x<0$ is, to our purposes, negligible.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Zhong Wang and Shaolong Wan[^1]\
Institute for Theoretical Physics and Department of Modern Physics\
University of Science and Technology of China, Hefei, 230026, [**P. R. China**]{}
title: An exact spin liquid state on the kagome lattice with topological order
---
0.8in
[**Abstract**]{}
[We find an exact spin liquid state without time reversal symmetry on the kagome lattice with odd number of electrons per unit cell and explicit wave functions for all eigenstates. We also obtain that all spin-spin correlations are zero except trivial cases. We then show that there are anyonic excitations in our model. Finally, we demonstrate the existence of ground state degeneracy and gapless edge states indicating nontrivial topological order in our model. We also label all eiginstates on torus by local and global string operators.]{}
**PACS** number(s): 75.10.Jm, 75.50.Mm
*Introduction.* –There has been enormous interest recently in topological quantum phases [@wen1; @nayak] beyond the Landau paradigm [@sachdev], which is based on the idea of symmetry breaking and local order parameters. The newly discovered phases have many exotic phenomena such as electron fractionalization [@laughlin] and emergent gauge fields [@anderson]. The traditional order parameter descriptions are either inapplicable or insufficient for these new phases.
Spin liquid states in dimension two ($2$D) are examples of such exotic phases. In the spin liquid states, strong quantum fluctuation forbids existence of nonzero spin order parameter. Although many numerical and analytical evidences indicate the existence of such exotic state in $2$D, exact solution is absent until very recently [@kitaev1; @kitaev2; @wen2; @yao; @schroeter; @sun; @xiang]. These exact solutions provide us with a lot of intuition of the internal structure of spin liquid states.
Spin liquid states that spontaneously break time reversal symmetry (TRS) and spacial inversion symmetry are called chiral spin states [@kalmeyer]. An exact chiral spin state is proposed in Ref.[@yao] based on short-range anisotropic interaction [@kitaev2], but this model have even number of electrons per unit cell, and thus not deserve to be called “Mott insulators” or “spin liquid” states in the traditional sense. We consider it an interesting question to find whether there exist other exact spin liquid states with arbitrarily given symmetry properties and odd number of electrons per unit cell.
In this letter we generalize the toric code model [@wen2; @kitaev1] to the kagome lattice (Fig.1) and propose an exact spin liquid state that breaks TRS. This state is not chiral spin state in conventional sense because TRS is already broken in the Hamiltonian. The kagome lattice has odd number of spins per unit cell, and thus the proposed state is a true spin liquid state. Our model treat three spin components $\sigma^{x,y,z}$ democratically, which is more symmetrical than the $\sigma^{x,y}$ model [@wen2; @kitaev1]. We propose explicit wave functions for all the energy eigenstates and interpret these eigenstates as string-net [@wen1] coherent states. We also compute spin-spin correlation functions and find that all of them are zero except the trivial cases of on-site correlations. After a short discussion of anyonic excitation in our model, we proceed to discuss the ground state degeneracy and gapless edge states indicating topological order. We also find all the eigenstates of the model on torus using string operators.
![(a) The kagome lattice is composed of two types of elementary plaquettes: triangles and hexagons. The dashed lines show the unit cell. (b) Elementary plaquettes of Kagome lattice. The letters x,y,z indicate the factors of $F_{p}$.[]{data-label="fig.1"}](fig1.eps){width="8.5cm" height="6.0cm"}
*The model*. –Our spin-$\frac{1}{2}$ Hamiltonian on the kagome lattice is given as: $$\begin{aligned}
H = -\sum_{p}V_{p}F_{p}\end{aligned}$$ with summation over all elementary triangular and hexagonal plaquettes. We choose $F_{p}=\sigma_{1}^{x} \sigma_{2}^{y}
\sigma_{3}^{z}$ on the triangles and $ F_{p}=\sigma_{1}^{x}
\sigma_{2}^{z} \sigma_{3}^{y}\sigma_{4}^{x} \sigma_{5}^{z}
\sigma_{6}^{y}$ on the hexagons (see Fig.1), where $\sigma_{i}^{a} $ are the well-known Pauli matrices. It follows from the property of Pauli matrices that $F_{p}^2=1$. Another crucial property of $F_{p}$ is that $[F_{p},F_{q}]=0$ for any pairs $p$ and $q$. So all $F_{p}$ can have eigenvalues($\pm1$) simultaneously. Since $\sigma_{i}^{a}$ change sign under time reversal transformation, the Hamiltonian breaks TRS explicitly. The solubility of our model does not depend on signs of $V_{p}$, so we will choose all $V_{p}=V>0$ for simplicity.
Firstly we consider a planar region with $N$ sites, where $N$ is very large. The number of elementary triangle and hexagon are shown to be $2N/3$ and $N/3$, respectively. So there are $N$ plaquettes, and $F_{p}$ can label $2^{N}$ states. The number of physical spin states is also $2^{N}$. From this counting argument we know that eigenvalues of $F_{p}$ label all the states on an infinite plane and the model is exactly solvable.
We can write down the following wave unctions for all eigenstates:
$$\begin{aligned}
|\psi_{\lambda}\rangle =
\prod_{p}(1+\lambda_{p}F_{p})|\psi_{0}\rangle\end{aligned}$$
where $p$ exhaust all elementary triangular and hexagonal plaquettes. $|\psi_{0}\rangle$ is an arbitrary spin state which we choose as the state satisfying $\sigma_{i}^{z}=+1$ for all sites. The irrelevant normalization factors have been omitted. Different choices of $\lambda_{p}=\pm1$ label different states. The $F_{p}$ expectation in $|\psi_{\lambda}\rangle$ is:
$$\begin{aligned}
\frac{\langle\psi_{\lambda}|F_{p}|\psi_{\lambda}\rangle}{\langle\psi_{\lambda}|\psi_{\lambda}\rangle}
= \lambda_{p}\end{aligned}$$
so $|\psi_{\lambda}\rangle$ is an eigenstate of $H$ with eigenvalue $E = -\sum_{p}\lambda_{p}V_{p}$. The ground state is the state with all $\lambda_{p}=1$. The above simple wave functions have transparent explanation in the physical language of string-net condensation [@wen1]. We will demonstrate this in our model. A caution in order is that this wave function can completely determine the state only on an infinite plane. On systems with boundary like disk or closed like torus, there are other degrees of freedom missing. We will return to this question later.
We define two types of strings $T_{1}$ and $T_{2}$ [@kitaev1; @wen1]. $T_{1}$ strings walk on triangles while $T_{2}$ strings walk on hexagon (see Fig.2). We define the string operators as $P_{t} = \prod_{i \in t}\sigma_{i}^{a} $ for each string $t$. The choice of $\sigma_{i}^{a}$ depends on both the site $i$ and the string type. For $T_{1}$ strings, $\sigma^{a} =
\sigma^{y},\sigma^{z},\sigma^{x}$ at the three vertexes $1$,$2$,$3$ of each triangle, respectively. For $T_{2}$ strings, $\sigma^{a} =
\sigma^{z},\sigma^{y},\sigma^{x},\sigma^{z},\sigma^{y},\sigma^{x}$ at the six vertexes $1,2,3,4,5,6$ of the hexagon, respectively. An important property of this definition is that $T_{1}$ and $T_{2}$ string operators involve different Pauli matrices at any sites $i$, from which we can readily obtain the simple commutative relations among $P_{t}$. For arbitrary $T_{1}$ strings $s,t$ and $T_{2}$ strings $u,v$, we have $[P_{s},P_{t}]=0$, $[P_{u}, P_{v}]=0$; $[P_{s},P_{u}]=0$ if $s$ and $u$ intersect each other by even times, while $\{P_{s},P_{u}\}=0$ if $s$ and $u$ intersect each other by odd times.
![Strings on the kagome lattice. $a,b$ are open strings, while $c,d$ are closed strings. $a,c$ are $T_{1}$ strings, while $b,d$ are $T_{2}$ strings. The letters $x,y,z$ denote the specific Pauli matrices appearing in the string operators.[]{data-label="fig.2"}](fig2.eps){width="8.5cm" height="6.0cm"}
These strings are divided into two classes by their topology. A string is called open string if it has endpoints, otherwise is called closed string [@wen1]. From the property of Pauli matrices we can find $[P_{t},F_{p}]=0$ for any $p$ for a closed string $t$, and it follows from this that $[P_{t},H]=0$. For open string $s$ with two endpoints on plaquettes $p,q$, we have $\{P_{s},F_{a}\}=0$ if $a=p$ or $q$, and $[P_{s},F_{a}]=0$ otherwise. So $P_{s}$ changes the sign of $F_{p}$ at its two endpoints. If we regard negative $F_{p}$ as vortex at $p$, then $P_{s}$ applied to the ground state will creates two vortexes at its endpoints and increase the energy by $2V$.
It is interesting that $F_{p}$ are also closed strings operators defined above. Actually, for a $ T_{1}$ string $t_{1}$ surrounding a single hexagon $p$ (such as string $c$ in fig. 2), we have $P_{t_{1}}=F_{p}$. Similar relations also hold for any $T_{2}$ string surrounding a single triangle. So we conclude that $F_{p}$ are special cases of closed string operators. This observation gives physical meaning to the above wave function. We expand the wave function:
$$\begin{aligned}
|\psi_{\lambda}\rangle =(1 +
\sum_{p}\lambda_{p}F_{p}+\sum_{p_{1},p_{2}}\lambda_{p_{1}}\lambda_{p_{2}}F_{p_{1}}F_{p_{2}}
+ \cdots)|\psi_{0}\rangle\end{aligned}$$
each term is a closed string state. So $|\psi_{\lambda}\rangle$ are coherent states of closed strings.
We mention that there are also two types of strings in the models of Kitaev and Wen [@kitaev1; @wen1; @wen2], but strings in our model have the additional feature that $T_{1}$ and $T_{2}$ strings walk on triangles and hexagons, respectively. So their properties are different. One of the differences is that there are twice as many triangles as hexagons and thus more room for open $T_{1}$ string endpoints than for $T_{2}$ string endpoints.
*Exact spin correlation.* –The spin-spin correlation functions can be extracted from the exact wave functions. Since wave functions of all eigenstats are known, we can obtained all the spin-spin correlations by direct computation:
$$\begin{aligned}
\sigma_{i}^{a}\sigma_{j}^{b}|\psi_{\lambda}\rangle = \prod_{s
}(1-\lambda_{s}F_{s}) \prod_{t}(1+\lambda_{t}F_{t})
\sigma_{i}^{a}\sigma_{j}^{b}|\psi_{0}\rangle\end{aligned}$$
Where $s$ exhausts elementary plaquettes with the property $\{P_{s},\sigma_{i}^{a}\sigma_{j}^{b}\}=0$, and $t$ exhausts elementary plaquettes with the property $[P_{s},\sigma_{i}^{a}\sigma_{j}^{b}]=0$. We can check that such $s$ exist except in the trivial cases when $i=j$ and $a=b$. Since $
(1+\lambda_{s}F_{s})(1-\lambda_{s}F_{s})=0 $, we obtain the exact result: $$\begin{aligned}
\langle\psi_{\lambda}|\sigma_{i}^{a}\sigma_{j}^{b}|\psi_{\lambda}\rangle
= 0\end{aligned}$$ except the trivial cases $
\langle\psi_{\lambda}|\sigma_{i}^{a}\sigma_{i}^{a}|\psi_{\lambda}\rangle
= \langle\psi_{\lambda}|1|\psi_{\lambda}\rangle = 1$. This computation is analogous to the computation of spin-spin correlations in the Kitaev model [@baskaran], where they are found to be exactly zero beyond the nearest neighbor. The spin-spin correlations in our model are even more exotic. They are identically zero for two different sites including the case of nearest neighbors. We also mention that the spin correlations in the earlier models [@wen1; @kitaev1] can also be computed using explicit wavefucntions. Our method can also be generalized to computation of multi-spin correlations. Another feature is that the result is valid for all energy eigenstates, besides the ground state. This remarkable feature is also shared by the Kitaev’s model on honeycomb lattice [@kitaev2; @baskaran].
The zero spin correlations indicate very strong quantum fluctuation in our model. So the conventional idea of using spins as local order parameter will not give meaningful result here. But there is topological order [@wen1] in this model manifesting in ground state degeneracy on closed manifold and gapless edge states at the boundary of open systems.
*Anyonic excitation.* –The anyonic excitation in our model has similar properties to earlier models [@kitaev1; @kitaev2; @wen2]. The energy of an open string is not changed if we move the two endpoints far apart, i.e. the endpoints are deconfined. We call these endpoints $T_{1}$ and $T_{2}$ vortexes. We also mention that these vortexes can be interpreted as $Z_{2}$ vortex excitations of an emergent $Z_{2}$ gauge field [@wen1; @senthil].
The vortex hopping operators is very simple, they are the same matrices appearing in the definition of $P_{t}$. The statistical angles of $T_{1}$ and $T_{2}$ vortexes can be shown to be both zero by the statistical algebra method [@levin]. So the same type of vortexes see each other as bosons. The mutual statistical angle of $T_{1}$ and $T_{2}$ vortexes can be computed in a different way. Following Kitaev’s work [@kitaev1], we consider the physical process of moving $T_{2}$ vortex $a$ around $T_{1}$ vortex $b$ along the path of string $u$ (shown in Fig.3). The initial state is $|\psi_{i}\rangle =P_{s}P_{t}|\psi\rangle$, where $|\psi\rangle$ is a state with all vortexes far away from $a,b$. The final state is:
$$\begin{aligned}
|\psi_{f}\rangle = P_{u}|\psi_{i}\rangle =
P_{u}P_{s}P_{t}|\psi\rangle =
-P_{s}P_{t}P_{u}|\psi\rangle =
-|\psi_{i}\rangle\end{aligned}$$
the result does not depends on detail shape of $u$. The only important thing is that $b$ is enclosed by $u$. So the phase factor $-1$ is purely topological. The mutual statistical angle $\theta$ is defined by $e^{2i\theta}=-1$. So we have $\theta=\pi /2$ which is half of fermion’s. This anyonic mutual statistic is a signal of topological order in our model.
![Statistic of open string endpoints. $s$ is an open $T_{2}$ string with one endpoint $a$ and the other endpoint far away, while $t$ is an open $T_{1}$ string with one endpoint $b$ and the other endpoint also far away. $u$ is a closed string enclosing $b$.[]{data-label="fig.3"}](fig3.eps){width="8.5cm" height="6.0cm"}
*Ground state degeneracy and gapless edge states.* –The ground state degeneracy on a closed manifold often indicate topological order hidden in a phase [@wen3; @wen1; @niu]. To study the topological order of our model, we consider a periodic system, i.e. we put the system on a torus. For a $m\times n$ system (as Fig.4), there are $2mn$ triangles, $mn$ hexagons, and $3mn$ sites (the sites at opposite boundaries are actually the same site and should not be re-counted). Now the mutual independency of $F_{p}$ is destroyed by the periodic boundary condition. Actually, one can readily check that $ \prod_{p \in triangle}F_{p}=1$ and $ \prod_{p \in
hexagon}F_{p}=1$, so the number of independent $F_{p}$ are $3mn-2$.
It has been mentioned that closed string operators $P_{t}$ commute with $H$ and thus do not change the energy eigenvalue. But does $P_{t}$ have chance to transform a given energy eigenstate to another eigenstate that degenerate with it? For the cases when $t$ is homotopically trivial, i.e. $t$ is able to be deformed to a point, the answer is no. Actually, when $t$ is a homotopically trivial $T_{1}$ string, it can be checked that $P_{t}=\prod_{p}F_{p}$, with $p$ exhausting the hexagons enclosed by $t$. So $P_{t}$ cannot transform a given energy eigenstate to another. Similar result holds for $T_{2}$ strings.
But there are also homotopically nontrivial strings called global strings on the torus. Consider four global strings $t_{1}$, $t_{2}$, $t_{3}$ and $t_{4}$ shown in Fig.4. We have $\{P_{t_{1}},
P_{t_{3}}\}=0$ and $\{P_{t_{2}}, P_{t_{4}}\}=0$ by their intersecting times. All other pairs of string operators are commutative. Since all these operators commute with $F_{p}$, we can choose the common eigenstates of $P_{t_{1}}$, $P_{t_{2}}$ and $F_{p}$ as the basis of the Hilbert space. To find whether eigenvalues of $P_{t_{1}}$, $P_{t_{2}}$ and $F_{p}$ can label all states completely, we just need to do a simple counting: All of $P_{t_{1}}$, $P_{t_{2}}$ and $F_{p}$ have two eigenvalues $\pm1$, so they can label $2^{3mn-2}\times2\times2=2^{3mn }$ states. This is just the dimension of physical Hilbert space. So $P_{t_{1}}$, $P_{t_{2}}$ and $F_{p}$ are complete labels and we have found all energy eigenstates on the torus. From the above discussion we conclude that states with given $F_{p}$ has additional 4-fold degeneracy coming from two global string operators $P_{t_{1}}$ and $P_{t_{2}}$. It follows from this that states with all $F_{p}=1$ are 4-fold degenerate, i.e. the ground state degeneracy is $4$.
![Kagome lattice on torus. $m=n=3$ in this figure. The effective region is within the dark bold lines. The four copies of the point “$A$” should be regarded as one and the same point because of periodic boundary condition. $t_{1}$, $t_{2}$, $t_{3}$ and $t_{4}$ are all closed strings on the torus. $t_{1}$, $t_{2}$ are $T_{1}$ strings, while $t_{3}$, $t_{4}$ are $T_{2}$ strings. None of them can be deformed to a point and thus they are called global strings.[]{data-label="fig.4"}](fig4.eps){width="8.5cm" height="6.0cm"}
Furthermore, we can figure out how these eigenstates transform under $P_{t_{3}}$ and $P_{t_{4}}$. Since $\{P_{t_{3}}, P_{t_{1}}\}=0$ and $[P_{t_{3}}, P_{t_{2}}]=0$, we conclude that $P_{t_{3}}$ change the sign of $P_{t_{1}}$ but not of $P_{t_{2}}$. By similar argument, we can also show that $P_{t_{4}}$ change the sign of $P_{t_{2}}$ but not of $P_{t_{1}}$. This argument also shows that both $\pm1$ eigenvalues of $P_{t_{1}}$ and $P_{t_{2}}$ are accessible by some physical states.
We mention that $t_{1}$ and $t_{2}$ can be chosen rather arbitrarily, with only requirements that they circle the torus in $m$ and $n$ directions respectively, and they are of the same string type. The physics here will not change with our choices because it is topological.
We also find the existence of gapless edge states in our model, which is another manifestation of topological order. If we cut the torus open along the bold lines in Fig.4, the dimension of Hilbert space is larger by a factor $2^{L}$ than the case of torus, where $L\approx m+n$ when $m,n$ are large. But the number of $F_{p}$ labeling energis is unchanged when we cut the torus open. So there is a large degeneracy proportional to $2^{L}$, which we can interpret as various edge excitations with zero energy [@wen1; @wen2].
*Conclusion.* –An exact spin liquid state has been given on the kagome lattice based on multi-spin interactions. All spin-spin correlations are found to be zero except trivial cases. There are also anyonic excitations and ground state degeneracy indicating nontrivial topological order. We think that this exact spin liquid state with odd number of spins per unit cell might be helpful to improve our understanding of $2$D spin liquid states. Furthermore, the existence of such exact states suggests that there maybe exist other interesting states based on multi-spin interactions.
The work was supported by National Natural Science Foundation of China under Grant No. 10675108.
[99]{}
X.G. Wen, *Quantum Field Theory of Many-Body Systems* (Oxford University, New York, 2004). C. Nayak, S. Simon, A. Stern, M. Freedman and S. Sarma, Rev. Mod. Phys. 80, 1083 (2008). S. Sachdev, *Quantum Phase Transitions* (Cambridge University Press, Cambridge, England, 1999). R. B. Laughlin, Phys. Rev. Lett. 50, 1395 (1983). G. Baskaran and P. W. Anderson, Phys. Rev. B 37, 580 (1988). A. Kitaev, Ann. Phys. 303, 2 (2003). A. Kitaev, Ann. Phys. 321, 2 (2006). X.G. Wen, Phys. Rev. Lett. 90, 016803 (2003). H. Yao and S. Kivelson, Phys. Rev. Lett. 99, 247203 (2007). D.F. Schroeter, E. Kapit, R. Thomale, and M. Greiter, Phys. Rev. Lett. 99, 097202 (2007). S. Yang, D.L. Zhou and C.P. Sun, Phys. Rev. B 76, 180404(R) (2007). X.Y. Feng, G.M. Zhang and T. Xiang, Phys. Rev. Lett. 98, 087204 (2007). V. Kalmeyer and R. B. Laughlin, Phys. Rev. Lett. 59, 2095 (1987). G. Baskaran, S. Mandal and R. Shankar, Phys. Rev. Lett. 98, 247201 (2007). T. Senthil and M. Fisher, Phys. Rev. B 62, 7850 (2000). M. Levin and X.G.Wen, Phys. Rev. B 67, 245316 (2003). X.G. Wen, Phys. Rev. B 40, 7387 (1989). X.G. Wen and Q. Niu, Phys. Rev. B 41, 9377 (1990).
[^1]: Corresponding author. Electronic address: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We propose and discuss a new computational method for the numerical approximation of reachable sets for nonlinear control systems. It is based on the support vector machine algorithm and represents the set approximation as a sublevel set of a function chosen in a reproducing kernel Hilbert space. In some sense, the method can be considered as an extension to the optimal control algorithm approach recently developed by Baier, Gerdts and Xausa. The convergence of the method is illustrated numerically for several examples.'
author:
- 'Martin Rasmussen[^1]'
- 'Janosch Rieger[^2]'
- 'Kevin N. Webster[^3]'
title: |
Approximation of reachable sets\
using optimal control and support vector machines
---
Introduction {#sec:introduction}
============
The numerical computation of reachable sets is a crucial topic in nonlinear control theory and the quantification of deterministic uncertainty in dynamical systems. Collision avoidance of manned and unmanned vehicles is one particular application that currently attracts a lot of attention (see e.g. [@Gerdts_12] and the references therein). Standard techniques such as the set-valued Euler method [@Dontchev_89_1; @Beyn_07_1] evolve a grid-based approximation of the reachable set along the relevant time interval. They are typically very slow, because there is a high degree of redundancy in the computations they carry out.
Recently, a version of the set-valued Euler method was presented in [@Rieger_Unpub_2] that tracks the boundaries of the reachable sets and uses only the boundaries of the right-hand side of the differential inclusion. With this approach, the complexity of the Euler scheme is reduced drastically in the low-dimensional setting, but only marginally in higher dimensions.
The DFOG optimal control algorithm [@Baier_13_1], which will be discussed in more detail in Section \[section:dfog\], is another recent attempt to reduce the proportion of irrelevant computations. Every point of a grid in the relevant region of the phase space is projected to the reachable set by solving a Mayer problem. From this data, one can derive – at least theoretically – an accurate description of the reachable set. In contrast to traditional methods, there is no guarantee that the numerical optimisation routine finds a global minimum, and therefore, the algorithm is, strictly speaking, unstable. Numerical studies, however, support the usefulness of this method.
In this paper, we propose a new approach to the calculation and representation of a reachable set approximation, motivated as an extension to the DFOG algorithm. The extension consists of using the results of these optimal control problems to search for a function in a particular function space, so that the reachable set is represented as a sublevel set of this function. The function space under consideration is a reproducing kernel Hilbert space (RKHS), and the algorithm to search for this function is a modified support vector machine (SVM) algorithm.
Our algorithm has the advantage that it is robust to a small number of errors made by the optimisation routines from the DFOG method. In addition, the function used for the reachable set approximation has a sparse representation in terms of the optimal control results, and the algorithm focuses on information provided by points that are close to the boundary of the reachable set.
Reachable sets and known techniques for their approximation
===========================================================
In the following, we give a condensed overview over basic properties of reachable sets (see Section \[sec:reachable\]), the currently most common numerical methods for approximating them (see Section \[sec:Runge:Kutta\]) and the DFOG method (see Section \[section:dfog\]), which is the basis of our new method.
We recall some standard definitions with regard to set representations.
Let $A,B\subset\mathbb{R}^d$ be compact sets, and $x\in\mathbb{R}^d$. The distance of a point $x$ to the set $A$ is defined by $$\mbox{\normalfont dist}(x,A):= \mbox{\normalfont inf}_{a\in A} \|x-a\|\,.$$ For any $r>0$, the $r$-neighbourhood of $A$ is the set $$B(A,r):=\{z\in\R^d: \mbox{\normalfont dist}(z,A)\le r\}.$$ The projection of $x$ to $A$ is the set of points in $A$ that realise the infimum distance to $x$, i.e. $$\operatorname{Proj}(x,A) :=\{a\in A:\|x-a\| = \mbox{\normalfont dist}(x,A)\}\,.$$ The Hausdorff semi-distance between sets $A$ and $B$ is given by $$\mbox{\normalfont d}(A,B):= \mbox{\normalfont sup}_{a\in A}\mbox{\normalfont dist}(a,B),$$ and the Hausdorff distance between $A$ and $B$ is given by $$\mbox{\normalfont d}_H(A,B):= \mbox{\normalfont max}\{\mbox{\normalfont d}(A,B), \mbox{\normalfont d}(B,A)\}\,.$$
Throughout this paper, the symbol $\|\cdot\|$ denotes the Euclidean norm. The symbols $\|\cdot\|_\infty$, $\mbox{\normalfont dist}_\infty(x,A)$ etc. denote the corresponding concepts based on the maximum norm.
Reachable sets {#sec:reachable}
--------------
Let $U$ be a nonempty convex and compact subset of $\mathbb{R}^d$ and $$\mathcal{U}:=\big\{u\in L^\infty([t_0,T],\mathbb{R}^d):u(t)\in U \text{ for almost all } t\in [t_0,T]\big\}$$ for fixed times $t_0 < T$. We consider the nonlinear control problem
\[gesamt\] $$\begin{aligned}
\dot{x}(t) &= g(t,x(t),u(t))\,,\quad u \in \mathcal{U}, \label{eqn:controlproblem1}\\
x(t_0) &= x_0, \label{eqn:controlproblem2}\end{aligned}$$
for some $x_0\in\mathbb{R}^d$, where holds for almost every $t\in[t_0,T]$ and $x(\cdot)\in W^{1,\infty}([t_0,T],\mathbb{R}^d)$ is absolutely continuous. We are interested in the *reachable set* at time $T$, given by $$\mathcal{R}(T,t_0,x_0) := \{x(T) : x(\cdot) \text{ solves \eqref{gesamt}}\} \,.$$ Problem is equivalent to the differential inclusion
\[eqn:controlprobleminclusion\] $$\begin{aligned}
\dot{x}(t) & \in G(t,x(t))\label{eqn:controlprobleminclusion1}\\
x(t_0) & = x_0, \label{eqn:controlprobleminclusion2}\end{aligned}$$
with valid for almost all $t\in [t_0,T]$, and $G(t,x) := \bigcup_{u\in U} \{g(t, x, u)\}$.
Reachable sets of nonlinear control systems, or, equivalently, nonlinear differential inclusions, are, in general, nonconvex. It is, however, well-known, that they enjoy several favourable properties under mild assumptions imposed on the right-hand side (see e.g. [@Deimling_92_1 Corollary 7.1]):
Let $G:[t_0,T]\times\R^d\rightrightarrows\R^d$ have closed and convex images, and assume that
- the mapping $t\mapsto G(t,x)$ is measurable for all $x\in\R^d$,
- the mapping $x\mapsto G(t,x)$ is upper semicontinuous for all $t\in[t_0,T]$,
- there exists $c\in L^1([t_0,T])$ such that $\|G(t,x)\|\le c(t)(1+\|x\|)$ for all $t\in[t_0,T]$ and $x\in\R^d$.
Then the mapping $x_0\mapsto\mathcal{R}(T,t_0,x_0)$ is upper semicontinuous, and the reachable set $\mathcal{R}(T,t_0,x_0)$ is nonempty and compact for all $x_0\in\R^d$.
Runge-Kutta methods {#sec:Runge:Kutta}
-------------------
Reachable sets may be approximated numerically using set-valued Runge-Kutta methods. Given a time discretisation $t_n = t_0 + nh$ with $n\in\{0,\dots, N\}$ and $h=T/N$, the iterations $$\label{svrkm}
x_{n+1} \in x_n + h\Phi(t_n,x_n),\quad n=0,\ldots,N,$$ with initial value $x_0$ from and $$\begin{aligned}
\mathcal{R}_h(t_{n+1},t_0,x_0) &=& \bigcup_{x\in\mathcal{R}_h(t_n,t_0,x_0)}\,x+h\Phi(t_n,x),\quad n=0,\ldots,N-1 \label{eqn:multivaluedRK1}\\
\mathcal{R}_h(t_0,t_0,x_0) &=& \{x_0\}, \label{eqn:multivaluedRK2}\end{aligned}$$ with a suitable increment function $\Phi: [t_0,T]\times\R^d\rightrightarrows\R^d$ define the trajectories and the reachable sets of the numerical scheme. The simplest example of such a Runge-Kutta scheme is the set-valued Euler method with increment function $\Phi(t,x)=G(t,x)$, which has been studied in [@Dontchev_89_1; @Dontchev_00_1] and several other contributions. The central result for our purposes is published in [@Dontchev_89_1].
\[semi:discrete:Euler\] Let $G:\R^d\rightrightarrows\R^d$ be Lipschitz continuous w.r.t. $\mbox{\normalfont d}_H$ with convex and compact values. Then the reachable sets of the Euler scheme satisfy $$d_H(\mathcal{R}(T, t_0, x_0), \mathcal{R}_h(T, t_0, x_0)) \le \text{const}\cdot h.$$
In practice, to compute approximations of the reachable set it is necessary to spatially discretise these schemes. The most natural approach is to introduce a grid $\rho\mathbb{Z}^d$ in the phase space $\R^d$ with grid size $\rho>0$, and to define a fully discretized scheme with trajectories $$x_{n+1} \in B_\infty(x_n + h\Phi(t_n,x_n),\rho/2)\cap\rho\Z^d,\quad n=0,\ldots,N,$$ yielding discrete reachable sets
$$\begin{aligned}
\mathcal{R}_{h,\rho}(t_{n+1},t_0,x_0)
&=& \bigcup_{x\in\mathcal{R}_{h,\rho}(t_n,t_0,x_0)}\,B_\infty(x+h\Phi(t_n,x),\rho/2)\cap\rho\Z^d,\quad n=0,\ldots,N-1, \label{union}\\
\mathcal{R}_{h,\rho}(t_0,t_0,x_0) &=& B_\infty(x_0,\rho/2)\cap\rho\Z^d.\end{aligned}$$
The blowup of the images is necessary in order to obtain subsets of the grid that are close to the original Euler sets in Hausdorff distance. Error estimates corresponding to spatial discretisation have been studied in [@Beyn_07_1]. Some results are subsumed in the following statement.
\[fully:discrete:thm\] Let $G:\R^d\rightrightarrows\R^d$ be Lipschitz continuous w.r.t. $\mbox{\normalfont d}_H$ with convex and compact values. Then the reachable sets of the fully discrete Euler scheme satisfy $$\mbox{\normalfont d}_H(\mathcal{R}(T,t_0,x_0),\mathcal{R}_{h,\rho}(T,t_0,x_0))
\le \text{const}\cdot(\rho+h+\rho/h).$$
The error term $\text{const}\cdot\rho/h$ forces the user to choose a very fine spatial discretisation, causing a high computational complexity, which is worsened by the high level of redundancy incurred by computing parts of the reachable set over and over again in the union . Higher-order Runge–Kutta methods are practically infeasible when directly transferred from ordinary differential equations to inclusions, and hence do not seem to be a cure for the complexity problem, because the computational costs of a successive evaluation of multifunctions undo all positive effects of the higher order time-discretisation.
Therefore, Euler’s method has been the main focus of study in this area. A method that is based on the Euler scheme and tracks the boundary instead of the complete reachable set for reducing the computational cost has been studied in [@Rieger_Unpub_2]. Variations of the implicit Euler scheme that are superior to the explicit Euler scheme when applied to stiff differential inclusions have been analysed in [@Beyn_10_1] and [@Rieger_14_1].
The DFOG method {#section:dfog}
---------------
In this section we review a different approach that has been recently proposed by Baier, Gerdts and Xausa [@Baier_13_1], and which they call the DFOG method (short for *distance fields on grids*). Exploratory work has been published earlier in [@BaiBusChaGer07] and [@Baier_09_1]. This method exploits the representation $$A = \mathbb{R}^d \ \backslash\ \bigcup_{x\in \mathbb{R}^d} \mbox{\normalfont int } B(x,\operatorname{dist}(x,A))\,.$$ of a closed set $A$ as the complement of the union of all open balls contained in $A^c$. Note that $x\in A$ implies $\operatorname{int} B(x,\mbox{\normalfont dist}(x,A)) = \emptyset$.
\[DFOGmethod\] Let $x_0\in\mathbb{R}^d$ and $t_0<T$ be as in . Let $\rho>0$, let $\Omega\subset\mathbb{R}^d$ be a bounded region with $B_\infty(\mathcal{R}(T,t_0,x_0),\rho/2)\subset\Omega$, and define a grid $\tilde\Omega:=\Omega\cap\rho\mathbb{Z}^d$.
For every $z\in \tilde\Omega$, solve the optimisation problem $$\operatorname{Minimise} \frac{1}{2}\|x_N - z\|^2$$ over trajectories $(x_n)_{n=0}^N$ of the Runge-Kutta scheme . Let $(x_n^*(z))_{n=0}^N$ be a solution to this problem, and denote $x^*(z):=x_N^*(z)$ and $\theta(z) := \frac{1}{2}\|x_N^*(z) - z\|^2$. Then the set $$\label{eqn:DFOGrepresentation}
\mathcal{R}_{DFOG}(T,t_0,x_0):= \Omega \ \backslash \ \bigcup_{z\in\tilde\Omega}
\mbox{\normalfont int } B(z,\sqrt{2\theta(z)})\,.$$ is an approximation to the reachable set $\mathcal{R}(T,t_0,x_0)$.
Algorithm \[DFOGmethod\] requires the solution of gobal optimisation problems in a very high-dimensional state space, to which standard tools for local optimisation such as the SQP method may be applied (see for example [@Nocedal_99_1]). Due to the high state-space dimension, it is impossible to use global optimisation routines. Therefore, one potential pitfall is the existence of local minima leading to incorrect results for $x^*(z)$ and $\theta(z)$. This means that the representation may potentially cut away large parts of the reachable set.
One heuristic solution to this problem offered in [@Baier_13_1] is ‘ball-checking’: For any $z,z'\in\tilde\Omega$, check whether the computed $x^*(z)$ and $\theta(z')$ satisfy $x^*(z)\in B(z',\sqrt{2\theta(z')})$. In that case, the optimisation routine failed to compute $\theta(z')$ properly, so that the ball $B(z',\sqrt{2\theta(z')})$ is incorrect and must be ignored. However, this strategy does not necessarily detect erroneous results.
The following statement is a consequence of Theorem \[semi:discrete:Euler\] and the representation .
\[dfog:estimate\] Let $G:\R^d\rightrightarrows\R^d$ be Lipschitz continuous w.r.t. $\mbox{\normalfont d}_H$ with convex and compact values. If the optimisation problems in Algorithm \[DFOGmethod\] are solved correctly, then $$\mbox{\normalfont d}_H(\mathcal{R}(T,t_0,x_0),\mathcal{R}_{DFOG}(T,t_0,x_0)) \le \text{const}\cdot(h+\rho)\,.$$
In contrast to Theorem \[fully:discrete:thm\], the error estimate does not contain the critical term $\text{const}\cdot\rho/h$, which indicates that the DFOG method should be substantially faster than the Euler scheme. Nevertheless, it is difficult to compare the performance of both schemes on a formal level, because their design and their behaviour are too different.
The DFOG method can be accelerated using the method of maximal gains published in [@Rieger_13_1], which does not use a spatial a grid, but chooses optimal test points with respect to all information that is available at runtime.
SVM Algorithm
=============
In this section, we develop step by step the algorithm the present paper is concerned with. As the DFOG method, it requires the results of the optimal control routine given in Algorithm \[DFOGmethod\], but it uses the well-known SVM algorithm from machine learning to represent the approximation of the reachable set as a sublevel set of a smooth function chosen from a reproducing kernel Hilbert space (RKHS). This representation provides a smoother boundary for the reachable set approximation and some robustness against a small number of errors corresponding to the global optimisation routine finding local minima.
In Section \[sec:rkhs\], we give a brief introduction to reproducing kernel Hilbert spaces, and in Section \[sec:SVM\], we adapt the classical SVM algorithm to our particular problem. In Section \[sec:incremental\], we discuss the possibility of an adaptive enlargement of the dataset and computational implications. A brief comment on the ball-checking procedure in the SVM context is given in Section \[sec:ball:check\].
Reproducing Kernel Hilbert Space {#sec:rkhs}
--------------------------------
We will define our RKHS in terms of a Mercer kernel.
Let $X\subset\mathbb{R}^d$. A Mercer kernel is a function $K: X \times X \rightarrow\mathbb{R}$ satisfying
- $K(x,x')=K(x',x)$ for all $x,x'\in X$,
- $\sum^n_{i,j=1} c_ic_jK(x_i,x_j)\ge0$ for any $n\in\mathbb{N}$, any $c_1,\ldots,c_n\in\mathbb{R}$ and any $x_1,\ldots,x_n\in X$.
Typical examples of kernel functions include the Gaussian kernel $$K_G(x,y) = \textrm{exp}\big(-\textstyle\frac{1}{\sigma}\|x-y\|^2\big)\qquad \text {with } \sigma >0\,, \label{eqn:Gaussiankernel}$$ and the *degree-*$p$ polynomial kernel $$K_P(x,y) = (x^T y +\tau)^p \qquad \text{with } \tau\ge 0\,.$$
Given a kernel function, we define $K_x:=K(x,\cdot)$. The following theorem states how a Mercer kernel uniquely defines a reproducing kernel Hilbert space [@Aronszajan_50_1]:
\[thm:Moore-Aronszajn\] Given a Mercer kernel $K$, there exists a unique Hilbert space $\mathcal{H}_K$ of functions on $X$ with associated inner product $\langle\cdot,\cdot \rangle_K$ satisfying the following conditions:
- $K_x\in\mathcal{H}_K$ for all $x\in X$,
- $\mbox{\normalfont span}\{K_x : x\in X\}$ is dense in $\mathcal{H}_K$,
- $f(x) = \langle K_x,f\rangle_K$ for all $f\in\mathcal{H}_K$ and all $x\in X$.
The inner product in the RKHS is defined by $\langle K_x,K_y\rangle_K = K(x,y)$ and extending linearly. $\mathcal{H}_K$ is then taken as the completion of the linear span of $\{K_x : x\in X\}$ with respect to this inner product. The third property in Theorem \[thm:Moore-Aronszajn\] is the *reproducing property*.
In this paper, we choose to work with the RKHS corresponding to the Gaussian kernel , although the algorithm is also viable with other choices of kernels. The Gaussian RKHS has been well studied and is a very rich function space to work in, which is illustrated by the following result from [@Steinwart_01_1].
Let $X\subset\mathbb{R}^d$ be compact. Then the Gaussian RKHS $\mathcal{H}_K$ on $X$ is dense in the space $C(X)$ of continuous functions on $X$.
For a detailed coverage of the Gaussian RKHS, we refer to [@Minh_10_1; @Steinwart_06_1]. In practice, we will be working with finite dimensional RKHS, and in particular, the set $X$ will be chosen according to the grid points and results from the DFOG optimal control method.
Support Vector Machine {#sec:SVM}
----------------------
Support vector machines (SVMs) are well-known supervised learning algorithms frequently used for classification problems, a common task in machine learning problems. The soft-margin SVM algorithm was first proposed by Cortes and Vapnik [@Cortes_95_1] and is now a popular choice for machine learning problems. Applications of the SVM algorithm include handwriting recognition, image classification and text categorisation [@DeCoste_02_1; @Joachims_97_1; @Schoelkopf_01_1; @Tong_01_1].
We apply the SVM algorithm to a labelled training set $\mathcal{D}$, which contains all relevant information encoded in the output data $(z,\theta(z),x^*(z))_{z\in\tilde\Omega}$ of the DFOG method, in order to recognise the shape of the reachable set. The training set $\mathcal{D}$ and index sets $\mathcal{I}$, $\mathcal{E}$ and $\mathcal{B}$ that partition $\mathcal{D}$ into interior, exterior and boundary points of the reachable set, respectively, are constructed as follows.
\[construct:training:set\] First run Algorithm \[DFOGmethod\] to obtain the data $(z,\theta(z),x^*(z))_{z\in\tilde\Omega}$. Set $\mathcal{I},\mathcal{E},\mathcal{B}:=\emptyset$ and $m:=0$. Fix $\epsilon>0$.\
[for]{} $z\in\tilde\Omega$\
$\theta(z)\le\epsilon$ [then]{}\
$x_{m+1}:=z$\
$\mathcal{D}:=(x_i)_{i=1}^{m+1}$, $\mathcal{I}:=\mathcal{I}\cup\{m+1\}$\
$m:=m+1$\
\
$x_{m+1}:=z$, $x_{m+2}:=x^*(z)$\
$\mathcal{D}:=(x_i)_{i=1}^{m+2}$, $\mathcal{E}:=\mathcal{E}\cup\{m+1\}$, $\mathcal{B}:=\mathcal{B}\cup\{m+2\}$\
$m:=m+2$.\
\
[end]{}
The idea behind this algorithm is simple. For any $z\in\tilde\Omega$, the fact that $\theta(z)=0$ implies $z\in\mathcal{R}(T,t_0,x_0)$. If, on the other hand, $\theta(z)>0$, then $z\notin\mathcal{R}(T,t_0,x_0)$ and $x^*(z)\in\partial\mathcal{R}(T,t_0,x_0)$, assuming no error has been made in the global optimisation routine. This way, Algorithm \[construct:training:set\] constructs a training set $\mathcal{D}:=(x_i)_{i=1}^m$ of points $(x_i)_{i=1}^m\subset\Omega$ with index set partitions $\mathcal{I}$, $\mathcal{E}$ and $\mathcal{B}$. The small parameter $\epsilon>0$ is introduced to compensate for numerical precision errors. By construction, we have $|\mathcal{I}| + |\mathcal{E}| + |\mathcal{B}| = m$.
The support vector machine algorithm is designed to find a function from an RKHS (along with its sublevel set) which best fits a labelled training set such as this, in the sense of minimising a suitable loss function. However, the context here differs from the usual setting in which the SVM is applied (to a set of randomly generated data potentially subject to noise) in two main ways.
Firstly, the training set is not just labelled according to whether a sample point belongs to the reachable set or not, but also has the possible label of being on the boundary, as indicated by the three index sets $\mathcal{I}$, $\mathcal{E}$ and $\mathcal{B}$.
Secondly, the standard soft-margin SVM classifier allows for statistical errors in the labelling of data. Here, only specific errors can occur: a point that is labelled as an interior or boundary point must belong to the reachable set as the optimal control routine finds an admissible path to reach that point. So we do not want to allow a point with index $i\in\mathcal{I}\cup\mathcal{B}$ to be on the exterior of our reachable set approximation. However, a point labelled as an exterior point could well belong to the reachable if the optimal control routine failed to find the global minimum.
We present the following adapted SVM algorithm in order to account for these differences:
\[AdaptedSVM\] First run Algorithm \[construct:training:set\] to obtain the set $\mathcal{D} = (x_i)_{i=1}^m$ and index set partitions $\mathcal{I}$, $\mathcal{E}$ and $\mathcal{B}$. Fix regularisation parameters $C_1,C_2>0$, and let $K(\cdot,\cdot)$ be a Mercer kernel with corresponding finite-dimensional RKHS $\mathcal{H}_K$ on $X:=\{x_1,x_2,\dots,x_m\}$. We search for a function $f=\sum_{i=1}^m a_i K_{x_i}$ in $\mathcal{H}_K$ by solving the following optimisation problem over the optimisation variables $(a, b, \xi, \eta) \in \mathbb{R}^m\times\mathbb{R} \times \mathbb{R}^{|\mathcal{E}|} \times \mathbb{R}^{|\mathcal{B}|}$: $$\begin{aligned}
\operatorname{Minimise }_{a,b,\xi,\eta} \quad \frac{1}{2}\|f\|_{\mathcal{H}_K}^2
+ C_1\sum_{i\in\mathcal{E}} \xi_i + C_2\sum_{i\in\mathcal{B}}\eta_i, & \label{AdaptedSVM:cost}\\
\text{subject to }\qquad \sum_{k=1}^m a_k K(x_k,x_i) + b \ge 1,& \qquad i\in\mathcal{I}, \label{AdaptedSVM:Iconstraints}\\
-\sum_{k=1}^m a_kK(x_k,x_i) - b \ge 1 - \xi_i,& \qquad i\in\mathcal{E}, \label{AdaptedSVM:Econstraints}\\
\sum_{k=1}^ma_kK(x_k,x_i) + b = \eta_i,& \qquad i\in\mathcal{B}, \label{AdaptedSVM:Bconstraints}\\
\xi_i\ge 0,& \qquad i\in\mathcal{E}, \label{AdaptedSVM:Eslack}\\
\eta_i \ge 0,& \qquad i\in\mathcal{B}. \label{AdaptedSVM:Bslack}\end{aligned}$$ The approximation of the reachable set $\mathcal{R}(T,t_0,x_0)$ is given by $$\label{eqn:SVMrepresentation}
\mathcal{R}_{SVM}(T,t_0,x_0):=\{x\in\mathbb{R}^d : f(x) + b \ge 0\}.$$
The labelled training set generated by Algorithm \[construct:training:set\], which contains all available knowledge about the reachable set, is incorporated in constraints , and . The constraint ensures that the function value is at least $1$ on the points that are labelled as interior points. Note that there is no slack variable appearing in this constraint, according to our observation that points labelled as interior points must lie within the reachable set. In contrast, contains the non-negative slack variable $\xi_i$ (see also ), which allows for the possibility of an error being made on a point labelled as an exterior point. Where the slack variable is zero, the function value is less than or equal to $-1$ on exterior points. The constraint tries to place boundary points on the level set $\{x\in\mathbb{R}^d:f(x)+b=0\}$. Here the non-negativity condition follows front the fact that points labelled as boundary points are the endpoints of orbits of and so cannot be on the exterior of the reachable set.
The first term of the cost function controls the complexity of the function $f\in\mathcal{H}_K$ (and hence the sub level set) to avoid overfitting the training set $\mathcal{D}$. This is contrasted with the following two terms, which control the penalty due to errors in classification. This bias-variance trade-off is managed through the regularisation coefficients $C_1$ and $C_2$. As these coefficients approach infinity, the function $f$ is allowed to become more and more complex, and the solution to the optimisation problem approaches the hard-margin solution where no errors are permitted on the training set.
The optimisation problem – is a convex optimization problem, and in particular all the constraints are affine. In this case Slater’s Theorem guarantees strong duality if the problem is feasible [@Boyd_04_1]. In the case of the Gaussian kernel, feasibility is guaranteed by the following theorem [@Micchelli_86_1].
\[thm:Gaussianfullrank\] Let $x_1,\ldots,x_m\subset X$ be distinct points, and $\sigma >0$. The matrix $\mathbb{K}$ given by $$\mathbb{K}_{ij} = \exp\big({-\textstyle\frac{1}{\sigma}\|x_i-x_j\|^2}\big)$$ has full rank.
Therefore, Algorithm \[AdaptedSVM\] can be recast into the dual problem using the KKT conditions. This is the problem that is generally solved in practice.
We introduce the variables $y_i \in\{-1,1\}$, $i=1,\ldots,m$ by defining $y_i = 1$ for $i\in\mathcal{I}\cup\mathcal{B}$ and $y_i=-1$ for $i\in\mathcal{E}$.
Under the same conditions as in Algorithm \[AdaptedSVM\], solve the following minimisation problem over the variables $\alpha \in \mathbb{R}^m$: $$\begin{aligned}
\operatorname{Minimise }_{\alpha} \quad \frac{1}{2}\sum_{i,j=1}^m y_i y_j \alpha_i \alpha_j K(x_i,x_j) - \sum_{i\in \mathcal{I}\cup\mathcal{E}}\alpha_i & \label{AdaptedSVMdual:cost}\\
\text{subject to }\qquad \sum_{i=1}^m y_i\alpha_i = 0, \label{AdaptedSVMdual:offsetconstraints}\\
\alpha_i\ge 0,\qquad i\in\mathcal{I} \label{AdaptedSVMdual:Iconstraints}\\
0\le\alpha_i \le C_1, \qquad i\in\mathcal{E} \label{AdaptedSVMdual:Econstraints}\\
\alpha_i \ge -C_2,\qquad i\in\mathcal{B} \label{AdaptedSVMdual:Bconstraints}\end{aligned}$$
The solution to the problem – provides the function $f=\sum_{i=1}^m a_i K_{x_i}$, where the $a_i$ are given by $a_i = y_i \alpha_i$ for $i\in\{1,\ldots, m\}$. The points $x_i$ for which the corresponding constraint , or are strictly satisfied are called the *support vectors* in the literature. For the support vectors the corresponding constraints , and are satisfied as equalities, and in addition $\xi_i$ or $\eta_i$ is equal to zero. The offset $b$ can therefore be computed from , and for the support vectors.
Accordingly, points $x_i$ ($i\in\mathcal{E}$) for which $\alpha_i = C_1$ and points $x_i$ ($i\in\mathcal{B}$) for which $\alpha_i = C_2$ are the so-called *error vectors*. These are the points for which $\xi_i$ and $\eta_i$ may be nonzero, and for which the reachable set approximation may misclassify. A boundary point $x_i$ for which $\eta_i >0$ will still be classified as being in the reachable set, but will not be on the boundary of the set approximation. However an exterior point $x_i$ for which $\xi_i>1$ will be misclassified by . If $0<\xi_i<1$ then $x_i$ will be still be on the exterior of $\mathcal{R}_{SVM}$ but will be inside the ‘margin’ $\{x\in\mathbb{R}^d : |f(x)+b| <1\}$ and so it is still called an error vector.
Finally, points $x_i$ $(i\in\mathcal{I}\cup\mathcal{E})$ for which $\alpha_i=0$ are *ignored vectors*. It is not hard to see that these points have no influence on the solution to the above optimisation problem, and could as well have been left out of the data set. In addition, the property of being an ignored vector is robust with respect to perturbation of the support and error vectors. Note that the set of boundary points $x_i$ ($i\in\mathcal{B}$) by definition does not contain any ignored vectors, since the property that $\alpha_i=0$ is not robust with respect to such a perturbation due to . Roughly speaking, points that are far away from the boundary of the reachable set (and are correctly classified) will be ignored vectors. However it is not practically possible to tell in advance which data points will be ignored vectors, or even if ignored vectors will remain ignored with the addition of new points.
Incremental updates {#sec:incremental}
-------------------
It is possible to increase the accuracy of the SVM approximation step by step, until a desired precision is reached. In that case, the optimisation problem – (or –) needs to be solved after each addition of a batch of new points. This optimisation problem runs over all points in the training set, so as this set becomes larger, this may become costly.
Fortunately it is possible to solve the SVM optimisation problem by means of incremental updates [@Cauwenberghs_01_1]. This procedure consists of deriving equations to keep the KKT conditions satisfied, as a new dual variable $\alpha_i$ is incremented from zero. The procedure ends when a new point becomes either a support vector or error vector. For details, we refer to [@Cauwenberghs_01_1]. Here, we outline the procedure for our adapted version of the SVM algorithm.
The cost function in the dual formulation of the optimisation problem may trivially be rewritten in the more convenient form $$\frac{1}{2}\sum_{i,j=1}^m y_i y_j \alpha_i \alpha_j K(x_i,x_j) - \sum_{i\in \mathcal{I}\cup\mathcal{E}}\alpha_i + b\sum_{i=1}^m y_i\alpha_i \label{AdaptedSVMdual:costinc}$$ retaining the constraints –, with the offset $b$ re-introduced as a Lagrange multiplier. The necessary and sufficient KKT conditions for this problem may be written as follows: $$\begin{aligned}
i\in\mathcal{I}: & g_i:= f(x_i) + b -1 &
\begin{cases}
\ge 0 \quad \textrm{when } \alpha_i =0\\
=0 \quad \textrm{when } \alpha_i >0
\end{cases} \label{eqn:KKTI}\\
i\in\mathcal{E}: & g_i:= -f(x_i) -b -1 &
\begin{cases}
\ge 0 \quad \textrm{when } \alpha_i =0\\
=0 \quad \textrm{when } 0 < \alpha_i < C_1\\
\le 0 \quad \textrm{when } \alpha_i = C_1
\end{cases} \label{eqn:KKTE}\\
i\in\mathcal{B}: & g_i:= f(x_i) + b &
\begin{cases}
\ge 0 \quad \textrm{when } \alpha_i =-C_2\\
=0 \quad \textrm{when } \alpha_i > -C_2
\end{cases} \label{eqn:KKTB}\end{aligned}$$ $$\sum_{i=1}^m y_i\alpha_i = 0 \label{eqn:KKTb}$$
The conditions – are satisfied with equality for the support vectors. Given a new labelled point $(x_c,y_c)$ with dual variable $\alpha_c$ initially set to zero, we need to ensure that these equality conditions (as well as ) continue to be satisfied for the support vectors as we increment $\alpha_c$ from zero. Following [@Cauwenberghs_01_1], we define the *coefficient sensitivities* $\beta_i$ by $$\label{eqn:betadef}
\left[
\begin{matrix}
\beta_0\\
\beta_{s_1}\\
\vdots\\
\beta_{s_{\mathcal{N}(S)}}
\end{matrix}
\right] = - \left[
\begin{matrix}
0 & y_{s_1} & \cdots & y_{s_{\mathcal{N}(S)}}\\
y_{s_1} & Q_{s_1s_1} & \cdots & Q_{s_1s_{\mathcal{N}(S)}}\\
\vdots & \vdots & \ddots & \vdots\\
y_{s_{\mathcal{N}(S)}} & Q_{s_{\mathcal{N}(S)}s_1} & \cdots &Q_{s_{\mathcal{N}(S)}s_{\mathcal{N}(S)}}
\end{matrix}
\right]^{-1}
\left[
\begin{matrix}
y_c\\ Q_{s_1c}\\
\vdots \\ Q_{s_{\mathcal{N}(S)}c}
\end{matrix}
\right]$$ where $Q_{ij}=y_iy_jK(x_i,x_j)$ and $\{s_1,\ldots,s_{\mathcal{N}(S)}\}$ is the index set corresponding to the support vectors ($\mathcal{N}(S)$ is the number of support vectors). We define $\beta_i=0$ for indices $i$ corresponding to ignored and error vectors. Then, the KKT conditions – will continue to be satisfied as $\alpha_c$ is incremented from zero provided the existing dual coefficients are also incremented according to $$\begin{aligned}
\Delta b & = & \beta_0 \Delta\alpha_c \label{eqn:deltab}\\
\Delta \alpha_i & = & \beta_i \Delta\alpha_c, \label{eqn:deltaalpha}\end{aligned}$$ where $\Delta\alpha_c$ is a small increment in $\alpha_c$. The *margin sensitivities* $\gamma_i$ are likewise defined by $$\gamma_i = Q_{ic} + \sum_{j=s_1}^{s_{\mathcal{N}(S)}}Q_{ij}\beta_{j} + y_i\beta_0$$ and give the variation of the margins $g_i$ in –: $$\Delta g_i = \gamma_i\Delta\alpha_c \label{eqn:deltag}$$ Equations – ensure that $\gamma_i = 0$ for support vectors.
Now, for each new point $(x_c,y_c)$ the corresponding $g_i$ is first computed. If the new point automatically satisfies the KKT conditions then it is an ignored vector and $\alpha_c$ is left at zero. Otherwise we use , and to compute the largest possible increment of $\alpha_c$ so that – continue to be satisfied, at which point either $x_c$ becomes a support or error vector, or else another point in the data set migrates between the sets of support, error or ignored vectors. Then the coefficient and margin sensitivities must be recomputed and the procedure continues.
Ball checking {#sec:ball:check}
-------------
The procedure outlined in the previous can also naturally be reversed in order to remove a point from the training set. In practical implementation we have included the ball checking routine as described in the end of section \[section:dfog\] in which case it is sometimes necessary to remove an exterior point from the data set. Note however that in these cases the SVM algorithm allows to keep the corresponding boundary point in the data set.
In practice, where the optimisation routine fails to find the global minimum, it is usually the case that the computed optimal trajectory still terminates at a boundary point of the reachable set. However, note that the SVM algorithm does allow for points in the index set $\mathcal{B}$ to actually be interior points. In this case, some information on the reachable set is still retained in the case of an error due to the global optimisation routine of the DFOG method.
Examples
========
We illustrate the qualities of our method by applying it to two examples from the literature. In each example we compare its performance in the reachable set representation $\mathcal{R}_{SVM}(T,t_0,x_0)$ with that of the DFOG method and its reachable set approximation $\mathcal{R}_{DFOG}(T,t_0,x_0)$ as given in .
A bilinear control system
-------------------------
The following example is taken from [@Baier_13_1; @Hajek_08_1] as a model system that exhibits convexity of the reachable set for small times, but nonconvexity for larger times. Both the DFOG and SVM methods work for either case. We consider the two-dimensional control system $$\begin{aligned}
\dot{x}_1 & = \pi x_2\,, \label{eqn:bilinear1}\\
\dot{x}_2 & = -\pi u(t)x_1\,, \label{eqn:bilinear2}\\
x_1(0) & = -1\,, \label{eqn:bilinear3}\\
x_2(0) & = 0\,, \label{eqn:bilinear4}\\
u(t) \in U & = [0,1]\,. \label{eqn:bilinear5}\end{aligned}$$
We are interested in approximating the reachable set $\mathcal{R}(1,0,{x}_0)$ for ${x}_0:=(x_1(0),x_2(0))$. The reachable set is shown in Figure \[fig:bilinearDFOG\]. In this computation the time interval $[0,1]$ has been discretised with $N=30$ steps. We note that the error due to time discretisation is the same for both the DFOG and modified SVM methods. This is because both methods use the same time discretisation in the constraints for the optimal control problem from Algorithm \[DFOGmethod\]. The difference between the two methods is the spatial representation of the reachable set. In order to compare the methods we leave the time discretisation at $N=30$ and vary the spatial grid size $\rho$.
![Reachable set $\mathcal{R}(1,0,(-1,0))$ for the bilinear control system.[]{data-label="fig:bilinearDFOG"}](BL-boundarytracking_NC.pdf){width="70.00000%"}
Figure \[fig:bilinearDFOG\_SVM\] shows successive approximations of both the DFOG and SVM methods for the reachable set $\mathcal{R}_h(1,0,{x}_0)$, where $h=\frac{1}{N}$ and $N=30$ is fixed. For both algorithms, the set of grid points $\tilde{\Omega}$ was defined as a restriction of $\rho\mathbb{Z}^2$, and the approximations are made for varying spatial discretizations $\rho$ independently. In this example there are very few errors made by the global optimisation routine. The Hausdorff distances calculated between the true reachable set and the numerical approximations from both the DFOG and SVM methods are shown in Figure \[fig:BLCS\_HD\].
--------------- -------- --------
$\rho$ DFOG SVM
\[0.5ex\] 1.0 0.3794 0.1889
0.9 0.3738 0.1177
0.8 0.3780 0.1034
0.7 0.2373 0.0792
0.6 0.2165 0.0919
0.5 0.1542 0.0803
0.4 0.1113 0.0296
0.3 0.0604 0.0412
0.2 0.0222 0.0190
\[1ex\]
--------------- -------- --------
{width="75.00000%"}
\[fig:BLCS\_HD\]
[0.45]{}
[0.45]{}
[0.45]{}
[0.45]{}
[0.45]{}
[0.45]{}
[0.45]{}
[0.45]{}
A nonlinear control system
--------------------------
The following example was presented in [@Rieger_Unpub_2] as an example of a reachable set that may change its topology for different times $T$. $$\begin{aligned}
\dot{x}_1 & = x_1(1-|x_1|) - x_1x_2 + u_1\,,\\
\dot{x}_2 & = x_1^4 - \tfrac{1}{2} + u_2\,,\\
x_1(0) & = 0\,,\\
x_2(0) & = 0\,,\\
(u_1(t), u_2(t)) \in U & = [-\tfrac{1}{5},\tfrac{1}{5}]\times[-\tfrac{1}{5},\tfrac{1}{5}]\,.\end{aligned}$$
In this example we will approximate the reachable set $\mathcal{R}(3.5,0,{x}_0)$, where ${x}_0 := (x_1(0),x_2(0))$. This reachable set is shown in Figure \[fig:topDFOG\].
![Reachable set $\mathcal{R}(3.5,0,(0,0))$ for the nonlinear control system.[]{data-label="fig:topDFOG"}](Top_boundary-tracking_NC.pdf){width="40.00000%"}
Figure \[fig:RiegerDFOG\_SVM\] shows successive approximations of the DFOG and SVM methods for the reachable set $\mathcal{R}_h(3.5,0,{x}_0)$, for $N=50$. As before, the grid points are defined as a restriction of the grid $\rho\mathbb{Z}^2$. The approximations are shown for varying spatial discretizations $\rho$, and each figure is produced by an independent run of the algorithms. Figure \[fig:NLCS\_HD\] provides the Hausdorff distances calculated between the true reachable set and the approximations made by both the SVM and DFOG methods.
This example contains more errors made by the global optimisation routine than the previous example, due to the highly non-convex topology of the reachable set. We can see that the SVM algorithm appears to converge faster to a good approximation of the reachable set. Again, the SVM algorithm is somewhat robust to these global optimisation errors. Note that information is still added to the SVM algorithm even in the case of a global optimisation error, since the computed optimal point will still be in the reachable set, and this point is still added to the algorithm.
--------------- -------- --------
$\rho$ DFOG SVM
\[0.5ex\] 1.0 0.6452 0.3629
0.9 0.7868 0.1862
0.8 0.3910 0.1994
0.7 0.6550 0.1128
0.6 0.6378 0.1755
0.5 0.5578 0.1256
0.4 0.4039 0.0321
0.3 0.1089 0.0443
0.2 0.0335 0.0288
\[1ex\]
--------------- -------- --------
{width="75.00000%"}
\[fig:NLCS\_HD\]
[0.28]{}
[0.28]{}
[0.28]{}
[0.28]{}
[0.28]{}
[0.28]{}
[0.28]{}
[0.28]{}
[0.28]{}
Conclusions
===========
The modified Support Vector Machine algorithm provides an alternative representation of the reachable set, based on the results gained from a set of global optimisation problems provided by the DFOG algorithm. This new approach has the advantage that it is robust to a small number global optimisation errors, and appears to benefit from faster convergence for particular examples. Several specialised algorithms exist for efficiently solving the standard SVM optimisation problem [@Joachims_98_1; @Platt_98_1], which could be adapted to the modified SVM algorithm we have presented here. The global optimal control problems are also particularly expensive when the dimension of the control variable is large, or when a fine time discretisation is used. Therefore for many real-world problems only few optimal control problems can be solved in practice. In these cases where relatively few data from the optimal control routine are available, the SVM algorithm performs significantly better. In addition, the sublevel set representation of the SVM approach is more handy for many applications than the DFOG representation.
As is always the case with algorithms of this type, there are several parameters in the algorithm that need to be tuned for optimal performance. The tolerance $\epsilon$ as described in Section \[sec:SVM\] is important to distinguish interior and exterior points, and affects the approximation for both the DFOG and SVM algorithms. Within the SVM algorithm, the parameters $C_1$ and $C_2$ control the regularity of the solution as described earlier. Also for radial basis functions such as the Gaussian kernel used in our examples, the scaling parameter $\sigma$ is an additional parameter, related to the regularisation parameters $C_1$ and $C_2$. These parameters can be chosen using standard validation techniques such as hold-out testing (see [@Cauwenberghs_01_1; @Vapnik_95_1]), but the precise effect of these parameters on the regularity of the solution (and how the parameters relate to each other) is not yet well understood.
A final important problem is that of choosing the best points on which to run the global optimisation routine in order to improve the current approximation. In our problem setting, we are in the fortunate position of being able to choose any point to run the algorithm on at each step. This is in contrast to many applications of the Support Vector Machine, where the data is randomly generated from an unknown underlying distribution. In our example applications we have run the algorithm on a regular grid, however it is clear that this is not the optimal strategy. The question of how to choose the best point is likely to be related to problem of understanding the effect of the parameters in the algorithm, and again is a worthwhile subject of future work. The framework provided in this paper to incrementally update the SVM algorithm is also a precursor to such a strategy.
A further benefit of our proposed methodology is that it may also in principle be used to compute invariant sets for random dynamical systems [@Homburg_06_1; @Lamb_Unpub_1], as well as invariant sets for control systems [@Colonius_00_1].
#### Acknowledgements.
The first and the third author were supported by the UK Engineering and Physical Sciences Research Council (EPSRC).
[CYLW98]{}
N. Aronszajn, *Theory of reproducing kernels*, Transactions of the American Mathematical Society **68** (1950), 337–404.
R. Baier, C. Büskens, I.A. Chahma, and M. Gerdts, *Approximation of reachable sets by direct solution methods for optimal control problems*, Optim. Methods Softw. **22** No. 3 (2007), 433–452.
R. Baier and M. Gerdts, *A computational method for non-convex reachable sets using optimal control*, Proceedings of the European Control Conference (ECC) 2009, Budapest (Hungary), August 23–26, 2009 (Budapest), EUCA, 2009, pp. 97–102.
R. Baier, M. Gerdts, and I. Xausa, *Approximation of reachable sets using optimal control algorithms*, Numerical Algebra, Control and Optimization **3** (2013), no. 3, 519–548.
W.-J. Beyn and J. Rieger, *Numerical fixed grid methods for differential inclusions*, Computing **81** (2007), no. 1, 91–106.
[to3em]{}, *The implicit Euler scheme for one-sided [L]{}ipschitz differential inclusions*, Disc. Cont. Dyn. Sys. B **14** (2010), no. 2, 409–428.
S. Boyd and L. Vandenberghe, *Convex optimization*, Cambridge University Press, Cambridge, 2004.
G. Cauwenberghs and T. Poggio, *Incremental and decremental support vector machine learning*, Advances in Neural Information Processing Systems, vol. 13, MIT Press, 2001, pp. 409–415.
F. Colonius and W. Kliemann, *The Dynamics of Control*, Birkhäuser, 2000.
C. Cortes and V. Vapnik, *Support-vector networks*, Machine Learning **20** (1995), no. 3, 273–297.
F.H. Clarke, R.J. Stern Y.S. Ledyaev, and P.R. Wolenski, *Nonsmooth analysis and control theory*, Graduate Texts in Mathematics, vol. 178, Springer, New York, 1998.
K. Deimling, *Multivalued [D]{}ifferential [E]{}quations*, de Gruyter Series in Nonlinear Analysis and Applications, vol. 1, Walter de Gruyter & Co., Berlin, 1992.
A.L. Dontchev and E.M. Farkhi, *Error estimates for discretized differential inclusion*, Computing **41** (1989), no. 4, 349–358.
A.L. Dontchev, W.W. Hager, and V.M. Veliov, *Second-order [R]{}unge–[K]{}utta approximations in control constrained optimal control*, SIAM Journal on Numerical Analysis **38** (2000), no. 1, 202–226.
D. DeCoste and B. Schölkopf, *Training invariant support vector machines*, Machine Learning **46** (2002), 161–190.
M. Gerdts, R. Henrion, D. Hömberg and C. Landry, *Path planning and collision avoidance for robots*, Numerical Algebra, Control and Optimization **2** (2012), no. 3, 437–463.
O. Hájek, *Control theory in the plane*, second ed., Lecture Notes in Control and Information Sciences, vol. 153, Springer, Berlin, 2008.
A.J. Homburg and T. Young, *Hard bifurcations in dynamical systems with bounded random perturbations*, Regular & Chaotic Dynamics **11** (2006), no. 2, 247–258.
T. Joachims, *Text categorisation with support vector machines*, Technical report, LS VIII No. 23, University of Dortmund, 1997.
[to3em]{}, *Making large-scale support vector machine learning practical*, in Advances in Kernel Methods - Support Vector Learning, Cambridge MA: MIT Press, Schölkopf, Burges and Smola, Eds., 1998, 169–184.
J.S.W. Lamb, M. Rasmussen and C.S. Rodrigues, *Topological bifurcations of minimal invariant sets for set-valued dynamical systems*, to appear in: Proceedings of the American Mathematical Society.
C.A. Micchelli, *Algebraic aspects of interpolation*, Approximation theory ([N]{}ew [O]{}rleans, [L]{}a., 1986), Proceedings of Symposia in Applied Mathematics, vol. 36, American Mathematical Society, Providence, RI, 1986, pp. 81–102.
H.Q. Minh, *Some properties of [G]{}aussian reproducing kernel [H]{}ilbert spaces and their implications for function approximation and learning theory*, Constructive Approximation **32** (2010), no. 2, 307–338.
J. Nocedal and S.J. Wright, *Numerical optimization*, Springer Series in Operations Research, Springer, New York, 1999.
J.C. Platt, *Fast training of support vector machines using sequential minimum optimization*, in Advances in Kernel Methods - Support Vector Learning, Cambridge MA: MIT Press, Schölkopf, Burges and Smola, Eds., 1998, 185–208.
J. Rieger, *Non-convex systems of sets for numerical analysis*, Computing **95** (2013), no. 1, suppl., S617–S638.
[to3em]{}, *Semi-implicit [E]{}uler schemes for ordinary differential inclusions*, SIAM J. Numer. Anal. **52** (2014), no. 2, 895–914.
[to3em]{}, *Robust boundary tracking for reachable sets of nonlinear differential inclusion*, to appear in: Foundations of Computational Mathematics, DOI: 10.1007/s10208-014-9218-8.
I. Steinwart, D. Hush, and C. Scovel, *An explicit description of the reproducing kernel [H]{}ilbert spaces of [G]{}aussian [RBF]{} kernels*, IEEE Transactions on Information Theory **52** (2006), no. 10, 4635–4643.
B. Schölkopf and A. Smola, *Learning with kernels: Support vector machines, regularization, optimization and beyond*, MIT Press, 2001.
I. Steinwart, *On the influence of the kernel on the consistency of support vector machines*, Journal of Machine Learning Research **2** (2001), no. 67–93.
S. Tong and E. Chang, *Support vector machine active learning for image retrieval*, Proceedings of the Ninth ACM International Conference on Multimedia, ACM, 2001, pp. 107–118.
V. Vapnik, *The Nature of Statistical Learning Theory*, New York: Springer-Verlag, 1998.
[^1]: [email protected]
[^2]: [email protected]
[^3]: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: '[In this paper, a finite volume element (FVE) method is considered for spatial approximations of time-fractional diffusion equations involving a Riemann-Liouville fractional derivative of order $\alpha \in (0,1)$ in time. Improving upon earlier results (Karaa [*et al.*]{}, IMA J. Numer. Anal. 2016), optimal error estimates in $L^2(\Omega)$- and $H^1(\Omega)$-norms for the semidiscrete problem with smooth and middly smooth initial data, i.e., $v\in H^2(\Omega)\cap H^1_0(\Omega)$ and $v\in H^1_0(\Omega)$ are established. For nonsmooth data, that is, $v\in L^2(\Omega)$, the optimal $L^2(\Omega)$-error estimate is shown to hold only under an additional assumption on the triangulation, which is known to be satisfied for symmetric triangulations. Superconvergence result is also proved and as a consequence, a quasi-optimal error estimate is established in the $L^\infty(\Omega)$-norm. Further, two fully discrete schemes using convolution quadrature in time generated by the backward Euler and the second-order backward difference methods are analyzed, and error estimates are derived for both smooth and nonsmooth initial data. Based on a comparison of the standard Galerkin finite element solution with the FVE solution and exploiting tools for Laplace transforms with semigroup type properties of the FVE solution operator, our analysis is then extended in a unified manner to several time-fractional order evolution problems. Finally, several numerical experiments are conducted to confirm our theoretical findings.]{}'
author:
- 'Samir Karaa[[^1] ]{} and Amiya K. Pani[[^2]]{}'
title: Error analysis of a finite volume element method for fractional order evolution equations with nonsmooth initial data
---
[ fractional order evolution equation, subdiffusion, finite volume element method, Laplace transform, backward Euler and second-order backward difference methods, convolution quadrature, optimal error estimate, smooth and nonsmooth data. ]{}
Introduction
============
Let $\Omega$ be a bounded, convex polygonal domain in $\mathbb{R}^2$ with boundary $\partial \Omega$, $T>0,$ and let $v$ be a given function (initial data) defined on $\Omega$. We now consider the following time-fractional diffusion problem: find $u$ in $ \Omega\times (0,T] $ such that
\[main\] $$\begin{aligned}
{2}\label{a1}
&u'(x,t)+{\partial_t^{1-\alpha}}{A}u(x,t)=0 &&\quad\mbox{ in }\Omega\times (0,T],
\\ \label{a2}
&u(x,t)= 0 &&\quad\mbox{ on }\partial\Omega\times (0,T],
\\ \label{a3}
&u(x,0)=v(x) &&\quad\mbox{ in }\Omega,\end{aligned}$$
where ${A}u=-\Delta u$, $u'$ is the partial derivative of $u$ with respect to time, and ${\partial_t^{1-\alpha}}:={^R}{\rm D}^{1-\alpha}$ is the Riemann-Liouville fractional derivative in time defined for $0<\alpha<1$ by: $$\label{Ba}
{\partial_t^{1-\alpha}}\varphi(t):=\frac{d }{d t}{\mathcal{I}}^{\alpha}\varphi(t):=\frac{d }{d t}\int_0^t\omega_{\alpha}(t-s)\varphi(s)\,ds\quad\text{with} \quad
\omega_{\alpha}(t):=\frac{t^{\alpha-1}}{\Gamma(\alpha)}.$$ Here, ${\mathcal{I}}^{\alpha}$ denotes the temporal Riemann-Liouville fractional integral operator of order $\alpha$. This class of problems describes the model of an anomalous subdiffusion, see [@GMMP], [@HW] and [@MK]. Over the last two decades, considerable attention from both practical and theoretical point of views has been given to fractional diffusion models due to their various applications. Several numerical techniques for the problem have been proposed with different types of spatial discretizations. The finite element (FE) method has, in particular, been given a special attention in approximating the solution of the problem , see [@McLeanThomee2004; @MT2010; @McLeanThomee2010; @MustaphaMcLean2011; @JLZ2013; @JLPZ2015; @JLZ2016; @EJLZ2016] and references, there in. Most recently, a FVE method is analyzed in [@KMP2015] and [*a prior*]{} error estimates with respect to data regularity have been derived.
Although the numerical study of has been discussed in a large number of papers, optimal error estimates with respect to the smoothness of the solution expressed through initial data have been established only in few papers recently. This is due to the presence of time-fractional derivative, and hence, deriving sharp error bounds under reasonable regularity assumptions on the exact solution has become a challenging task.
To motivate our results, we begin by recalling some facts on the spatially semidiscrete standard Galerkin FE method for the problem in the piecewise FE element space $$V_h=\{\chi\in C^0(\overline {\Omega})\;:\;\chi|_{K}\;\mbox{is linear for all}~ K\in {\mathcal{T}}_h\;
\mbox{and} \; \chi|_{\partial \Omega}=0\},$$ where $\{{\mathcal{T}}_h\}_{0<h<1}$ is a family of regular triangulations ${\mathcal{T}}_h$ of the domain $\Omega$ into triangles $K$ with $h$ denoting the maximum diameter of the triangles $K\in {\mathcal{T}}_h$. With $a(\cdot,\cdot)$ denoting the bilinear form associated with the operator $A$, and $(\cdot,\cdot)$ the inner product in $L^2(\Omega)$, the semidiscrete Galerkin FE method is to seek $u_h(t)\in V_h$ satisfying $$\label{semi-FE}
( u_h',\chi)+ a({\partial_t^{1-\alpha}}u_h,\chi)= 0\quad
\forall \chi\in V_h,\quad t\in (0,T], \quad u_h(0)=v_h,$$ where $a(v,w):= (\nabla v, \nabla w)$ and $v_h\in V_h$ is an approximation of the initial data $v$. Upon introducing the discrete operator $A_h:V_h\rightarrow V_h$ defined by $$(A_h\psi,\chi)=(\nabla \psi,\nabla \chi) \quad \forall \psi,\chi\in V_h,$$ the semidiscrete FE scheme is rewritten in an operator form as $$\label{FEP}
u_h'(t)+{\partial_t^{1-\alpha}}A_h u_h(t)= 0, \quad t>0, \quad u_h(0)=v_h.$$ In [@MT2010], McLean and Thomée have established the following estimate for the Galerkin FE approximation to : with $v_h=P_hv$, there holds for $t>0$ $$\label{FE-1}
\|u_h(t)-u(t)\|\leq Ch^2t^{-\alpha(2-q)/2}|v|_{q}, \quad 0\leq q\leq 2,$$ where $\|v\|$ is the $L^2(\Omega)$-norm of $v$ and $|v|_{q}=\|A^{q/2}v\|$ is a weighted norm defined on the space $\dot{H}^q(\Omega)$ to be described in Section 2. Here, $P_h:L^2(\Omega)\rightarrow V_h$ is the $L^2$-projection given by : $(P_hv-v,\chi)=0$ for all $\chi\in V_h$. For a smooth initial data, that is, $v\in \dot{H}^2(\Omega)$, the estimate is still valid for the initial approximation $v_h=R_hv,$ where $R_h:H_0^1(\Omega)\rightarrow V_h$ is the standard Ritz projection defined by the relation: $a(R_hv-v,\chi)=0$ for all $\chi\in V_h$. The estimate extends results obtained for the standard parabolic problem, i.e, $\alpha=1$, which has been thoroughly studied, see [@thomee1997]. In the recent work [@EJLZ2016], an approach based on Laplace transform and semigroup type theory has been exploited to derive [*a priori*]{} error estimates of the type , and most recently, a delicate energy analysis has been developed in [@KMP2016] to obtain similar estimates.
Regarding the optimal estimate in the gradient norm, the following result $$\label{FE-2}
\|\nabla(u_h(t)-u(t))\|\leq Cht^{-\alpha(2-q)/2}|v|_{q}, \quad 0\leq q\leq 2,$$ holds with $v_h=P_hv$ on quasi-uniform meshes. For the cases $q=1,2$, one can also choose $v_h=R_hv$. However, without the quasi-uniformity assumption on the mesh, the estimate remains valid only for $0\leq q\leq 1$, see [@KMP2016].
Optimal convergence rate up to a logarithmic factor in the stronger $L^\infty(\Omega)$-norm has been derived in [@McLeanThomee2010; @KMP2016]. While in [@McLeanThomee2010], Laplace transform technique combined with semigroup type theoretic approach is used to derive maximum norm estimates, in [@KMP2016] a novel energy argument combined with Sobolev inequality for $2$D-problems is employed to establish, under quasi-uniformity assumption on the mesh, the following $L^{\infty}(\Omega)$-error estimate for $v \in \dot H^q(\Omega)\cap L^\infty(\Omega)$ and $v_h=P_h v$ $$\label{FE-3}
\|u(t)-u_h(t)\|_{L^{\infty}(\Omega)} \leq
C|\ln h|^{\frac{5}{2}} h^2 t^{-\alpha(3-q)/2}(|v|_q+\|v\|_{L^\infty(\Omega)}), \quad 1\le q\le 2.$$
In this article, we discuss the error analysis of the approximate solution $\bar{u}_h$ satisfying the following FVE method: $$\label{semi-FV}
(\bar{u}_h',\chi)_h+ a({\partial_t^{1-\alpha}}\bar{u}_h,\chi)= 0\quad
\forall \chi\in V_h,\quad t\in (0,T], \quad \bar{u}_h(0)=v_h,$$ where $(\cdot,\cdot)_h$ is a discrete inner product on $V_h$ to be defined in Section \[sec: FVM\]. Here, one of our objective is to establish the analogous of estimates (\[FE-1\]) and (\[FE-2\]) for the solution of the FVE semidiscrete problem , namely; with the appropriate choices of $v_h$, $$\label{FVE}
\|{\bar u}_h(t)-u(t)\|+h \|\nabla({\bar u}_h(t)-u(t))\|\leq Ch^2t^{-\alpha(2-q)/2}|v|_{q}, \quad 0\leq q\leq 2.$$ We shall derive this estimate for $q=1,\,2$ in Section \[sec:H2\] and for $q=0$ in Section \[sec:L2\]. For the latter case, we are only able to prove the [*a priori*]{} estimate under an additional hypothesis on ${\mathcal T}_h$, which is known to be satisfied for symmetric triangulations. Without any such condition, only sub-optimal order convergence is obtained, which is similar to the result proved in [@CLT-2013] for linear parabolic problems. For the stronger $L^\infty(\Omega)$-norm, a quasi-optimal error estimate analogous to (\[FE-3\]) is established for $1\leq q\leq 2$.
Our analysis provides improvements of earlier results in [@KMP2015], where the initial data $v$ is required to be in $\dot H^q(\Omega)$ with $q\geq 3$. Unlike the classical FE error analysis in which an intermediate projection, usually, a Ritz projection, is introduced to derive optimal error estimates, our approach, here, shall combine the error estimates for the standard Galerkin FE solution stated above with new bounds for the difference $\xi(t)=\bar u_h(t)-u_h(t)$. A similar idea has been used in [@CLT-2012] and [@CLT-2013] for the approximation of the standard parabolic problem by the lumped mass FE method and the FVE method, respectively, leading to an improvement of their earlier results in [@CLT-2004].
Our second objective is to analyze two fully discrete schemes for the semidiscrete problem based on convolution quadrature in time generated by the backward Euler and the second-order backward difference methods. Error estimates with respect to the data regularity are provided in Theorems \[thm:BE\] and \[thm:SBD\]. For instance, it is shown that the discrete solution $U_h^n$ obtained by the backward Euler method with a time step size $\tau$ satisfies the following [*a priori*]{} error estimate $$\|U_h^n-{\bar u}_h(t_n)\|\leq C(\tau^{-1+\alpha q/2}+ h^2t_n^{-\alpha(1-q/2)})|v|_{q}, \quad q=0,1,2.$$ When $q=0,$ an additional restriction on the triangulation is imposed. A similar type of error bound is shown to hold for the second-order backward difference scheme in Subsection 5.2.
Our third objective is to generalize our results on FVE method for both smooth and nonsmooth initial data to other classes of fractional order evolution equations in Section 6. Say for example, we can extend our FVE analysis to the following class of time-fractional problems: $$\label{J-alpha-class}
u'(x,t)+ {\mathcal{J}}^{\alpha} {A}u(x,t)=0 \quad\mbox{ in }\Omega\times (0,T],$$ with homogeneous Dirichlet boundary conditions and initial condition $u(x,0)= v(x)$ for $x\in \Omega.$ When ${\mathcal{J}}^{\alpha}= {\mathcal{I}}^{\alpha},$ this class of problems is known as fractional diffusion-wave equation or evolution equation with positive memory, see [@LST-1996; @MT2010] and references, therein. The case ${\mathcal{J}}^{\alpha}=
I+{\mathcal{I}}^{\alpha}$ corresponds to the PIDE with singular kernel, refer to [@MST2006]. Now if ${\mathcal{J}}^{\alpha} = I + {\partial_t^{1-\alpha}},$ then this class of problems is known as the Rayleigh-Stokes problems for generalized second grade fluid, see [@EJLZ2016]. Even our FVE analysis can be directly applied to the following time-fractional order diffusion problem: $$\label{caputo}
{{^C}\partial_t^{\alpha}}u(x,t) + {A}u(x,t)= 0,$$ where ${{^C}\partial_t^{\alpha}}v(t):={\mathcal{I}}^{1-\alpha}v'(t)$ is the fractional Caputo derivative of order $0<\alpha<1.$ For the semidiscrete FE analysis of (\[caputo\]), we refer to Jin [*et al.*]{} [@JLZ2013]. The unifying analysis of all these classes of evolution problems is based on comparing the FVE solution with the corresponding FE solution and exploiting the Laplace transform technique along with semigroup type properties of the FVE solution operator.
The rest of the paper is organized as follows. In the next section, we introduce notation, recall the solution representation for the continuous problem and some smoothing properties of the solution operator, which play an important role in our subsequent error analysis. Section 3 deals with a brief description of the spatially semidiscrete FVE scheme and their properties. In Section \[sec:error\], we derive error estimates for the semidiscrete FVE scheme for smooth and nonsmooth initial data $v\in \dot H^q$, $q=0,1,2$ in Subsections \[sec:H2\] and \[sec:L2\]. For $q=0$, i.e., $v\in L^2(\Omega)$, we show an optimal error bound under an additional assumption on the triangulation. Superconveregence result is proved in Subsection \[sec:Linfty\] and as a consequence, a quasi-optimal error estimate is established in the $L^\infty(\Omega)$-norm. In Section \[sec:discrete\], two fully discrete schemes based on convolution quadrature approximation of the fractional derivative are presented and error estimates are established. Section \[sec:extensions\] focuses on possible generalization of the present FVE error analysis to various types of time-fractional evolution problems. Finally, in Section \[sec:NE\], we present numerical results to confirm our theoretical findings.
Throughout the paper, $C$ denotes a generic positive constant that may depend on $\alpha$ and $T$, but is independent of the spatial mesh element size $h$.
Representation of exact solution and properties {#sec:notation}
================================================
We first introduce some notations. Let $\{(\lambda_j,\phi_j)\}_{j=1}^\infty$ be the Dirichlet eigenpairs of the selfadjoint and positive definite operator ${A}$, with $\{\phi_j\}_{j=1}^\infty$ being an orthonormal basis in $L^2(\Omega)$. For $r\geq 0$, we denote by $\dot H^r(\Omega)\subset L^2(\Omega)$ the Hilbert space induced by the norm $$|v|_r^2 =\|{A}^{r/2}v\|^2 =\sum_{j=1}^\infty \lambda_j^r (v,\phi_j)^2,$$ with $(\cdot,\cdot)$ being the inner product on $L^2(\Omega)$. Then, it follows that $
\dot{H}^r(\Omega)=\{\chi \in H^r(\Omega);\, {A}^j\chi=0 \text{ on } \partial \Omega, \text{ for } j<s/2\}$, see [@thomee1997 Lemma 3.1]. In particular, $|v|_0=\|v\|$ is the norm on $L^2(\Omega)$, $|v|_1=\|\nabla v\|$ is also the norm on $H_0^1(\Omega)$ and $|v|_2=\|A v\|$ is the equivalent norm in $H^2(\Omega)\cap H^1_0(\Omega)$. Note that $\{\dot H^r(\Omega)\}$, $r\geq 0$, form a Hilbert scale of interpolation spaces. Motivated by this, we denote by $\|\cdot\|_{H^r_0(\Omega)}$ the norm on the interpolation scale between $H^2(\Omega)\cap H^1_0(\Omega)$ and $L^2(\Omega)$ for $r$ in the interval $[0,2]$. Then, the $\dot H^r(\Omega)$ and $H^r_0(\Omega)$ norms are equivalent for any $r\in (1/2,2]$ for $r\in [0,1/2],$ $\dot H^r(\Omega)= H^r(\Omega)$ by interpolation.
For $\delta>0$ and $\theta\in (\pi/2,\pi)$, we introduce the contour $\Gamma_{\theta,\delta}\subset \mathbb{C}$ defined by $$\Gamma_{\theta,\delta}=\{\rho e^{\pm i\theta}:\rho\geq \delta\}\cup\{\delta e^{i\psi}: |\psi|\leq \theta\},$$ oriented with an increasing imaginary part. Further, we denote by $\Sigma_{\theta}$ the sector $$\Sigma_{\theta}=\{z\in \mathbb{C}, \,z\neq 0,\, |\arg z|< \theta\}.$$ For $z\in \Sigma_\theta$, it is clear that $z^\alpha \in \Sigma_\theta$ as $\alpha \in (0,1)$. Since the operator $A$ is selfadjoint and positive definite, its resolvent $(z^\alpha I+A)^{-1}:L^2(\Omega)\rightarrow L^2(\Omega)$ satisfies the bound $$\label{res1}
\|(z^\alpha I+A)^{-1}\|\leq M_\theta |z|^{-\alpha} \quad \forall z\in \Sigma_\theta,$$ where $M_\theta=1/\sin(\pi-\theta)$. We now make use of the Laplace transform $\hat{u}:=\mathcal{L}(u)$ of the solution $u$ defined by $$\hat{u}(z,x)=\int_0^\infty e^{-zt}u(t,x)\,dt.$$ The boundary condition $u(x,t)=0$ on $\partial \Omega$ transforms into $\hat u(x,z)=0$ on $\partial \Omega$. Taking Laplace transforms in (\[a1\]), we, then, arrive at $$\label{m1}
(zI+z^{1-\alpha}A)\hat{u}(z)=v,$$ and hence, $$\label{m2}
\hat{u}(z)= \hat{E}(z)v, \quad \hat{E}(z)= z^{\alpha-1}(z^{\alpha}I+A)^{-1}.$$ In view of (\[res1\]) and (\[m2\]), $\hat{E}(z)$ satisfies the following bound $$\label{m3}
\|\hat{E}(z)\|\leq M_\theta |z|^{-1} \quad \forall z\in \Sigma_\theta.
$$ From , the Laplace inversion formula yields an integral representation for the solution of (\[main\]) as $$\label{mm}
u(t)=\frac{1}{2\pi i}\int_{\mathcal C} e^{zt}\hat{E}(z)v\,dz, \quad t>0,
$$ where the contour of integration ${\mathcal C}$, known as Bromwich contour, is any line in the right-half plane parallel to the imaginary axis and with Im$z$ increasing. Since $\hat{E}(z)$ is analytic in $\Sigma_\theta$ and satisfies the bound , the path of integration may, therefore, be deformed into the curve $\Gamma_{\theta,\delta}$ so that the integrand has an exponential decay property.
In the next lemma, we present some smoothing properties of the operator $\hat{E}(z)$ which play a key role in our error analysis. The estimates are proved for instance in [@Lubich-2006 Lemma 2.2]. Note that the first estimate given below is obtained by interpolation technique.
\[lem:Ah\] The following estimates hold: $$\label{00}
\|A\hat{E}(z)\chi\|\leq C_{\theta} |z|^{\alpha(1-p/2)-1} |\chi|_p \quad \forall z\in\Sigma_\theta,\quad 0\leq p\leq 2,$$ $$\label{11}
\|\nabla\hat{E}(z)\chi\|\leq C_{\theta} |z|^{\alpha/2-1} \|\chi\| \quad \forall z\in\Sigma_\theta,$$ where $C_{\theta}$ depends only on $\theta$.
In the next section, we introduce the semidiscrete finite volume element scheme.
Semidiscrete FVE scheme and its properties {#sec: FVM}
==========================================
To describe the finite volume element formulation, we first introduce the dual mesh on the domain $\Omega$. Let $N_h$ be the set of nodes or vertices, that is, $$N_h :=\left\{P_i:P_i~~\mbox{ is a vertex of the element }~K \in
{\mathcal{T}}_h~\mbox{and}~P_i\in \overline{\Omega}\right\}$$ and let $N_h^0$ be the set of interior nodes in ${\mathcal{T}}_h.$ Further, let ${\mathcal{T}}_h^*$ be the dual mesh associated with the primary mesh ${\mathcal{T}}_h,$ which is defined as follows. With $P_0$ as an interior node of the triangulation ${\mathcal{T}}_h,$ let $P_i\;(i=1,2\cdots
m)$ be its adjacent nodes (see, Figure \[fig:mesh\] with $m=6$ ). Let $M_i,~i=1,2\cdots
m$ denote the midpoints of $\overline{P_0P_i}$ and let $Q_i,~i=1,2\cdots
m,$ be the barycenters of the triangle $\triangle P_0P_iP_{i+1}$ with $P_{m+1}=P_1$. The [*control volume*]{} $K_{P_0}^*$ is constructed by joining successively $ M_1,~ Q_1,\cdots
,~ M_m,~ Q_m,~ M_1$. With $Q_i ~(i=1,2\cdots
m)$ as the nodes of $control~volume~$ $K^*_{p_i},$ let $N_h^*$ be the set of all dual nodes $Q_i$. For a boundary node $P_1$, the control volume $K_{P_1}^*$ is shown in Figure \[fig:mesh\]. Note that the union of the control volumes forms a partition ${\mathcal{T}}_h^*$ of $\overline{\Omega}$.
![Control volume for interior node[]{data-label="fig:mesh"}](figure1.eps){width="8.0cm" height="4.0cm"}
The dual volume element space ${V_h^*}$ on the dual mesh ${\mathcal{T}}^*_h$ is defined as $${V_h^*}=\{\chi\in L^2(\Omega)\;:\;\chi|_{K_{P_0}^*}\;\mbox{is constant for all}\; K_{P_0}^*\in {\mathcal{T}}_h^*\; \mbox{and}\; \chi|_{\partial \Omega}=0\}.$$ The semidiscrete FVE formulation for is to seek $\bar u_h(t) \in V_h$ such that $$\label{FV}
(\bar u_h',\chi)+ a_h({\partial_t^{1-\alpha}}\bar u_h,\chi)=0 \quad
\forall \chi\in {V_h^*},\quad t>0, \quad \bar u_h(0)=v_h,$$ where the bilinear form $a_h(\cdot, \cdot): V_h\times {V_h^*}\longrightarrow \R$ is defined by $$\label{Ah}
a_h(\psi,\chi)= -\sum_{P_i\in N_h^0} \chi(P_i) \int_{\partial K_{P_i}^*}
\nabla \psi\cdot{\bf n}\,ds\quad \forall \psi \in V_h,\; \chi\in {V_h^*}$$ with ${\bf n}$ denoting the outward unit normal to the boundary of the control volume $K_{P_i}^*$. For $w\in H^2(\Omega)$ and $\chi\in {V_h^*}$, a use of Green’s formula yields $$\label{eq: h2 identity}
({A}w, \chi)= a_h(w,\chi).$$ To rewrite the Petrov-Galerkin method (\[FV\]) as a Galerkin method in $V_h$, we introduce the interpolation operator $\Pi_h^*:C^0(\bar{\Omega})\longrightarrow {V_h^*}$ by $$\label{naa}
\Pi_h^*\chi=\sum_{P_i\in N_h^0}\chi(P_i)\eta_i(x),$$ where $\eta_i$ is the characteristic function of the control volume $K_{P_i}^*$. The operator $\Pi_h^*$ is selfadjoint and positive definite, see [@ChouLi2000], and hence, the following relation $$\label{inner-1}
(\psi,\chi)_h=(\psi,\Pi_h^*\chi)\quad \forall \psi, \chi\in V_h$$ defines an inner product on $V_h$. Also, the corresponding norm $(\chi,\chi)_h^{1/2}$ is equivalent to the $L^2(\Omega)$-norm on $V_h$, uniformly in $h$, see [@LiChenWu2000]. Furthermore, from the following identity [@BR-1987; @EwingLazarovLin2000] $$a_h(\chi,\Pi_h^* v)=(\nabla \chi,\nabla v)\quad \forall \chi, v\in V_h,$$ the bilinear form $a_h(.,.)$ is symmetric and $a_h(\chi,\Pi_h^*\chi)=\|\nabla\chi\|^2$ for $\chi\in V_h$.
We now introduce the discrete operator $\bar A_h:V_h\rightarrow V_h$ corresponding to the inner product $(\cdot,\cdot)_h$ by $$\label{Dh}
(\bar A_h\psi,\chi)_h=(\nabla \psi,\nabla \chi) \quad \forall \psi,\chi\in V_h.$$ Then, the FVE method is written in an operator form as $$\label{FVP}
\bar u_h'(t)+{\partial_t^{1-\alpha}}\bar A_h \bar u_h(t)= 0, \quad t>0, \quad \bar u_h(0)=v_h.$$ An appropriate modification of arguments in [@CLT-2013; @JLZ2013] yields the following discrete analogous of Lemma \[lem:Ah\] and therefore, we skip the proof.
\[lem:Ah2\] Let $\hat E_h(z)=z^{\alpha-1}(z^\alpha I+\bar A_h)^{-1}$. With $\chi\in V_h$, the following estimates hold: $$\label{0}
\|{\bar A_h\hat{E}_h(z)\chi}\|\leq C_{\theta} |z|^{\alpha(1-p/2)-1} \;\|\bar {A}_h^{p/2}{\chi}\|
\quad \forall z\in \Sigma_\theta,\quad 0\leq p\leq 2,$$ $$\label{1}
|{\hat{E}_h(z)\chi}|_1\leq C_{\theta} |z|^{\alpha/2-1} \|{\chi}\|\quad \forall z\in \Sigma_\theta,$$ where $C_{\theta}$ is independent of the mesh size $h$.
Moreover, an analogous of Lemma \[lem:Ah2\] holds for $\hat F_h(z)=z^{\alpha-1}(z^\alpha I+A_h)^{-1}$, when we replace ${\hat{E}}_h(z)$ in Lemma \[lem:Ah2\] by $\hat F_h(z).$
Error analysis {#sec:error}
==============
This section deals with [*a priori*]{} optimal error estimates for the semidiscrete FVE scheme with initial data $v\in \dot{H}^q(\Omega)$, $q=0,1,2$. To do so, we first introduce the quadrature error $Q_h:V_h\rightarrow V_h$ defined by $$\label{m8}
(\nabla Q_h\chi,\nabla\psi)=\epsilon_h(\chi,\psi):=(\chi,\psi)_h-(\chi,\psi)\quad \forall \psi \in V_h.$$ The operator $Q_h$, introduced in [@CLT-2012] for the lumped mass FE element, represents the quadrature error in a special way. It satisfies the following error estimates, see [@CLT-2012; @CLT-2013].
\[lem:Qh\] Let $Q_h$ be defined by . Then, there holds $$\label{QQ}
\|\nabla Q_h\chi\|+h\|\bar{A}_hQ_h\chi\|\leq Ch^{p+1}\|\nabla^{p}\chi\|\quad \forall \chi\in V_h, \quad p=0,1.$$
Note that, by Lemma \[lem:Qh\], and without additional assumptions on the mesh, the following estimate holds: $$\|Q_h\chi\| \leq C \|\nabla Q_h\chi\| \leq Ch \|\chi\|\quad \forall \chi\in V_h.$$ This estimate cannot be improved in general, see [@CLT-2012; @CLT-2013] for some counter examples. However, on some special meshes, one can derive a better approximation. For instance, if the mesh is symmetric (see [@CLT-2012; @CLT-2013] for the definition and examples), the operator $Q_h$ is shown to satisfy $$\label{sym}
\|Q_h\chi\|\leq Ch^2 \|\chi\| \quad \forall \chi\in V_h.$$ To derive optimal error estimates for the FVE solution $\bar u_h$, we split the error $\bar{e}(t):=\bar{u}_h(t)-u(t)$ into $\bar{e}(t):=(u_h(t)-u(t))+\xi(t)$, where $\xi(t)=\bar{u}_h(t)-u_h(t)$ and $u_h$ being the standard Galerkin FE solution. Then, from the definitions of $u_h(t)$, $\bar{u}_h(t)$ and $Q_h$, $\xi(t)$ satisfies $$\label{m10}
\xi_t(t)+\partial_t^{1-\alpha}\bar{A}_h\xi(t)=-\bar{A}_hQ_hu_{ht}(t),\quad t>0,\quad \xi(0)=0.$$
Error estimates for smooth initial data {#sec:H2}
---------------------------------------
In the following theorem, optimal error estimates are derived for smooth initial data $v\in \dot{H}^q(\Omega)$ with $q\in [1,2].$
\[thm:H2\] Let $u$ and $\bar{u}_h$ be the solutions of $\eqref{main}$ and $\eqref{semi-FV}$, respectively, with $v\in \dot{H}^q(\Omega)$ for $q\in [1,2]$ and $v_h=R_h v$. Then, there is a positive constant $C$, independent of $h,$ such that $$\label{e1}
\|\bar{u}_h(t)-u(t)\|+h\|\nabla(\bar{u}_h(t)-u(t))\|\leq C\;t^{-\alpha(2-q)/2}\;h^2|v|_q,\qquad t>0.$$
Since the estimates for $u_h-u$ are given in (\[FE-1\]) and (\[FE-2\]), it is sufficient to show $$\label{0e2}
\|\xi(t)\|+h\|\nabla\xi(t)\|\leq C \;t^{-\alpha(2-q)/2}\; h^2 |v|_q,\;\; q\in [1,2].$$ By taking Laplace transforms in and following the analysis in Section \[sec:notation\], we represent $\xi(t)$ by $$\label{m013}
\xi(t)=-\frac{1}{2\pi i}\int_{\Gamma} e^{zt}\hat{E}_h(z)\bar{A}_hQ_h\widehat{u_{ht}}(z)\,dz.$$ Here and also throughout this article, $\Gamma$ is the particular contour chosen as $\Gamma = \Gamma_{\theta,\delta}$ with $\delta=1/t$. From , it follows that $$\begin{aligned}
\label{xi-estimate-1}
\|\xi(t)\| + h \|\nabla \xi(t)\| \leq \frac{1}{2\pi}\int_{\Gamma} |e^{zt}|
\Big(\|\hat{E}_h(z)\bar{A}_hQ_h\widehat{u_{ht}}(z)\| + h \|\nabla \hat{E}_h(z)\bar{A}_hQ_h\widehat{u_{ht}}(z)\|
\Big) \,|dz|.\end{aligned}$$ To complete the proof of the estimate, we need to compute the terms under the integral sign on the right of side of . Now, we discuss two cases for $q=2$ and $q=1$ separately.
When $q=2$, that is, $v\in \dot{H}^2(\Omega),$ apply (\[0\]) with $p=1$ and (\[1\]) in Lemma \[lem:Ah2\] to obtain $$\label{m014}
\|\hat{E}_h(z)\bar{A}_hQ_h\widehat{u_{ht}}(z)\|\leq C |z|^{\alpha/2-1}\|\nabla Q_h \widehat{u_{ht}}(z)\|,$$ and $$\label{g2}
\|\nabla \hat{E}_h(z)\bar{A}_hQ_h\widehat{u_{ht}}(z)\|\leq C |z|^{\alpha/2-1}\|\bar A_hQ_h\widehat{u_{ht}}(z)\|.$$ Then, by (\[QQ\]), it follows that $$\label{Q1aa}
\|\hat{E}_h(z)\bar{A}_hQ_h\widehat{u_{ht}}(z)\|+h\|\nabla \hat{E}_h(z)\bar{A}_hQ_h\widehat{u_{ht}}(z)\|\leq Ch^2
|z|^{\alpha/2-1}\|\nabla \widehat{u_{ht}}(z)\|.$$ Since $$\widehat{u_{ht}}(z)=-z^{1-\alpha}A_h\hat{{u}}_{h}(z)= -z^{1-\alpha}A_h\hat{F}_{h}(z)v_h,$$ an estimate analogous to (\[1\]) yields $$\label{Q1}
\|\nabla \widehat{u_{ht}}(z)\|=|z|^{1-\alpha}\|\nabla \hat{F}_{h}(z)A_hv_h\|\leq C |z|^{1-\alpha}|\,|z|^{\alpha/2-1}
\|A_h v_h\|\leq C|z|^{-\alpha/2} \|A_h v_h\|.$$ On substitution of (\[Q1aa\]) and (\[Q1\]) in , we use (\[m013\]) to obtain $$\begin{aligned}
\label{xi-estimate-2}
\|\xi(t)\|+h\|\nabla\xi(t)\|&\leq & Ch^2 \left(\int_{\Gamma} |e^{zt}|\, |z|^{-1}\,|dz|\right)\|A_h v_h\| \nonumber \\
&\leq & Ch^2 \left(\int_{1/t}^\infty e^{\rho t\cos\theta}\rho^{-1}d\rho
+\int_{-\theta}^{\theta}e^{\cos\psi}d\psi\right)\|A_h v_h\| \nonumber \\
&\leq & C h^2 \|A_hv_h\|.\end{aligned}$$ Now, by the identity ${A}_hR_h=P_hA$, we have $$\|A_h R_h v\| = \|P_hA v\| \leq \|A v\|=|v|_2,$$ which shows the estimate for $q=2.$ For the case $q=1$, that is, $v\in \dot{H}^1(\Omega),$ consider (\[Q1aa\]) and the identity $$\widehat{u_{ht}}(z)=z\hat{u}_{h}(z)-v_h$$ to obtain using (\[m3\]) $$\label{m0156}
\|\nabla \hat{u}_{ht}(z)\|=\|\nabla(z\hat{F}_{h}(z)v_h-v_h)\|\leq (M+1) \|\nabla v_h\|.
$$ From the estimate , using and with $\|\nabla v_h\|=\|\nabla R_h v\|\leq\|\nabla v\|,$ we deduce that $$\begin{aligned}
\|\xi(t)\|+h\|\nabla\xi(t)\|&\leq & Ch^2 \left(\int_{\Gamma} |e^{zt} |z|^{\alpha/2-1}\,|dz|\right)|v|_1 \\
&\leq & Ch^2 \left(\int_{1/t}^\infty e^{\rho t\cos\theta}\rho^{\alpha/2-1}d\rho+\int_{-\theta}^{\theta}e^{\cos\psi}t^{-\alpha/2}d\psi\right)|v|_1 \\
&\leq & Ct^{-\alpha/2}h^2 |v|_1.\end{aligned}$$ This completes the proof for the case $q=1.$
Since estimates for $q=1$ and $q=2$ are known, then interpolation technique provides result for $q\in [1,2].$ This concludes the rest of the proof.
\[r2\] Note that the estimate in Theorem \[thm:H2\] remains valid when $v_h= P_h v$. Indeed, for $q=2$, let $\tilde{u}_h$ denote the solution of with $v_h=P_h v.$ Then $\zeta:=\tilde{u}_h-\bar{u}_h$ satisfies $$\label{e7}
\zeta_t+\partial_t^{1-\alpha}\bar{A}_h\zeta=0,\quad t>0,\quad \zeta(0)=P_h v-R_h v.$$ Since $$\zeta(t)=-\frac{1}{2 \pi i} \int_{\Gamma} e^{zt}{\hat{E}}_h(z)(P_h v-R_h v) \; dz,$$ we deduce $$\|\zeta(t)\|\leq C\;\| P_h v-R_h v\| \int_{\Gamma} |e^{zt}|\; |z|^{-1} |d\;z|\leq C h^2 | v|_2.$$ Thus, the estimate with $q=2$ follows by the triangle inequality. If the inverse inequality $\|\nabla \chi\|\leq Ch^{-1}\|\chi\|$ holds, which is the case if the mesh is quasi-uniform, then the estimate in the gradient norm follows directly for $v_h=P_h v$. If the $L^2(\Omega)$-projection operator $P_h$ is stable in $\dot H^1(\Omega)$, i.e., $\|\nabla P_hw\|\leq C |w|_1$, then the estimate holds for the case $q=1$ and the choice $v_h=P_hv$. A sufficient condition for such stability of $P_h$ is the quasi-uniformity of the mesh. Now, by interpolation the estimate holds for $q\in [1,2]$ and $v_h=P_hv$.
Error estimates for nonsmooth initial data {#sec:L2}
------------------------------------------
In this subsection, we establish optimal error estimates for the semidiscrete FVE scheme (\[semi-FV\]) for nonsmooth initial data $v\in L^2(\Omega)$.
\[thm:L2\] Let $u$ and $\bar{u}_h$ be the solution of $(\ref{main})$ and $(\ref{semi-FV})$, respectively, with $v\in L^2(\Omega)$ and $v_h =P_hv$. Then, there exists a positive constant $C$, independent of $h,$ such that $$\label{m12a}
\|\bar{u}_h(t)-u(t)\| + \|\nabla(\bar{u}_h(t)-u(t))\|\leq Cht^{-\alpha}\|v\|, \quad t>0.$$ Furthermore, if the quadrature error operator $Q_h$ satisfies , then the following optimal error estimate holds: $$\label{m12b}
\|\bar{u}_h(t)-u(t)\| \leq Ch^2t^{-\alpha}\|v\|, \quad t>0.$$
As before, it is sufficient to prove estimates for $\xi$. We first apply (\[0\]) with $p=0$ to arrive at $$\|\hat{E}_h(z)\bar{A}_hQ_h\widehat{u_{ht}}\|\leq C|z|^{\alpha-1}\|Q_h\widehat{u_{ht}}\|.$$ Then, the following bound follows from the integral representation (\[m013\]): $$\label{0m13}
\|\xi(t)\|\leq C\int_{\Gamma} |e^{zt}||z|^{\alpha-1}\|Q_h\widehat{u_{ht}}(z)\|\,|dz|.$$ To estimate the gradient of $\xi$, we note that $$\|\nabla \hat{E}_h(z)\bar{A}_hQ_h\widehat{u_{ht}}\|\leq C|z|^{\alpha-1}\|\nabla Q_h\widehat{u_{ht}} \|,$$ and hence, $$\label{0m14}
\|\nabla\xi(t)\|\leq C\int_{\Gamma} |e^{zt}||z|^{\alpha-1}\|\nabla Q_h\widehat{u_{ht}}(z)\|\,|dz|.$$ Note that $\|Q_h\widehat{u_{ht}}\|\leq Ch \|\widehat{u_{ht}} \|$ holds on a general mesh, and $\|\nabla Q_h\widehat{u_{ht}}\|\leq Ch \|\widehat{u_{ht}} \|$ by . Since $\|\widehat{u_{ht}}(z)\|=\|z\hat{F}_{h}(z)v_h-v_h\|\leq C \|v_h\|$ by (\[m3\]), a substitution into and yields the first estimate . Finally, if (\[sym\]) holds, then follows immediately from , which completes the proof.
$L^\infty(\Omega)$-error estimates {#sec:Linfty}
----------------------------------
In the following, we obtain a superconvergence result for the gradient of $\xi$ in the $L^2(\Omega)$-norm. As a consequence, assuming $v\in L^{\infty}(\Omega)$ and the quasi-uniformity on the mesh, a quasi-optimal error estimate in the stronger $L^\infty(\Omega)$-norm is derived for the semidiscrete FVE solution $\bar u_h$. We first prove the following Lemma by refining some of the estimates derived in the proofs of Theorem \[thm:H2\].
\[thm:Linfty-1\] For $1\leq q\leq 2$, and with $v_h =R_hv$, there is a positive constant $C,$ independent of $h,$ such that $$\label{12a}
\|\nabla \xi(t)\|\leq Ch^2t^{-\alpha(3-q)/2}|v|_q, \quad t>0.$$ The estimate is still valid for $v_h =P_hv,$ but with quasi-uniform assumption on the mesh.
By using bounds and , we obtain instead of the following estimate $$\|\nabla \hat{E}_h(z)\bar{A}_hQ_h\widehat{u_{ht}}(z)\|\leq C |z|^{\alpha-1}\|\nabla Q_h\widehat{u_{ht}}(z)\|
\leq C h^2 |z|^{\alpha-1}\|\nabla \widehat{u_{ht}}(z)\|.$$ Since $\|\nabla\widehat{u_{ht}}(z)\|\leq c|z|^{-\alpha/2}\|A_hv_h\|$ by , we note from the representation that $$\label{0gg2}
\|\nabla\xi(t)\| \leq Ch^2 |v|_2 \int_{\Gamma} |e^{zt} |z|^{\alpha/2-1}\,|dz|\leq Ct^{-\alpha/2}h^2 |v|_2.$$ Similarly, taking into account , we obtain $$\label{0gg2-b}
\|\nabla\xi(t)\| \leq Ch^2 |v|_1 \int_{\Gamma} |e^{zt} |z|^{\alpha-1}\,|dz|\leq Ct^{-\alpha}h^2 |v|_1.$$ Now, the desired estimate for $q\in [1,2]$ follows by interpolation which completes the proof.
Note that for 2D-problems, the Sobolev inequality $$\|\chi\|_{L^{\infty}(\Omega)} \leq C\; |\ln h| \; \|\nabla\chi\| \;\;\;\forall \chi\in V_h,$$ and Lemma 4.2 imply for $q\in [1,2]$ that $$\label{max-norm}
\|\xi(t)\|_{L^{\infty}(\Omega)} \leq C\,|\ln h| \, \|\nabla\xi(t)\| \leq
C|\ln h| \,h^2t^{-\alpha(3-q)/2}|v|_q.
$$ As a consequence, we obtain the following quasi-optimal $L^\infty(\Omega)$-error estimate by combining the results in and .
\[thm:Linfty\] Let $u$ and $\bar{u}_h$ be the solution of $(\ref{main})$ and $(\ref{semi-FV})$, respectively, with $v_h =P_hv$. Assume that $v \in \dot H^q(\Omega)\cap L^\infty(\Omega)$ for $1\le q\le 2$. Then, under the quasi-uniformity condition on the mesh, there holds $$\label{FV-3}
\|\bar u_h(t) - u(t)\|_{L^{\infty}(\Omega)} \leq
C|\ln h|^{\frac{5}{2}} h^2 t^{-\alpha(3-q)/2}\Big(|v|_q+\|v\|_{L^\infty(\Omega)}\Big),\qquad 1\le q\le 2.$$
Fully discrete schemes {#sec:discrete}
======================
In this section, we analyze two fully discrete schemes for the semidiscrete problem using the framework of convolution quadrature developed in [@LST-1996; @Lubich-2006], which has been initiated in [@Lubich-1986; @Lubich-1988]. To describe this framework, we first divide the time interval $[0,T]$ into $N$ equal subintervals with a time step size $\tau=T/N$, and let $t_j=j\tau$. Then, the convolution quadrature [@Lubich-1986] refers to an approximation of any function of the form $k\ast\varphi$ as $$(k\ast\varphi) (t_n):=\int_0^{t_n}k(t_n-s)\varphi(s)\,ds\approx \sum_{j=0}^n \beta_{n-j}(\tau) \varphi(t_j),$$ where the convolution weights $\beta_j=\beta_j(\tau)$ are computed from the Laplace transform $\hat{k}(z)$ of $k$ rather than the kernel $k(t)$. This method provides, in particular, an interesting tool for approximating the Riemann-Liouville fractional integral of order $\alpha$, $\partial_t^{-\alpha}\varphi := \omega_\alpha\ast \varphi$, where $\omega_\alpha(t)=t^{\alpha-1}/\Gamma(\alpha)$. Here, $\hat{k}(z)=\hat{\omega}_\alpha(z)=z^{-\alpha}$.
With $\partial_t$ being time differentiation, we define $\hat{k}(\partial_t)$ as the operator of (distributional) convolution with the kernel $k$: $\hat{k}(\partial_t)\varphi=k\ast \varphi$ for a function $\varphi(t)$ with suitable smoothness. A convolution quadrature approximates $\hat{k}(\partial_t)\varphi$ by a discrete convolution $\hat{k}(\bar \partial_\tau)\varphi$ at $t=t_n$ as $$\hat{k}(\bar \partial_\tau)\varphi(t_n) = \sum_{j=0}^n \beta_{n-j}(\tau) \varphi(t_j),$$ where the quadrature weights $\{\beta_j(\tau)\}_{j=0}^{\infty}$ are determined by the generating power series $$\sum_{j=0}^\infty \beta_j(\tau) \xi^j=\hat{k}(\delta(\xi)/\tau)$$ with $\delta(\xi)$ being a rational function, chosen as the quotient of the generating polynomials of a stable and consistent linear multistep method. In this paper, we consider the Backward Euler (BE) and the second-order backward difference (SBD) methods, for which $\delta(\xi)=1-\xi$ and $\delta(\xi)=(1-\xi)+ (1-\xi)^2/2$, respectively. For the BE method, the convolution quadrature formula for approximating the fractional integral $\partial_t^{-\alpha}\varphi$ is given by $$\bar \partial_\tau^{-\alpha}\varphi(t_n) = \sum_{j=0}^n \beta_{n-j} \varphi(t_j), \text{ where }
\sum_{j=0}^\infty \beta_{j} \xi^j=[(1-\xi)/\tau]^{-\alpha}, \quad \beta_j=\tau^\alpha(-1)^{j}
\left(\begin{array}{c}
-\alpha\\
j
\end{array}\right),$$ while for the SBD method, the quadrature weights are provided by the formula [@Lubich-1986]: $$\beta_j=\tau^\alpha(-1)^{j}\left(\frac{2}{3}\right)^\alpha\sum_{l=0}^j
3^{-l}
\left(\begin{array}{c}-\alpha\\j-l\end{array}\right)
\left(\begin{array}{c}-\alpha\\l\end{array}\right).$$ An important property of the convolution quadrature is that it maintains some relations of the continuous convolution. For instance, the associativity of convolution is valid for the convolution quadrature [@Lubich-2004] such as $$\label{s1}
\hat{k}_1(\bar \partial_\tau) \hat{k}_2(\bar \partial_\tau)=\hat{k}_1 \hat{k}_2(\bar \partial_\tau)\quad
\text{ and }\quad
\hat{k}_1(\bar \partial_\tau)(k\ast\varphi)=(\hat{k}_1(\bar \partial_\tau)k)\ast\varphi.$$
In the following lemma, we state an interesting result on the error of the convolution quadrature, see [@Lubich-1988 Theorem 4.1] and [@Lubich-2004 Theorem 2.2].
\[lem:Lubich\] Let $G(z)$ be analytic in the sector $\Sigma_\theta$ and such that $$\label{s3}
\|G(z)\|\leq M|z|^{-\mu}\quad \forall z\in \Sigma_\theta,$$ for some real $\mu$ and $M$. Assume that the linear multistep method is strongly $A$-stable and of order $p\geq 1$. Then, for $\varphi(t)=ct^{\nu-1}$, the convolution quadrature satisfies $$\label{s4}
\|G(\partial_t)\varphi(t) - G(\bar \partial_\tau)\varphi(t)\| \leq \left\{
\begin{array}{ll}
C t^{\mu-1+\nu-p} \tau^p, & \nu\geq p\\
C t^{\mu-1} \tau^\nu, & 0< \nu\leq p.
\end{array} \right.$$
Error analysis for the BE method
--------------------------------
In this subsection, we specify the construction of a fully discrete scheme based on the BE method for the semidiscrete problem . Then, we derive $L^2(\Omega)$-error estimates for smooth and nonsmooth initial data.
After integrating in time from $0$ to $t$, the semidiscrete scheme takes the form $$\label{s5}
\bar{u}_h+\partial_t^{-\alpha} \bar A_h\bar{u}_h=v_h.$$ The second term on the left-hand side is a convolution, and then, it can be approximated at $t_n=n\tau$ with $U_h^n$ by $$\label{s6}
U_h^n+\bar\partial_\tau^{-\alpha} \bar A_h U_h^n=v_h.$$ The symbol $\bar\partial_\tau^{-\alpha}$ refers to the relevant convolution quadrature generated by the BE method.
Thus, with $U_h^0=v_h$, the fully discrete solution can be represented by $$\label{BE}
U_h^n=\left(I+\beta_0\bar A_h\right)^{-1} \left(U_h^0-\sum_{j=0}^{n-1}\beta_{n-j}\bar A_h U^j\right) \quad \mbox{ for } n\geq 1.$$ We notice that the term corresponding to $j=0$ in the formula can be omitted without affecting the convergence rate of the scheme [@LST-1996].
In view of and , we can write the error $U^n_h-\bar{u}_h(t_n)$ at $t=t_n$ as $$\label{s7}
U^n_h-\bar{u}_h(t_n)= \left( G(\bar\partial_\tau)- G(\partial_t)\right)v_h,$$ where $G(z)=(I+z^{-\alpha}\bar A_h)^{-1}$. Using the identity $$(I+z^{-\alpha}\bar A_h)^{-1} = I-(z^{\alpha}I+\bar A_h)^{-1}\bar A_h,$$ and denoting $\bar G(z)=-(z^{\alpha}I+\bar A_h)^{-1}$, the error can be represented as $$\label{s8}
U^n_h-\bar{u}_h(t_n)= \left( \bar G(\bar\partial_\tau)- \bar G(\partial_t)\right) \bar A_hv_h.$$ Using Lemma \[lem:Lubich\], we now derive the following error estimates.
\[lem:BE\] Let $\bar{u}_h$ and $U^n_h$ be the solutions of problems $(\ref{semi-FV})$ and $(\ref{s6})$, respectively, with $U^0_h=v_h$. Then, the following estimates hold:
\(a) If $v\in \dot{H}^2(\Omega)$ and $v_h =R_hv$, then $$\label{s10}
\|U^n_h-\bar{u}_h(t_n)\|\leq C \tau t_n^{\alpha-1}|v|_2.$$
\(b) If $v\in L^2(\Omega)$ and $v_h =P_hv$, then $$\label{s9}
\|U^n_h-\bar{u}_h(t_n)\|\leq C \tau t_n^{-1}\|v\|.$$
For the estimate , we recall that, by , $\|\bar G(z)\|\leq M_\theta |z|^{-\alpha}\; \forall z\in \Sigma_\theta.$ An application of Lemma \[lem:Lubich\] (with $\mu=\alpha$, $\nu=1$ and $p=1$) to yields $$\|U^n_h-\bar{u}_h(t_n)\|\leq C \tau t_n^{\alpha-1}\|\bar A_hv_h\|.$$ Now, we introduce a projection operator $\bar P_h:L^2(\Omega)\rightarrow V_h$ defined by $$(\bar P_hw,\chi)_h=(w,\chi) \quad \forall \chi\in V_h.$$ Then, $\bar P_h$ is stable in $L^2(\Omega)$ and the identity $\bar A_hR_h=\bar P_hA$ holds, since $$(\bar A_h R_h w,\chi)_h=(\nabla R_h w,\nabla \chi)=(\nabla w, \nabla\chi)=(A w,\chi)=(\bar P_h A w,\chi)_h \quad \forall \chi \in V_h.$$ As $v_h=R_h v$, it follows that $$\|\bar A_h v_h\| =\|\bar A_h R_h v\| = \|\bar P_h A v\| \leq C \|A v\|=C |v|_2,$$ which shows .
For the estimate , we notice that $ \|{G}(z)\|=|z|^\alpha\|(z^{\alpha}I+\bar A_h)^{-1}\|\leq M_\theta\; \forall z\in \Sigma_\theta.$ Then, by applying Lemma \[lem:Lubich\] (with $\mu=0$, $\nu=1$ and $p=1$) to , we obtain $$\|U^n_h-\bar{u}_h(t_n)\|\leq C \tau t_n^{-1}\|v_h\|.$$ Now, the estimate follows from the $L^2(\Omega)$-stability of $P_h$. This completes the rest of the proof.
\[rem:BE\] For $v\in \dot{H}^2(\Omega)$, we can choose $v_h =P_hv$. Let $\tilde U_h^n$ be the solution of the fully discrete scheme with $v_h =P_hv$. Then, by the stability of the scheme, a direct consequence of Lemma \[lem:BE\], we have $\|U^n_h-\tilde U^n_h\|\leq \|R_hv-P_hv\|\leq C h^2|v|_2,$ showing that $$\label{s11}
\|U^n_h-\bar{u}_h(t_n)\|\leq C (\tau t_n^{\alpha-1}+h^2)|v|_2.$$ Hence, by interpolating and it follows that for $v_h =P_hv$, $$\label{s11a}
\|U^n_h-\bar{u}_h(t_n)\|\leq C (\tau t_n^{-1})^{1/2} (\tau t_n^{\alpha-1}+h^2)^{1/2}|v|_1.$$
As a consequence of Lemma \[lem:BE\], we obtain error estimates for the fully discrete scheme with smooth and nonsmooth initial data.
\[thm:BE\] Let $u$ and $U^n_h$ be the solutions of problems $(\ref{main})$ and $(\ref{s6})$, respectively, with $U^0_h=v_h$. Then, the following error estimates hold:
\(a) If $v\in \dot{H}^2(\Omega)$ and $v_h =R_hv$, then $$\label{s12}
\|U^n_h-u(t_n)\|\leq C (h^2+\tau t_n^{\alpha-1})|v|_2.$$
\(b) If $v\in \dot{H}^1(\Omega)$, $v_h =P_hv$ and the mesh is quasi-uniform, then $$\label{s13}
\|U^n_h-u(t_n)\|\leq C (h^2t_n^{-\alpha/2}+\tau t_n^{-1+\alpha/2})|v|_1.$$
\(c) If $v\in L^2(\Omega)$, $v_h =P_hv$ and $Q_h$ satisfies , then $$\label{s14}
\|U^n_h-u(t_n)\|\leq C (h^2t_n^{-\alpha}+\tau t_n^{-1})\|v\|.$$
The first estimate follows from , and the triangle inequality, while the third estimate follows from and . By combining (with $q=1$) which holds for $v_h =P_hv$ and , we deduce $$\label{s15}
\|U^n_h-u(t_n)\|\leq C (h^2t_n^{-\alpha/2}+\tau t_n^{-1+\alpha/2}+\tau^{1/2}t_n^{-1/2}h)|v|_1.$$ An inspection of the three terms between brackets shows that the square of the third term equals the product of the first two terms, which proves the estimate . This concludes the proof.
Error analysis for the SBD method
---------------------------------
Now we consider the time discretization of $(\ref{semi-FV})$ constructed with the convolution quadrature based on the second-order backward difference formula. From Lemma \[lem:Lubich\], it is obvious that one can get only a first-order error bound if, for instance, $\varphi$ is constant (i.e., $\nu=1$). In order to overcome this difficulty, a correction of the scheme is needed. Below, we present modifications of the convolution quadrature based on the strategy in [@LST-1996] and [@Lubich-2006]. By noting the identity $$\label{k5}
(I+\partial_t^{-\alpha} \bar A_h)^{-1}=I-(I+\partial_t^{-\alpha} \bar A_h)^{-1}\partial_t^{-\alpha} \bar A_h,$$ it turns out from that the semidiscrete solution $\bar{u}_h$ can be rewritten as $$\label{k5}
\bar{u}_h=v_h-(I+\partial_t^{-\alpha} \bar A_h)^{-1}\partial_t^{-\alpha} \bar A_h v_h.$$ This leads to the modified convolution quadrature [@Lubich-2006] $$\label{k6}
U_h^n= v_h-(I+\bar\partial_\tau^{-\alpha} \bar A_h)^{-1}\partial_t^{-\alpha} \bar A_h v_h,$$ where the exact contribution $\partial_t^{-\alpha} A_h v_h = \omega_{\alpha+1}(t)A_h v_h$ is kept in the new formula (\[k6\]) in order to improve the time accuracy. The symbol $\bar\partial_\tau^{-\alpha}$ refers to the convolution quadrature generated by the SBD method. Unfortunately, this correction would not yield optimal time accuracy. A second choice for the modified convolution quadrature which will be considered here is based on the approximation [@LST-1996] $$\label{k6n}
U_h^n= v_h-(I+\bar\partial_\tau^{-\alpha} \bar A_h)^{-1}\bar \partial_\tau^{1-\alpha} \partial_t^{-1}\bar A_h v_h,$$ where the term $\partial_t^{-1}$ is kept to achieve second-order time accuracy. The advantages of both numerical methods and are described in [@Lubich-2006]. For the numerical implementation, it is essential to write as a time stepping algorithm. Let $1_\tau=(0,3/2,1,\cdots)$ so that $1_\tau=\bar\partial_\tau\partial_t^{-1}1$ at grid point $t_n$. Then by applying the operator $(I+\bar\partial_\tau^{-\alpha} \bar A_h)$ to both sides of and using the associativity of convolution in , we arrive at the equivalent form $$\label{k6n2}
(I+\bar\partial_\tau^{-\alpha}\bar A_h)(U_h^n-v_h)=-\bar\partial_\tau^{-\alpha} \bar A_h 1_\tau v_h.$$ By applying again the operator $\bar\partial_\tau$, we obtain $$\label{SBD}
\bar\partial_\tau(U_h^n-v_h)+\bar\partial_\tau^{1-\alpha} \bar A_h(U_h^n-v_h)=-\bar\partial_\tau^{1-\alpha}\bar A_h 1_\tau v_h.$$ By noting that $1v_h- 1_\tau v_h=(v_h,-1/2v_h,0,\cdots)$, we thus define the time stepping scheme as: with $U^0_h=v_h$, find $U_h^n$ such that $$\frac{3}{2}\tau^{-1}(U_h^1-U_h^0)+\tilde\partial_\tau^{1-\alpha} \bar A_hU_h^1=0,$$ and for $n\geq 2$ $$\bar\partial_\tau U_h^n+\tilde\partial_\tau^{1-\alpha} \bar A_hU_h^n=0,$$ where the modified convolution quadrature $\tilde\partial_\tau^{1-\alpha}$ is given by [@LST-1996] $$\tilde\partial_\tau^{1-\alpha}\varphi^n=
\left( \sum_{j=1}^n \beta_{n-j}^{(1-\alpha)}\varphi^j+ \frac{1}{2}\beta_{n-1}^{(1-\alpha)}\varphi^0\right),$$ with the weights $\{\beta_j^{(1-\alpha)}\}$ being generated by the SBD method.
Now using Lemma \[lem:Lubich\], we derive the following error bounds for smooth and nonsmooth initial data.
\[lem:SBD\] Let $\bar{u}_h$ and $U^n_h$ be the solutions of problems $(\ref{semi-FV})$ and $(\ref{SBD})$, respectively, and set $U^0_h=v_h$. Then, the following estimates hold:
\(a) If $v\in \dot{H}^2(\Omega)$ and $v_h =R_hv$, then $$\label{k10}
\|U^n_h-\bar{u}_h(t_n)\|\leq C \tau^2 t_n^{\alpha-2}|v|_2.$$
\(b) If $v\in L^2(\Omega)$ and $v_h =P_hv$, then $$\label{k9}
\|U^n_h-\bar{u}_h(t_n)\|\leq C \tau^2 t_n^{-2}\|v\|.$$
For the estimate , we set $$\bar G(z)=z^{1-\alpha}(I+z^{-\alpha}\bar A_h)^{-1}$$ and write the error as $$\label{k9b}
U^n_h-\bar{u}_h(t_n)= \left( \bar G(\bar\partial_\tau)- \bar G(\partial_t)\right)\partial_t^{-1} \bar A_h v_h.$$ Since $\|\bar G(z)\|\leq M_\theta |z|^{1-\alpha}\; \forall z\in \Sigma_\theta$ by , and Lemma \[lem:Lubich\] (with $\mu=\alpha-1$, $\nu=2$ and $p=2$) imply $$\|U^n_h-\bar{u}_h(t_n)\|\leq c\tau^2 t_n^{\alpha-2}\|\bar A_hv_h\|.$$ Then, the desired estimate follows from the identity $\bar A_hR_h=\bar P_hA$.
For the estimate , we note with $$\bar{G}(z)=z^{1-\alpha}(I+z^{-\alpha}\bar{A}_h)^{-1}\bar A_h$$ and using that $$\label{k9a}
U^n_h-\bar{u}_h(t_n)= \left( \bar {G}(\bar\partial_\tau)- \bar{G}(\partial_t)\right)\partial_t^{-1} v_h.$$ Since $\|\bar{G}(z)\|\leq M_\theta |z|\; \forall z\in \Sigma_\theta$, a use of , Lemma \[lem:Lubich\] (with $\mu=-1$, $\nu=2$ and $p=2$) and the $L^2(\Omega)$ stability of $P_h$ yield the estimate . This completes the rest of the proof
By the stability of the scheme, a direct consequence of Lemma \[lem:SBD\], and the arguments in Remark \[rem:BE\], the following error estimate holds for $v_h= P_h v$ $$\label{k11}
\|U^n_h-\bar{u}_h(t_n)\|\leq C (\tau^2 t_n^{\alpha-2}+h^2)|v|_2.$$ Then, by interpolation of and we get for $v_h =P_hv$ $$\label{k11a}
\|U^n_h-\bar{u}_h(t_n)\|\leq C (\tau^2 t_n^{-2})^{1/2} (\tau t_n^{\alpha-2}+h^2)^{1/2}|v|_1.$$
Using the estimates derived in Sections \[sec:H2\] and \[sec:L2\] for the semidiscrete problem, and following the arguments in the proof of Theorem \[thm:BE\], we can now state the error estimates for the fully discrete scheme with smooth and nonsmooth initial data.
\[thm:SBD\] Let $u$ and $U^n_h$ be the solutions of problems $(\ref{main})$ and $(\ref{SBD})$, respectively, with $U^0_h=v_h$. Then, the following error estimates hold:
\(a) If $v\in \dot{H}^2(\Omega)$ and $v_h =R_hv$, then $$\label{k12}
\|U^n_h-u(t_n)\|\leq C (h^2+\tau^2 t_n^{\alpha-2})|v|_2.$$
\(b) If $v\in \dot{H}^1(\Omega)$, $v_h =P_hv$ and the mesh is quasi-uniform, then $$\label{k13}
\|U^n_h-u(t_n)\|\leq C (h^2t_n^{-\alpha/2}+\tau^2 t_n^{\alpha/2-2})|v|_1.$$
\(c) If $v\in L^2(\Omega)$, $v_h =P_hv$ and $Q_h$ satisfies , then $$\label{k14}
\|U^n_h-u(t_n)\|\leq C (h^2t_n^{-\alpha}+\tau^2 t_n^{-2})\|v\|.$$
On extensions {#sec:extensions}
=============
In this section, we discuss the extension of our analysis to other type of problems including those with more general linear elliptic operator and other time-fractional evolution problems. We only concentrate on the error analysis of the semidiscrete FVE method. Completely discrete schemes can be discussed in a similar way by choosing appropriate convolution quadratures and following the analysis in Section \[sec:discrete\].
Problems with more general elliptic operators
---------------------------------------------
More precisely, we consider problem with $$A u =-\nabla\cdot(\kappa(x) \nabla u) + c(x) u,$$ where $\kappa(x)$ is a symmetric, positive definite $2\times 2$ matrix function on $\bar\Omega$ with smooth entries and $c(x)\in L^\infty(\Omega)$ and $c(x)\geq c_0 >0.$ The corresponding bilinear form $a(\cdot,\cdot):H_0^1(\Omega)\times H_0^1(\Omega)\rightarrow \mathbb{R}$ becomes $$\label{bilinear-n}
a (w,\chi) = (\kappa(x) \nabla w,\nabla \chi)+(c(x)w, \chi)\;\;\;\forall \chi\in H^1_0(\Omega).$$ The natural generalization of the finite volume element method yields $$\label{Ah-n}
a_h(w,\chi)= \sum_{P_i\in N_h^0} \chi(P_i)\left(- \int_{\partial K_{P_i}^*}
(\kappa\nabla w)\cdot{\bf n}\,ds +\int_{ K_{P_i}^*}c(x)w\, dx\,ds\right)
\quad \forall w \in V_h,\, \chi\in {V_h^*}.$$ In general, the bilinear form $a_h(w,\Pi_h^\ast\chi)$, $\chi \in V_h$, is not symmetric on $V_h$. However, if $\kappa$ and $c$ are constant over each element of the triangulation $ \mathcal{T}_h,$ then the bilinear form takes the form, see [@BR-1987], $$a_h(w,\Pi_h^\ast\chi) = (\kappa(x) \nabla w,\nabla \chi)+(c(x)w, \Pi_h^\ast\chi)\quad \forall w,\chi\in V_h,$$ which is symmetric since $(c(x)w, \Pi_h^\ast\chi)=(c(x)\chi, \Pi_h^\ast w)$. As symmetry is important in our analysis, we shall consider the modified bilinear form, see [@CLT-2013], $$\label{Ah-2}
\tilde a_h(w,\chi)= \sum_{P_i\in N_h^0} \chi(P_i)\left(- \int_{\partial K_{P_i}^*}
(\tilde\kappa(x)\nabla w)\cdot{\bf n}\,ds +\int_{ K_{P_i}^*}\tilde c(x) w\, dx\,ds\right)
\quad \forall w \in V_h,\, \chi\in {V_h^*},$$ where, for each $x\in K$, $K\in \mathcal{T}_h$, $\tilde\kappa(x)=\kappa(x_K)$ and $\tilde c(x)=c(x_K)$, with $x_K$ being the barycenter of the element $K$. Now, the FVE method reads: find $\tilde u_h(t)\in V_h$ such that $$\label{semi-FV-n}
(\tilde{u}_h',\chi)_h+ \tilde a_h({\partial_t^{1-\alpha}}\tilde{u}_h,\Pi_h^\ast\chi)= 0\quad
\forall \chi\in V_h,\quad t\in (0,T], \quad \tilde{u}_h(0)=v_h.$$ Introducing the discrete operator $\tilde A_h:V_h\rightarrow V_h$ by $$\label{Dh-n}
(\tilde A_h w,\chi)_h=\tilde a_h(w,\Pi_h^\ast\chi) \quad \forall w,\chi\in V_h,$$ we rewrite as $$\label{FVP-n}
\tilde u_h'(t)+{\partial_t^{1-\alpha}}\tilde A_h \tilde u_h(t)= 0, \quad t>0, \quad \tilde u_h(0)=v_h.$$
Following our analysis in Section \[sec:error\], with $\xi(t)=\tilde{u}_h(t)-u_h(t)$, we split the error $\tilde{u}_h(t)-u(t)= (u_h(t)-u(t))+\xi(t)$, where it is well known that $u_h(t)-u(t)$ and $\nabla(u_h(t)-u(t))$ are estimated by the analogues of -. It is, therefore, sufficient to derive estimates for $\xi$, which satisfies for $t\geq 0$ $$\label{semi-FV-nn}
(\xi',\chi)_h+ \tilde a({\partial_t^{1-\alpha}}\xi,\Pi_h^\ast\chi)= -\epsilon_h(u_{ht},\chi)-\tilde\epsilon_h(u_h,\chi)\quad
\forall \chi\in V_h,\quad \tilde{u}_h(0)=v_h,$$ where $\epsilon_h(\cdot,\cdot)$ is defined in and $\tilde \epsilon_h(\cdot,\cdot)$ is given by $$\label{m8-n}
\tilde\epsilon_h(w,\chi)= \tilde a_h(w,\Pi_h^\ast\chi)-a(w,\chi) \quad \forall w,\chi \in V_h.$$ Upon introducing the quadrature error operators $Q_h:V_h\rightarrow V_h$ and $\tilde Q_h:V_h\rightarrow V_h$ defined by $$\label{m8-nn}
\tilde a_h(Q_h w,\Pi_h^\ast\chi)=\epsilon_h(\chi,\psi) \quad \text{and} \quad
\tilde a_h(\tilde Q_h w,\Pi_h^\ast\chi)=\tilde\epsilon_h(\chi,\psi)\quad \forall w,\chi \in V_h,$$ the equation can be rewritten in the operator form as $$\label{m10-n}
\xi_t(t)+\partial_t^{1-\alpha}\tilde{A}_h\xi(t)=-\tilde{A}_hQ_hu_{ht}(t)- \tilde{A}_h\tilde Q_hu_{h}(t),\quad t>0,\quad \xi(0)=0.$$ To derive estimates for $\xi$, we need the following bound, see [@CLT-2013] for a proof.
\[lem:Qh-n\] Let $\tilde A_h$, $Q_h$ and $\tilde Q_h$ be the operators defined in and . Then $$\label{QQ-n}
\|\nabla Q_h\chi\|+h\|\tilde{A}_hQ_h\chi\|\leq Ch^{p+1}\|\nabla^{p}\chi\|\quad \forall \chi\in V_h, \quad p=0,1,$$ and similar result holds for the operator $\tilde Q_h$.
Now, we show the following estimates.
\[thm:H2-1\] For the error $\xi$ defined by $(\ref{m10-n})$, there is a positive constant $C,$ independent of $h,$ such that for $t>0$, $$\label{e1-m}
\|\xi(t)\|+h\|\nabla\xi(t))\|\leq C\max\{t^{1-\alpha/2},t^{1-\alpha}\}h^2\|A_hv_h\|,$$ $$\label{e1-mm}
\|\xi(t)\|+h\|\nabla\xi(t))\|\leq Ct^{1-\alpha/2}h^2\|\nabla v_h\|,$$ and $$\label{e1-mmm}
\|\xi(t)\|+h\|\nabla\xi(t))\|\leq Ct^{1-\alpha}h\|v_h\|.$$ If $\tilde Q_h$ satisfies $\|\tilde Q_h\chi\|\leq Ch^2\|\chi\|$ $\forall \chi \in V_h$, then $$\label{e1-4m}
\|\xi(t)\|\leq Ct^{1-\alpha}h^2\|v_h\|.$$
By taking Laplace transforms in , we represent $\xi(t)$ by $$\label{m10-nn}
\begin{aligned}
\xi(t)=-\frac{1}{2\pi i}\int_{\Gamma} e^{zt} & \hat{E}_h(z)\tilde{A}_hQ_h\hat{u}_{ht}(z)\,dz \\
&-\frac{1}{2\pi i}\int_{\Gamma} e^{zt}\hat{E}_h(z)\tilde{A}_h\tilde Q_h\hat{u}_{h}(z)\,dz =: \xi_1+\xi_2,
\end{aligned}$$ where $\hat{E}_h(z)=z^{\alpha-1}(z^\alpha I+\tilde{A}_h)^{-1}$. The first term $\xi_1$ is bounded as in the proofs of Theorems \[thm:H2\] and \[thm:L2\] using Lemma \[lem:Qh-n\] instead of Lemma \[lem:Qh\]. To bound the second term $\xi_2$, we notice that, similar to , we arrive at $$\label{Q1aa-m}
\|\hat{E}_h(z)\tilde{A}_h\tilde Q_h\hat u_h(z)\|+h\|\nabla \hat{E}_h(z)\tilde{A}_h\tilde Q_h\hat u_h(z)\|\leq Ch^2
|z|^{\alpha/2-1}\|\nabla \hat{u}_h(z)\|.$$ Using the identity $$\hat E_h(z)=z^{-1}[I-\hat E_h(z)\tilde A_h]$$ and , it follows that $$\begin{aligned}
\label{Q2-a}
\|\nabla\hat E_h(z)v_h\|&\leq& |z|^{-1}[\|\nabla v_h\|+\|\nabla\hat E_h(z)\tilde A_hv_h\|]\nonumber\\
&\leq & C |z|^{-1}[\|\tilde A_hv_h\|+|z|^{\alpha/2-1}\|\tilde A_hv_h\|].\end{aligned}$$ Substituting in and using the integral representation of $\xi_2$ in , we obtain the estimate . To derive , a use of yields $$\|\nabla\hat E_h(z)v_h\|\leq C |z|^{-1} \|\nabla v_h\|.$$ Then, the bound follows immediately. For the last cases and , we apply to get $$\|\hat{E}_h(z)\bar{A}_h\tilde Q_h\hat{u}_h\|_p\leq C|z|^{\alpha-1}\|\tilde Q_h\hat{u}_h\|_p,\quad p=0,1.$$ Then, the left-hand side in is bounded by $$\label{Q3a-n}
C|z|^{\alpha-1}( \|\tilde Q_h\hat{u}_h(z)\| + h \|\nabla \tilde Q_h\hat{u}_h(z)\|).$$ Using Lemma \[lem:Qh-n\] and the fact that $\|\hat{u}_h(z)\|\leq |z|^{-1}\|v_h\|$, we obtain the desired results by following the arguments in the proof of Theorem \[thm:L2\]. This completes the proof of the theorem.
Other time-fractional evolution problems
----------------------------------------
Our analysis can be applied to obtain optimal FVE error estimates for other type of time-fractional evolution problems. This may include, for instance, evolution equations with memory terms of convolution type: $$\label{EE}
u'(x,t)+\mathcal{I}^\alpha A u(x,t)=0, \quad \alpha\in(0,1),$$ see [@LST-1996], which is also called fractional diffusion-wave equation, the following parabolic integro-differential equation with singular kernel of the type $$\label{PIDE-S}
u'(x,t)+ (I + \mathcal{I}^\alpha) A u(x,t)=0, \quad \alpha\in(0,1),$$ see, [@MST2006], and the Rayleigh-Stokes problem described by the time-fractional differential equation $$\label{RS}
u'(x,t)+(I+\gamma\partial_t^\alpha) A u(x,t)=0,\quad \alpha\in(0,1),$$ which has been considered in [@EJLZ2016]. Here $\gamma$ is a positive constant. In order to unify problems -, we define $\mathcal{J}^\alpha$ denoting a time integral/differenial operator and consider the unified problem by $$\label{J}
u'(x,t)+\mathcal{J}^\alpha A u(x,t)=0.$$ Now an application of Laplace transforms in yields $$z\hat{u}+h(z)A\hat{u}=v,$$ with some function $h(z)$ depending on $\alpha$. Hence, we formally have, $\hat{u}= (z+ h(z) A)^{-1} v =: \hat{E}_h(z)v.$
Let $\bar{A}_h$ and $Q_h$ be the operators defined in Section \[sec: FVM\]. Then, the FVE method reads: find $\bar u_h(t)\in V_h$ such that $$\label{semi-GG1}
\bar{u}_h'+ \mathcal{J}^\alpha \bar A_h\bar{u}_h= 0\quad
\quad t\in (0,T], \quad \bar{u}_h(0)=v_h.$$ Again using the corresponding FE solution $u_h,$ we split $\bar{u}_h- u := (u_h-u) + (\bar{u}_h- u_h)
=:(u_h-u) + \xi,$ where $\xi$ satisfies the similar representation formula $$\label{m013-nnn}
\xi(t)=-\frac{1}{2\pi i}\int_{\bar{\Gamma}_\theta} e^{zt}\hat{E}_h(z)\bar{A}_hQ_h\hat{u}_{ht}(z)\,dz.$$ Note that in this case the operator $\hat{E}_h(z)$ is given by $$\label{sm1}
\hat{E}_h(z)=\beta(z)(z\beta(z)I+\bar A_h)^{-1},$$ and $\beta(z)=h(z)^{-1}$. For the problem , we observe that $\beta(z)=z^{\alpha},$ for the problem , $\beta(z)= z^\alpha/(1+z^{\alpha}),$ and for the problem , $\beta(z)=1/(1+\gamma z^\alpha)$. We assume that one can properly choose $\theta$ in $(\pi/2,\pi)$ such that $z\beta(z)\in \Sigma_{\theta'}$ for all $z\in \Sigma_{\theta}$ where the angle $\theta'\in(\pi/2,\pi)$. This is indeed possible in all given examples. With this, the resolvent estimate yields $$\label{sm2}
\|(z\beta(z)I+\bar A_h)^{-1}\|\leq \frac{M_{\theta'}}{|z\beta(z)|}\quad \forall z\in \Sigma_{\theta},$$ where $M_{\theta'}=1/\sin(\pi-\theta')$. Therefore, from , $$\label{sm3}
\|\hat{E}_h(z)\|\leq M_{\theta'}|z|^{-1} \quad \forall z\in \Sigma_{\theta}.$$ Following arguments from [@LST-1996], we deduce that $$\label{sm2}
\|\bar A_h\hat{E}_h(z)\|\leq C_{\theta'}|\beta(z)|\,\quad \forall z\in \Sigma_{\theta}.
$$ Now, we can prove the analogous of Lemma \[lem:Ah\].
\[lem:Ah2-n\] Let $\hat E_h(z)$ be given by . With $\chi\in V_h$, the following estimates hold: $$\label{0-n}
\|{\bar A_h\hat{E}_h(z)\chi}\|\leq C_{\theta'} |\beta(z)|^{1-p/2}|z|^{-p/2} \;\|\bar {A}_h^{p/2}{\chi}\|
\quad \forall z\in \Sigma_\theta,\quad 0\leq p\leq 2,$$ $$\label{1-n}
|{\hat{E}_h(z)\chi}|_1\leq C_{\theta'} |\beta(z)|^{1/2}|z|^{-1/2} \|{\chi}\|\quad \forall z\in \Sigma_\theta,$$ where $C_{\theta'}$ is independent of the mesh size $h$.
We obtain the first estimate by interpolating and . The second estimate follows from the fact that $$\|\nabla (z\beta(z)I+\bar A_h)^{-1}\chi\|\leq C |z\beta(z)|^{-1/2} \|\chi\|\quad \forall \chi \in V_h,$$ see (2.13) in [@Lubich-2006].
In the following theorem, optimal error estimates are obtained for smooth and nonsmooth initial data $v\in \dot{H}^q(\Omega)$, $q=0,1,2.$
\[thm:H2-n\] For the error $\xi$ defined by $\eqref{m013-nnn}$, there is a positive constant $C$, independent of $h,$ such that $t>0$, $$\label{e1-n}
\|\xi(t)\|+h\|\nabla\xi(t))\|\leq Ch^2\|\bar A_hv_h\|.$$ If $|\beta(z)|\leq C |z|^\mu$ $\forall z\in\Sigma_\theta$ for some real $\mu<1$, then $$\label{e1-nn}
\|\xi(t)\|+h\|\nabla\xi(t))\|\leq Ct^{-(\mu+1)/2}h^2\|\nabla v_h\|.$$ If $|\beta(z)|\leq C |z|^\mu$ $\forall z\in\Sigma_\theta$ and $\bar Q$ satisfies , then $$\label{e1-nnn}
\|\xi(t)\|+h\|\nabla\xi(t))\|\leq Ct^{-(\mu+1)}h^2\| v_h\|.$$
We will only prove the estimate in the $L^2(\Omega)$-norm. The estimate in the gradient norm is derived in a similar way. We shall make use of the estimate obtained in the proof of Theorem \[thm:H2\].
When $q=2$, that is, $v\in \dot{H}^2(\Omega),$ apply (\[0-n\]) with $p=1$ and (\[1-n\]) in Lemma \[lem:Ah2-n\] to get $$\label{m014-n}
\|\hat{E}_h(z)\bar{A}_hQ_h\widehat{u_{ht}}(z)\|\leq
C |\beta(z)|^{1/2} |z|^{-1/2}\|\nabla Q_h \widehat{u_{ht}}(z)\|,$$ and $$\label{g2-n}
\|\nabla \hat{E}_h(z)\bar{A}_hQ_h\widehat{u_{ht}}(z)\|\leq C |\beta(z)|^{1/2} |z|^{-1/2}\|\bar A_hQ_h\widehat{u_{ht}}(z)\|.$$ Then, by (\[QQ\]) in Lemma \[lem:Qh\], we deduce $$\label{Q1aa-n}
\|\hat{E}_h(z)\bar{A}_hQ_h\widehat{u_{ht}}(z)\|+h\|\nabla \hat{E}_h(z)\bar{A}_hQ_h\widehat{u_{ht}}(z)\|\leq
Ch^2 |\beta(z)|^{1/2} |z|^{-1/2}\|\nabla \widehat{u_{ht}}(z)\|.$$ Since $$\widehat{u_{ht}}(z)=-h(z)\bar A_h\hat{{u}}_{h}(z)= -h(z)\bar A_h\hat{F}_{h}(z)v_h,$$ an estimate analogous to (\[1-n\]) yields $$\begin{aligned}
\|\nabla \widehat{u_{ht}}(z)\|&= &|h(z)|\|\nabla \hat{F}_{h}(z)\bar A_hv_h\|\nonumber\\
&\leq & C|h(z)|\,|\beta(z)|^{1/2} |z|^{-1/2}\|\bar A_hv_h\|\nonumber\\
&\leq & C |\beta(z)|^{-1/2} |z|^{-1/2} \|\bar A_h v_h\|.\nonumber\end{aligned}$$ Thus, the left-hand side in (\[Q1aa-n\]) is bounded by $|z|^{-1} \|\bar A_h v_h\|$. Now, substitution in gives the desired estimate.
For $q=1$, we notice that in view of , the bound (\[m0156\]) holds, and therefore substitution in (\[Q1aa-n\]) gives the new upper bound $Ch^2 |z|^{\mu/2-1/2} \|\nabla v_h\|$ in (\[Q1aa-n\]). The estimate follows then by integration.
Finally, for $q=0$, we have by , $$\|\hat{E}_h(z)\bar{A}_hQ_h\widehat{u_{ht}}\|\leq C|\beta(z)|\,\|Q_h\widehat{u_{ht}}\|
\leq C|z|^{\mu}\|Q_h\widehat{u_{ht}}\|.$$ In view of , we have $\|\widehat{u_{ht}}(z)\|=\|z\hat{F}_{h}(z)v_h-v_h\|\leq C \|v_h\|$. Therefore, if is satisfied then $\|\hat{E}_h(z)\bar{A}_hQ_h\widehat{u_{ht}}\|
\leq C h^2|z|^{\mu}\|\widehat{u_{ht}}\|\leq C h^2|z|^{\mu}\|v_h\|.$ Now, follows by integration and this concludes the rest of the proof.
By interpolating and we obtain for $q\in [0,2]$ $$\|\xi(t)\|+h\|\nabla\xi(t))\|\leq Ct^{-(\mu+1)(1-q/2)}h^2\|\bar A_h^{q/2} v_h\|,\quad t>0.$$ Notice that $\mu=\alpha$ for problems and , while $\mu=-\alpha$ for the Rayleigh-Stokes problem . Hence, for the Rayleigh-Stokes problem the previous estimate reads: $$\|\xi(t)\|+h\|\nabla\xi(t))\|\leq Ct^{-(1-\alpha)(1-q/2)}h^2\|\bar A_h^{q/2} v_h\|,\quad t>0,$$ provided is satisfied.
We finally consider the following class of time-fractional order diffusion problems: $$\label{eq-caputo}
^C\partial_t^{\alpha}u (x,t)+{A}u(x,t)=0,$$ where $^C\partial_t^{\alpha}$ is the fractional Caputo derivative of order $\alpha\in(0,1)$. For this class of equations, optimal error estimates for the semidiscrete FE method have been established in [@JLZ2013]. The FVE method applied to is to seek $\bar{u}_h\in V_h$ such that $$\label{semi-GG}
{{^C}\partial_t^{\alpha}}\bar{u}_h + \bar A_h\bar{u}_h= 0\quad
\quad t\in (0,T], \quad \bar{u}_h(0)=v_h.$$ Again a comparison between the FE solution and FVE solution along with Laplace techniques and semigroup type properties as has been done in Section \[sec:error\] yields [*a priori*]{} FVE error estimates for the fractional order evolution problem for both smooth and nonsmooth initial data. Since the proof technique is similar to the tool used in Section $4$, we skip the details.
![Triangular meshes with $M=8$, (a) symmetric mesh (b) nonsymmetric mesh.[]{data-label="Fig:meshes"}](figureSymMesh "fig:"){width="6cm" height="6cm"} $\;$ ![Triangular meshes with $M=8$, (a) symmetric mesh (b) nonsymmetric mesh.[]{data-label="Fig:meshes"}](figureNonsymMesh "fig:"){width="6cm" height="6cm"}
Derivation by the lumped mass FE method
---------------------------------------
In this subsection, we extend our analysis to the lumped mass FE method applied to the time-fractional diffusion problem . For completeness, we briefly describe, below, this approximation. For $K\in\mathcal{T}_h$ with vertices $P_i$, $i=1,2,3$, consider the quadrature formula $$\label{QF}
Q_{K,h}(f)=\frac{|K|}{3}\sum_{i=1}^3f(P_i)\approx\int_Kf\,dx.$$ Then, we define an approximation of the $L^2$-inner product on $V_h$ by $$\label{QF}
\langle w,\chi\rangle=\sum_{K\in\mathcal{T}_h} Q_{K,h}(w\chi).$$ The lumped mass Galerkin FE method reads: find $\bar u_h(t)\in V_h$ satisfying $$\label{semi-LM}
\langle\bar{u}_h',\chi\rangle+ a({\partial_t^{1-\alpha}}\bar{u}_h,\chi)= 0\quad
\forall \chi\in V_h,\quad t\in (0,T], \quad \bar{u}_h(0)=v_h.$$ In operator form, the method can be written as $$\label{LM-OF}
\bar u_h'(t)+{\partial_t^{1-\alpha}}\bar A_h \bar u_h(t)= 0, \quad t>0, \quad u_h(0)=v_h,$$ where $\bar A_h:V_h\rightarrow V_h$ is the discrete Laplacian corresponding to the inner product $\langle\cdot,\cdot\rangle$ given by $$\label{LM-OF-2}
\langle\bar A_h w,\chi\rangle=(\nabla w,\nabla \chi) \quad \forall w,\chi\in V_h.$$ Now, introduce $\xi(t)=\tilde{u}_h(t)-u_h(t)$ with $u_h(t)$ being the Galerkin FE solution. Then $\xi$ satisfies $$\label{LM-OF-1}
\xi'(t)+{\partial_t^{1-\alpha}}\bar A_h \xi(t)= -\bar A_hQ_hu_{ht}, \quad t>0, \quad \xi(0)=0,$$ where $Q_h:V_h\rightarrow V_h$ is the quadrature error defined by $$\label{m8-LM}
(\nabla Q_h\chi,\nabla\psi)=\epsilon_h(\chi,\psi):=\langle\chi,\psi\rangle-(\chi,\psi)\quad \forall \psi \in V_h.$$ Since the operators $\bar{A}_h$ and $Q_h$ defined by and have properties similar to the corresponding operators in the FVE method in Section \[sec:error\], (see also [@CLT-2012]), then the error estimates for the lumped mass FE method and their proofs are quite analogous to the results proved in Sections \[sec:error\] and \[sec:discrete\] for the FVE method. Therefore, we can easily derive optimal error estimates and we shall not pursue it further.
Numerical Experiments {#sec:NE}
======================
In this section, we present some numerical tests to validate our theoretical results. We choose $\Omega=(0,1)\times (0,1)$ and perform the computation on two families of symmetric and nonsymmetric triangular meshes. The symmetric meshes are uniform with mesh size $h=\sqrt{2}/M$, where $M$ is the number of equally spaced subintervals in both the $x$- and $y$-directions, see Figure \[Fig:meshes\](a). For the nonsymmetric meshes, we choose $M$ subintervals in the $x$-direction and $3M/4$ equally spaced subintervals in the $y$-direction with the assumption that $M$ is divisible by 4. The intervals in the $x$-direction are of lengths $4/3M$ and $2/3M$ and distributed such that they form an alternating series as shown in Figure \[Fig:meshes\](b). One can notice that the nonsymmetric mesh defines a triangulation that is not symmetric at any vertex, see [@CLT-2013 Section 5] for more details.
We consider three numerical examples with smooth and nonsmooth initial data. By separation of variables, the exact solution of problem can represented by a rapidly converging Fourier series $$\label{eq: u series}
u(x,y,t)=2\sum_{m,n=1}^\infty (v, \phi_{mn})
E_{\alpha}(-\lambda_{mn} t^{\alpha})\phi_{mn}(x,y),$$ where $E_{\alpha}(t):=\sum_{p=0}^\infty\frac{t^p}{\Gamma(\alpha p+1)}$ is the Mittag-Leffler function and $$\phi_{mn}(x,y)=2\sin(m \pi x)\sin(n \pi y)
\quad\text{and}\quad
\lambda_{mn}=(m^2+n^2)\pi^2\quad{\rm for}~~ m\,,n=1, 2, \ldots$$ are the orthonormal eigenfunctions and corresponding eigenvalues of $-\Delta$ subject to homogeneous Dirichlet boundary conditions. In our computation, we evaluate the exact solution by truncating the Fourier series in after $60$ terms.
[|r|cccc|]{} $N$& BE & rate& SBD & rate\
\
5& 4.8880e-003 & & 1.3161e-003 &\
10& 2.1844e-003 & 1.16 & 3.1530e-004 & 2.06\
20& 1.0367e-003 & 1.08 & 7.2627e-005 & 2.12\
40& 5.0547e-004 & 1.04 & 1.6922e-005 & 2.10\
80& 2.4952e-004 & 1.02 & 3.6949e-006 & 2.18\
\
5 &4.8270e-003 & &1.3857e-003 &\
10 &2.1578e-003 &1.16 &3.3341e-004 &2.06\
20 &1.0247e-003 &1.07 &7.7019e-005 &2.11\
40 &5.0021e-004 &1.03 &1.7736e-005 &2.19\
80 &2.4751e-004 &1.02 &3.6842e-006 &2.27\
\
5 & 2.9708e-003 & & 8.2449e-004 &\
10 & 1.3300e-003 & 1.16 & 2.0483e-004 & 2.01\
20 & 6.3206e-004 & 1.07 & 4.7324e-005 & 2.11\
40 & 3.0862e-004 & 1.03 & 1.0961e-005 & 2.11\
80 & 1.5275e-004 & 1.01 & 2.4291e-006 & 2.17\
We consider the following initial data to illustrate the convergence theory.
- With $v=xy(1-x)(1-y)$, its Fourier sine coefficients become $$(v,\phi_{mn})=8(1-(-1)^m)(1-(-1)^n)(mn\pi^2)^{-3},\quad {\rm for}~~m,n=1,\,2,\ldots.$$ This example represents the smooth case as $v \in \dot H^{2}(\Omega)$.
- For this example, choose $v=xy\chi_{(0,1/2]\times(0,1/2]} +(1-x)y\chi_{(1/2,1)\times(0,1/2]}+
x(1-y)\chi_{(0,1/2]\times(1/2,1)}+(1-x)(1-y)\chi_{(1/2,1)\times(1/2,1)}$, where $\chi_D$ denotes the characteristic function on the domain $D$. This initial data is less smooth compared to the previous case. One can verify that its Fourier coefficients are given by $$(v,\phi_{mn})=2(1-(-1)^m)(1-(-1)^n)(mn\pi^2)^{-2}(-1)^{mn},\quad {\rm for}~~m,n=1,\,2,\ldots.$$ Note that $v \in \dot H^{1+\epsilon}(\Omega)$ for $0\le \epsilon<1/2$.
- With $v=\chi_{(0,1/2[\times(0,1)}(x,y)$, its Fourier sine coefficients become $$(v,\phi_{mn})=2(1-\cos(m\pi/2))(1-(-1)^n)(mn\pi^2)^{-1},\quad {\rm for}~~m,n=1,\,2,\ldots.$$ Here, $v \in \dot H^{\epsilon}(\Omega)$ for $0\le \epsilon<1/2$.
[|r|cccc|]{} $M$& $L^2$-norm error & rate&$L^\infty$-norm error& rate\
\
8 & 1.4556e-003 & & 1.0596e-004 &\
16 & 3.7356e-004 & 1.96 & 2.7366e-005 & 1.95\
32 & 9.3259e-005 & 2.00 & 6.8602e-006 & 2.00\
64 & 2.2546e-005 & 2.05 & 1.6792e-006 & 2.03\
128 & 4.8155e-006 & 2.23 & 3.8055e-007 & 2.14\
\
8 & 8.9301e-004 & & 2.0405e-004 &\
16 & 2.2952e-004 & 1.96 & 5.5397e-005 & 1.88\
32 & 5.7285e-005 & 2.00 & 1.4340e-005 & 1.95\
64 & 1.3820e-005 & 2.05 & 3.5649e-006 & 2.01\
128 & 2.9842e-006 & 2.21 & 8.0446e-007 & 2.15\
\
8 & 7.1870e-004 & & 2.7011e-003 &\
16 & 1.8148e-004 & 1.99 & 8.7438e-004 & 1.63\
32 & 4.5181e-005 & 2.01 & 2.7169e-004 & 1.69\
64 & 1.1033e-005 & 2.03 & 7.6187e-005 & 1.83\
128 & 2.6557e-006 & 2.05 & 2.0470e-005 & 1.90\
To examine the temporal accuracy of the proposed schemes, we employ a uniform temporal mesh with a time step $\tau=T/N$, where $T=0.5$ is the time of interest in all numerical experiments. We fix the mesh size $h$ at $h = 1/400$ so that the error incurred by spatial discretization is negligible, which enable us to examine the temporal convergence rate. The computation is performed on symmetric meshes. We measure the error $e^n =: u(t_n)-U^n$ by the normalized $L^2(\Omega)$-norm $\| e^n\|_{L^2(\Omega)}/\|v\|_{L^2(\Omega)}$. The numerical results are presented in Table \[table:1\] for the three proposed cases (a)-(c). In the table, BE and SBD denote the convolution quadrature generated by the backward Euler and the second-order backward difference methods, respectively. The [rate]{} refers to the empirical convergence rate, when the time step size $\tau$ halves. From the Table \[table:1\], a convergence rate of order $O(\tau)$ and $O(\tau^2)$ is observed for the BE and SBD schemes, respectively, and clearly both schemes exhibit a very steady behavior for both smooth and nonsmooth data, which agree well with our convergence theory. Additional numerical experiments with different values of fractional order $\alpha$ have shown similar convergence rates. It was, in particular, observed that the error decreases as the fractional order $\alpha$ increases. More details on the behaviour of errors from BE and SBD methods combined with a Galerkin FE discretization in space can be found in [@JLZ2016].
To check the spatial discretization error, we fix the time step $\tau=1/500$ and use the SBD scheme so that the temporal discretization error is negligible. We carry out the computation on symmetric meshes. In Table \[table:2\], we list the normalized $L^2(\Omega)$-norm and $L^\infty(\Omega)$-norms of the error for the cases (a)-(c). The numerical results show a convergence rate $O(h^2)$ for the $L^2(\Omega)$-norm of the error for smooth and nonsmmoth initial data. A similar convergence rate is obtained in the $L^\infty(\Omega)$-norm (ignoring a logarithmic factor). The results fully confirm the predicted rates on symmetric meshes. They also show the validity of the convergence rate in Theorem \[thm:Linfty\] for case (c) where $0<q<1$.
----- ------------------ ------ ----------------------- ------
$M$ $L^2$-norm error rate $L^\infty$-norm error rate
8 1.1209e-003 4.1704e-003
1.6 2.7755e-004 2.01 1.3697e-003 1.61
32 6.8036e-005 2.03 4.1953e-004 1.71
64 1.6529e-005 2.04 1.1120e-004 1.92
128 3.9610e-006 2.06 3.0306e-005 1.88
8 1.1627e-003 4.1512e-003
16 3.1215e-004 1.90 1.3697e-003 1.60
32 8.2238e-005 1.92 4.1472e-004 1.72
64 2.1382e-005 1.94 1.1120e-004 1.90
128 5.8007e-006 1.88 3.3495e-005 1.73
----- ------------------ ------ ----------------------- ------
: Errors for case (c) on nonsymmetric meshes, $\alpha=0.75$, $\tau=1/500$.[]{data-label="table:Nonsym-1"}
For nonsymmetric meshes, we are especially interested in spatial errors for nonsmooth initial data as the convergence theory suggests. In Table \[table:Nonsym-1\], we display the $L^2(\Omega)$- and $L^\infty(\Omega)$-norms of the error for case (c) using the FVE and the lumped mass FE discretizations on nonsymmetric meshes. The numerical results reveal that both discretizations exhibit a convergence rate of order $O(h^2)$, which may be seen as an unexpected result. However, as the initial data $v\in \dot H^{1/2-\epsilon}(\Omega)$ for any $\epsilon>0$, $v$ has some smoothness, and hence, the numerical results do not contradict our theoretical findings. In addition, we notice that as the convergence rate is $O(h^2)$ for initial data in $\dot H^{1}(\Omega)$, by interpolation in $[0,1]$, a convergence rate of order $O(h^{3/2})$ is expected for $v\in \dot H^{1/2}(\Omega)$. In our case, the smoothness of the particular initial data $v$ could then have a positive effect on the convergence rate.
In [@CLT-2013], the authors considered the nonsymmetric partition shown in Figure \[Fig:meshes\](b) and provided an initial data for which the optimal $L^2$-convergence does not hold. They proved that the best possible error bound in this case is of order 1, see Proposition 5.1 of [@CLT-2013]. Earlier in [@CLT-2012], the same authors have established a one-dimensional example for which the $O(h^2)$ nonsmooth data error does not hold for the lumped mass FE method. We, then, carried out our computation based on the example in [@CLT-2013 Proposition 5.1]. The numerical results are presented in Table \[table:Nonsym-2\] using the SBD scheme. The error reported in the table represents the quantity $\xi(t)$ which measures the difference between the Galerkin FE solution and the FVE solution for the first set of numerical results and between the Galerkin FE solution and the lumped mass FE solution for the second set. As the nonsmooth data error from the standard Galerkin FE is always $O(h^2)$, the error from the considered methods is dominated by $\xi(t)$. From the Table \[table:Nonsym-2\], an order $O(h)$ of convergence rate is observed for both methods, which agrees well with the results in [@CLT-2013] and confirms our theoretical analysis.
----- ------------------ ------ ----------------------- ------
$M$ $L^2$-norm error rate $L^\infty$-norm error rate
8 9.9247e-005 3.8454e-004
16 2.3133e-005 2.10 1.1238e-004 1.77
32 1.1497e-005 1.01 4.8469e-005 1.21
64 5.1181e-006 1.17 2.0545e-005 1.24
128 2.5579e-006 1.00 9.7156e-006 1.08
8 4.4924e-004 1.7395e-003
16 1.0429e-004 2.11 5.0652e-004 1.78
32 5.1762e-005 1.01 2.1821e-004 1.21
64 2.3035e-005 1.17 9.2466e-005 1.24
128 1.1511e-005 1.00 4.3722e-005 1.08
----- ------------------ ------ ----------------------- ------
: Errors for case (d) on nonsymmetric meshes, $\alpha=0.75$, $\tau=1/500$.[]{data-label="table:Nonsym-2"}
For completeness, we extend our numerical study to examine some of the problems presented in Section 6, namely; the subdiffusion problem with a fractional Caputo derivative and the wave-diffusion problem . The numerical solution in each case is obtained by using the FVE method in space and a convolution quadrature in time generated by the second-order backward difference method. We run both examples with the initial data $v$ given in case (c).
$M$ $L^2$-norm error rate $L^\infty$-norm error rate
----- ------------------ ------ ----------------------- ------
8 7.1870e-004 2.7011e-003
16 1.8148e-004 1.99 8.7438e-004 1.63
32 4.5181e-005 2.01 2.7169e-004 1.69
64 1.1033e-005 2.03 7.6187e-005 1.83
128 2.6557e-006 2.05 2.0470e-005 1.90
: Numerical results for problem , $\alpha=0.75$, $\tau=1/500$.[]{data-label="table:5"}
For the first problem, we employ the second-order time discretization scheme derived in [@JLZ2016 formula (2.16)]. The computed errors are presented in Table 5 and are clearly identical to the results in Table 2. Even though it is known that the two representations and are equivalent, the numerical methods obtained for each representation are in general different. However, in the current case, the fact that the time discrete schemes are equivalent is due to the feature of the convolution quadrature, in particular, to the properties given in .
For the wave-diffusion problem, the numerical results are listed in Table 6 for $\alpha = 0.5$. We observe a $O(h^2)$ convergence for the $L^2(\Omega)$- and $L^\infty(\Omega)$-norm of the errors which confirms our predictions. It is known that the model interpolates the heat and wave equations when the fractional order $\alpha$ increases from zero to one. This transition is observed numerically. In Figure \[fig:transition\], we display the profile of the numerical solutions to case (c) at time $t=0.1$ with different values of $\alpha$. We observe that, the closer $\alpha$ is to zero, the slower is the decay. Furthermore, the oscillations in Figure \[fig:transition\](a) are inherited from the $L^2$-projection $P_hv$ which is oscillatory. This reflects, in particular, the wave feature of the model .
$M$ $L^2$-norm error rate $L^\infty$-norm error rate
----- ------------------ ------ ----------------------- ------
8 5.7494e-003 1.0952e-002
16 1.4393e-003 2.00 2.7976e-003 1.97
32 3.5725e-004 2.01 7.2567e-004 1.94
64 8.5491e-005 2.06 1.9564e-004 1.89
128 1.9769e-005 2.11 5.1351e-005 1.93
: Numerical results for problem , $\alpha=0.5$, $\tau=1/500$.[]{data-label="table:6"}
[0]{}
R. E. Bank and D. J. Rose, Some error estimates for the box method, [*SIAM J. Numer. Anal.*]{}, 24 (1987), 777–-787.
E. Bazhlekova, B. Jin, R. Lazarov and Z. Zhou, An analysis of the Rayleigh-Stokes problem for a generalized second-grade fluid, [*Numer. Math.*]{}, 131 (2016), 1–31.
, Error estimates for a finite volume element method for parabolic equations in convex polygonal domains, [*Numer. Meth. PDEs*]{}, 20 (2004), 650–674.
, Some error estimates for the lumped mass finite element method for a parabolic problem, [*Math. Comp.*]{}, 81 (2012), 1–-20.
, Some error estimates for the finite volume element method for a parabolic problem, [*Comput. Meth. Appl. Math.*]{}, 13 (2013), 251–-279.
, Error estimates in $L^2$, $H^1$ and $L^\infty$ in covolume methods for elliptic and parabolic problems: A unified approach, [*Math. Comp.*]{}, 69 (2000), 103–120.
E. Cuesta, C. Lubich and C. Palencia, Convolution quadrature time discretization of fractional diffusion-wave equations, [*Math. Comp.*]{}, 75 (2006), 673–-696.
, Finite volume element approximations of nonlocal reactive flows in porous media, [*Numer. Methods Partial Differential Equations*]{}, 16 (2000), 285–311.
, Time fractional diffusion: a discrete random walk approach, [*Nonlinear Dynamics*]{}, 29 (2002), 129–143.
, Fractional reaction-diffusion, [*Physica A*]{}, 276 (2000), 448–455.
B. Jin, R. Lazarov and Z. Zhou, Two fully discrete schemes for fractional diffusion and diffusion-wave equations with nonsmooth data, [*SIAM J. Sci. Comput.*]{}, 38 (2016), 146–-170.
B. Jin, R. Lazarov and Z. Zhou, Error estimates for a semidiscrete finite element method for fractional order parabolic equations, [*SIAM J. Numer. Anal.*]{}, 51 (2013), 445-–466.
B. Jin, R. Lazarov, J. Pascal and Z. Zhou, Error analysis of semidiscrete finite element methods for inhomogeneous time-fractional diffusion, [*IMA J. Numer. Anal.*]{}, 35 (2015), 561–-582.
S. Karaa, K. Mustapha and A. K. Pani, Finite volume element method for two-dimensional fractional subdiffusion problems, [*IMA J. Numer. Anal.*]{}, (2016), to appear.
S. Karaa, K. Mustapha and A. K. Pani, [*A priori*]{} estimates of a finite element method for fractional diffusion problems by energy arguments, arXiv:1605.09104.
R. H. Li, Z. Y. Chen and W. Wu, Generalized Difference Methods for Differential Equations, Marcel Dekker, New York, 2000.
C. Lubich, Discretized fractional calculus, [*SIAM J. Math. Anal.*]{}, 17 (1986), 704–-719.
C. Lubich, Convolution quadrature and discretized operational calculus-I, [*Numer. Math.*]{}, 52 (1988), 129-–145.
C. Lubich, Convolution quadrature revisited, [*BIT Numerical Mathematics*]{}, 44 (2004), 503–-514.
C. Lubich, I. H. Sloan and V. Thomée, Nonsmooth data error estimates for approximations of an evolution equation with a positive-type memory term, [*Math. Comp.*]{}, 65 (1996), 1–-17.
W.McLean, I. H. Sloan, and V. Thomée, Time discretization via Laplace transformation of an integrodifferential equation of parabolic type, [*Numer. Math.*]{} 102 (2006), 497–522.
W. McLean and V. Thomée, Numerical solution via Laplace transforms of a fractional order evolution equation, [*J. Integral Equations Appl.*]{} 22 (2010), 57-–94.
W. McLean and V. Thomée, Maximum-norm error analysis of a numerical solution via Laplace transformation and quadrature of a fractional order evolution equation, [*IMA J. Numer. Anal.*]{}, 30 (2010), 208–230.
W. McLean and V. Thomée, Time discretization of an evolution equation via Laplace transforms, [*IMA J. Numer. Anal.*]{}, 24 (2004), 439–-463.
, The random walk’s guide to anomalous diffusion: a fractional dynamics approach, [*Physics Reports*]{}, 339 (2000), 1–77.
K. Mustapha and W. McLean, Piecewise-linear, discontinuous Galerkin method for a fractional diffusion equation, [*Numer. Algorithms*]{}, [56]{} (2011), 159–184.
V. Thomée. Galerkin finite element methods for parabolic problems, Springer-Verlag, Berlin , 2006.
[^1]: Department of Mathematics and Statistics, Sultan Qaboos University, P. O. Box 36, Al-Khod 123, Muscat, Oman. Email: [email protected]. The research of this author is supported by Sultan Qaboos University under Grant IG/SCI/DOMS/16/01.
[^2]: Department of Mathematics, Industrial Mathematics Group, Indian Institute of Technology Bombay, Powai, Mumbai-400076.
| {
"pile_set_name": "ArXiv"
} |
Since the discovery of the $d$-wave pairing symmetry in high-$T_c$ cuprates, there has been increased interest in low-energy quasiparticle properties in unconventional superconductors. In such a $d$-wave superconductor, quasiparticles are gapless along the four nodal directions on the essentially cylindrical Fermi surface, in contrast to the conventional $s$-wave superconductors, where quasiparticles are gapped everywhere. The issues of how the disorder affects the low energy quasiparticle excitations and whether these quasiparticles are localized remain unresolved. Some perturbative self-consistent $T$-matrix (SCTM) calculations [@GK85; @SMV86; @HVW86; @Lee93; @HL93; @LW95; @XW95; @Joynt97] and a non-perturbative one [@ZHH96] predicted a nonzero constant density of states (DOS) in low energy region in the presence of weak disorder. However, most of non-perturbative calculations [@NT97; @Mudry96] and one numerical study [@AHM00] showed that the DOS at zero energy vanishes. With this constant or even vanishing DOS, some groups [@Lee93; @SFBN98] suggested that all quasiparticle states are localized. On the other hand, it has been shown [@BSR95] that a single unitary scattering impurity produces a zero-energy quasiparticle resonant state while the long-range overlap between these impurity states [@BS96] may lead to extended quasiparticle band near zero energy. More recently, a singularity in the DOS at zero energy was obtained by non-perturbative $T$-matrix method [@PL00] for the random distributed unitary impurities. It is noteworthy that in one dimension, there is a direct relation between the localization length and the DOS [@Thouless72]; thus a singularity in zero-energy DOS signals the delocalization in the system. However, in two dimensions, this theorem does not hold generally [@PL00]. For example, in two dimensional integer quantum Hall systems, it has been shown that the delocalization property at the quantum critical point is not changed by the changing of the DOS due to strong electron-electron interaction [@YMH95]. Therefore, the localization of quasiparticles in a $d$-wave superconductor in the presence of nonmagnetic unitary impurities is still an open question.
In this Letter, we numerically examine the disorder effect in $d$-wave superconductors with nonmagnetic impurities in the unitary limit. The quasiparticle DOS is calculated by exact diagonalization and the spin conductance is computed by the transfer matrix method. It is found that, depending on the particle-hole symmetry of the Hamiltonian, a sharp DOS peak can occur at zero energy and correspondingly the spin conductance is strongly enhanced at finite sample size. However, using one parameter scaling analysis we show that all the quasiparticle states are always localized regardless of the existence of the zero-energy peak in the DOS. In weak disorder limit, the localization length is so long that the spin conductance at small sample size remains close to the universal value $2\xi_0/a $ ($\xi_0$ is the coherence length of the superconductor and $a$ is the lattice constant) in agreement with the theoretical prediction by Lee [@Lee93]. A non-universal feature in one-parameter scaling of conductance related to the resonant DOS peak is also discussed.
We begin with a lattice Hamiltonian for the $d$-wave superconductor [@FB99] $$\begin{aligned}
H&=&-\sum_{\langle ij\rangle,\sigma }
c_{i\sigma}^{\dagger}c_{j\sigma}
+\sum _{i,\sigma} (U_i-\mu) c^{\dagger}_{i\sigma}
c_{i\sigma} \nonumber \\
&&+\sum_{\langle ij\rangle}
[\Delta _{ij} c_{i\uparrow}^{\dagger}c_{j\downarrow}^{\dagger}
+ \mbox{h.c.}]\;,
\label{EQ:Hamil}\end{aligned}$$ where $\langle ij\rangle $ refers to two nearest neighboring sites with the hopping integral taken as the unit, $\mu$ is the chemical potential and $U_{i}$ is the impurity potential. We mainly consider the unitary limit where $U_i$ takes a nonzero value $U_0$ only at a fraction $n_i$ of the sites which are randomly distributed in space. The $d$-wave symmetry is imposed by choosing order parameters: $\Delta_{i,i\pm \hat {x}}=-\Delta_{i,i\pm \hat{y}}=\Delta_{d}$, which yields the excitation spectrum $E_{k}=\sqrt{\epsilon_{k}^{2}+\Delta_{k}^{2}}$, with $\epsilon_{k}=-2(\cos k_{x} +\cos k_{y})-\mu$ and $\Delta_{k}=2\Delta_{d}
(\cos k_{x}-\cos k_{y})$. Therefore, gapless quasiparticle states exist along the direction $\vert
k_{x}\vert =\vert k_{y}\vert$ in the momentum space. Unless otherwise stated, $\mu=0$, $\Delta_d=0.1$, and $U_0=100$ are taken throughout the work. In the presence of a single impurity in the unitary scattering limit, earlier study has shown [@ZLTH00] that the order parameter is strongly suppressed near the impurity site on a scale of a few lattice constant. To take into account this effect, the order parameters on bonds connecting with the strong impurity sites are taken as zero. By exactly diagonalizing the Hamiltonian (1), one can calculate the quasiparticle DOS, which is defined as [@AHM00; @note1] $$\rho(E)=\frac{1}{N_L}\sum_{n}\delta(E-E_{n})\;,$$ where $N_L=L\times L$ with $L$ the linear dimension of the system in units of lattice constant ($a=1$). In Fig. \[FIG:DOS\](a), the DOS is plotted as a function of quasiparticle energy $E$ with impurity density $n_i=0.1$ ($\mu=0$ and $\Delta_{d}=0.1$) at $N_L=90\times 90$. In the calculation the periodic boundary condition is used and it has been checked that the results do not depend on the boundary condition for large $L$ considered here. As shown in Fig. \[FIG:DOS\](a), we find that a sharp zero-energy peak shows up in the DOS, which can be fitted by the analytical form $c_0 n_i/2|E|(\ln^{2}|E/E_g|+\pi^2 /4)$ from Ref. [@PL00] with $c_0=0.66$ (the superconducting gap $E_g \sim 4
\times \Delta _d$) as shown in the inset of Fig. \[FIG:DOS\](a). The fitting breaks down at an energy scale close to $1/N_L$ [@note2]. The strength of the zero-energy peak is reduced when $n_i$ is changed to $0.04$ as shown in $E<0$ part of Fig. \[FIG:DOS\](b). The overall shape of the peak is sample size independent (from $N_{L}=25\times 25$ to $120\times 120$ as well as a strip system $30\times 300$) with the DOS value at the peak position increasing with $N_L$ very slowly. Note that in the presence of disorder, the order parameter $\Delta _{ij}$ is in principle subject to the self-consistency condition: $\Delta_{ij}=-g_{ij}\langle c_{j \downarrow}c_{i\uparrow}\rangle $ where $g_{ij}=g_0$ is the attracting interaction for $d$-wave pairing. We have also calculated the DOS by diagonalizing the Hamiltonian Eq. (\[EQ:Hamil\]) self-consistently for each disorder configuration with $N_{L}=25 \times 25$. The obtained DOS is also shown in Fig.1(b) ($E>0$ part), which is averaged over 50 impurity configurations and $8\times 8$ wavevectors in the supercell Brillouin zone. As can be seen, the DOS value at $E_g$ is reduced due to the suppression of $\Delta_{ij}$ around each impurity site. However, all other features remain essentially unchanged (compared to the $E<0$ part of Fig. 1(b)).
The presence of zero-energy peak in the DOS crucially depends on the symmetry of the Hamiltonian. Since a repulsive or attractive impurity center with infinite strength (i.e., unitary limit) is equivalent to the exclusion of a lattice site, both the local and global particle-hole symmetry remain if the chemical potential $\mu=0$, which then produces a resonant peak at $E=0$ [@PL00]. This feature was not exhibited in earlier works because of either the breaking of local particle-hole symmetry by the soft impurity scattering [@NT97; @Mudry96; @SFBN98] or the breaking of the band particle-hole symmetry by considering $\mu\neq 0$ [@AHM00], which indicates the importance of the realization of disorder model. In Fig. \[FIG:DOS\](c), the DOS is presented for random on-site disorders with $U_i$ uniformly distributed between \[-1,1\] with a homogeneous order parameter at all sites. Due to the absence of the local particle-hole symmetry in this case, $\rho(E)$ has a finite value at low energy region down to a mesoscopic scale $E\sim 1/N_L$, below which it shows a zero-energy dip, in agreement with those of the non-perturbative calculations [@NT97; @Mudry96; @SFBN98]. We have also relaxed the unitary scattering limit by taking $U_{0}=10$ (comparable to the band width) or (and) broken the band particle-hole symmetry by taking $\mu=-1$, and found that the DOS at $E=0$ is always strongly suppressed, which is similar to the results obtained in Ref. [@AHM00]. There is a smooth crossover from the zero-energy peak to dip in the DOS with the varying of model parameters as long as the Hamiltonian is driven away from the perfect particle-hole symmetry. Therefore, our numerical result indicates that different realizations of disorder give rise to different profiles of the DOS as the energy approaches to the Fermi level.
In the absence of the zero-energy resonant peak in DOS due to the breaking of band particle-hole symmetry by $\mu\neq 0$, Franz [*et al.*]{} [@FKB96] have studied the similar problem by examining the sensitivity of the wave function to the boundary conditions and by analyzing the finite-size dependence of inverse participation ratios, and presented a strong evidence for the localization of low energy quasiparticles. Here we are concerned with the question of quasiparticle localization or delocalization in the disordered $d$-wave superconductor when the zero-energy peak appears in the DOS. We employ the transfer-matrix method to calculate the finite-size localization length and the longitudinal conductance. We consider the quasi-one-dimensional strip sample with the length $L\ge 10^5$ and width $M$. The quasiparticle wavefunction amplitudes in the $ix$-th and $(ix+1)$-th slices satisfy the following equation: $$\left( \begin{array}{c}
\hat{\phi}_{ix+1} \\
\hat{\phi}_{ix}
\end{array}
\right) = T_{ix}
\left( \begin{array}{c}
\hat{\phi}_{ix} \\
\hat{\phi}_{ix-1}
\end{array}
\right)
\;,$$ where $\hat{\phi}_{ix}$ is a $2M$-component vector of the Bogoliubov amplitudes for quasiparticle states, and $T_{ix}$ is a $4M\times 4M$ transfer matrix. The transfer matrix through the whole system, $P_{L}=\prod_{ix=1}^{L} T_{ix}$, has a set of $2M$ pairs of Lyapunov exponents, which determine the inverse of length $\lambda_{i}$ ($i=1,2,\dots,2M$). The usual orthonormalization procedure is taken [@MK83] in our calculation. Correspondingly, the longitudinal conductance $g_s$ extrapolated for square sample with width $M$ is given by [@Beenakker97]: $$g_{s}(M,n_{i})=\sum_{j=1}^{2M} \cosh^{-2}\Lambda_{j}\;,$$ where $\Lambda_{j}=\lambda_{j}/M$. Note that $g_s$ corresponds to the spin conductance as the spin carried by quasiparticle is conserved [@SFBN98]. As shown in Fig. \[FIG:SCALING\](a), $g_s$ as a function of $M$ monotonically decreases with the increase of $M$ at $E=0$ for each selected impurity density from $n_i=0.01$ to $0.16$, consistent with localization in the large $M$ limit. All the data between $M=32$ and $M=120$ at different $n_i$ can be collapsed onto a single curve: $$g_{s}(M,n_i)=f\left(\frac{\xi(n_i)}{M}\right)\;,$$ as shown in Fig. \[FIG:SCALING\](b) in accordance with the one-parameter scaling law. Here $\xi(n_i)$ is the thermodynamic localization length which only depends on $n_i$ as shown in the inset of Fig. \[FIG:SCALING\](b), and it remains finite for all the disorder density $n_i$, suggesting that all the states are localized even in the unitary limit with the presence of the zero energy resonant peak. In addition, we display in Fig. \[FIG:SPC\] the conductance $g_s$ as a function of quasiparticle energy $E$ at sample width $M=48$ and $96$ with $n_i=0.1$. It is found that $g_s$ is strongly enhanced as $E\rightarrow 0$ in the region $E\sim 0.01$ corresponding to the width of the DOS peak while away from this region $g_s$ generally increases with the increase of $E$. For quasiparticle at low energy with $E<0.1 $, it has been found that $g_s$ always decreases with the increase of $M$ and all the states are localized in large $M$ limit. However, the scaling curve $g_s(M,n_i)=f(\frac{\xi(n_i)}{M})$ found at $E$ away from the DOS peak ($E>0.01$) is different from that for $E=0$ at the peak of the DOS, indicating the breaking-down of the universal one-parameter scaling law due to the presence of the resonant peak in the DOS. In addition, for all values of $\Delta_d$, $g_s$ decreases with $M$, implying localization in all the parameter region. At fixed $M$, $g_s$ always decreases monotonically with the increasing $\Delta_d$, in agreement with the general argument that quasiparticle states in a superconducting phase are always more localized [@Lee93] than the corresponding normal state. Given the localization of the quasiparticle states, the spin conductance at the longest length scales must vanish. However, in the weak disorder (the impurity density $n_i \le 4\%$), i.e., the Born limit, we found that the localization length is so long that at small sample size $M\sim 32$, the spin conductance $g_s^0$ follows the universal form [@Lee93] $g_s^0=2\xi_0/a$ as the coherence length $\xi_0 $ is changed from $1.6 $ to $6.4$ by changing $\Delta_d$ between $0.05$ and $0.20$. For the strong disorder, the conductance is generally smaller than the Born limit value due to the onset of localization effect.
In conclusion, we have studied the quasiparticle states in 2D $d$-wave superconductors with randomly distributed strong impurities in the unitary scattering limit. As the particle-hole symmetry holds, a very sharp DOS peak is obtained at zero energy. Such a DOS peak enhances the finite-size conductance. However, using one-parameter scaling analysis we have shown that all the quasiparticle states at low energy are still localized and the localization effect is generally enhanced with the increase of superconducting order parameter.
[**Note added**]{}: After the submission of this paper, we received a preprint from Atkinson [*et al.*]{} [@AHMZ00] where similar results for the DOS in the unitary limit of the symmetric band were obtained.
[**Acknowledgments**]{} - The authors would like to acknowledge A. V. Balatsky, M. P. A. Fisher, P. A. Lee, X.-G. Wen, and Z. Y. Weng for helpful and stimulating discussions. JXZ also thanks W. A. Atkinson for patiently explaining their work to him. This work is supported by the State of Texas through ARP Grant No. 3652707, the Texas Center for Superconductivity at University of Houston, and the Robert A. Welch Foundation.
L. P. Gor’kov and P. A. Kalugin, Pis’ma Zh. Eksp. Teor. Fiz. [**41**]{}, 208 (1985) \[JETP Lett. [**41**]{}, 253 (1985)\].
S. Schmitt-Rink, K. Miyake, and C. M. Varma, Phys. Rev. Lett. [**57**]{}, 2575 (1986).
P. J. Hirschfeld, D. Vollhardt, and P. Wölfe, Solid State Commun. [**59**]{}, 111 (1986); Phys. Rev. B [**37**]{}, 83 (1988).
P. A. Lee, Phys. Rev. Lett. [**71**]{}, 1887 (1993).
Y. Hatsugai and P. A. Lee, Phys. Rev. B [**48**]{}, 4204 (1993).
D.-H. Lee and Z. Wang, Phil. Mag. Lett. [**73**]{}, 145 (1995).
T. Xiang and J. M. Wheatley, Phys. Rev. B [**51**]{}, 11721 (1995).
R. Joynt, J. Low Temp. Phys. [**109**]{}, 811 (1997).
K. Ziegler, M. H. Hettler, and P. J. Hirschfeld, Phys. Rev. Lett. [**77**]{}, 3013 (1996).
A. A. Nersesyan and A. M. Tsvelik, Phys. Rev. Lett. [**78**]{}, 3981 (1997); A. A. Nersesyan, A. M. Tsvelik, and F. Wenger, Nucl. Phys. B [**438**]{}, 561 (1995); Phys. Rev. Lett. [**72**]{}, 2628 (1994).
C. Mudry, C. Shamon, and X.-G. Wen, Nucl. Phys. B [**446**]{}, 382 (1996).
W. A. Atkinson, P. J. Hirschfeld, and A. H. MacDonald, cond-mat/0002333; A. Ghosal, M. Randeria, and N. Trivedi, cond-mat/0004481.
T. Senthil [*et al.*]{}, Phys. Rev. Lett. [**81**]{}, 4704 (1998); T. Senthil and M. P. A. Fisher, Phys. Rev. B [**60**]{}, 6893 (1999).
A. V. Balatsky, M. I. Salkola, and A. Rosengren, Phys. Rev. B [**51**]{}, 15547 (1995).
A. V. Balatsky and M. I. Salkola, Phys. Rev. Lett. [**76**]{}, 2386 (1996).
C. Pépin and P. A. Lee, cond-mat/0002227; Phys. Rev. Lett. [**81**]{}, 2779 (1998).
D. J. Thouless, J. Phys. C [**5**]{}, 77 (1972).
S.-R. E Yang, A. H. MacDonald, and B. Huckestein, Phys. Rev. Lett. [**74**]{}, 3229 (1995).
M. E. Flatté and J. M. Byers, in [*Solid State Physics*]{} [**52**]{}, ed. H. Ehrenreich and F. Spaepen (Academic Press, New York, 1999).
Jian-Xin Zhu [*et al.*]{}, Phys. Rev. B [**61**]{}, 8667 (2000).
Notice that the band DOS defined here is related to the local density of states as, $\rho(E)=\frac{1}{2N_{L}}\sum_{\bf i}
[{\cal N}_{\bf i}(E)
+{\cal N}_{\bf i}(-E)]$, where ${\cal N}_{\bf i}=2\sum_{n}(\vert
u_{\bf i}^{n}\vert^{2} \delta(E-E_{n})
+\vert v_{\bf i}^{n}\vert^{2} \delta(E+E_{n})$ with $(u_{\bf i}^{n},v_{\bf i}^{n})$ the Bogoliubov amplitudes corresponding to the $n$-th quasiparticle eigenstate.
Due to the finite-size level repulsion effect, the DOS vanishes exactly at $E=0$. However, the integrated DOS at low energy region $|E|\sim 0.0005$ saturates to a constant in this case, suggesting the unimportance of the level-repulsion gap in the large $L$ limit.
M. Franz, C. Kallin, and A. J. Berlinsky, Phys. Rev. B [**54**]{}, R6897 (1996).
A. MacKinnon and B. Kramer, Z. Phys. B [**53**]{}, 1 (1983).
C. W. J. Beenakker, Rev. Mod. Phys. [**69**]{}, 731 (1997).
W. A. Atkinson [*et al.*]{}, cond-mat/0005487.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'HTTPS enhances end-user privacy and is often preferred or enforced by over-the-top content providers, but renders inoperable all intermediate network functions operating above the transport layer, including caching, content/protocol optimization, and security filtering tools. These functions are crucial for the optimization of integrated satellite-terrestrial networks. Additionally, due to the use of end-to-end and per-session encryption keys, the advantages of a satellite’s wide-area broadcasting capabilities are limited or even negated completely. This paper investigates two solutions for authorized TLS interception that involve TLS splitting. We present how these solutions can be incorporated into integrated satellite-terrestrial networks and we discuss their trade-offs in terms of deployment, performance, and privacy. Furthermore, we design a solution that leverages satellite broadcast transmission even in the presence of TLS (i.e. with the use of HTTPS) by exploiting application layer encryption in the path between the satellite terminal and the TLS server. Our findings indicate that even if no other operation than TLS splitting is performed, TLS handshake time, which involves roundtrips through possibly a Geosynchronous satellite, can be reduced by up to 94%. Moreover, by combining an application layer encryption solution with TLS splitting, broadcast transmissions can be exploited as well as proactive caching, content pushing, request aggregation, and other optimizations.'
author:
-
-
-
bibliography:
- 'globecom.bib'
title: Exploiting Satellite Broadcast despite HTTPS
---
Introduction
============
It is evident that the Web is becoming encrypted. Initiatives such as [*Let’s Encrypt*]{} allow hassle free and no-cost HTTPS services. Therefore, it comes to no surprise that according to Google, Chrome users spend two-thirds of their time in HTTPS pages [@goo]. Similarly, HTTPS pages receive higher ranking by search engines and browsers already mark plain HTTP pages as non-secure. At the same time, privacy concerns have led to the use of TLS even for transferring DRM protected content (e.g., the case of Netflix [@net]). Although HTTPS improves significantly end-user security and privacy, it comes with a cost: it prevents in-network functions, such as network-based security services, application-level gateways and fine-level differentiation of services, session controllers, transcoders, proxies, and caches.
In-network content manipulation is not uncommon in wireless and mobile networks and it is mainly used for optimizing network performance (as perceived by both the network operators and the end-users). Naylor et al. [@Nay2014] reported that a transparent proxy used by a major European mobile carrier, serving more than 20 million subscribers, contributes to a 2TB/day decrease of upstream traffic using caching and 28.5% decrease of last-mile downstream traffic using compression. Woo et al. [@Woo2013] reported that standard Web caching can reduce download bandwidth consumption up to 27.1%. Sivakumar et al. [@Siv2014] developed a proxy–code named PARCEL–for mobile networks that pre-fetches and pre-processes Web content: browsing 34 Web pages from the top 500 Alexa global pages using PARCEL resulted in a 49.6% reduction in page load time and 65% reduction in energy consumption. Similarly, various publications (e.g., [@Sek2011], [@Sher2012]) report that the number of middleboxes that manipulate network traffic in big enterprise networks is almost equal to the number of L2 switches and L3 routers. All these functions can be used for improving the performance of integrated satellite-terrestrial networks, as well as for decreasing the latency introduced by the satellite part of those networks. However, the use of TLS/HTTPS affects all of them. Additionally, end-to-end encryption, as well as the use of per-session encryption keys (all imposed by TLS) render broadcast communication useless, since a content encrypted for a specific session is just “junk” for all other sessions. This has a huge impact to satellite-terrestrial communications, where broadcast is widely used for delivering content.
End-to-end encryption has been a problem for Content Distribution Networks (CDNs) for a long time now. For this reason CDN providers are using solutions, such as *Custom Certificates* and *Certificate sharing* [@Lia2014] that allow them to intercept TLS connections. However, these solutions require a long-term trust relationship between the content owner and the CDN provider. Furthermore, these solutions assume that CDN providers are trusted to stop intercepting TLS traffic whenever requested by content owners. These requirements can be hardly satisfied when it comes to TLS interception in an integrated satellite-terrestrial network since satellite terminals (i.e., the location where TLS split should take place) are not so well protected (compared to a CDN node), and their operators cannot always be trusted by content owners. Therefore, using these solutions in such an architecture would create intolerable security and privacy risks.
In this paper we leverage two solutions that allow authorized TLS splitting by in-network devices and we use them in the context of integrated satellite-terrestrial networks. These solutions, namely Keyless TLS [@Ste2015] and DANE with delegation semantics [@Lia2014], allow a content owner to temporarily authorize a device to (lawfully) intercept a specific TLS session. In essence, these solutions *split* a TLS connection into two parts: one between the TLS client and the device and another between the device and the TLS server; we use these solutions in order to split a TLS connection at the satellite terminals. Additionally, we leverage the fact that due to this split in the TLS connection, the algorithms and protocols for securing the path between the satellite terminal and the TLS server are hidden from the TLS client; hence, we can apply application layer security solutions in that part: these solutions, if configured properly, do not impede satellite’s broadcasting capabilities. The contributions of our work presented in this paper are:
- We design an integrated satellite-terrestrial network architecture which incorporates TLS splitting mechanisms.
- We analyze the performance and security properties of our architecture through analysis and simulation.
- We design an extension to our architecture that uses application layer encryption in the path between the satellite terminal and the service provider, enabling solutions that take advantage of the broadcast capabilities of the satellite network.
The structure of the remainder of this paper is as follows. In Section 2 we present background information and we introduce the selected solutions. In Section 3 we present how the selected solutions can be used with integrated satellite-terrestrial networks and we evaluate our approach in Section 4. Finally, in Section 5 we provide our conclusions and plans for future work.
Background
==========
Transport Layer Security
------------------------
The primary goal of the Transport Layer Security (TLS) protocol is to provide privacy and data integrity between two communicating endpoints; a client and a server [@Die2008]. TLS enables the establishment of a secure connection that protects the confidentiality and the integrity of the transmitted data. TLS is composed of two protocols: the Handshake Protocol and the Record Protocol. The Handshake Protocol allows a client and a server to authenticate each other and to negotiate security algorithms, as well as the corresponding cryptographic keys. The TLS Record Protocol is then used for securing application layer data using the agreed keys and algorithms. The Handshake protocol is critical when it comes to TLS splitting; for this reason we provide some more details about it next.
A TLS Handshake is completed in 3 steps. The first step is the cipher suite negotiation. In this step the client and server exchange “Hello” messages and choose the cipher suite that will be used throughout a session. The second step is authentication. In TLS, a server proves its identity to the client and a client may also prove its identity to the server. Digital certificates (and their corresponding private keys) are the basis of this authentication whereas the exact method used for authentication is determined by the cipher suite negotiated. In any case, the authentication process involves the private key of a public-private key pair, owned by the authenticating entity. The final step is key exchange where the client and server exchange random numbers which combined with additional data permit the secure calculation of the session-specific shared keys.
Keyless TLS
-----------
Keyless TLS allows authorized devices to intervene in a secured connection, but without having access to any private key, hence they cannot be authenticated without the “help” of the TLS server. In particular, the intercepting device performs the TLS handshake and responds on behalf of the server, but all handshake operations requiring the private key (of the server) are relayed to the server over a dedicated secured channel: the server authenticates the intercepting device, performs the private key operations, and returns the result through the same secured channel. This process is illustrated in Figure \[keyless\]. Cloudflare [@Sou2014] and Akamai [@ger2017] are offering keyless TLS as a service.
Keyless TLS has two significant advantages: intercepting devices do not learn private keys and TLS servers participate in all session establishments, hence they can prevent at any time a device from intervening in an encrypted connection. Moreover, keyless TLS requires no modification to TLS clients. Of course, keyless TLS does not come without disadvantages. Its main drawback is the weakening of end-to-end security since, in reality, a TLS connection is split into two independent connections that involve two different TLS handshakes; one of them may result in a weak cipher which reduces the security of the end-to-end connection. According to a study [@Lia2014] performed across many major CDN services that use similar techniques this phenomenon is common and there were even cases where the connection between the intercepting CDN node and the server was not secured at all: sensitive information was transferred using plain HTTP, yet clients were under the impression they were using an end-to-end encrypted connection.
![Keyless TLS.[]{data-label="keyless"}](keyless.pdf){width="0.90\linewidth"}
DANE with delegation semantics
------------------------------
DNS-Based Authentication of Named Entities (DANE) (originally proposed in RFC 6698 [@Sch2012] and further refined in RFC 7671 [@Har2015]), allows binding a domain name to a certificate. This binding is implemented by including certificates in DNS records.[^1] In particular, a special type of DNS record, referred to as TLSA DNS, is used for associating certificates with domain names, allowing DANE-enabled TLS clients to validate TLS server certificates. The resolution of TLSA records is secured using DNSSEC. DANE with delegation semantics [@Lia2014] leverages DANE, allowing servers to add intermediate certificates to their TLSA records, which can be used by intercepting devices. This way a client may obtain a list of certificates that can be used for a particular TLS connection. Figure \[dane\] illustrates an example of a TLS handshake interception supported by DANE. As it can be observed, the intercepting device responds to the client “Hello” message with its own certificate (as opposed to the Keyless TLS case, where the intercepting device responds with the certificate of the server). Subsequently, the client performs a TLSA record resolution and validates that the received certificate is “pinned” to the server’s domain name, therefore it is approved. Finally, the client proceeds with the subsequent TLS handshake messages.
With this solution, certificate revocation is completely controlled by the origin server and can be performed by simply altering the corresponding TLSA record. Furthermore, an interesting property of this solution is that a handshake can be completed without any involvement of the original TLS server. Deploying this solution requires DNSSEC along with modifications to the certificate validation process performed by TLS clients. Of course, this approach suffers from the DNSSEC inherent problems, for example, an attacker may replay a TLSA record response related to a certificate that is not valid any more. Overhead is also added due to the extra DNS round trip.
![DANE with delegation semantics.[]{data-label="dane"}](dane.pdf){width="0.90\linewidth"}
{width="0.68\linewidth"}
TLS interception for Integrated Satellite-Terrestrial Networks
==============================================================
In this section we describe our solution for supporting TLS interception in integrated satellite-terrestrial networks. We consider a typical integrated satellite-terrestrial network architecture where a satellite terminal and a satellite gateway are responsible for connecting client applications with server applications over a satellite network (see for example Figure \[arch\]). In the following we consider that clients and servers wish to communicate over TLS and we propose two approaches that allow terminals to perform TLS interception so as (i) to accelerate the TLS handshake–which under normal circumstances has to be performed over the satellite network, and (ii) to perform content transmission optimizations (e.g., content caching, aggregation of content requests, etc.).
Our two approaches are based on the TLS interception solutions presented in the previous sections. With both solutions, whenever a terminal receives a client “Hello” TLS message (i.e., the first message of the TLS handshake) it should be able to tell (i) if it is allowed to intercept this TLS handshake, and (ii) which certificate to use. Both these problems can be solved using the Server Name Indication TLS extension [@Eas2011]. With this extension a TLS client includes in its “Hello” message the domain name of the service with which it wants to interact. We consider that satellite terminals are pre-configured with the domain names of the services they are authorized to intercept, therefore they can examine the client “Hello” message and decide whether to intercept the handshake or to forward the message to the intended recipient.[^2]
Integration with Keyless TLS
----------------------------
Keyless TLS is integrated in our architecture by implementing the Keyless TLS protocol at the terminals and the TLS servers; no modification is required to TLS clients (these components are depicted with a green box in Figure \[arch\]). Furthermore, a secure communication channel is established between the terminal and the server. This channel, which in our implementation is secured using TLS, can only be used by terminals authorized to intervene in a TLS handshake and its purpose is to protect the confidentiality of the Keyless TLS specific messages transmitted between the terminal and the server. In order to assure that only authorized terminals can access this channel TLS client authentication with certificates [@Sou2014] is used i.e., for each authorized terminal, a TLS server generates a certificate, which is installed using out-of-band mechanisms and it is used by the terminal to authenticate itself to the server when setting up the channel. The channel setup takes place only once and the same channel is used for forwarding Keyless TLS specific messages for all subsequent handshakes intercepted by the same terminal. With all these components in place, a TLS client initiates the TLS handshake, which is intercepted by the terminal and whenever the terminal requires to perform an operation using the private key of the server, it sends all necessary information to the server through the secure channel; then the server performs the necessary actions and responds back to the terminal, again through the secure channel.
An interesting property of Keyless TLS is that the server participates in all TLS handshakes. Although, this property enhances the security of the scheme (we discuss security properties in the following sections), it adds latency since the satellite network has to be used once per TLS handshake. In some cases, it is possible to compensate for this delay by “abusing” the SNI TLS extension: since the domain of the service in which the client is interested in is known, it may be possible to *push* content to a terminal together with the first Keyless TLS message. Then, when requested, this content can be served to a client directly by the terminal (which in this case acts as a transparent cache), hence the satellite link does not have to be used.
Integration with DANE with delegation semantics
-----------------------------------------------
For the integration of DANE with delegation semantics, clients implement TLS with DANE assisted certificate verification, i.e., they are able to verify the validity of a digital certificate by retrieving the corresponding TLSA DNS record (using DNSSEC). The components of this solution are depicted with a blue box in Figure \[arch\]. With this approach a secure communication channel between the terminal and the server is not required in order to complete the TLS handshake.
As already discussed, during the TLS handshake clients should perform a DNS resolution in order to validate the certificate of the terminal. In the general case, this resolution will cross the satellite network. However, this can be easily mitigated by installing DNS forwarders in satellite terminals and by configuring clients to use these forwarders as the default DNS server. The forwarders, which should implement the DNSSEC validation processes, can then cache the corresponding DNS replies. Furthermore, and since these replies concern the terminal operation, the forwarders have incentives to periodically perform DNS requests to the authoritative DNS server (even if they are not instructed by a client) in order to keep their cache fresh and up to date.
Combination with application layer encryption
---------------------------------------------
With both approaches the client ends up establishing a TLS connection with the satellite terminal (the difference lies in the fact that with Keyless TLS the client “thinks” that it is communicating to the server, whereas with DANE with delegation semantics the client knows that it communicates with the terminal, acting on behalf of the server). Clients are oblivious to the security mechanisms used for securing the communication between the terminal and the server. A typical approach would have been to use another TLS connection between these two entities; alternatively an application layer encryption mechanism can be used so that solutions that leverage a satellite’s wide area broadcasting capabilities will be able to function properly. The components of this approach are illustrated with purple colors in Figure \[arch\].
Our application encryption approach assumes that servers and (authorized) terminals share a secret key. This key is used for periodically exchanging content encryption keys. The latter keys are used for encrypting transmitted content and are common for all terminals. This is a typical mechanism used for broadcasting protected content over satellites, e.g., as used by the Conditional Access system of the Digital Video Broadcasting (DVB) standard. However, the content encryption key is not related to the encryption key used in the TLS connection between the client and the terminal. For this reason the terminals have to decrypt the received items and re-encrypt them with the corresponding TLS key.
Evaluation
==========
TLS Handshake speed improvements
--------------------------------
Compared to “vanilla” TLS, the discussed TLS interception solutions require fewer message exchanges over the satellite network. Therefore, even if the satellite terminals implement no additional content transmission optimization function, significant gains can be achieved. Figure \[evaluation\] illustrates the time required to complete a TLS handshake in a network where a terminal and a gateway are connected through a Geosynchronous satellite. The roundtrip delay of the path between the terminal and the gateway has been calculated to be 500 ms using the OpenSAND satellite network emulator [@open]. Furthermore, the roundtrip delay of a terrestrial path has been set to 20 ms. It can be observed that when a TLS interception solution is used, the TLS handshake is completed much faster. Especially, when DANE with delegation semantics is used combined with cached DNS records, the satellite network does not have to be used during a TLS handshake, hence the TLS handshake time is reduced by 94%.
![Time required to complete a TLS handshake.[]{data-label="evaluation"}](results1.pdf){width="0.90\linewidth"}
Content transmission optimizations
----------------------------------
Here we discuss some content transmission optimizations that become possible if the application layer encryption approach is used.
The application layer encryption solution used in our system is Encrypted Content-Encoding for HTTP, specified in RFC 8188 [@Tho2017]. This solution enables HTTP messages (request or response) to be encrypted using a symmetric encryption key distributed using out-of-band mechanisms. In the following we give an example that illustrates how the combination of application layer encryption with TLS interception can benefit integrated satellite-terrestrial networks. In this example we assume the architecture of Figure \[arch\]. In this architecture the terminals and the servers share a symmetric encryption key. Furthermore, the terminals include a cache and support TLS interception.
Suppose an end-user requests a piece of static content over TLS (e.g., a video file), its terminal intercepts the TLS session and forwards the request to the server. The server encrypts the requested content using the shared encryption key and transmits it over plain HTTP. The terminal re-encrypts the received content (to match the TLS encryption key) and transmits it back to the client. All other terminals can cache this piece of content (since it is being broadcast). Then, if another client, connected to (the same or) another terminal, requests the same piece of content (using TLS), its terminal can re-encrypt the cached content using the established (between them) TLS key and send it back to the client. Therefore, although there are two clients in this case, connected to (potentially) different terminals, requesting the same piece of content, there is a single content request, as well as, a single content transmission (the bulk of the data) traversing the satellite link. There are of course two separate (potentially asynchronous) data transmissions traversing short terrestrial links (using HTTPS). Nevertheless, and depending on the TLS interception approach, the satellite link may have to be used during the TLS handshake. Figure \[results2\] shows the time required to load a simple text-based Web page using HTTPS. In this figure, the bars corresponding to the case where Encrypted Content-Encoding (ECE) is used (i.e., the application layer encryption approach), the page is retrieved from a cache, otherwise it is retrieved directly from the server.
![Time to load a Web page using HTTPS.[]{data-label="results2"}](results2.pdf){width="0.90\linewidth"}
Additional advantages that can be gained, include: (i) if terminals are connected using a terrestrial network, collaborative caching solutions can be deployed, (ii) popular content (such as OS updates) can be pushed to terminals, (iii) simultaneous content requests can be batched or even aggregated, (iv) solutions that achieve better performance by manipulating the transmitted content (such as network coding) can be easily deployed. Quantifying these advantages has been left as future work.
Security considerations
-----------------------
The presented solutions are more secure compared to the certificate-based solutions used by CDNs. With the solutions presented in this paper TLS connections are intercepted only for a specific session (with Keyless TLS), or only for a specific period of time (with DANE with delegation semantics). Furthermore, with the presented solutions it is easier for a content provider to revoke the access rights of an intercepting device. Nevertheless, TLS splitting and the use of application layer encryption security solutions create some security and privacy concerns.
Intercepting devices are authorized to access content items stored under a specific domain, hence they can easily modify them. The integrity of the received items should be verified at the application layer using tools such as the “Subresource Integrity” HTML tag. Furthermore, application designers should make sure that sensitive content, including user specific information, session identifiers, cookies, etc. cannot be accessed by third parties, even if the TLS session is intercepted. One should make sure that sensitive information and less sensitive content items are stored under different domains.
When it comes to DANE with delegation semantics, there is a time frame during which a “de-authorized” device can intercept a TLS session. This duration depends on the time-to-live (TTL) of the corresponding TLSA record. Application designers should consider this performance-security trade-off and adapt TTL accordingly. Note that there is also such a time frame with Keyless TLS, related to the TTL of a TLS session, but in this case “de-authorized” devices can only intercept already established sessions and not new ones.
Application layer encryption on the other hand creates privacy concerns. Since two HTTP messages, between two entities sharing the same key, are encrypting the same (secret) information, it is possible for a malicious user that observes the network traffic to discern if the content of these messages is the same or not. Furthermore, all network fields below the application layer are transmitted in plaintext. Application designers should take special precautions in order to properly anonymize transmitted messages.
Conclusions
===========
Although end-to-end encryption enhances greatly end-user security and privacy, there are cases where access to the plaintext of the transmitted content by in-network devices is beneficial. For this reason, various solutions for enabling encrypted connection interception have been proposed. In this paper, and in the context of integrated satellite-terrestrial networks, we considered two of them, namely Keyless TLS and DANE with delegation semantics. The former solution is “pushed” by big CDN providers and requires no modification to TLS clients, whereas the latter solution is based on a promising standard and it can achieve TLS session establishment without any communication with the origin server. Both solutions can be deployed using readily available software and can co-exist with legacy TLS implementations. Furthermore, both solutions exhibit better security properties compared to the certificate-based solutions currently used by CDN providers.
The presented transport layer solutions can be combined with application layer encryption between a satellite terminal and the origin TLS server (i.e., the network path that includes the satellite network). The combination of the two approaches, creates interesting opportunities. In particular by sharing the application layer encryption key among the application server and the satellite terminals, it becomes possible to broadcast content to multiple terminals, enabling this way existing satellite-based content distribution solutions to operate with TLS clients and facilitating the deployment of novel optimization solutions, such as opportunistic caching.
Acknowledgment {#acknowledgment .unnumbered}
==============
Partial support for this work has been received by ESA ARTES FP Project “SATNEX-IV COO2.” This paper does not necessarily reflect ESA views.
The work has been performed while Luca Boero was employed at the University of Genoa.
[^1]: In particular digitally signed hashes of the certificates.
[^2]: This extension is well supported by all browsers since it is used for connecting to a TLS server hosted in a shared (e.g., Cloud) environment.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
In this paper a fully Eulerian solver for the study of multiphase flows for simulating the propagation of surface gravity waves over submerged bodies is presented. We solve the incompressible Navier–Stokes equations coupled with the volume of fluid technique for the modeling of the liquid phases with the inteface, an immersed body method for the solid bodies and an iterative strong–coupling procedure for the fluid–structure interaction. The flow incompressibility is enforced via the solution of a Poisson equation which, owing to the density jump across the interfaces of the liquid phases, has to resort to the splitting procedure of @Dodd2014 [@Dodd2014].
The solver is validated through comparisons against classical test cases for fluid–structure interaction like migration of particles in pressure–driven channel, multiphase flows, ‘water exit’ of a cylinder and a good agreement is found for all tests. Furthermore, we show the application of the solver to the case of a surface gravity wave propagating over a submerged reversed pendulum and verify that the solver can reproduce the energy exchange between the wave and the pendulum. Finally the three–dimensional spilling breaking of a wave induced by a submerged sphere is considered.
author:
- Francesco De Vita
- Filippo De Lillo
- Roberto Verzicco
- Miguel Onorato
bibliography:
- 'biblio.bib'
title: 'A fully Eulerian solver for the simulation of multiphase flows with solid bodies: application to surface gravity waves'
---
Introduction
============
Multiphase flows with fluid–structure interaction (FSI) are found in many physical and engineering areas. The computational modeling of these systems poses several challenges and different methods have been developed in the past to solve the two problems, separately. Multiphase flows involve two immiscible fluids separated by an interface; an accurate description of the motion of this surface is necessary to correctly predict the overall flow dynamics; examples of these flows are droplets and bubbles but also ocean waves.
On the other hand, in FSI problems the motion of a body determines the flow, via the boundary conditions, and, in turn, the hydrodynamic loads cause the motion of the body. Both classes of problems imply a time–dependent configuration and possibly changes of flow topology; in order to avoid the generation of body–fitted meshes, a common approach has been used which requires the solution on a Cartesian grid of the Naviers—Stokes equations coupled with additional models to describe the presence of the interfaces and solid bodies.
Immersed boundary methods (IBM) [@peskin1972] are an efficient solution for flows around moving bodies since the governing equations are solved on a fixed Cartesian grid and an additional forcing term in the momentum equation imposes the no–slip boundary conditions at the immersed surface. The direct forcing approach [@Fadlun2000] is particularly attractive due to the small limitation on the timestep and computational overhead although it suffers from spurious pressure oscillations [@Lee2011], which can be alleviated by mass corrections [@Kim2001] or extrapolation of the fluid field inside the solid [@Yang2006]. Alternatively, an Eulerian/Lagrangian approach can be used [@uhlmann2005] by describing the immersed surface through a network of Lagrangian markers and transfering the forcing between the Lagrangian and Eulerian grid by a discrete delta function; recently this approach has been modified introducing a moving–least–square interpolation for the spreading of the forcing [@Vanella2014]. A drawback of this method is that, differently from in [@Fadlun2000], it requires an iteration of the forcing step since the force spreading modifies the Eulerian points associated to each Lagrangian marker. The direct forcing approach of [@Fadlun2000] has been effectively used for simulations of flows around complex rigid bodies [@deTullio2009; @DeVita2016; @Viola2020] because in this case the rigid motion depends only on the net hydrodynamic force and the oscillations smooth out in the surface integration. On the contrary, for deformable bodies it is necessary to use a Lagrangian approach to avoid local large displacements induced by pointwise force fluctuations [@DeTullio2016].
Regarding multiphase flows, one of the most used approache is the one–fluid formulation. Within this method the two fluid phases are considered as a single contiuum with variable material properties which are discontinuous across the interface; a source term is then introduced in the momentum equation to take into account the presence of surface tension (see for example Chapter 2 of [@tryggvason2011direct]). An additional advection equation is however required to track in time the motion of the interface. The front–tracking method, first developed by [@unverdi1992], is an Eulerian/Lagrangian method used to solve the Navier–Stokes equations on a Cartesian grid and it employs a moving mesh for the interface separating the fluids; since the additional grid deforms, remeshing is necessary during the calculation. To avoid this computationally expensive step, an alternative approach is the front–capturing method, which is fully Eulerian and handles topological changes without the need of additional grids. This yields a more flexible numerical method that allows for a more efficient parallelization with respect to the Eulerian/Lagrangian counterpart. The Eulerian methods are basically the volume–of–fluid (VoF) [@scardovelli1999] and the level–set (LS) methods [@sethian1996]. VoF identifies the different phases introducing a color function; its main advantage is the intrinsic mass conservation but the geometric representation of the interface (the normal vector and the curvature) is not trivial. LS, instead, identifies the fluids using a signed distance function which allows an efficient and accurate evaluation of normal vector and curvature but it suffers from mass loss.
A common approach to advance in time the incompressible Navier–Stokes equations is the fractional–step method [@Kim1985] which consist of two main steps: the first computes a provisional velocity field by the momentum equation with the pressure at the old timestep and then projects the velocity onto a solenoidal field to enforce mass conservation. This correction step requires the solution of a Poisson equation for the pressure with the density as coefficient. Unfortunately, in all the methods previously described for multiphase flows the density is a function of the interface position, hence the Poisson equation has space–dependent coefficients which makes impossible the use of fast direct solvers (FDS) [@swarztrauber1977]. This issue is usually coped by the use of multigrid methods [@Chen1999; @Popinet2009; @Yang2009] that, however, are computationally expensive and whose convergence rate is problem dependent. Recently, @Dodd2014 [@Dodd2014] proposed a new procedure to approximate the variable coefficient Poisson equation by a constant coefficient counterpart: this is accomplished by introducing an approximation of the unknown new pressure at timestep $n+1$ which allows to split the updated pressure and density in the Poisson equation. They have shown that, by computing an approximated pressure as an extrapolation from the previous timestep, the overall $2^{nd}$ order accuracy of the method is preserved. Nevertheless, the combination of this splitting technique with the IBM is non trivial owing to the presence of the pressure in the right hand side of the Poisson equations. @Frantzis2019 [@Frantzis2019] have proposed, for a stationary solid, a local reconstruction of the pressure gradient at the boundary in order to avoid the use of solid nodes in the Poisson equations and to yield a decoupled solution of the fluid and solid domains.
To the best of the authors’ knowledge there are no solvers in literature which simulate the Navier–Stokes equations for multiphase flows with moving solid boundaries employing FDS for the solution of the Poisson equation. In order to fill this gap, the aim of this work is to present an efficient fully Eulerian solver for the simulations of multiphase flows with moving solid bodies. To achieve this goal we couple several ingredients: the presence of solid bodies in a fluid phase is described by the direct forcing approach [@Fadlun2000] combined by an interpolation scheme similar to that proposed by @Balaras2004 [@Balaras2004]; the solid phase dynamics is integrated by a $4^{th}$ order predictor–corrector scheme, based on @Hamming1959 [@Hamming1959] method. For the multiphase flow, we use the Multi-dimensional Tangent of Hyperbola Interface Capturing (MTHINC) VoF developed by [@Ii2012]. Although the method can be used for generic geometries by the ray–tracing procedure described in [@Iacca], here the analysis is limited to solid bodies which can be described by analytical formulas (*e.g.* spheres, ellipsoids *etc*): the proposed method is optimal for simulations of suspensions in multiphase flows or surface gravity waves with submerged bodies.
The paper is structured as follows: in Section \[sec:formulation\] we introduce the mathematical framework for the physical system and in Section \[sec:numerics\] we describe all the details of the numerical solver. In section \[sec:results\] we present some validation tests and two applications for a surface gravity wave propagating over a submerged reversed pendulum in two and three dimensions.
Formulation {#sec:formulation}
===========
![Sketch of a three–phase flow: a surface gravity wave propagation with a submerged body.\[fig:sketch\]](disegno){width="50.00000%"}
The basic features of the problem are sketched in figure \[fig:sketch\] with two immiscible fluids separated by an interface and one or more solid bodies. We use the one–fluid formulation of the Navier–Stokes equations and introduce an indicator (or color) function $\mathcal{H}$ to identify the fluids such that $\mathcal{H}$ is $1$ in one fluid and $0$ in the other; the interface is tracked by an additional advection equation for the indicator function. To account for the solid phase, we use the IBM which adds a forcing term to the momentum equation to impose the no–slip boundary condition at the fluid/solid boundary. The full system of equations reads:
\[eqn:fullsystem\]
[align]{} & = 0 \[eqn:mass\]\
&( + ) = -p + + + + \_ \[eqn:momentum\]\
& + = 0 \[eqn:advection\]\
&m\_n = \[eqn:veltra\]\
& = \[eqn:postra\]\
&\_n = \_n \[eqn:velrot\]\
& = \_n \[eqn:posrot\]
where $\vec{u}$ is the velocity vector, $p$ the pressure, $\rho$ and $\mu$ the density and dynamic viscosity of the fluid, $\mathbf{\tau}$ the viscous stress tensor that for a Newtonian fluid in indicial form is $\tau_{ij} = 2\mu D_{ij} = \mu(\partial u_i/\partial x_j +
\partial u_j/\partial x_i)$. $\vec{g}$ is the gravity, $\vec{f}$ the forcing term which enforces the no–slip boundary condition at the immersed boundaries (described in section \[sec:sbt\]) and $\mathcal{H}$ the indicator function. The force $\vec{f}_{\sigma}$ is the surface tension force acting at the interface between the two fluids and directed normal to the interface. The vectors $\vec{X}_n$ and $\vec{\Theta}_n$ are the linear and angular positions of the center of mass of the $n^{th}$ solid body, while $\vec{V}_n$ and $\vec{\Omega}_n$ its linear and angular velocity, $\vec{F}_n$ and $\vec{M}_n$ the external force and moment acting on the body. $m_n$ is the mass and $\mathbf{I}_n$ the moment of inertia tensor; the subscript $n$ spans all the solid bodies ($1 \leq n \leq N$). Note that lower case letters ($u$, $p$, *etc.*) are used for Eulerian fluid variables while upper case letters ($V$, $X$, *etc.*) for Lagrangian solid variables.
The fluid properties are constant within each phase and therefore are determined only by the function $\mathcal{H}$, as specified below.
The system has $5$ unknowns for the fluid phases (three velocity components, the pressure and the interface function $\mathcal{H}$) and $12N$ unknowns for the solid phase ($6$ positions and $6$ velocities). The momentum equation depends on the interface location via the material properties $\rho$ and $\mu$ and on the location of the solid phase via the forcing term $\vec{f}$. The solid body equations -, instead, depend on the flow field via the loads $\vec{F}_n$ and $\vec{M}_n$:
\[eqn:loads\] $$\begin{aligned}
&\vec{F}_n = \int_{S_n} \left(\mathbf{\tau}\cdot\vec{n} - p\vec{n}\right)\,dS \\
&\vec{M}_n = \int_{S_n} \left[\left(\mathbf{\tau}\cdot\vec{n} - p\vec{n}\right)\times \vec{r}\right]\,dS
\end{aligned}$$
where $S_n$ is the surface of the $n^{th}$ solid body, $\vec{n}$ the local outward normal and $\vec{r}$ the local distance from its center of mass. This makes the equations of the system strongly coupled and some instabilities can arise when integrating them numerically [@Borazjani2008]. In the next section we describe the steps to solve the full system of equations.
Numerical solver {#sec:numerics}
================
Flow solver
-----------
We advance in time the momentum equation by a fractional–step method [@Chorin1967]: first a provisional non–soleinodal velocity field $\vec{u}^*$ is computed using an explicit $2^{nd}$ order Adams–Bashfort scheme for the non–linear and viscous terms, yielding the semi–discrete equation for the timestep $l$: $$\label{eqn:projection}
\frac{\vec{u}^*-\vec{u}^l}{\Delta t} = -\frac{1}{\rho^{l+1}}\nabla p^l
+ \frac{3}{2}\vec{H}^l - \frac{1}{2}\vec{H}^{l-1} + \vec{g}^l + \vec{f}^l,$$ where the term $\vec{H}^l$ is given by $$\label{eqn:rhs}
\vec{H}^l = -\vec{u}^l\cdot \nabla \vec{u}^l +
\frac{1}{\rho^{l+1}}\nabla \cdot \left(2\mu^{l+1}\mathbf{D}^l\right)$$ with $\mathcal{D}^l$ the deformation rate tensor at timestep $l$ $D_{ij}^l = (\partial u_i^l/\partial x_j + \partial u_j^l/\partial x_i)$ and the material properties $\rho^{l+1}$ and $\mu^{l+1}$ are obtained after the advection of the interface, as described in the next section. The correct velocity field at timestep $l+1$ is obtained by computing a scalar function $\phi$ such that $$\label{eqn:correction}
\vec{u}^{l+1} = \vec{u}^* - \frac{\Delta t}{\rho^{l+1}}\nabla \phi^l;$$ by summing equation and it can be seen that the new pressure is given by $$\label{eqn:newp}
p^{l+1} = p^{l} + \phi^l,$$ hence $\phi^l$ represents the pressure increment at the timestep $l$, that can be computed taking the divergence of equation : $$\label{eqn:variablepoisson}
\nabla \cdot \left( \frac{\Delta t}{\rho^{l+1}}\nabla \phi^l \right) =
\nabla \cdot \vec{u}^*.$$ This is a variable coefficient Poisson equation because of the presence of the density $\rho^{l+1}$ inside the divergence operator. In order to use FDS, following @Dodd2014[@Dodd2014], the pressure gradient in the momentum equation is expressed as $$\label{eqn:splitting}
\frac{\nabla p^{l+1}}{\rho^{l+1}} \rightarrow \frac{\nabla p^{l+1}}{\rho_0}
+ \left(\frac{1}{\rho^{l+1}}-\frac{1}{\rho_0}\right)\nabla \hat{p} =
\frac{\nabla p^{l}}{\rho_0} + \frac{\nabla \phi^l}{\rho_0} +
\left(\frac{1}{\rho^{l+1}}-\frac{1}{\rho_0}\right)\nabla \hat{p},$$ where $\hat{p}$ is an approximation of the pressure $p^{l+1}$ and $\rho_0 = \min\left(\rho_1,\rho_2\right)$ is a constant density. Using this surrogate for the pressure gradient the momentum equation yields the following relation for the provisional velocity field $$\label{eqn:ustar}
\frac{\vec{u}^*-\vec{u}^l}{\Delta t} = -\frac{1}{\rho_0}\nabla p^l - \left(\frac{1}{\rho^{l+1}}-\frac{1}{\rho_0}\right)\nabla \hat{p} + \frac{3}{2}\vec{H}^l - \frac{1}{2}\vec{H}^{l-1} + \vec{g}^l + \vec{f}^l,$$ with the new Poisson equation for the pressure increment $$\label{eqn:poisson}
\nabla^2 \phi^l = \frac{\rho_0}{\Delta t}\nabla \cdot \vec{u}^*$$ and the new velocity correction $$\label{eqn:newcorrection}
\vec{u}^{l+1} = \vec{u}^* - \frac{\Delta t}{\rho_0}\nabla \phi^l.$$ The splitting procedure decouples the pressure gradient from the variable density and allows the use of FDS schemes for the solution of the Poisson equation. @Frantzis2019 [@Frantzis2019] proposed a local reconstruction of the pressure gradient at the immersed boundaries to avoid the use of points inside the solid regions. However, with the formulation –, this is not effective since, while the reconstruction proposed in [@Frantzis2019] provides a solenoidal velocity field only in the fluid domain, here the solid phase is moving and information from the solid region would enter anyway the fluid domain because of the dynamics. On the other hand we will see that, because of the fluid/structure interaction, at least two steps (predictor and corrector) and often some iterations are necessary to advance one time interval $\Delta t$, therefore the errors at the immersed boundaries are largely reduced by the multi–step procedure.
It is worth noticing that in the phase with lower density there is no approximation (the last term on the RHS of cancels out) whereas in the phase with the higher density the quality of the approximation depends on how close $\hat{p}$ is to $p^{l+1}$: in the limit $\hat{p} \equiv p^{l+1}$ again there is no approximation. @Dodd2014[@Dodd2014] have shown that extrapolating $\hat{p}$ from the steps $l$ and $l-1$, that for a constant time step integration corresponds to $\hat{p} = 2p^l - p^{l-1}$, the approximation preserves the second–order time accuracy of the scheme. Equations , and are the steps needed to advance the flow field from timestep $l$ to $l+1$. All the spatial derivatives are approximated by second–order accurate finite–difference scheme except for the viscous terms of equation where the fifth–order WENO is used [@Shu2009]; this is necessary to avoid oscillations since viscosity jumps are localized across the interface over three grid nodes. Due to the explicit treatment of the non–linear and viscous terms, the timestep is restricted for stability reasons $$\Delta t = \min\left(CFL\frac{\Delta}{|u_i|_{\max}}, C_{\nu}\frac{\Delta ^2}{\nu}\right),$$ where $\Delta$ is the grid size, $CFL$ the Courant–Friedrichs–Lewy parameter and $C_{\nu}$ a coefficient depending on the number of spatial dimensions ($1/4$ in 2D and $1/6$ in 3D). Note that even though the theoretical $CFL$ condition for the Adams–Bashfort scheme is $CFL \leq 1$, in practice the maximum is $CFL \approx 0.3$ [@VanderPoel2015]. As it will be shown in the section of the results, this condition can be further restricted for large density ratios bacause of the approximation .
Interface tracking
------------------
The method employed for the interface tracking is that proposed in [@Rosti2018b], based on the MTHINC originally developed by [@Ii2012]. Here we briefly describe only the key features of the method and the reader is referred to [@Rosti2018b] for a detailed description of the scheme and the relevant validation tests. The cell averaged value of the indicator function $\mathcal{H}$ is defined as the volume fraction of a fluid phase (or volume of fluid, VoF) within a control volume $\delta V$ $$\mathcal{F}\left(\vec{x},t\right) \equiv \frac{1}{\delta V}
\int_{\delta V}\mathcal{H}\left(\vec{x},t\right) dV,$$ being $0 \le \mathcal{F}\left(\vec{x},t\right) \le 1$. The advection equation for the VoF function is then $$\label{eqn:vof}
\frac{\partial \mathcal{F}}{\partial t} +
\nabla \cdot \left(\vec{u}\mathcal{H}\right) = \mathcal{F}\vec{u}.$$ The key point of the MTHINC method is the approximation of the color function by an hyperbolic tangent $$\mathcal{H}(\vec{X}) \approx \hat{\mathcal{H}}(\vec{X}) =
\frac{1}{2}(1 + \text{tanh}(\beta(P(\vec{X})+q))),$$ where $\vec{X} \in [0,1]$ is a local coordinate system, $\beta$ a sharpness parameter, $q$ a shift and $P$ a three–dimensional surface function which can be either linear (plane) or quadratic (curved surface) at no additional cost. This discretization allows to solve the fluxes of equation by integration of the approximated color function in each computational cell. The material properties of the two fluids are connected to the VoF function $\mathcal{F}$ as follows: $$\label{eqn:properties}
\begin{aligned}
\rho(\vec{x},t) = \rho_1\mathcal{F}(\vec{x},t) + \rho_0(1-\mathcal{F}(\vec{x},t)), \\
\mu(\vec{x},t) = \mu_1\mathcal{F}(\vec{x},t) + \mu_0(1-\mathcal{F}(\vec{x},t)),
\end{aligned}$$ where the subscript $1$ stands for phase where $\mathcal{F}$ is equal to $1$ and the subscript $0$ for the other phase. The surface tension force $\vec{f}_{\sigma}$ can be numercially described usgin the Continuum Surface Force method [@Brackbill1992] for which $\vec{f}_{\sigma} = \sigma \kappa \nabla \mathcal{H}$, with $\sigma$ the surface tension coefficient and $\kappa$ the local curvature of the interface. The solver has been tested in [@Rosti2018b] and recently used in [@Rostihst2019; @devita2019].
Solid boundary treatment {#sec:sbt}
------------------------
The interaction between the fluid and the solid phases is dealt by the Immersed Boundary method [@peskin1972]. This technique avoids the use of body conforming meshes by introducing the term $\vec{f}$ in the momentum equation which enforces the fluid velocity to be equal to that of the solid (no–slip boundary condition). Essentially, the IBM consists of two steps: *i)* evaluation of the forcing term $\vec{f}$; *ii)* evaluation of the hydrodynamic loads . Several methods have been proposed to implement these two steps, see for example [@Fadlun2000; @Balaras2004; @uhlmann2005; @Vanella2009; @Breugem2012; @DeTullio2016]. In the direct forcing approach [@Fadlun2000], the term $\vec{f}$ is computed by imposing that the velocity $\vec{u}^*$ matches the solid velocity $\vec{V}$ at the solid nodes $$\frac{\vec{V}^{l+1}-\vec{u}^l}{\Delta t} = -\frac{1}{\rho^{l+1}}\nabla p^l - \left(\frac{1}{\rho^{l+1}}-\frac{1}{\rho_0}\right)\nabla \hat{p} + \frac{3}{2}\vec{H}^l - \frac{1}{2}\vec{H}^{l-1} + \vec{g}^l + \vec{f}^l$$ which yields for the force $$\vec{f}^l = \frac{\vec{V}^{l+1}-\vec{u}^l}{\Delta t} + \frac{1}{\rho^{l+1}}\nabla p^l + \left(\frac{1}{\rho^{l+1}}-\frac{1}{\rho_0}\right)\nabla \hat{p} - \frac{3}{2}\vec{H}^l + \frac{1}{2}\vec{H}^{l-1} - \vec{g}^l$$ with the velocity $\vec{V}^{l+1}$ computed by and . In practice this corresponds to locally reconstructing the velocity field at the nodes next to the solid boundary (forcing nodes in figure \[fig:forcing\]). It is worth noticing that, after the correction step , which enforces the free divergence of the velocity field, the no–slip boundary condition might be violated. However, it has been shown that, provided the first external node is within the linear part of the boundary layer profile, the error is small and it does not affect significantly the solution [@Fadlun2000]. It could be anyway further reduced by iterating between equations , and - until the desired convergence is achieved.
[0.4]{} ![a) Definition of point tags: solid triangles are velocity solid points, empty triangles are fluid velocity points, red and blue triangles are forcing points for the horizontal and vertical velocity, respectively; solid circles are solid pressure points and empty circles are fluid pressure points. b) Definition of the interpolation stencil: forcing point (blue triangles), auxiliary point (empty diamond), solid boundary point (black diamond).\[fig:punti\]](puntiforza.pdf "fig:"){width="\textwidth"}
[0.4]{} ![a) Definition of point tags: solid triangles are velocity solid points, empty triangles are fluid velocity points, red and blue triangles are forcing points for the horizontal and vertical velocity, respectively; solid circles are solid pressure points and empty circles are fluid pressure points. b) Definition of the interpolation stencil: forcing point (blue triangles), auxiliary point (empty diamond), solid boundary point (black diamond).\[fig:punti\]](interpolazione.pdf "fig:"){width="\textwidth"}
Since the Cartesian grid does not conform to the solid body, the forcing points do not lie exactly at the boundary, as shown in figure \[fig:forcing\]. For this reason a geometrical description of the solid is required to identify the forcing points and to interpolate the velocity in these points. The proposed scheme has been applied to bodies whose surface could be described by an analytical function since this allows a significant reduction of the computational cost. It is worth mentioning, however, that the method could be easily extended to general complex geometries by using a ray tracing procedure as described in [@Iacca].
If, as in the present case, the solid body is a cylinder or a sphere, the geometry is simply defined by the position of the center and a distance from it. We introduce a signed distance function $d$ (positive in the fluid region and negative in the solid) from the surface of the body for every computational node. In the same way we can define the local outward normal, by differentiation of the distance function $\vec{n} = \nabla d$. Starting from the distance field we can tag all the computational velocity points as follows: if $d < 0$ the point is solid and tagged with a $0$ flag (black triangles in figure \[fig:forcing\]). Points where $d \ge 0$ and at least one neighbor with a negative distance are tagged as interface points with a flag $1$ (red and blue triangles in figure \[fig:forcing\]); all other points with $d > 0$ are tagged as fluid with a flag $2$ (empty triangles in figure \[fig:forcing\]). Pressure points, at the cell center, are tagged only as solid or fluid without the interface flag. For all the interface points the velocity needs to be locally reconstructed in order to apply the no–slip boundary condition. This is done, similarly to [@Balaras2004], by interpolation between the velocity value on the solid body and a virtual point in the fluid domain, as shown in figure \[fig:interpolation\]. Starting from the forcing node, following the local normal, a support of $4$ total adjacent points ($8$ in 3D) is constructed. If all the points, except the forcing one, are fluid (as for $A_1$ in figure \[fig:forcing\]) then the auxiliary point is located at a distance $h$ equal to the distance between the forcing node and the solid boundary. The velocity is computed by bi–linear interpolation at the auxiliary point and the desired velocity at the forcing node by linear interpolation between this auxiliary and the boundary nodes (black diamond in figure \[fig:forcing\]). In case one of the adjacent points is another forcing node (for example $A_2$ has as neighbor $A_1$), then the auxiliary point is found by intersection of the local normal and the Cartesian grid. The velocity here is computed using the adjacent points and then the velocity on the forcing point is evaluated as done for the previous case. This method provides the value of the velocity $\vec{V}^{l+1}$ on the forcing node to impose the desired boundary condition, and it avoids the inversion of a matrix, as proposed in [@Yang2006]. Note that the interpolation is performed only for the interface points, while for solid ones the velocity is imposed directly as the solid body velocity $\vec{V}^{l+1}$.
Hydrodynamic forces and solid body dynamics
-------------------------------------------
The dynamics of the solid phase is governed by the hydrodynamic forces which are computed by integrating over the body surface the integral of viscous stress and pressure. Given that velocity gradients and pressure are not defined exactly at the immersed surface also in this case interpolation is necessary. This is accomplished by considering virtual probes in the fluid domain where $\tau$ and $p$ are computed and then projected along the normal direction onto the surface. Provided the first external point is within the linear part of the boundary layer velocity profile, the viscous stress on the surface is then equal to the value at the probe. Instead, the pressure on the surface is evaluated imposing the boundary condition $$\frac{\partial p}{\partial n} = -\rho \frac{D\vec{u}}{Dt}\cdot\vec{n} + \rho \vec{g}\cdot\vec{n},$$ which is obtained by projecting the momentum equation in the wall normal direction. The surface pressure is then $$p = p_s + h_s\left(\rho \frac{D\vec{u}}{Dt}\cdot\vec{n} - \rho \vec{g}\cdot\vec{n}\right),$$ being $p_s$ the pressure at the probe and $h_s$ the distance between the probe and the immersed surface. The probes are initially located at a distance $\Delta$ from the surface and a support of four nodes (eight in 3D) , as in figure \[fig:interpolation\], is built in the wall normal direction. If these points are all fluid nodes then $\tau$ and $p$ are computed in the probe by bilinear interpolation, otherwise the probe distance is progressively increased by steps of $0.2\Delta$ until the condition on the support is verified. The number of probes is computed by considering a spacing equal to $\Delta$. Once the forces are known, equations - are advanced in time using a Hamming $4^{th}$ order modified predictor–corrector method [@Hamming1959], described in appendix \[sec:A\].
Summary of the algorithm
------------------------
To summarize the algorithm let’s define $\vec{\mathcal{X}} = \left[\vec{X}, \vec{\Theta}\right]^T$ and $\vec{\mathcal{V}} = \left[\vec{V}, \vec{\Omega}\right]^T$ the generalized position and velocity for each solid body; for every timestep (after the third), starting from $\vec{u}^l$, $p^l$, $\mathcal{F}^l$, $\vec{\mathcal{X}}^l$ and $\vec{\mathcal{V}}^l$, the procedure is the following:
1. compute the predicted imporved solution $\left(\vec{\mathcal{X}_m}^{l+1},
\vec{\mathcal{V}_m}^{l+1}\right)$;
2. solve equation to compute $\mathcal{F}^{l+1}$;
3. update fluid properties with ;
4. compute the provisional velocity field ;
5. reconstruct the velocity field to impose the no–slip boundary condition given by $\left(\vec{\mathcal{X}_m}^{l+1},\vec{\mathcal{V}_m}^{l+1}\right)$ on the solid boundary;
6. solve Poisson equation to compute $\phi^l$;
7. compute $p^{l+1}$ by and $\vec{u}^{l+1}$ by ;
8. compute the hydrodynamic loads $\vec{F}^l_1$ and $\vec{M}^l_1$ by ;
9. perform the correction step of the Hamming solver and compute ; $\left(\vec{\mathcal{X}_1}^{l+1},\vec{\mathcal{V}_1}^{l+1}\right)$;
10. repeat steps from 2. to 9. with boundary condition $\left(\vec{\mathcal{X}_1}^{l+1},\vec{\mathcal{V}_1}^{l+1}\right)$ for step 5. and check for convergence on $\left(\vec{\mathcal{X}_k}^{l+1},\vec{\mathcal{V}_k}^{l+1}\right)$;
11. if not converged continue to iterate over steps 2. to 9.;
12. if converged, compute the final solution $\left(\vec{\mathcal{X}}^{l+1},\vec{\mathcal{V}}^{l+1}\right)$ and repeat steps 2. to 8. with this boundary condition for the velocity field.
Note that the strong coupling (*i.e.* the iterative solver) is necessary only for large density ratio between fluid and solid; in other cases it is possible to use a direct solver performing steps 1-8. When using the strong coupling, the first three timesteps are performed with a forward Euler, Adam-Bashfort 2nd order and Adam-Bashfort 3rd order method, respectively; a tolerance of $10^{-6}$ for convergence is usually small enough for an accurate solution. In case of large acceleration under-relaxation can be used to stabilize the simulation by replacing the force in the corrector step with $F = (1-\chi)F_k + \chi F_{k-1}$, with $\chi = [0.1, 0.3]$ [@deTullio2009]. For the descirption of the Hamming’s method see appendix A.
Results {#sec:results}
=======
Validations {#sec:validations}
-----------
We present here some classical tests either for the immersed boundary method and for its interaction with an interface separating two fluids. We also verify that the method is able to compute the proper energy decay of surface gravity waves and the oscillation period of a submerged pendulum. Note that in all simulations we have set surface tension coefficient to zero to avoid further restrictions on the timestep.
### Particle migration
[0.45]{} ![Lateral migration of a circular particle in a pressure–driven channel at $Re = 12.78$ and $96.74$: (a) vertical $vs$ horizontal position; solid line for the present results and symbols for [@Pan2002]. (b) Effect of the splitting : case without splitting (solid purple line), case with splitting and $\Delta t$ ten times smaller (dashed green line), case with splitting and the same $\Delta t$ (dotted blue line).\[fig:Pan\]](test1a.pdf "fig:"){width="\textwidth"}
[0.45]{} ![Lateral migration of a circular particle in a pressure–driven channel at $Re = 12.78$ and $96.74$: (a) vertical $vs$ horizontal position; solid line for the present results and symbols for [@Pan2002]. (b) Effect of the splitting : case without splitting (solid purple line), case with splitting and $\Delta t$ ten times smaller (dashed green line), case with splitting and the same $\Delta t$ (dotted blue line).\[fig:Pan\]](test1b.pdf "fig:"){width="\textwidth"}
In the first test case we consider the migration of a neutrally buoyant particle in a two–dimensional pressure–driven Poiseuille flow [@Pan2002]. The motion of the particle is a consequence of pressure distribution, inertia and rotational lift; an accurate evaluation of the forces is necessary to properly simulate the particle dynamics. We consider a square domain of size $L = 1$, periodic in the horizontal direction and vertically bounded by two solid walls. Initially the flow is at rest and a pressure gradient is applied in the horizontal direction. The circular particle has a radius of $0.125L$ and it is initially located at a vertical distance of $0.4L$ from the lower plate. Two different cases are considered by varying the viscosity of the fluid $\mu$, which correspond to two different Reynolds numbers, namely $Re = 12.78$ and $96.74$. The Reynolds number is based on the space–averaged inlet velocity $\tilde{u}$ and the size of the domain $L = 1$, *i.e.* $Re = \rho L\tilde{u}/\mu$, the density of the particle $\rho_p$ is the same as that of the fluid $\rho$. For the lower value of $Re$ we use a grid of $96\times96$ points while for the higher $Re$ the grid is $192\times192$. The particle migrates from the initial to an equilibrium position close to the lower wall, as shown in figure \[fig:Pancomp\]. A good agreement for both Reynolds numbers, with the results of [@Pan2002], is found.
To show the effect of the splitting procedure we simulate the case at $Re = 12.78$ by introducing a minimum density equal to $\rho_0 = 0.001$, which corresponds to a case with two fluids with density ratio 0.001, so that the term involving $\hat{p}$ is non-zero. Due to the small value of $\rho_0$, it is necessary to decrease the timestep in order to recover the approximation between $\hat{p}$ and $p^{l+1}$, as clearly shown in figure \[fig:Pandt\]. The timestep restriction depends on the density ratio and we have found that a reduction of $10$ times for a density ratio of $0.001$ is enough while for bigger density ratios, for example $0.01$, already halving the timestep yields the correct time evolution of the particle. This is consistent with the results of [@Frantzis2019]. For this test the direct solver is stable and it is not necessary to use the iterative solver.
### Water exit
[0.3]{} ![Water exit test case: a) sketch of the computational domain; b)-c) interface deformation with $\epsilon = 0.8$ and $Fr = 0.39$: present results (open circles), data from [@Greenhow1997] (solid circles); lengths are made non–dimensional by the initial distance $d$.\[fig:Greenhow\]](test20.pdf "fig:"){width="\textwidth"}
[0.3]{} ![Water exit test case: a) sketch of the computational domain; b)-c) interface deformation with $\epsilon = 0.8$ and $Fr = 0.39$: present results (open circles), data from [@Greenhow1997] (solid circles); lengths are made non–dimensional by the initial distance $d$.\[fig:Greenhow\]](test2a.pdf "fig:"){width="\textwidth"}
[0.3]{} ![Water exit test case: a) sketch of the computational domain; b)-c) interface deformation with $\epsilon = 0.8$ and $Fr = 0.39$: present results (open circles), data from [@Greenhow1997] (solid circles); lengths are made non–dimensional by the initial distance $d$.\[fig:Greenhow\]](test2b.pdf "fig:"){width="\textwidth"}
The next test case consists of a cylinder rising close to the interface separating water and air, as in [@Greenhow1997]. The cylinder has radius $a$, it is at a distance $d$ below the still water level, and rises with a constant velocity $V$. The inviscid problem is governed by two nondimensional parameters, $\epsilon = a/d = 0.8$ and the Froude number $Fr = V/\sqrt{gd} = 0.39$, with $g$ the acceleration of gravity. The material properties ratio is that of water and air, *i.e.* $\rho_w/\rho_a = 850$ and $\mu_w / \mu_a = 1.92\text{x}10^{-2}$. Note that the reference solution of [@Greenhow1997] has been derived from the potential theory for inviscid fluids; the present solver, instead, is designed for viscous flows therefore we have selected a Reynolds number high enough to minimize viscous effects but, at the same time, small enough to limit the computational effort. In the real air/water case it results $Re \approx {\cal O}(10^5)$; here we have set the water viscosity so to have a Reynolds number $Re = \rho_w V a / \mu_w = 1000$ which proved to be enough for our purposes. The domain is $40 a$ wide and $20 a$ high, with a water depth of $16 a$ and it is discretized by a grid of $4096\times2048$ points. We compare our results with the numerical ones proposed by [@Greenhow1997], reported in figure \[fig:Greenhow\] at fixed non–dimensional times of $T = tV/d = 0.4$ and $0.6$. The results, in terms of interface deformation, are in very good agreement with those of [@Greenhow1997]. Also for this test the direct solver is enough stable and it is not necessary to use the iterative solver.
### Viscous damping of surface gravity waves
Surface gravity waves propagating in inviscid fluids can be described using the potential theory which yields an irrotational velocity field. For viscous fluids, instead, there is an additional rotational flow given by the presence of the boundary layer at the free–surface. For waves of small amplitude, *i.e.* waves for which $a << \lambda$, with $a$ the wave amplitude and $\lambda$ its wavelength, the rotational flow is confined in a small layer across the surface while the rest of the flow can be still described by the potential theory. Under this assumption the total wave energy decay in time, given by the sum of the kinetic and potential contributions, can be computed from the potential theory (see @Landau [@Landau] chapter II section 25) and it results in an exponential decay of the form $E(t) = E(t=0)e^{-2\gamma t}$ with the coefficient $\gamma = 2 \nu k^2$, being $\nu$ the kinematic viscosity of the fluid and $k = 2 \pi /\lambda$ the wavenumber. The kinetic energy contribution is given by $$\label{eqn:kin}
E_k(t) = \int_0^{\lambda} \int_{-h}^{\eta} \frac{1}{2}\rho \vec{u}^2\,dxdy$$ with $h$ the fluid depth at rest and $\eta$ the free–surface level with respect to the equilibrium position ($z=0$). The potential energy is $$\label{eqn:pot}
E_p(t) = \int_0^{\lambda} \int_{-h}^{\eta} \rho gz\,dxdy - E_{p0}$$ with $E_{p0} = \int_0^{\lambda} \int_{-h}^{0} \rho gz\,dxdy = -\rho g\lambda h^2/2$ the potential energy of the still water level. We simulate the propagation of a surface gravity wave of wavelength $\lambda$ in a periodic square box of size $\lambda$ with a fluid depth of $h = \lambda / 2$ and a steepness $\varepsilon = a k = 0.05$. The density and viscosity ratios are those of water and air, $1/850$ and $1.96 \text{x} 10^2$, respectively, while the Reynolds number based on the phase velocity $c_f = \sqrt{g \lambda}$ and on the wavelength $\lambda$ is $Re = g^{0.5} \lambda^{3/2} / \nu = 1.0 \text{x} 10^4 $. The initial conditions for the surface elevation $\eta$ and the velocity field $\vec{u} = (u,w)$ are given by the linear theory:
$$\begin{aligned}
\eta(x,0) &= a \cos\left(kx\right), \\
u(x,y,0) &= \left(1-\mathcal{H}\right)a\omega e^{ky} \cos\left(kx\right) - \mathcal{H}a\omega e^{-ky} \cos\left(kx\right), \\
v(x,y,0) &= \left(1-\mathcal{H}\right)a\omega e^{ky} \sin\left(kx\right) + \mathcal{H}a\omega e^{-ky} \sin\left(kx\right),
\end{aligned}$$
with $\mathcal{H} = 0$ for the water and $\mathcal{H} = 1$ in the air. For stability reason the initial VoF function $\mathcal{H}$ is filtered one time using bilinear interpolation. In figure \[fig:decay\] (left) we report the vorticity contours after one period, which clearly shows its concentration in a small layer below the water surface; in the right panel we show the energy decay in time computed by equations - alongside the theoretical exponential decay with excellent agreement. The simulation has been performed on a grid of $256\times256$ points.
[0.49]{} ![(Left panel) Vorticity contours after one wave period; (Right panel) Comparison of the energy decay between the computed values from equations - and the theoretical exponential decay.\[fig:decay\]](vorticita.pdf "fig:"){width="80.00000%"}
[0.49]{} ![(Left panel) Vorticity contours after one wave period; (Right panel) Comparison of the energy decay between the computed values from equations - and the theoretical exponential decay.\[fig:decay\]](damping.pdf "fig:"){width="\textwidth"}
### Frequency oscillations of a submerged reversed pendulum
[0.49]{} ![Oscillation frequency of a submerged reversed pendulum: (left) time history of the horizontal position of the center of mass for one case with $\rho = 400$, symbols highlight the maxima and minima used for the computation of the period; (right) comparison of the computed frequency with equations and . \[fig:frequency\]](maximin.pdf "fig:"){width="\textwidth"}
[0.49]{} ![Oscillation frequency of a submerged reversed pendulum: (left) time history of the horizontal position of the center of mass for one case with $\rho = 400$, symbols highlight the maxima and minima used for the computation of the period; (right) comparison of the computed frequency with equations and . \[fig:frequency\]](frequency.pdf "fig:"){width="\textwidth"}
We considere here a 2D reversed pendulum, *i.e.* a cylinder lighter than ther surrounding fluid and anchored at the bottom. The natural frequency of a simple pendulum, for small oscillation amplitude, is given by $\omega = \sqrt{g/\ell}$, with $g$ the gravity and $\ell$ the distance of the center of rotation from the center of mass, and is obtained by a balance between the tension of the contraint and the weight. On account also of the buoyancy force the frequency becomes $$\label{eqn:buoy}
\omega = \sqrt{\frac{g(\rho-\rho_p)}{\ell \rho_p}},$$ where $\rho$ is the density of the surrounding fluid and $\rho_p$ that of the pendulum. This relation is appropriate when the pendulum is much denser than the fluid whereas in the general case it is necessary to take into account also the added mass: $$\label{eqn:added}
\omega = \sqrt{\frac{g}{\ell}\frac{\rho-\rho_p}{\rho_p + c\rho}}$$ with $c$ the added mass coefficient depending on the shape of the swinging mass (for example $c=0.5$ for spheres and $c=1$ for cylinders). Here we compute the oscillation frequency for a cylinder (2D case) for different values of the density $\rho$ and compare the results with equations and . To this end, we give an initial displacement to the center of mass of the cylinder from the equilibrium position and let the pendulum oscillate; after $10$ periods we compute the time difference between all maxima and minima of the horizontal position of the pendulum (as in figure \[fig:frequency\]). In figure \[fig:frequency\]b we compare the computed frequency of the pendulum with equations and : the results show that including the added mass term gives a good approximation of the evaluations of the pendulum frequency, with an error about 5%.
The simulations are performed with the following setup: radius of the pendulum $r = 1$, fluid density $\rho = 1000$, fluid viscosity $\mu = 2.6\text{x}10^{-3}$, pendulum length $\ell = 1.8r$, size of the domain $10r\text{x}10r$; the grid has $256\times256$ points. The iterative solver has been used with a tolerance of $10^{-6}$.
Applications
------------
### Surface gravity wave propagating over a submerged reversed pendulum
We study now the interaction between surface gravity waves propagating over a submerged reversed pendulum. The wave has wavelength $\lambda$ and propagates with a phase velocity $c_p$ from the left to the right. The pendulum of radius $r$, is located (at rest) at a distance $d$ from the still water level and it is anchored at a distance $\ell$ from its center of mass. The density of the pendulum $\rho_p$ is smaller than the water density $\rho_w$, the buoyancy pulls the cylinder upwards and the constraint is always under positive tension. From the linear theory, the frequency of the wave is $$\omega_w = \sqrt{gh\tanh{kh}},$$ with $g$ the gravitational acceleration, $k = 2\pi/\lambda$ the wavenumber and $h$ the depth of the still water level. The frequency of the pendulum can be estimated using equation .
The problem depends on several parameters, which make the analysis quite complicated; here we show one case with the following setup:
$\lambda/r$ $h/r$ $\ell/r$ $d/r$ $\rho_p/\rho_w$ $\rho_a/\rho_w$ $\mu_a/\mu_w$ $\varepsilon = ak$ $Re$
------------- --------- ---------- ------- ----------------- ----------------- --------------- -------------------- --------
33.615 8.40375 2.5 1.65 0.8 1/850 1.96e-2 0.05 $10^5$
with the Reynolds number defined as $Re = \rho_w\lambda^{3/2}g^{1/2}/\mu_w$. Note that with this choice of the parameters the period of the pendulum is $1.6$ times that of the wave. The domain is wide $\lambda$ and high $0.5\lambda$, the pendulum is located at the center of the domain in the horizontal direction; the initial wave profile is $$\eta(x,0) = a\cos(kx),$$ while the initial velocity field in the water is
$$\begin{aligned}
&u(x,y,0) = a\omega\frac{\cosh\left(k\left(y+h\right)\right)}{\sinh\left(kh\right)}\cos\left(kx\right), \\
&v(x,y,0) = a\omega\frac{\sinh\left(k\left(y+h\right)\right)}{\sinh\left(kh\right)}\sin\left(kx\right)
\end{aligned}$$
and in the air the velocity has the same expression except for a change of sign in $y$. To ensure that the cord is always at the maximum extension, *i.e.* $\ell$ is a rigid constraint; accordingly we compute the hydrodynamic forces and then solve for the rotation around the fulcrum of the pendulum; velocity and position of the center of mass are then computed from the angular quantities. Hence, there is no rotation of the pendulum around its center of mass.
The wave induces an oscillatory motion of the pendulum, as shown in figure \[fig:pendpos\]. Since the wave period is different from that of the pendulum, the dynamics is characterised by different oscillation frequencies. Figures \[fig:t1\]-\[fig:t2\]-\[fig:t3\] exhibit the contour of the horizontal velocity, the interface location and the pendulum position at four time instants, corresponding to the black dots in figure \[fig:pendpos\]. Based on the location of the pendulum with respect to the phase of the velocity field, pendulum and wave can be in phase or out of phase. The small peaks in the position, for example, correspond to instants in which the pendulum has a negative velocity and interacts with the positive phase velocity of the wave, as shown in figure \[fig:t2\].
[0.49]{} ![Surface gravity wave propagating over a submerged pendulum. (a) horizontal position of the pendulum; black dots are the instants for other panels. (b)-(d) color map of horizontal velocity, position of the pendulum and of the interface (the line has been made thicker in the sake of clarity), corresponding to the black dots in panel (a) with time increasing from top left to bottom right. The entire domain is represented. \[fig:results\]](pendpos.pdf "fig:"){width="\textwidth"}
[0.49]{} ![Surface gravity wave propagating over a submerged pendulum. (a) horizontal position of the pendulum; black dots are the instants for other panels. (b)-(d) color map of horizontal velocity, position of the pendulum and of the interface (the line has been made thicker in the sake of clarity), corresponding to the black dots in panel (a) with time increasing from top left to bottom right. The entire domain is represented. \[fig:results\]](t1.pdf "fig:"){width="\textwidth"}
[0.49]{} ![Surface gravity wave propagating over a submerged pendulum. (a) horizontal position of the pendulum; black dots are the instants for other panels. (b)-(d) color map of horizontal velocity, position of the pendulum and of the interface (the line has been made thicker in the sake of clarity), corresponding to the black dots in panel (a) with time increasing from top left to bottom right. The entire domain is represented. \[fig:results\]](t2.pdf "fig:"){width="\textwidth"}
[0.49]{} ![Surface gravity wave propagating over a submerged pendulum. (a) horizontal position of the pendulum; black dots are the instants for other panels. (b)-(d) color map of horizontal velocity, position of the pendulum and of the interface (the line has been made thicker in the sake of clarity), corresponding to the black dots in panel (a) with time increasing from top left to bottom right. The entire domain is represented. \[fig:results\]](t3.pdf "fig:"){width="\textwidth"}
One interesting aspect of the problem is the energy transfer between the wave and the pendulum. We report in figure \[fig:ekp\] the time history of the kinetic and potential energy, computed by equations and , for the cases with and without pendulum. For this case the wave potential energy at rest is computed only in the fluid domain, *i.e.* $E_{p0} = \rho_wg(-\lambda h^2/2+d\pi r^2)$, with $d$ varying in time. The presence of the pendulum induces strong oscillations in the wave energy components, both the kinetic and the potential, which implies a continuous energy transfer between the wave and the pendulum. When the wave is out of phase with the pendulum, the solid opposes to the wave and induces an increase in the wave height (as in figure \[fig:s1\]) which corresponds to an increase in the potential energy of the wave. This energy then goes into kinetic energy with a strong reduction of the potential contribution, which in some instant drops almost to zero, corresponding to a small wave amplitude. The energy is then dissipated in the fluid by vorticity production and the system has an overall higher energy dissipation in time, as clearly shown in figure \[fig:et\].
[0.49]{} ![Surface gravity wave propagating over a submerged pendulum. (a) time history of wave energy components for the case with and without pendulum: wave kinetic energy with pendulum (purple open square), wave potential energy with pendulum (blue open circles), wave kinetic energy without pendulum (green solid square), wave potential energy without pendulum (orange solid circles). (b) Total wave energy for the case with pendulum (black solid triangles) and without pendulum (red open triangles). All terms are normalized by the initial total energy. (c)–(d) comparison of the surface profile between the case with pendulum (solid line, labeled as wp) and without (dashed line, labeled np). Dotted line represents the still water level.\[fig:results2\]](ekp.pdf "fig:"){width="\textwidth"}
[0.49]{} ![Surface gravity wave propagating over a submerged pendulum. (a) time history of wave energy components for the case with and without pendulum: wave kinetic energy with pendulum (purple open square), wave potential energy with pendulum (blue open circles), wave kinetic energy without pendulum (green solid square), wave potential energy without pendulum (orange solid circles). (b) Total wave energy for the case with pendulum (black solid triangles) and without pendulum (red open triangles). All terms are normalized by the initial total energy. (c)–(d) comparison of the surface profile between the case with pendulum (solid line, labeled as wp) and without (dashed line, labeled np). Dotted line represents the still water level.\[fig:results2\]](efp.pdf "fig:"){width="\textwidth"}
[0.49]{} ![Surface gravity wave propagating over a submerged pendulum. (a) time history of wave energy components for the case with and without pendulum: wave kinetic energy with pendulum (purple open square), wave potential energy with pendulum (blue open circles), wave kinetic energy without pendulum (green solid square), wave potential energy without pendulum (orange solid circles). (b) Total wave energy for the case with pendulum (black solid triangles) and without pendulum (red open triangles). All terms are normalized by the initial total energy. (c)–(d) comparison of the surface profile between the case with pendulum (solid line, labeled as wp) and without (dashed line, labeled np). Dotted line represents the still water level.\[fig:results2\]](s1.pdf "fig:"){width="\textwidth"}
[0.49]{} ![Surface gravity wave propagating over a submerged pendulum. (a) time history of wave energy components for the case with and without pendulum: wave kinetic energy with pendulum (purple open square), wave potential energy with pendulum (blue open circles), wave kinetic energy without pendulum (green solid square), wave potential energy without pendulum (orange solid circles). (b) Total wave energy for the case with pendulum (black solid triangles) and without pendulum (red open triangles). All terms are normalized by the initial total energy. (c)–(d) comparison of the surface profile between the case with pendulum (solid line, labeled as wp) and without (dashed line, labeled np). Dotted line represents the still water level.\[fig:results2\]](s2.pdf "fig:"){width="\textwidth"}
3D wave breaking induced by a submerged body
--------------------------------------------
[0.49]{} ![3D surface wave breaking: interface location at $T = 0$ (a) and at $T = 0.5$ (b); the wave propagates from the left to the right. Detailed view of the breagkin front from top (c). Wave profile in a $x-y$ plane located at the center in the $z$ direction (d).\[fig:3D\]](3D_0.pdf "fig:"){width="\textwidth"}
[0.49]{} ![3D surface wave breaking: interface location at $T = 0$ (a) and at $T = 0.5$ (b); the wave propagates from the left to the right. Detailed view of the breagkin front from top (c). Wave profile in a $x-y$ plane located at the center in the $z$ direction (d).\[fig:3D\]](3D_1.pdf "fig:"){width="\textwidth"}
[0.49]{} ![3D surface wave breaking: interface location at $T = 0$ (a) and at $T = 0.5$ (b); the wave propagates from the left to the right. Detailed view of the breagkin front from top (c). Wave profile in a $x-y$ plane located at the center in the $z$ direction (d).\[fig:3D\]](wavefront.pdf "fig:"){width="\textwidth"}
[0.49]{} ![3D surface wave breaking: interface location at $T = 0$ (a) and at $T = 0.5$ (b); the wave propagates from the left to the right. Detailed view of the breagkin front from top (c). Wave profile in a $x-y$ plane located at the center in the $z$ direction (d).\[fig:3D\]](profili.pdf "fig:"){width="\textwidth"}
Finally we want to show a three–dimensional application of the solver. Wave breaking is an important phenomenon which occurs in ocean and heavily affects the air–water mass and energy exchanges. It has been extensively studied in the last decades and several criteria have been proposed in order to predict the onset of wave breaking [@Perlin2013]. Here we want to show the effect of a submerged body on the breaking of a surface gravity wave. To this aim we simulate a third–order Stokes wave $$\eta(x) = \frac{a}{\lambda}\left(\cos(kx) + \frac{1}{2}\varepsilon \cos(2kx) + \frac{3}{8}\varepsilon^2 \cos(3kx)\right),$$ which, for a high enough value of the initial steepness $\varepsilon = 2\pi a/\lambda$ (about 0.32), leadto wave breaking [@Chen1999; @Iafrati2009; @Deike2015]. The submerged body is a sphere of radius $r = 0.05 \lambda$ located at a distance $d = 0.1$ below the still water level; the initial velocity field is derived from the linear theory and the Reynolds number, based on the wavelength $\lambda$ and phase speed $c = \sqrt{g\lambda}$, is set to $Re = \rho \lambda \sqrt{g \lambda}/\mu = 40000$. In this simulation the initial steepness is $\varepsilon =0.25$ which corresponds to a non–breaking case [@Iafrati2009; @Deike2015]. However, the presence of the sphere produces a breaking front on the surface which is a spilling breaking, as found for larger values of the steepness. The sphere produce a disturbance on the surface and when the wave crest reaches the location of the sphere (figure \[fig:3D\] d), it induces an increase in the local height which corresponds to a steepness rise. If this perturbation is strong enough to overcome the limiting steepness, wave breaking occurs, as clearly shown in figure \[fig:3D\](b). The simulation has been performed in a domain of size $\lambda \times \lambda \times \lambda/4$ with a resolution $256$ nodes per wavelength and a computational time of 1 hour per wave period on 512 processors Intel Xeon E5-2697 v4 at 2.30 GHz.
Conclusions
===========
A fully Eulerian solver for the solution of multiphase flows with solid bodies is presented which allows an efficient parallelization of the solver. The solid phase is described by an immersed body formulation with a direct forcing approach and an interpolated procedure for the reconstruction of the velocity field close to the solid boundary. The rigid body motion is coupled with the fluid solver through a 4th order predictor corrector method which allows for stable simulations in presence of large oscillations or solids lighter than the surrounding fluid, for which the added mass play an important role. In this case also underelxation could be used to stabilize the simulation. The hydrodynamic loads are computed by means of probes located in the fluid domain close to the forcing nodes and their accurate computation is crucial for a proper evaluation of the solid body dynamics. The two fluids phase are simulated by a volume of fluid solver coupled with the splitting method for the constant coefficient Poisson solver. The immersed boundary method do not pose any restriction on the timestep by its coupling with the splitting procedure require a decrease of the timestep for large density ratio beween the fluid and the solid phase while no restriction is found for buoyancy free bodies. The solver is validated with some test cases available in literature as migration of particle in pressure–driven flow, water exit of a cylidner and the frequency oscillation of a sumberged reversed pendulum. Excellent agreement has been found for all test cases. The method is also applied to a case of a surface gravity wave propagating over a submerged pendulum and the 3D spilling breaking induced by a submerged sphere. The method can be applied to solid body of arbitraty shape by replacing the distance function for the solid geometry description with a ray tracing algorithm. Extension to deformable bodies could be done by replacing the direct forcing apporach with a moving least square method and wil be the subject of future work as for the contact angle between the solid and the liquid.
Acknowledgments {#acknowledgments .unnumbered}
===============
M. Onorato, F. De Lillo and F. De Vita have been funded by Progetto di Ricerca d’Ateneo CSTO160004, by the “Departments of Excellence 2018/2022” Grant awarded by the Italian Ministry of Education, University and Research (MIUR) (L.232/2016). M. Onorato and F. De Vita acknowledge also from EU, H2020 FET Open BOHEME grant No. 863179. The authors also acknowledge CINECA for the computational resources under the grant IscraC SGWA.
Appendix A: Hamming’s $4^{th}$ order modified predictor-corrector {#sec:A .unnumbered}
=================================================================
At each timestep, once the forces are known, equations - are advanced in time using a Hamming’s 4th order modified predictor-corrector method [@Hamming1959]. Defining $\vec{\mathcal{X}} = \left[\vec{X}, \vec{\Theta}\right]^T$ the unknown vector for the position and $\vec{\mathcal{V}} = \left[\vec{V}, \vec{\Omega}\right]^T$ the unknown vector for the velocity, the evolution is computed by the following steps
1. *Predictor step*: evaluate the predicted solution $\left(\mathcal{X}^{l+1}_p,\mathcal{V}^{l+1}_p\right)$ at timestep $n+1$
- $\vec{\mathcal{A}}^{l} = \vec{F}^l/m$
- $\vec{\mathcal{V}}^{l+1}_p = \vec{\mathcal{V}}^{l-3} + \frac{4}{3}\Delta t\left(2\vec{\mathcal{A}}^{l}-\vec{\mathcal{A}}^{l-1}+2\vec{\mathcal{A}}^{l-2}\right)$
- $\vec{\mathcal{X}}^{l+1}_p = \vec{\mathcal{X}}^{l-3} + \frac{4}{3}\Delta t\left(2\vec{\mathcal{V}}^{l}-\vec{\mathcal{V}}^{l-1}+2\vec{\mathcal{V}}^{l-2}\right)$
improve it with an estimation of the error at previous timestep
- $\vec{\mathcal{V}}^{l+1}_m = \vec{\mathcal{V}}^{l+1}_p - \frac{112}{121}\left(\vec{\mathcal{V}}^{l}_p-\vec{\mathcal{V}}^{l}\right)$
- $\vec{\mathcal{X}}^{l+1}_m = \vec{\mathcal{X}}^{l+1}_p - \frac{112}{121}\left(\vec{\mathcal{X}}^{l}_p-\vec{\mathcal{X}}^{l}\right)$
and solve flow equations using as boundary condition $\left(\mathcal{X}^{l+1}_m,\mathcal{V}^{l+1}_m\right)$ to evaluate $F^{l+1}_1$.
2. *Corrector step*: correct the solution iteratively (loop on $k$, starting from $k = 1$)
- $\vec{\mathcal{A}}^{l+1}_k = \vec{F}^{l+1}_k/m$
- $\vec{\mathcal{V}}^{l+1}_k = \frac{1}{8}\left(9\vec{\mathcal{V}}^{l}-\vec{\mathcal{V}}^{l-2}\right) +\frac{3}{8}\Delta t\left(2\vec{\mathcal{A}}^{l+1}_k+2\vec{\mathcal{A}}^{l}-\vec{\mathcal{A}}^{l-1}\right)$
- $\vec{\mathcal{X}}^{l+1}_k = \frac{1}{8}\left(9\vec{\mathcal{X}}^{l}-\vec{\mathcal{X}}^{l-2}\right) +\frac{3}{8}\Delta t\left(2\vec{\mathcal{V}}^{l+1}_k+2\vec{\mathcal{V}}^{l}-\vec{\mathcal{V}}^{l-1}\right)$
and check for convergence: $\min\left(\vec{|\mathcal{X}}^{l+1}_{k}-\vec{\mathcal{X}}^{l+1}_{k-1}|,|\vec{\mathcal{V}}^{l+1}_{k}-\vec{\mathcal{V}}^{l+1}_{k-1}|\right) < \epsilon$, where the values at $k=0$ are the modified value of the predictor step and $\epsilon$ is the tolerance. If converged, make the final correction to the solution:
- $\vec{\mathcal{V}}^{l+1} = \vec{\mathcal{V}}^{l+1}_k + \frac{9}{121}\left(\vec{\mathcal{V}}^{l+1}_p-\vec{\mathcal{V}}^{l+1}_k\right)$
- $\vec{\mathcal{X}}^{l+1} = \vec{\mathcal{X}}^{l+1}_k + \frac{9}{121}\left(\vec{\mathcal{X}}^{l+1}_p-\vec{\mathcal{X}}^{l+1}_k\right)$
and solve flow equations with boundary condition $\left(\mathcal{X}^{l+1},\mathcal{V}^{l+1}\right)$. If not converged, solve flow equations with boundary condition $\left(\mathcal{X}^{l+1}_k,\mathcal{V}^{l+1}_k\right)$, evaluate $F^{l+1}_k$ and then repeat the corrector step until convergence.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Since its discovery in 1993, we witness an intensive theoretical and experimental effort centered on teleportation. Very recently it was claimed in the press that “quantum teleportation has been achieved in the laboratory” (T. Sudbery, [*Nature*]{} [**390**]{}, 551). Here, I briefly review this research focusing on the connection to [ *nonlocal measurements*]{}, and question Sudbery’s statement. A philosophical inquiry about the paradoxical meaning of teleportation in the framework of the many-worlds interpretation is added.'
author:
- Lev Vaidman
title: 'Teleportation: Dream or Reality?'
---
15.2 pt
[*School of Physics and Astronomy\
Raymond and Beverly Sackler Faculty of Exact Sciences\
Tel Aviv University, Tel-Aviv 69978, Israel.\
*]{}
The meaning of the Word “Teleportation” {#Intro}
=======================================
Let me start with a citation from the Oxford English Dictionary [@OX]:
> [**teleportation**]{}. [*Psychics*]{} and [*Science Fiction*]{}. The conveyance of persons (esp. of oneself) or things by psychic power; also in futuristic description, apparently instantaneous transportation of persons, etc., across space by advanced technological means.
Recently, the word “teleportation” has appeared outside of the realm of mystical and science fiction literature: in science journals. Bennett, Brassard, Crepeau, Jozsa, Peres, and Wootters (BBCJPW) [@BBC] proposed a gedanken experiment they termed “quantum teleportation”.
Classically, to move a person is to move all the particles it is made of. However, in quantum theory particles themselves do not represent a person: all objects are made of the same elementary particles. An electron in my body is identical to an electron in the paper of the page you are reading now. An object is characterized by the [*quantum state*]{} of the particles it is made of. Thus, reconstructing the quantum state of these particles on other particles of the same kind at a remote location [*is*]{} “transportation” of the object.
The quantum state of the object to be transported is supposed to be unknown. Indeed, usually we do not know and cannot find out what the quantum state of an object is. Moreover, frequently an object is not in a pure quantum state, its particles may be correlated to other systems. In such cases the essence of the object is these correlations. In order to transport such correlations (even if they are known), without access to the systems which are in correlation with our system, a method for teleportation of an unknown quantum state is necessary.
Quantum teleportation [@BBC] transfers the quantum state of a system and its correlations to another system. Moreover, the procedure corresponds to the modern meaning of teleportation: an object is disintegrated in one place and a perfect replica appears at another site. The object or its complete description is never located between the two sites during the transportation. Note that “disintegration” of the quantum state is a necessary requirement due to the no-cloning theorem.
The teleportation procedure, apart from preparing in advance the quantum channels, requires telegraphing surprisingly small amounts of information between the two sites. This stage prevents “instantaneous” transportation. Indeed, because of special relativity, we cannot hope to achieve superluminal teleportation: objects carry signals.
Due to the arguments presented above, I find the BBCJPW procedure to be very close to the concept of “teleportation” as it is used in the science-fiction literature. However, the name teleportation is less justified for the recent implementations of this idea in the laboratory, as well as for some other proposals for experiments. For me, an experiment deserves the name “teleportation” if I can give to Alice (the sender) a system whose quantum state in unknown to her and that she can, without moving this system and without moving any other system which can carry the quantum state of the system, transport this state to Bob (the receiver) which is located at a remote location. In the next section I shall discuss, in the light of my definition, the usage of the word “teleportation”.
Recent Experiments and Proposals for Experiments Termed “Teleportation” {#examples}
=======================================================================
What I discuss in this section is essentially a semantic issue, but I feel that its clarification is important. I find the original teleportation paper to be one of the most important results in the field in the last ten years, and I think that it should be clearly distinguished from other interesting but less profound achievements.
Recently I heard the word “teleportation” in the context of NMR-type quantum computation experiments [@NMR-tele]. Using certain pulses, a spin state of a nucleus in a large molecule is transported to another nucleus in the same molecule. The main deficiency of this experiment as teleportation is that it does not allow to transport an [*unknown*]{} quantum state. Indeed, in the NMR experiments a macroscopic number of molecules have to be in a particular quantum state. If Alice receives a single quantum object in an unknown quantum state, she cannot duplicate it and in that manner prepare many copies in many molecules, due to the no-cloning theorem.
An apparent weakness of the NMR experiment is that the internal coupling which plays the role of the channel for classical information required for teleportation [*can*]{}, in principle, carry the quantum state. However, due to the strong interaction with the environment, the quantum state transmitted through such a channel is effectively measured by the environment. Only the eigenstates corresponding to the classical outcomes are stable under this interaction and, therefore, there is good reason to consider this channel to be classical.
Another place in which I encountered the word “teleportation” is the work on optical simulation of quantum computation [@cerf]. It includes a proposal for implementation of the idea to view “teleportation” as a particular quantum computation circuit [@Brass; @BBC]. The problem in the optical experiment is that instead of the classical channel which is supposed to transmit two bits of information, real photons are moving from Alice to Bob and these photons [*can*]{} transmit the whole quantum state of the polarization degree of freedom of the photon. This is exactly the apparent weakness of the NMR-teleportation experiments mentioned above, but in the present case the environment does not make the quantum channel to be effectively classical. Note that in the original proposal the quantum channel is explicitly replaced by a classical one to make the proposal akin to teleportation in the BBCJPW sense. It is the optical simulation of this proposal which is something less than teleportation. It seems that the authors [@cerf] were aware of this problem when they added a footnote: “The term teleportation is used in the literature to refer to the transfer of the state of a qubit to another”. I find this meaning to be too general. Many processes corresponding to this definition were proposed (and even implemented in laboratories) long before the teleportation paper has appeared.
Next, let me discuss “teleportation” in the Rome experiment [@Rome]. As I will explain in the next section, the main obstacle for successful reliable teleportation is the experimental difficulty to make one quantum object interact with another. In optical experiments quantum objects, photons, interact with classical objects such as beam splitters, detectors, etc. Popescu [@Pop] proposed a very elegant solution: two degrees of freedom of a [*single*]{} photon do interact effectively one with the other. This idea was successfully implemented in the Rome experiment in which the polarization state of the photon was transported to another photon. However, the weakness of this experiment is that the quantum state to be teleported has to be the state of (the second degree of freedom of) one of the members of the EPR pair which constitute the quantum channel of the teleportation experiment. Therefore, this method cannot be used for teleportation of an unknown quantum state of an external system. The authors [@Rome] view this experiment as “teleportation” because after the preparation Alice cannot find out the quantum state, which, nevertheless, is transported (always and with high fidelity) to Bob.
Finally, let me discuss the Innsbruck teleportation experiment [@Inn; @swap]. Although the word “teleportation” appears in the title of the first Letter [@Inn], the second experiment [@swap] is a much better demonstration of teleportation. I believe that the Innsbruck experiment deserves the name teleportation. It showed for the first time that an unknown state of an external photon can be teleported. It is not a reliable teleportation: the experiment has a theoretical success rate of 25% only, and the employed methods cannot, in principle, lead to reliable teleportation. For a system consisting of $N$ qubits the probability of successful teleportation is exponentially small.
Recently, Braunstein and Kimble [@BK-N] pointed out a weak point of the Innsbruck experiment. In the current version of the experiment one might know that teleportation has been successful only after the time Bob detects (and, therefore, destructs) the photon with the teleported state. Thus, the name given by Braunstein and Kimble for the Innsbruck experiment: “a posteriori teleportation” appears to be appropriate. However, as mentioned in the reply [@reply] and in the comment itself, it is feasible to solve this problem by a modification of the experiment and therefore it is not a conceptual difficulty.
Another possible improvement of the demonstration of teleportation in the Innsbruck experiment is using single input photons. In the current version of the experiment, the polarizer which controls the input quantum state is stationary, and, therefore, many photons are created in the same state. Thus, this state can hardly be considered an “unknown” quantum state. Low intensity of the input beam and frequent changing of the angle of the polarizer is a simple and effective solution of the problem. An ideal solution is using a “single-photon gun” [@SFG] which creates single-photon states.
Apart from the impossibility of performing a measurement of the nondegenerate Bell operator, there is another problem for achieving reliable teleportation of an unknown state of a single photon. Today, there is no source which creates a single EPR pair at will, something frequently called an “event-ready” source. The second Innsbruck experiment [@swap] is the best achievement in this direction: entanglement swapping may be viewed as creation of an entangled pair at the moment of the coincidence detection of the two photons coming from the beam splitter. What is missing is a “sophisticated detection procedure” [@swap] which rules out the creation of two pairs in a single crystal.
Bell-Operator Measurement and Teleportation {#proof}
===========================================
The original BBCJPW teleportation procedure consists of three main stages: (i) Preparation of an EPR pair, (ii) Bell-operator measurement performed on the “input” particle and one particle of the EPR pair, (iii) Transmission of the outcome of the Bell measurement and appropriate unitary operation on the second particle of the EPR pair (the “output” particle). Completing (i)-(iii) ensures transportation of the pure state of the input particle to the output particle. It also ensures transportation of correlations: if the input particle were correlated to other systems, then the output particle ends up correlated to these systems in the same way.
The main difficulty in this procedure is performing the Bell measurement. Recently it has been proved [@VY; @FIN] that without “quantum-quantum” interaction one cannot perform measurement of the nondegenerate Bell operator which is required for reliable teleportation. Using only “quantum-classical” interactions one can perform a measurement of a degenerate Bell operator [@Harold; @BrMa], thus allowing a teleportation which succeeds sometimes.
The size limitations of this paper allow only to outline the proof [@VY]. In order to prove that it is impossible to perform complete (nondegenerate) Bell-operator measurements without using interactions between quantum systems, I assume that any unitary transformation of single-particle states and any local single-particle measurement are allowed. There are four distinct (orthogonal) single-particle states involved in the definition of the Bell states: two channels, and a two-level system which enters into each channel. We name the channels left (L) and right (R), corresponding to the way the Bell states are written: $$\begin{aligned}
\label{Bell}
\nonumber |\Psi_{\pm}\rangle &=&{1\over\sqrt2}
(|{\uparrow}\rangle_L|{\downarrow}\rangle_R {\pm}
|{\downarrow}\rangle_L|{\uparrow}\rangle_R),
\\
|\Phi_ {\pm}\rangle &=&{1\over \sqrt2}(|{\uparrow}\rangle_L|{\uparrow}\rangle_R {\pm}
|{\downarrow}\rangle_L|{\downarrow}\rangle_R).\end{aligned}$$
The measurement procedure can be divided into two stages: the unitary linear evolution, and local detection. The general form of the unitary linear evolution of the four single-particle states can be written in the following manner: $$\label{basic}
|{\uparrow}\rangle_L \rightarrow \sum a_i |i\rangle,
~~ |{\downarrow}\rangle_L \rightarrow \sum b_i |i\rangle,~~
|{\uparrow}\rangle_R \rightarrow \sum c_i |i\rangle,
~~ |{\downarrow}\rangle_R \rightarrow \sum d_i |i\rangle,$$ where $\{|i\rangle\}$ is a set of orthogonal single-particle local states. The “linearity” implies that the evolution of the particle in one channel is independent on the state of the particle in another channel and, therefore, Eq. (\[basic\]) is enough to define the evolution of the Bell states: $$\begin{aligned}
\label{unit}
\nonumber
|\Psi_-\rangle
&\rightarrow &\sum_{i,j} \alpha_{ij} |i\rangle |j\rangle,
~~
|\Psi_+\rangle
\rightarrow \sum_{i,j} \beta_{ij} |i\rangle |j\rangle,
\\
|\Phi_-\rangle
&\rightarrow &\sum_{i,j} \gamma_{ij} |i\rangle |j\rangle,
~~
|\Phi_+\rangle
\rightarrow \sum_{i,j} \delta_{ij} |i\rangle |j\rangle.\end{aligned}$$ Proper symmetrization is required for identical particles.
I assume that there are only local detectors and, therefore, only product states $ |i\rangle |j\rangle$ (and not their superpositions) can be detected. Measurability of the non-degenerate Bell operator means that there is at least one nonzero coefficient of every kind $\alpha_{ij}, \beta_{ij}, \gamma_{ij}, \delta_{ij}$ and if, for a certain $i,j$, it is not zero, then all others are zero. This observation leads to numerous equations which, after some tedious algebra, yield the desired proof.
A somewhat different approach was taken in proof [@FIN]. This proof considers only photons, but it proves the impossibility of non-degenerate Bell-measurements for a more general case in which measurements in two stages are allowed. The procedure in which the choice of the measurements in the second stage depends on the results of the measurements in the first stage is an indirect quantum-quantum interaction: the state of one quantum system influences the result of the first measurement and the action on the second quantum system depends on this result.
If we allow direct quantum-quantum interactions, we can achieve reliable (theoretically 100% efficient) teleportation. In this case, we can perform a measurement of the non-degenerate Bell operator. Indeed, a quantum-quantum interaction such as a conditional spin-flip transforms the Bell states into product states which then can be measured using single-particle measuring devices.
An alternative method of teleportation [@tele-V] is based on [ *nonlocal*]{} measurements [@AAV86] “crossed” in space-time. In order to teleport a quantum state from particle 1 to particle 2 and, at the same time, the quantum state of particle 2 to particle 1, the following (nonlocal in space-time) variables should be measured (see Fig. 1): $$\label{swap}
{\cal Z} \equiv \Bigl({\sigma_1}_z(t_1) + {\sigma_2}_z(t_2)\Bigr){\rm
mod~4}, ~~
{\cal X} \equiv \Bigl({\sigma_1}_x(t_2) + {\sigma_2}_x(t_1)\Bigr){\rm
mod~4}.$$ For any set of outcomes of the nonlocal measurements (\[swap\]) the spin state is teleported; in some cases the state is rotated by $\pi$ around one of the axes, but the resulting rotation can be inferred from the nonlocal measurements. In order to perform nonlocal measurements (\[swap\]), correlated pairs of auxiliary particles located at the sites of particle 1 and 2 are required. For completing the whole procedure we need two singlets instead of the one required in the original teleportation procedure. The reason for requiring more resources is that two-way (rather than one-way) teleportation is achieved.
[ Space-time locations of local couplings are shown. When the nonlocal measurements (\[swap\]) are completed, the states of the two particles are interchanged up to local $\pi$ rotations, $\tilde \Psi_i$ signifies “rotated” $\Psi_i$.]{}
Towards experimental realization of reliable teleportation {#tele-exp}
===========================================================
Due to the lack of an effective photon-photon interaction, the currently available methods do not allow reliable teleportation of the photon polarization state. It seems that the most promising candidates for teleportation experiments which might have 100% success rate are proposals which involve atoms and electro-magnetic cavities. First suggestions for such experiments [@Slea; @David; @Cirac] were made shortly after publication of the original teleportation paper and numerous modifications appeared since. The implementation of these proposals seems to be feasible because of the existence of the “quantum-quantum” interaction between the system carrying the quantum state and a system forming the EPR pair. A dispersive interaction (DI) of a Rydberg atom passing through a properly tuned micro-wave cavity leads to a conditional phase flip depending on the presence of a photon in the cavity. A resonant interaction (RI-$\pi$) between the Rydberg atom and the cavity allows swapping of quantum states of the atom and the cavity. Thus, manipulation of the quantum state of the cavity can be achieved via manipulation of the state of the Rydberg atom. The atom’s state is transformed by sending it through an appropriately tuned microwave zone. Moreover, the direct analog of conditional spin-flip the interaction can be achieved through the Raman atom-cavity interaction [@ZG97]. No teleportation experiment has been performed as of yet using these methods, but it seems that the technology is not too far from this goal. Recent experiments on atom-cavity interactions [@Har] teach us about the progress in this direction.
Until further progress in technology is achieved, it is not easy to predict which proposal will be implemented first. Assuming that resonant atom-cavity interactions can be performed with very good precision and that a dispersive interaction is available with a reasonable precision, it seems that the following is the simplest proposal, see Fig. 2. The quantum channel consists of a cavity and a Rydberg atom in a correlated state. A particular resonant interaction, RI-$\pi / 2$, of an excited atom passing through an empty cavity, $$\label{resona}
\nonumber
|e\rangle |0\rangle \rightarrow {1\over \sqrt2}(|g\rangle |1\rangle +
|e\rangle |0\rangle) ,$$ prepares this quantum channel (Fig. 2a). The quantum state to be teleported is the state of another Rydberg atom. The Bell measurement is then performed on this atom and the cavity. To this end, the atom passes through the cavity interacting dispersively (Fig. 2b), induces the conditional phase flip, $$\label{pha-fli-ca}
|e\rangle|0\rangle \rightarrow
|e\rangle|0\rangle,
~~~
|e\rangle|1\rangle \rightarrow -
|e\rangle|1\rangle,
~~~
|g\rangle|0\rangle \rightarrow
|g\rangle|0\rangle,
~~~
|g\rangle|1\rangle \rightarrow
|g\rangle|1\rangle,$$ which disentangles the following Bell states: $$\begin{aligned}
\label{Bell-ca}
\nonumber |\Psi_{\pm}\rangle &=& {1\over 2}\Bigl(|e\rangle (|0\rangle
-|1\rangle) \pm |g\rangle (| 1\rangle +|0\rangle)\Bigr),
\\
|\Phi_{\pm}\rangle &=& {1\over 2}\Bigl(|e\rangle (|0\rangle
+|1\rangle) \pm |g\rangle (| 1\rangle -|0\rangle)\Bigr).\end{aligned}$$
[\
(a) Preparation of the quantum channel. An atom undergoes resonant interaction RI-$\pi$/2 with the cavity and moves to a remote site.\
(b) The atom, carrying the quantum state to be teleported, interacts with the cavity dispersively and its state is measured.\
(c) The state of the cavity is measured using an auxiliary atom.]{} .4cm
The Bell states (\[Bell-ca\]) have the form of Eq. \[Bell\] when the first $|{\uparrow}\rangle$ in the product is identified with $|e\rangle$, the second $|{\uparrow}\rangle$, with $(1/\sqrt2)(|0\rangle +
|1\rangle)$, etc. Measurement of the atom state and the cavity state completes the Bell measurement procedure. In order to make the measurement of the cavity state we perform another resonant interaction, RI-$\pi$, between the cavity and an auxiliary atom prepared initially in the ground state (Fig. 2c), $$\label{resona1}
|g\rangle |1\rangle \rightarrow |e\rangle |0\rangle, ~~~
|g\rangle |0\rangle \rightarrow |g\rangle |0\rangle.$$ This interaction transfers the quantum state of the cavity to this atom. The final measurements on the atoms distinguish between the states $(1/\sqrt2)(|g\rangle +
|e\rangle)$ and $(1/\sqrt2)(|g\rangle -
|e\rangle)$. Since detectors can distinguish between $|g\rangle$ and $|e\rangle$, the states of the atoms are rotated while passing through the appropriate microwave zones before detection. When the Bell measurement is completed, the quantum state is teleported up to the known local transformation determined by the results of the Bell measurement. (This final local transformation is not shown in Fig. 2.)
One relatively simple method for “two-way” teleportation of atomic states is a direct implementation of the crossed nonlocal measurement scheme presented in the previous section. This method is described in Ref. [@VY].
One difficulty with the teleportation of atomic states is that usually experiments are performed with atomic [*beams*]{} and not with individual atoms. Such experiments might be good for demonstration and studying experimental difficulties of teleportation, but they cannot be considered as implementation of the original wisdom of teleportation or used for cryptographic purposes. In fact, optical experiments have this difficulty too, unless “single-photon guns” will be used. Both for atomic and for optical experiments this difficulty does not seem to be unsolvable, but it certainly brings attention to experiments with trapped ions [@ion]. There are many similarities between available manipulations with atoms and with ions, so the methods discussed above might be implemented for ion systems too.
Note also another recent proposal for teleportation using quantum-quantum interaction [@chiral]. It is based on rotation of the photon polarization due to presence of a single chiral molecule in an optical cavity. I am, however, skeptical about the feasibility of such experiment due to difficulties in tuning the interferometer in which photons undergo multiple reflections in the cavity; the number of reflections has to be very large due to weakness of the interaction between the molecule and the photon.
Teleportation of continuous variables {#tele-cont}
=====================================
In the framework of nonlocal measurements there is a natural way of extending the teleportation scheme to systems with continuous variables [@tele-V]. Consider two similar systems located far away from each other and described by continuous variables $q_1$ and $q_2$ with corresponding conjugate momenta $p_1$ and $p_2$. In order to teleport the quantum state of the first particle $\Psi_1(q_1)$ to the second particle (and the state of the second particle $\Psi_2(q_2)$ to the first) we perform the following “crossed” nonlocal measurements (see Fig. 3), obtaining the outcomes $a$ and $b$:
$$\label{cross-conti}
q_1(t_1) - q_2(t_2) = a,~~~
p_1(t_2) - p_2(t_1) = b.$$
In Ref. [@tele-V] it is shown that these nonlocal “crossed” measurements “swap” the quantum states of the two particles up to the known shifts in $q$ and $p$. Indeed, the states of the particles after completion of the measurements (\[cross-conti\]) are $$\label{psi-shif}
\Psi_{f}(q_1)= e^{ibq_1} \Psi_2(q_1+a),~~~
\Psi_{f}(q_2)= e^{-ibq_2} \Psi_1(q_2-a) .$$ .4cm
[ Space-time locations of local couplings are shown. When the nonlocal measurements (\[cross-conti\]) are completed, the states of the two particles are interchanged up to the known shifts in $q$ and $p$. ]{}
The state of particle 2 after $t_2$ is the initial state of the particle 1 shifted by $-a$ in $q$ and by $-b$ in $p$. Similarly, the state of particle 1 is the initial state of particle 2 shifted by $a$ in $q$ and by $b$ in $p$. After transmitting the results of the local measurements, $a$ and $b$, the shifts can be corrected (even if the quantum state is unknown) by appropriate kicks and back shifts, thus completing a reliable teleportation of the state $\Psi_1(q_1)$ to $\Psi_1(q_2)$ and of the state $\Psi_2(q_2)$ to $\Psi_2(q_1)$.
Surprisingly, the implementation of the reliable teleportation of continuous variables is possible. Braunstein and Kimble made a realistic proposal for teleporting the quantum state of a single mode of the electro-magnetic field [@BK]. This remarkable result is an implementation of a variation of the scheme described above which achieves a one-way teleportation. In their method $q$ is “$x$”defined for a single mode of an electro-magnetic field, and correspondingly $p$ is the conjugate momentum of $x$. The analog of the EPR pair is obtained by shining squeezed light with a certain $x$ from one side and squeezed light with a certain $p$ from the other side onto a simple beam splitter. The analog of the local Bell measurement is achieved using another beam splitter and homodyne detectors. The shifts in $x$ and $p$ which complete the teleportation procedure can be done by combining the output field with the coherent state of appropriate amplitude fixed by the results of the homodyne measurements. Note also a related proposal [@Mol1; @Mol2] for teleporting a single-photon wave packet.
Very recently the Braunstein-Kimble proposal for implementation of continues variable teleportation [@tele-V] has been performed in California Institute of Technology [@Furu]. This is the first reliable teleportation experiment. The meaning of “reliable” (“unconditional” in [@Furu]) is that theoretically it is always successful. It is the first experiment in which the final stage of teleportation, i.e., transmission of the classical information to Bob and the appropriate transformation which results in the appearance of the teleported state in the Bob’s site, has been implemented. The weakness of this experiment is that the teleported state is significantly distorted. The main reason for low fidelity is the degree of squeezing of the light which controls the quality of the EPR pairs, the quantum channel of the teleportation. Significant improvement of the squeezing parameters is a very difficult technological problem. Thus, in this type of experiment one cannot reach the high fidelity of (conditional) teleportation experiments of photon polarization states.
One may note an apparent contradiction between the proof of Section \[proof\] that 100% efficient teleportation cannot be achieved using linear elements and single-particle state detectors and the successful [*reliable*]{} teleportation experiment of the state of the electro-magnetic field which involved only beam-splitters and local measuring devices reported above. Indeed, it is natural to assume that if reliable teleportation of a quantum state of a two-level system is impossible under certain circumstances, it should certainly be impossible for quantum states of systems with continuous variables. However, although it is not immediately obvious, the circumstances are very different. There are numerous differences. The analog of the Bell operator for continuous variables does not have among its eigenvalues four states of the general form (\[Bell\]) where $|{\uparrow}\rangle$ and $|{\downarrow}\rangle$ signify some orthogonal states. Another problem is that one cannot identify “the particles”: In the beam-splitter [*one*]{} input port goes to [ *two*]{} output ports. One can see a “quantum-quantum” interaction: the variable $x$ of one of the output ports of the beam splitter becomes equal to $(1/ \sqrt 2)(x_1 + x_2)$, essentially, the sum of the quantum variables of the input ports. The absence of such “quantum-quantum” interactions is an essential ingredient in the proof of Section \[proof\]. If, however, we consider the “particles” to be photons (which do not interact with one another) then the homodyne detectors which measure $x$ are not single-particle detectors—another constraint used in the proof. Note also that the Braunstein-Kimble method is not applicable directly for teleporting $\Psi (x)$ where $x$ is a spatial position of a quantum system. An additional quantum-quantum interaction which converts the continuous variable of position of a particle to the variable $x$ of the electro-magnetic mode is required.
Is There a Paradox with Teleportation?
======================================
My complaints about the (mis)interpretation of the word “teleportation” in Section II shows that I am (over)sensitive about this issue. This is because I was thinking a lot about it, resolving for myself a paradox [@psa] which I, as a believer in the Many-Worlds Interpretation (MWI) [@MWI-V] had with this experiment.
Consider teleportation, say in the BBCJPW scheme. We perform some action in one place and the state is immediately teleported, up a local transformation (“rotation”), to an arbitrary distant location. But relativity theory teaches us that anything which is physically significant cannot move faster than light. Thus it seems that it is the classical information (which cannot be transmitted with superluminal velocity) about the kind of back “rotation” to be performed for completing the teleportation which is the only essential part of the quantum state. However, the amount of the required classical information is very small. Is the essence of a state of a spin-1/2 particle just 2 bits?
I tend to attach a lot of physical meaning to a quantum state. For me, a proponent of the MWI, everything is a quantum state. But I also believe in relativistic invariance, so only entities which cannot move faster than light have physical reality. Thus, teleportation poses a serious problem to my attitude. I was ready to admit that “I” am just a quantum state of $N\sim 10^{30}$ particles. This is still a very rich structure: a complex function on ${\cal R}^N$. But now I am forced to believe that “I” am just a point in the ${\cal
R}^{2N}$ ?!
The resolution which I found for myself is as follows: In the framework of the MWI, the teleportation procedure does not move the quantum state: the state was, in some sense, in the remote location from the beginning. The correlated pair, which is the necessary item for teleportation, incorporates all possible quantum states of the remote particle, and, in particular, the state $\Psi$ which has to be teleported. The local measurement of the teleportation procedure splits the world in such a manner that in each of the worlds the state of the remote particle differs form the state $\Psi$ by some known transformation. The number of such worlds is relatively small. This explains why the information which has to be transmitted for teleportation of a quantum state—the information which world we need to split into, i.e., what transformation has to be applied—is much smaller than the information which is needed for the creation of such a state. For example, for the case of a spin-1/2 particle there are only 4 different worlds, so in order to teleport the state we have to transmit just 2 bits. As for teleporting myself, the number of worlds is the number of distinguishable (using measuring devices and our senses) values of $x_i$ and $p_i$ for all continues degrees of freedom of my body.
Teleportation of people will remain a dream for the foreseeable future. First, we have to achieve the reliable teleportation of an unknown quantum state of an external system with reasonable fidelity which is also only a dream today. Although the teleportation of an unknown quantum state has not yet been achieved, the current experiments clearly demonstrate that it can be done. I urge the experimenters to perform a persuasive teleportation experiment: Carol gives to Alice (single) particles in different states (unknown to Alice), Alice teleports the states to Bob, Bob gives them back to Carol who tests that what she gets is what she has sent before.
I am grateful for very useful correspondence with Chris Adami, Gilles Brassard, Samuel Braunstein, John Calsamiglia, Lior Goldenberg, Daniel Lidar, Sergey Molotkov, Harald Weinfurter, Asher Peres, Sandu Popescu, and Anton Zeilinger. The research was supported in part by grant 471/98 of the Basic Research Foundation (administered by the Israel Academy of Sciences and Humanities).
[99]{}
, 2nd Ed. (Clarendon Press, Oxford 1989) XVII p. 730.
C. H. Bennett, G. Brassard, C. Crepeau, R. Jozsa, A. Peres, W. K. Wootters, Phys. Rev. Lett. [**70**]{}, 1895 (1993).
R. Laflamme, Talk given at the Benasque Workshop on Quantum Computation, July 1998.
N. J. Cerf, C. Adami, P. G. Kwiat, Phys. Rev. A [**57**]{}, R1477 (1998).
G. Brassard, e-print quant-ph/9605035 (1996).
G. Brassard, S. L. Braunstein and R. Cleve, Physica D [**120**]{}, 43 (1998).
D. Boschi, S. Branca, F. De Martini, L. Hardy and S. Popescu, Phys. Rev. Lett. [**80**]{}, 1121 (1997).
S. Popescu, e-print quant-ph/9501020.
D. Bouwmeester, J. W. Pan, K. Mattle, M. Eibl, H. Weinfurter, A. Zeilinger, Nature [**390**]{}, 575 (1997).
J. W. Pan, D. Bouwmeester, H. Weinfurter, and A. Zeilinger Phys. Rev. Lett. [**80**]{}, 3891 (1998).
S. L. Braunstein, H. J. Kimble, Nature [**394**]{}, 840 (1998).
D. Bouwmeester, J. W. Pan, K. Mattle, M. Eibl, H. Weinfurter, A. Zeilinger, Nature [**394**]{}, 840 (1998).
L. Vaidman, in [*The Geometric Universe*]{}, S. Huggett [*et al.*]{} eds., Oxford University Press, Oxford, p. 349 (1998).
L. Vaidman and N. Yoran, Phys. Rev. A, to be published, quant-ph/9808043.
N. Lütkenhaus, J. Calsamiglia, K-A. Suominen, e-print quant-ph/9809063.
H. Weinfurter, Europhys. Lett. 559 (1994).
S. L. Braunstein and A. Mann, Phys. Rev. A [**51**]{}, R1727 (1955).
L. Vaidman, Phys. Rev. [**A 49**]{}, 1473 (1994).
Y. Aharonov, D. Albert, and L.Vaidman, Phys. Rev. D [**34**]{}, 1805 (1986).
T. Sleator and H. Weinfurter, in [*IQEC Technical Digest 1994*]{}, Vol. 9, 1994, OSA Technical Digest Series (OSA, Washington, D.C., 1994), p. 140.
L. Davidovich, N. Zagury, M. Brune, J. M. Raimond and S. Haroche, Phys. Rev. A, [**50**]{}, 895, (1994).
J. I. Cirac and A. S. Parkins, Phys. Rev. A [**50**]{} R4441 (1994).
S. B. Zheng, G. C. Guo, Phys. Lett. A [**232**]{}, 171 (1997).
S. Haroche, M. Brune, J. R. Raimond, Philos. T. Roy. Soc. A [**355**]{} (1733) 2367 (1997).
R. J. Hughes, Philos. T. Roy. Soc. A [**356**]{} (1743) 1853 (1998).
C. S. Maierle, D. A. Lidar, R. A. Harris, e-print quant-ph/9807020.
S. L. Braunstein, H. J. Kimble, Phys. Rev. Lett. [**80**]{}, 869 (1998).
A. Furusawa, J. S[ø]{}renson, S. L. Braunstein, C. A. Fuchs, H. J. Kimble and E. S. Polzik, Science, [**282**]{}, 706 (1998).
S. N. Molotkov, Phys. Lett. A [**245**]{}, 339 (1998).
S. N. Molotkov, JETP Lett. [**68**]{}, 263 (1998).
L. Vaidman, in [*PSA 1994*]{}, D. Hull, M. Forbes and R. M. Burian eds., (East Lansing, MI 1994) vol. 1, p. 211.
L. Vaidman, Int. Stud. Phil. Sci. [**12**]{}, (1998).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The observed abundance of very red galaxies at high redshift has been recognized as a long standing problem for hierarchical models of galaxy formation. Here we investigate the effect of AGN feedback associated to the bright QSO phase onto the color distribution of galaxies from $z=0$ up to $z\approx 3$, and on the abundance of extremely red objects (EROs, with $R-K>5$) and distant red galaxies (DRGs, with $J-K>2.3$) up to $z\approx 4$; to this aim, we insert a blast-wave model of AGN feedback in our semi-analytic model of galaxy formation, which includes the growth of supermassive black holes and the AGN activity triggered by mergers and interactions of the host galaxies. In such a model, the AGN feedback is directly related to the impulsive, luminous quasar phase. We test our model by checking the consistency of its results against i) the QSO luminosity functions from $z=0$ to $z=4$; ii) the observed local relation between the black hole mass $m_{BH}$ and the mass (or velocity dispersion) of the host galaxy; iii) the color-dependent galaxy luminosity functions up to $z=3$. We then show how the efficiency of the AGN feedback associated to QSO phase increases with $z$. At low redshift it enhances the number of red bright galaxies, so that the color distribution of $M_r<-22$ objects is entirely dominated by red ($u-r>1.5$) galaxies; at $0.5\leq z\leq 2$ it yields rest-frame $U-V$ color distribution in agreement with existing observations. In the range $z\approx 1.5-2.5$, we find that $\approx 31 \%$ of galaxies contribute to the EROs population with $m_K<20$, in good agreement with the observed fraction $35\%$. In particular, at magnitudes $m_K=20$ (Vega system) the model yields an EROs surface density of $6.3\,10^{3}$ deg$^{-2}$ matching existing data. Extending our analysis to $z\approx 4$, the model matches the observed surface density $1.5\,10^3$ deg$^{-2}$ of DRGs at $m_K=20$; such a population is predicted to be dominated by galaxies with old stellar for $z\gtrsim 2.5$.'
author:
- 'N. Menci, A. Fontana, E. Giallongo, A. Grazian, S. Salimbeni'
title: 'THE ABUNDANCE OF DISTANT AND EXTREMELY RED GALAXIES: THE ROLE OF AGN FEEDBACK IN HIERARCHICAL MODELS'
---
INTRODUCTION
============
The color distribution of galaxies constitutes a key observable to constrain models of galaxy formation in a cosmological context. In fact, the most delicate sector of such models is that linking the age of stellar populations (closely related to the color) to the depth of the growing dark matter (DM) potential wells (hence the luminosity) of the galaxies hosting them. Indeed, while hierarchical models predict massive objects to be assembled at later cosmic times, they also predict their progenitor clumps to be formed in biased-high density regions of the primordial density field, where the enhanced density allowed early star formation.
Thus, the abundance of red, massive galaxies resulting from such models results from the balance between the later epoch of assembly predicted for the most massive objects and the older stellar population predicted to be in place in the progenitors of such objects. Such a delicate balance makes the comparison with the observations concerning the abundance and the color distributoin of massive galaxies a sensible probe for such models.
Indeed, recent models, including the enhancement in the star formation of progenitors of massive galaxies triggered by interactions, are able to match many global properties of the evolving galaxy population, like the observed decline of the global stellar mass density from $z=0$ to $z=2$ (see, e.g., Fontana et al. 2004) and the evolution of the B and UV luminosity functions (Somerville , Primack & Faber 2001; Menci et al. 2004a; Dahlen et al. 2005) from $z=0$ to $z=4$. The bimodal feature of the observed local color distribution (Strateva et al. 2001; Baldry et al. 2004; Bell et al. 2004; Giallongo et al. 2005) is also obtained in recent versions of hierarchical models (Kang et al. 2005; Menci et al. 2005; Bower et al. 2006; Croton et al. 2006), so that the existence of two well-defined populations appears to stem from the interplay between the biasing properties of the primordial density field – originating the DM condensations constituting the progenitors of galaxies – and the star formation and feedback processes driving the evolution of baryons in such progenitors.
Despite the above successes in matching the global behaviour of galaxy evolution, when one focuses on the observed proportion of red/blue galaxies, it is found that current hierarchical models underestimate the number of luminous/massive [*red*]{} galaxies. The models (Menci et al. 2005; Croton et al. 2006 with no AGN feedback) addressing the bimodal color distribution of local galaxies are characterized by an excess of massive ($M_r<-22$) galaxies in the blue ($u-r<1.5$) branch of the local color distribution compared to the red one; a similar deficit of luminous red galaxies is present in other semi- analytic models of galaxy formation (Somerville 2004). At higher redshift the problem is even more severe: the abundance of extremely red objects (EROs, with optical infra-red colors redder than a passively evolving elliptical, $R-K>5$) at $z\approx 1-2.5$ is underestimated by present hierarchical models (see McCarthy et al 2004; Cimatti et al. 2004; Daddi et al. 2005; Somerville et al. 2004) by factors up to ten (depending on the exact color and magnitude cut). This suggests that an additional process must be at work in determining the observed properties of galaxies; the overall agreement of hierarchical model predictions with the global (not color-selected) evolutionary properties of galaxies discussed above suggests that such a process should not constitute the main driver, but rather a complementary physical mechanism affecting mainly the evolution of the bright galaxy population.
In this respect, a possible solution has been proposed in terms of the feedback from Active Galactive Nuclei (AGNs) hosted at the center of galaxies (see, e.g., Ciotti & Ostriker 1997; Silk & Rees 1998, Haehnelt, Natarajan & Rees 1998, Fabian 1999). Indeed, the role of the energy injection from the AGNs into the interstellar medium of the host galaxies is at present one of the most pressing issues in the study of galaxy evolution, both on the observational and on the theoretical side.
On the one hand, blueshifted absorption lines in the UV and X-ray spectra of active galaxies reveal the presence of massive outflows of ionized gas from their nuclei; they are characterized by high-velocities (up to 0.4 c) indicating mass flows of 1-10 $M_{\odot}$/yr (see Crensahw, Kramer, George 2003; Chartas 2002, 2004; Pounds et al. 2003). Further evidences based on the radio and X-ray observations of galaxies (Böringer et al. 1993, Fabian et al. 2000; McNamara et al. 2000) indicate that bubbles of radio-emitting plasma are present in elliptical galaxies containing supermassive black <holes (SMBHs). On larger scales, fast (1000 km/s) massive (10-50 $M_{\odot}/yr$) flows of neutral gas are observed through 21-cm absorption of radio-loud AGNs (see Morganti, Tadhunter & Oosterloo 2005), indicating that AGNs have a major effect on the circumnuclear gas in the central kiloparsec region of AGNs. Since the powers of outflows are similar to (or even excede) the observed bolometric luminosities and the cooling losses, these observations indicate that the feedback from AGN has to be considered among the processes which regulate the evolution of baryons in the galactic potential wells, and, of course, the growth of the SMBHs in the host galaxy.
On the thoretical side, the impact of AGN feedback on galaxy formation has been investigated in a number of papers (Murray, Quataert & Thompson, Monaco & Fontanot 2005, Begelman & Nath 2005; Scannapieco, Silk & Bowens 2005); all the different assumed mechanisms for the energy injection can produce the expulsion of a significant fraction of the interstellar gas; the effectiveness of the AGN feedback in quenching the black hole growth and the subsequent star formation has been confirmed by recent aimed simulations of galaxy collisions triggering AGN activity (Di Matteo , Springel & Hernquist 2005) and of AGN feeding activated by gas infall (Kawata & Gibson 2005). However, inserting such a mechanism into a cosmological framework of galaxy formation constitues a challenging task. The model by Granato et al. (2004) successfully uses the shining of QSOs as a clock for the formation of elliptical galaxies, but does not include spiral galaxies, nor it provides predictions for the lower- energy AGNs at $z<1$. In ”ab initio” hierarchical models of galaxy formation in a cosmolgical context (Bower et al. 2005, Cattaneo et al. 2006; Croton et al. 2006; Kang, Jing, Silk 2006) SMBHs are assumed to grow during galaxy mergers both by merging with other BHs and through gas accretion, the latter being described in terms of tunable scaling laws. The inclusion of AGN feeedback in these models allows to suppress the cooling in massive haloes (a long-standing problem of hierarchical models); however, such a feedback must be still at work at low redshift to continuously suppress star formation in massive haloes at $z\lesssim 1$. Since the QSO activity drops sharply for $z\lesssim 2.5$, these authors assume the energy feedback responsible for the suppression of the cooling to be effective only in halos that undergo quasi-static cooling, and is associated to a continual and quiescent accretion of hot gas onto the SMBHs. In these models the feedback mechanism is thus associated only to a smooth accretion process which is not the main driver of BH growth, so they do not focus on the evolution of the luminous properties of QSOs and of the bright AGN sources.
Here we aim to investigate the effect on galaxy formation of the feedback directly associated with the observable QSOs and luminous AGNs. To this aim we include the AGN feedback model developed by Lapi, Cavaliere, Menci (2005) in our semi-analytic model of galaxy formation (Menci et al. 2004a; 2005), which includes the growth of SMBHs through gas accretion triggered by galaxy encounters. The AGN feedback we include in the present paper is substantially different from the implementations in Croton et al. (2006), Bower et al. (2005) and Cattaneo et al. (2006), since it is produced during the short AGN active phase, sweeping the cold gas content of the galaxies in a way similar to that resulting in the simulations by Di Matteo et al (2005); in addition, since it is associated with the observable active phase of AGNs, the feedback effect we investigate is mainly produced at high redshift. Before exploring the impact of such an impulsive form of AGN feedback on galaxy formation, we first test against observations the predicted evolution of the luminosity functions of QSOs from $z=4$ to the present, since the feedback we consider is directly related to their emission. After checking the consistency with the observed local $m_{BH}- \sigma$ relation between the BH mass $m_{BH}$ and the one-dimensional velocity dispersion $\sigma$ of the host galactic bulges, we compute the galaxy colour distribution at low ($z\lesssim 0.1$) and high (up to $z=3$) redshifts, discussing the effect of the AGN feedback on such distributions. Since the distinctive feature of our AGN feedback model (associating the feedback to the brigh QSO phase) is constituted by an effect peaking at redshift $1.5<z<3.5$, we particulary focus on the properties of high-redshift galaxies, and we compare with the observed number density of EROs up to $z=2.5$, and of Distant red Galaxies (DRGs), selcted by observed-frame $J-K>2.3$ (Franx et al. 2003; Van Dokkum et al. 2003), up to $z\approx 4$. In the final section we discuss our results.
THE MODEL
=========
We adopt the semi-analytic model (SAM) of galaxy formation described in Menci et al. (2005); this connects the baryonic processes (gas cooling, star formation, Supernovae feedback) to the merging histories of the DM haloes (with mass $M$, virial radius $R$ and circular velocity $V$) and of the galactic sub-haloes (with mass $m$, radius $r$ and circular velocity $v$) following the canonical recipes adopted by SAMs of galaxy formation (Kauffman, White & Guiderdoni 1993; Cole et al. 2000; Somerville, & Primack 1999; Somerville, Primack & Faber 2001; Menci et al. 2002; 2004a; Okamoto & Nagashima 2003).
The host DM haloes contain hot gas (with mass $m_h$) at the virial temperature $T$. The fraction of such a gas which is able to radiatively cool settles into a rotationally supported disk with gas mass $m_c$; the disk radius $r_d$ and rotation velocity $r_d$ are computed after the model by Mo, Mao & White (1998). The merging histories of the host haloes, and the dynamical friction and binary aggregations acting on the included sub-halos, are computed adopting the canonical Monte Carlo technique as in Menci et al. (2005). We keep the implementation of gas cooling, star formation and feedback described in the above paper with the same choice of free parameters (the normalization of the star formation timescale and of the feedback efficiency).
The integrated stellar emission from galaxies is computed convolving the star formation rate resulting from our model with the synthetic spectral energy distributions from Bruzual & Charlot (1993), adopting a Salpeter IMF. The dust extinction affecting the intrinsic galactic luminosities is computed assuming the dust optical depth to be proportional to the metallicity of the cold phase (computed assuming a constant effective yield) and to the disk surface density (see Menci et al. 2002 for further details). When no otherwise specified, we assume a Galaxy extinction curve. We have shown in our previous paper that the model is able to produce Tully-Fisher relation, cold gas and disk size distributions, and B-band galaxy luminosity functions (from $z=0$ to $z\approx 4$) in good agreement with observations. As for the growth of SMBHs with the associated the AGN emission, and for the corresponding AGN feedback, we adopt the models by Cavaliere & Vittorini (2000) and by Lapi, Cavaliere & Menci (2005), respectively. Here we recall their basic points.
The Growth of SMBHs
-------------------
The growth of SMBH is implemented as described in Menci et al. (2003); the accretion of cold gas is triggered by galaxy encounters (both fly-by and merging), which destabilize part of the avaiable cold gas. The rate of interactions is (Menci et al. 2003) $$\tau_r^{-1}=n_T\,\Sigma (r_t,v,V_{rel})\,V_{rel}.$$ Here $n_T$ is the number density of galaxies in the same halo and $V_{rel}$ is their relative velociyt. The encounters effective for angular momentum transfer require 1) the interaction time to be comparable with the internal oscillation time in the involved galaxies (resonance), ii) the orbital specific energy of the partners not to exceed the sum of their specific internal gravitational energies. The cross section $\Sigma$ for such encounters is given by Saslaw (1995) in terms of the distance $r_t\approx 2r$ for a galaxy with given $v$ (see Menci et al. 2003, 2004).
The fraction of cold gas accreted by the BH in each interaction event is computed in terms the variation $\Delta j$ of the specific angular momentum $j\approx Gm/v_d$ of the gas, to read (Menci et al. 2003) $$f_{acc}\approx {1\over 8}\,
\Big|{\Delta j\over j}\Big|=
%{1\over 6}\Big\langle {Gm'\,r_d\over V\,b}/{G\,m\over v_d}\Big\rangle =
{1\over 8}\Big\langle {m'\over m}\,{r_d\over b}\,{v_d\over V_{rel}}\Big\rangle\, .$$ Here $b$ is the impact parameter, evaluated as the average distance of the galaxies in the halo. Also, $m'$ is the mass of the partner galaxy in the interaction, and the average runs over the probability of finding such a galaxy in the same halo where the galaxy $m$ is located.
The average cold gas accreted during an accretion episode is thus $\Delta
m_{acc}=f_{acc}\,m_c$, and the duration of an accretion episode, i.e., the timescale for the QSO to shine, is assumed to be the crossing time for the destabilized cold gas component, $\tau=r_d/v_d$.
The bolometric luminosity so produced by the QSO hosted in a given galaxy is then given by $$L(v,t)={\eta\,c^2\Delta m_{acc}\over \tau} ~.$$ We adopt an energy-conversion efficency $\eta= 0.1$ (see, e.g., Yu & Tremaine 2002). The SMBH mass $m_{BH}$ grows through the accretion described above and by coalescence with other SMBHs during galaxy merging. As initial condition, we assume small seed BHs of mass $10^2\,M_{\odot}$ (Madau & Rees 2001) to be initially present in all galaxy progenitors; our results are insensitive to the specific value as long as it is smaller than $10^5\,M_{\odot}$.
In our Monte Carlo model, at each time step we assign to each galaxy the probability to interact after the rate given in eq. (1). According to such a probability, we assign an active BH accretion phase (with duration $\tau$) to the considered galaxy, and compute the accreted cold gas and associated QSO emission through equations (2) and (3).
The Feedback from AGNs
----------------------
To explore the dynamical effect of feedback occurring during the active AGN phase on the interstellar medium we adopt the model by Lapi, Cavaliere & Menci (2005). They compute the effect of an energy injection $\Delta E$ by AGNs on the surrounding gas by solving, in the ”shell approximation”, the equations for a blast-wave propagating outwards in the interstellas gas, including the effects of gravity and of the gas density gradient. The perturbed gas is confined to a shell with outer (shock) radius $R_s(t)$ which sweeps the gas around the AGN; the effect is similar to that resulting in the simulations by Di Matteo et al. (2005), although the latter have been performed only in selected cases of major mergers, while our treatment applies also to less energetic events (the accretion rate, and hence the energy $\Delta E$ injected by the AGN, is determined by eq. 2).
The mass $\Delta m$ expelled out of the virial radius by the blowout is computed as a function of the ratio $\Delta E/E$ where $E$ is the gas binding energy; The values of $\Delta m$ and $\beta$ for any given ratio $\Delta E/E$ are tabulated in Cavaliere, Lapi & Menci (2002), the former being well approximated (to better than 10 %) by $\Delta m/m\approx 0.5 \Delta E/E$ for $\Delta E/E<1.4$.
In our semi-analytic model, for any galaxy undergoing an active AGN phase (see sebsection 2.1), we compute $\Delta E=f\,L\,\tau$ assuming the AGN feedback efficiency $10^{-2}\lesssim f\lesssim 10^{-1}$ as a free parameter, the lower value being more appropriate to radio-quiet AGNs because of the flat spectrum and the low photon momenta; observations of wind speeds up to $v_W\approx 0.4 c$ suggest values around $v_w/2c\approx 10^{-1}$ associated with covering factors of order $10^{-1}$ (see Chartas et al. 2003; Pounds et al. 2003). We than compute the fraction $\Delta m$ of cold gas expelled by AGNs; the galactic cold gas expelled by the AGN feedback enriches the hot gas phase which fills the dark matter potential wells of the structure (group or cluster) hosting the galaxy. Expanding out of the galaxy, the blastwave also expels a tiny fraction of the hot gas in the host structure and re-sets the hydrostatic equilibrium of such a hot gas to a new temperature $T+\Delta T$ larger than the initial temperature $T$ by a factor $1/\beta$ (with values in the range 1-1.1 in most cases), also computed after Cavaliere, Lapi & Menci (2002).
RESULTS
=======
Here we present our results for a $\Lambda$-CDM cosmology with $\Omega_0=0.3$, $\Omega_{\lambda}=0.7$, a baryon fraction $\Omega_b=0.05$, and Hubble constant $h=0.7$ in units of 100 km s$^{-1}$ Mpc$^{-1}$. The star formation and stellar feedback parameters are the same as in Menci et al. (2005), while the AGN feedback efficiency (see sect. 2.2) is set to $f=0.05$.
Testing the model: The AGN evolution and the Intergalactic Gas
--------------------------------------------------------------
Since our feedback is tightly related to the active AGN phase, we require the model to match not only the observables concerning the galaxy population, but also those concerning the AGN population, in order to have a reliable modelization for the interplay between the cooling and star formation processes and the growth of SMBH with the ensuing AGN emission and feedback.
\[2.8\]
[Fig. 1. - The relation between the black hole mass $M_{BH}$ and the one-dimensional velocity dispersion $\sigma$ of the bulges of the host model galaxies at $z=0$ is compared with data from Ferrarese & Merritt (2000, filled squares) and Gebhardt et al. (2000, circles); the color code refers to logarithm of the abundance normalized to its maximum value. ]{}
Thus, we first show in fig. 1 the $m_{BH}-\sigma$ relation, to check that, when integrated over time, the BH accretion that we implement in our model is consistent with the local observations for the whole range of galactic mass spanned by the model. Note that the spread in the $m_{BH}-\sigma$ increases for decreasing galactic mass, a feature shared by all hierarchical models and related to the increasing number of merging histories involved the formation of massive haloes, which results in a lower statistical deviation. Note also that our average relation remains close to a power-law, at variance, e.g., with the Bower et al. (2005) model, which predicts a steepening of the relation for large $\sigma$; in fact, such a steepening is due to the second ”mode” of BH growth implemented in that model, associated with the smooth accretion of hot gas and which ensures that the luminosity of the BH is sufficient to quench cooling in the most massive haloes.
Since the time behavior of the energy injection from AGNs is crucial in determining the effects on the galaxy properties, and in our model such injection is related to the active QSO phase, we test in fig. 2 the evolution of the QSO luminosity distribution by comparing it with the observations from $z=0$ up to $z=4$. Note that our model is able to match the observed decline in the population of bright sources from $z=2$ to the present; this is due to 1) the decline of merging events refueling the cold gas reservoir of the galaxies; 2) the exhaustion of such a reservoir, due to its earlier conversion in stars; 3) the decline in the rate of interactions (triggering the active accretion phase); 4) the decrease of the effectiveness of interactions in destabilizing the cold gas (see eq. 1 and the discussion in Menci et al. 2003).
\[0.36\]
[ Fig. 2. - The QSO B-band luminosity functions from our model (solid lines) are shown for $z=0.55$ (lower curve), $z=1.2$ (middle curve) and $z=2.2$ (uppermost curve), and are compared with the data. These are taken from Hartwick & Shade (1990, solid squares) and Boyle et al. (2000, crosses), and rescaled to our cosmology with $\Omega_0=0.3$, $\Omega_{\lambda}=0.7$, $h=0.7$. We also show the model results for $z=3$ (dashed line) and $z=4$ (dotted line), compared to the Sloan data at $z04.2$ from Fan et al. (2001, diamonds). The blue luminosity $L_B$ has been obtained by applying a bolometric correction of 13 (Elvis et al. 1994) to the bolometric luminosity in eq. (3). ]{}
As a final test, we also probe the model predictions for the hot gas phase which is enriched by the gas expelled from the galaxies by the AGN feedback (see sect. 2.2). In particular we show in fig. 3 how the baryon fraction and the mass temperature relation resulting from the model compares with present observations. This shows that: i) that the amount of hot gas (enriched by the blowout of cold galactic gas due to the AGN feedback) is not enhanced over the baryon fraction in groups and clusters; ii) the temperature of the hot gas is not enhanced by the contribution from AGNs up to values exeeding the observations for a given cluster mass. Indeed, the value of $\beta$ (sect. 2.2) stays close to unity for typical values of the energy $\Delta E$ injected by AGNs.
\[3.2\]
The agreement with observations is not unexpected: indeed, while the AGNs are expected to produce significant effects on the cold gas contained in the shallow potential wells of galaxies (and hence to appreciably affect the star formation history of galaxies which entirely depends on the cold gas phase), the relative variation in the hot gas mass (which is an order of magnitude larger than the $m_c$) remains however small, and has little effect on the baryon fraction in groups and clusters of galaxies. In fact, for such structures the ratio $\Delta E/E$ (see sect. 2.2) is much less than unity due to their deep potential wells (large gas binding energy $E$), at least for our assumed value of the AGN feedback efficiency $f\approx 10^{-2}$, so that relative changes in the temperature ($\Delta T/T\lesssim 0.1$) and hot gas mass ($\Delta m/m_h \lesssim 0.05$) remain much smaller than one.
The Bimodal Color Distribution of Galaxies
------------------------------------------
The effect of the energy injection from AGNs on the color distribution of local galaxies is shown in fig. 4 for three luminosity bins, and compared with the data from the Sloan survey (SDSS) given in Baldry et al. (2004) for the $u-r$ colors; details on the SDSS $u$ and $r$ bands are given by the above authors. As expected, the AGN feedback affects only the color distribution of luminous galaxies; this is due to their larger cross section for interactions (triggering the AGN activity) and to the larger fraction of destabilized gas accreted by the SMBH (powering the AGN activity). Note also that both the [*bimodal partition*]{} and the [*downsizing*]{} (i.e., the correlation bewetten the age of the stellar population, or the color, and the galaxy luminosity) are not determined by the AGN feedback, as we showed in our previuous paper and as confirmed by the recent results by Croton et al.(2006). Rather, the AGN feedback affects the partition of galaxies between the blue and the red population, enhancing the fraction of red galaxies. In particular, the distribution of very luminous ($M_r<-22$) galaxies is dominated by red objects when AGN feedback is included, in agreement with observations.
\[0.35\]
[Fig. 4. - Predicted rest-frame $u-r$ color distributions (heavy lines) for different dust extiction laws are compared with the Gaussian fit to the SDSS data (from Baldry et al. 2004, dots) for different magnitude bins. The left column refers to the model without the inclusion of AGN feedback, while right column shows the results when AGN feedback is turned on. The distributions are normalized to the total number of galaxies in the magnitude bin. ]{}
At higher redshift, we expect the effect of the AGN feedback directly related to the bright QSO phase to be even more important, due to the enhanced AGN activity at such cosmic epochs. This is first shown in fig. 5, where we show the color-magnitude relation at various redshifts for the model with no AGN energy injection and the model where the AGN feedback is included.
Note that the energy injection from AGN affects appreciably the number of red galaxies at redshifts $z\gtrsim 2$. In particular, at such redshifts, bimodality only appears in the model including the AGN feedback, while at lower redshifts $z=1-2$ the bimodal partition of the color distribution is appreciably enhanced by the effect of AGNs.
\[3.5\]
To perform a more quantitative comparison between the predicted and the observed bimodal properties of the galaxy population at intermediate and high redshifts, we compare in figs. 6 and 7 with the observational results from the Great Observatories Origins Deep Survey (GOODS) database. We show in fig. 6 the B-band luminosity functions of the red and blue components at different redshifts and compare them with the GOODS data (Salimbeni et al. 2006). According to the selection criterium discussed in Giallongo et al. (2005), in this plot we define the blue and the red population as those separated by the rest-frame color $U-V=\alpha(19.9+M_B)+(U-V)_0$ (Vega system), where the parameters $ \alpha$ and $(U-V)_0$ are given in the caption for the three redshift bins. At high redshifts ($z\gtrsim 2$) the model with no AGN feedbak substantially underestimates the number of red objects, while the model including the AGN energy injection produces a reasonable agreement at all redshifts for both the blue and the red population, although the long-standing problem of the overprediction of faint objects is still present in the model.
Note also how the galaxies effectively reddened by the inclusion of the AGN feedback constitute a minor fraction of the galaxies contributing to the B and UV band luminosity; although the abundance of bright galaxies in the UV and B bands is indeed decreased by the inclusion of the AGN feedback (see top-left and bottom-left panels of fig. 6), the resulting luminosity functions are still consistent with the data.
\[0.43\]
A direct comparision of the predicted and the observed color distribution of galaxies at intermediate redshifts $z=1-2$ is perfomed in fig. 7 for three magnitude bins, to show how the model matches the observed early (at $z\approx
2$) appearence of the same downsizing effect which marks the color distribution at low redshifts (see text below fig. 4), indicating that the model captures the observed [*early*]{} decline of the star formation rate in massive, bright galaxies.
\[0.4\]
The Color Distribution and Abundance of EROs
--------------------------------------------
From the discussions above, and from inspection of fig. 5, we expect that the inclusion of the AGN feedback will solve the long-standing problem of semi- analytic models related to their severe underestimate of EROs (with observed $R-K>5$) at $z\approx 1.5-2.5$ (see McCarthy et al 2004; Cimatti et al. 2004; Daddi et al. 2005; Somerville et al. 2004) that we have recalled in the Introduction. Such a point is addressed in fig. 8, where we plot the observed frame $R-K$ color distribution of $K<20$ (Vega system) galaxies in the redshift range $1.7<z<2.5$ and compare it with the observed distribution. Note how the inclusion AGN feedback strongly enhances the number of predicted EROs (see also fig. 5, bottom panel), to yield a fraction 0.31 of objects with $R-K>5$ close to the observed value 0.35. The normalization of both data and model predictions is provided by the redshift distribution of $K<20$ galaxies shown in the bottom panel. The same panel also shows how the model matches the observed $z$-distribution of $m_K<20$ EROs, and (in the inset) the distribution of $I-K>4$ galaxies down to $m_K=21.5$ to probe the predictions of the model for the most luminous galaxies up to redshift $z=3$.
Note that the matching between the model [*global*]{} (not color selected) redshift distributions and the observations is not due to the effect of AGN feedback. Indeed, this is due to the effect of starbursts triggered not only by margers (like, e.g., in the model by Somerville et al. 2004) but also by galaxy fly-by (see Menci et al. 2003); the latter produce starbursts with a lower efficiency $\approx 0.1-0.4$ (at $z\gtrsim 3$) for a single event, but dominate the encounter statistics due to the high rate of fly-by events at such redshifts; the point is discussed in detail in Menci et al. (2003).
\[0.4\]
The predictions in fig. 8 constitute a distinctive feature of our model, where the AGN feedback is associated with the bright AGN (QSO) phase (at variance with the other existing semi-analytic model, where it is associated to a smooth accretion phase, continuing down to low redshift), thus strongly acting on the properties of galaxies at $1.5\lesssim z<3.5$, when the AGN activity reaches its maximum. Thus we expect that in our model the density of EROs at magnitudes fainter than $m_K=20$ (contributed by high-redshift objects) to be still rising. Such a specific prediction of our model is shown in fig. 9 where we plot the differential magnitude counts of EROs resulting from our model. The model is able to match the observed value $\approx 6.3\,10^3$ deg$^{-2}$ for the surface density of EROs with magnitude $m_K\approx 20$, and predicts a continuous increase in the EROs surface density to reach values $\approx
3\,10^4$ deg$^{-2}$ at $m_K=22$.
The agreement of the model with observations holds not only for all EROs, but also in detail for density of dust-free passive EROs, which constitute about 80 % of all EROs in our model. This is shown in the bottom panel of fig. 9, where we compare the model predictions for the cumulative number counts of EROs without dust extinction with the data concerning passive ù dust-free EROs presented in by Miyazaki et al. (2003).
\[0.35\]
The Abundance and Redshift Distribution of DRGs
-----------------------------------------------
A more sensible probe for the model predictions at high redshifts is constituted by the observed surface density and redshift distribution of Distant Red Galaxies selected by the criterion $J-K>2.3$ (Franx et al. 2003; Van Dokkum et al. 2003), as shown in fig. 10. The predicted surface density of such objects (Top panel) matches the observed value $1.5\,10^3$ deg$^{-2}$ at $m_K=20$, while at $m_K=22$ it reaches values $\approx 10^4$ deg$^{-2}$. Since the $J-K$ color cut allows for selection of red galaxies up to higher redshift compared to EROs, we can probe the model predictions concerning their redshift distribution up to $z\approx 4$ (bottom panel); the redshift distribution of the total DRG population is peaked at $z\approx 2.5$, in good agreement with the observed distributions, and shows a minor peak at $z\approx 1$. To investigate the nature of the DRG galaxies in our model, we also show the model predictions when no dust extinction is included (dashed line in the bottom panel of fig. 10); in such a case, the low-redshift peak disappears, showing that the low-redshift part of the distribution is entirely contributed by heavily extincted galaxies peaking at $z\approx1$, while at $z\gtrsim 1.5$ the distribution is dominated by galaxies with old stellar population.
\[0.4\]
Summary and Discussion
======================
We have included the energy injection from AGN feedback in our semi- analytic model of galaxy formation, which self-consistently includes the growth of SMBHs and the corresponding AGN emission. We have focussed on the effects of such a feedback on the evolution of the galaxy color distribution, and found that the inclusion of the energy feedback from AGNs enhances the number of red galaxies at low, intermediate and high redshifts, as to match existing observations up to $z\approx 4$. In particular, we find: $\bullet$ at low redshifts, the color distribution of bright ($M_r<-22$) galaxies is entirely dominated by red ($u-r>1.5$) objects (fig. 2). On the other hand, the color distribution of faint galaxies with ($M_r>-18$) remains peaked at blue ($u-r<1.5$) colors, and is not affected by the AGN feedback. $\bullet$ the effect of AGN feedback increases at higher redshifts, enhancing the fraction of red galaxies (figs. 4, 5, 6); a bimodal color distribution at $z\gtrsim 1.7$ is only obtained when the AGN feedback is considered (see the bottom panels in fig. 5). $\bullet$ at $1.5<z<2.5$ the model predicts a fraction 0.31 of EROs (with $R-K>5$) close to the observed value 0.35 (see fig. 8); the predicted surface density at $m_K=20$ is $6.3\,10^{3}$ deg$^{-2}$ while at $m_K=22$ it increases to $\approx 3\,10^{4}$ deg$^{-2}$ (fig. 9). $\bullet$ at higher redshift $2<z<4$ the surface density of predicted DRGs (with $J-K>2.3$) is $1.5\,10^3$ deg$^{-2}$ at $m_K=20$ in agreement with observations (fig. 10), while at $m_K=22$ the model yields values $\approx 10^4$ deg$^{-2}$, even slightly larger than current estimates $6\,10^3$ deg$^{-2}$ based on HST data. $\bullet$ the redshift distribution of $m_K<22$ DRGs is characterized by a major peak at $z\approx 2.5$ and a lower peak at $z\approx 1$ (see fig. 10); we find that the latter is contributed only by heavily absorbed galaxies, while at $z\gtrsim 1.5$ the distribution is dominated by galaxies with old stellar populations. Such a finding is in agreement with the recent results by Papovich et al. (2006), who measured the same partition in the redshift distribution of DRGs, although in their sample the peak of the lower-redshift, extincted galaxies is at a slightly larger $z\approx 1.5$. Our results are also consistent with the observed stronger clustering of high-redshift DRGs compared their lower-redshift counterpart (Grazian et al. 2006).
We stress that in our model the downsizing (i.e., the older ages of stellar population observed in the most massive galaxies) and the appearence of a bimodal color distribution at $z\lesssim 1.5$ are not caused by the effect of AGNs. Indeed, hierarchical models predict massive objects to be assembled from progenitor clumps to be formed in biased-high density regions of the primordial density field, where the enhanced density allowed early star formation, so they naturally predict older stellar populations to be present in massive galaxies. In our model, the bimodality observed in the galaxy color distribution is the final results of the interplay between the above biasing properties of the primordial density field from which the progenitor of local galaxies have formed, and the dependence of feedback/star formation processes on the depth of the DM potential wells, as we showed in our previous paper (Menci et al 2005), and also obtained in independent works (Croton et al. 2006, case with no heating source). Rather, the AGN feedback strongly enhances the proportion of galaxies populating the red (or extremely- red) branch of the color distribution, and – in our model – such an enhancement is particulary effective at high-$z$ where the AGN activity is much larger than the present (see bottom panels in fig. 5).
Note that the gas expelled from galaxies by the AGN feedback in our model enriches the hot gas content of the host halo (group or cluster); thus, the AGN feedback is partially counteracted by the cooling of such a hot gas which may re-convert part of it back to the cold gas at the center of galaxies. The final effect of the AGN feedback depends on the balance of the above two effects.
At high redshifts, the AGN activity is so frequent, and the AGN luminosities so high, that AGN feedback rapidly counteracts such a cooling. The net result, is that galaxies retain - on average - a lower amount of cold gas (and hence show redder colors) compared to the case with no AGN feedback (see fig. 5, bottom panels).
At lower redshift, the AGN activity decreases and the effectiveness of AGN feedback drops. Nevertheless, the effects of the AGN feedback (mainly generated at high redshifts) are not completely erased even at low redshift, since the reconversion of the hot gas in massive haloes into the cold phase is suppressed by i) the lower gas densities in the larger host haloes (the radius of groups and clusters has grown); ii) the lower value of the cooling function (which drops for host halo virial temperatures $T>10^5$ K) haloes; iii) the reheating of cold galactic gas during major mergers as described in Menci et al. (2005). These processes (also present in our previous papers) allow for the persistence at low $z$ of effects (like the gas consumption in starbursts of by the AGN feedback) mainly active at high redshifts.
Such a latter feature is peculiar of our model for AGN feedback, directly associated with the impulsive, luminous quasar phase (the main phase of BH growth) triggered by galaxy interactions. During such a phase, a small coupling to the gas is expected for the radiative output; estimates based on the observed wind speeds up to $v_W\approx 0.4 c$ suggest values $v_W/2c\approx 10^{-1}$ associated with covering factors of order $10^{-1}$, so that efficiencies close to our adopted value $f\approx 5\, 10^{-2}$ are expected. In this phase gas is swept out of the galactic potential wells in a way similar to that resulting from the simulation of Di Matteo et al. (2005). Such a mode of AGN feedback differs from that implemented in the other recent semi-analytic models, where the outflows are associated (with an efficiency close to 100 %) to a lower-accretion phase of AGN, corresponding to accretion rates so low that the AGN is not optically luminous, and which is triggered by gas shock-heating (Cattaneo et al. 2006) or by smooth hot gas accretion within a static hot gas halo (Croton et al. 2006; Bower et al. 2005). The feedback associated to such a smooth mode of AGN accretion may start already at $z\approx 3$ but increases with time down to low redshifts. In this respect, the color distribution of galaxies at high-redshifts constitutes an important probe to discriminate between the two scenarios of AGN accretion and feedback pending a definite observational evidence for a statistically significant occurence of strong outflows in massive galaxies.
We thank the referee for helpful comments.
Baldry, I.K., Glazebrook, K., Brinkmann, J., Zeljko, I., Lupton, R.H., Nichol, R.C., Szalay, A.S. 2004, ApJ, 600, 681 Begelman, M.C., Nath, B.B, 2005, MNRAS, 361, 1387 Bell, E., Wolf, C., Meisenheimer, K., Rix, H.-W., Borch, A., Dye, S., Kleineinrich, M., McIntosh, D. 2004, ApJ, 608, 752 Böringer, H., Voges, W., Fabian, A.C., Edge, A.C. Neumann, D.M., 1993, MNRAS, 264, L25 Boyle, B.J., Shanks, T., Croom, S.M., Smith, R.J., Miller, L., Loaring, N., Heymans, C. 2000, MNRAS, 317, 1014 Bower, R.G., Benson, A.J., Malbon, R., Helly, J.C., Frenk, C.S., Baugh, C.M., Cole, S., Lacey, C.G. 2006, astro-ph/0511338 Cattaneo, A., Dekel, A., Devriendt, J., Guiderdoni, B., Blaizot, J., preprint \[astro-ph/0601295\] Calzetti, D., 1997, in Proc. of “The Ultraviolet Universe at Low and High Redshift : Probing the Progress of Galaxy Evolution”, AIP Conference Proceedings, 408, p.403 Cavaliere, A., Vittorini, V., 2000, ApJ, 543, 599 Cavaliere, A., Lapi, A., Menci, N., 2002, ApJ, 581, L1 Cimatti, A. et al. 2002, , 381, L68 Cimatti, A. et al. 2004, Nature, 430, 184 Chartas, G., Brandt, W. N., Gallagher, S. C., Garmire, G. P., 2002, ApJ, 579, 169 Chartas, George; Brandt, W. N.; Gallagher, S. C. 2004, Proceedings of IAU Symposium, No. 222, Edited by T. Storchi-Bergmann, L.C. Ho, and Henrique R. Schmitt. Cambridge, UK: Cambridge University Press, 2004., p.411-414 Ciotti, L. & Ostriker, J.P., 1997ApJ, 487, L105 Cole, S., Lacey, C.G., Baugh, C.M., Frenk, C.S., 2000, MNRAS, 319, 168 Croton, D.J, Springel, V., White, S.D.M., De Lucia, G., Frenk, C.S., Gao, L., Jenkins, A., Kauffmann, G., Navarro, J.F., Yoshida, N., 2006, MNRAS, 365, 11 Crenshaw, D.M., Kraemer, S.B.. George, I.M. 2003, ARA&A, 41, 11 Daddi, E. et a. 2002, , 384, L1 Daddi, E. et al. 2005, ApJ, 626, 680 Dahlen, T. et al. 2005, ApJ, 631, 126 Di Matteo, T., Springel, V., Hernquist, L., 2005, Nature, 433, 604 Elvis M. et al. 1994, ApJS, 95, 68 Fabian, A. 1999, MNRAS, 308, 39 Fabian, A.C., et al., 2000, MNRAS, 318, L65 Fan, X. et al., 2001, ApJ, 121, 54 Ferrarese, L. Merritt, D., 2000, ApJ, 539, L9 Figuenov, A., Reiprich, T.H., Böringer H., 2001, A&A, 368, 749 Fontana, A. et al., 2004, A&A, 424, 23 Franx, M. et al. 2003, ApJ, 587, L79 Gebhardt, K. et al., 2000, ApJ, 539, L13 Giallongo, E., Salimbeni, S., Menci, N., Zamorani, G., Fontana, A., Dickinson, M., Cristiani, S., Pozzetti, L. 2005, ApJ, 622, 116 Granato, G.L., De Zotti, G., Silva, L., Bressan, A., Danese, L., 2004, ApJ, 600, 580 Grazian, A. et al., 2006, A&A, accepted \[astro-ph/0603095\] Haehnelt, M.J., Natarajan, P., & Rees, M.J., 1998, MNRAS, 300, 817 Hartwick, F.D.A., Shade, D., 1990, ARA[&]{}A, 28, 437 Kang, X., Jing, Y.P., Mo, H.J., Börner, G. 2005, ApJ, 631, 21 Kang, X., Jing, Y.P., Silk, J. 2006, preprint \[astro-h/0601685\] Kauffmann, G., White, S.D.M., & Guiderdoni, B., 1993, MNRAS, 264, 201 Kawata, D., Gibson 2006, MNRAS, 358, L16 Lapi, A., Cavaliere, A., Menci, N., 2005, ApJ, 619, 60 Lin, Y.-T., Mohr, Y.J., Stanford, S.A. 2003, ApJ, 591, 749 Madau, P., & Rees, M.J., 2000, ApJ, 551, L27 McCarthy, P.J. 2004, ARA&A, 42, 477 McNamara, B.R., et al., 2000, ApJ, 534, L135 Menci, N., Cavaliere, A., Fontana, A., Giallongo, E., Poli, F., 2002, ApJ, 578, 18 Menci, N., Cavaliere, A., Fontana, A., Giallongo, E., Poli, F., Vittorini, V. 2003, ApJ, 587, L63 Menci, N., Cavaliere, A., Fontana, A., Giallongo, E., Poli, F., Vittorini, V. 2004a, ApJ, 604, 12 Menci, N., Fiore, F., Perola, G.C., Cavaliere, A. 2004b, ApJ, 606, 58 Menci, N., Fontana, A., Giallongo, E., Salimbeni, S. 2005, ApJ, 632, 49 Mirabel, I.F., Sanders, D.B., 1996, ARA&A, 34, 749 Miyazaki, M. et al., 2003, PASJ, 55, 1079 Mo, H.J, Mao S., & White, S.D.M., 1998, MNRAS, 295, 319 Monaco, P., Fontanot, F., 2005, MNRAS, 359, 283 Morganti, R., Tadhunter, C. N., Oosterloo, T. A., 2005, A&A, 444, L9 Murray, N., Quataert, E., & Thompson, T.A. 2005, ApJ, 618, 569 Okamoto, T., Nagashima, M. 2003, ApJ, 587, 500 Papovich, C. et al., 2006, ApJ, accepted \[astro-ph/0511289\] Pounds, K., King, A.R., Page, K.L., O’Brien, P.T 2003, MNRAS, 346, 1025 Roche, N.D., Dunlop, J., Almaini, O. 2003, MNRAS, 346, 803 Salimbeni, S., et al. 2006, in preparation Scannapieco, E., Silk, J., & Bouwens 2005, ApJ, in press Silk, J. & Rees, M.J. 1998, A&A, 331, 1 Somerville, R.S., & Primack, J.R., 1999, MNRAS, 310, 1087 Somerville, R.S., Primack, J.R., & Faber, S.M., 2001, MNRAS, 320, 504 Somerville, R.S., et al. 2004, ApJ, 600, L135 Strateva, I. et al. 2001, AJ, 122, 1861 Van Dokkum, P.G. et al. 2003, ApJ, 587, L83 Yu, Q., & Tremaine, S., 2002, MNRAS, 335, 965
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper we define a class of combinatorial structures the instances of which can each be thought of as a model of directed hypergraphs in some way. Each of these models is uniform in that all edges have the same internal structure, and each is simple in that no loops or multiedges are allowed. We generalize the concepts of Turán density, blowup density, and jumps to this class and show that many basic extremal results extend naturally in this new setting. In particular, we show that supersaturation holds, the blowup of a generalized directed hypergraph (GDH) has the same Turán density as the GDH itself, and degenerate GDHs (those with Turán density zero) can be characterized as being contained in a blowup of a single edge. Additionally, we show how the set of jumps from one kind of GDH relates to the set of jumps of another. Since $r$-uniform hypergraphs are an instance of the defined class, then we are able to derive many particular instances of jumps and nonjumps for GDHs in general based on known results.'
author:
- Alex Cameron
bibliography:
- 'GDH.bib'
title: Extremal Problems on Generalized Directed Hypergraphs
---
Introduction
============
This paper explores Turán-type problems for a class of relational structures that can each be thought of as generalized directed hypergraphs. This class includes the standard undirected $r$-uniform hypergraphs that have been extensively studied in combinatorics as well as totally directed $r$-uniform hypergraphs where each edge is a set of $r$ vertices under a linear ordering. Instances of this latter structure have been studied in the extremal setting by Erdős, Brown, Simonovits, Harary, and others [@brown1973; @brown1969; @brown2002; @brown1984]. Other instances of this class are uniform versions of the model used to represent definite Horn formulas in the study of propositional logic and knowledge representation [@angluin1992; @russell2002]. The combinatorial properties of this model have been recently studied by Langlois, Mubayi, Sloan, and Gy. Turán in [@langlois2009] and by this author in [@cameron2015] and [@cameron2015deg]. Other structures in this class are slight variations on the $d$-simplex structures studied by Leader and Tan in [@leader2010].
Turán-type extremal problems for uniform hypergraphs make up a large and well-known area of research in combinatorics that ask the following: “Given a family of forbidden $r$-uniform hypergraphs $\mathcal{F}$ what is the maximum number of edges an $r$-uniform hypergraph on $n$ vertices can have without containing any member of $\mathcal{F}$ as a (not necessarily induced) subgraph?" Such problems were named after Paul Turán due to his important early results and conjectures concerning forbidden complete $r$-graphs [@turan1941; @turan1954; @turan1961].
A related question for undirected hypergraphs was proposed by Erdős known as the jumping constant conjecture. A real number $\alpha \in [0,1)$ is called a jump for an integer $r \geq 2$ if there exists some positive constant $c$ which depends only on $\alpha$ such that for any $\epsilon > 0$ and positive integer $l$ there exists a positive integer $N$ for which any $r$-uniform hypergraph on $n \geq N$ vertices which has edge density at least $\alpha + \epsilon$ contains a subgraph on $l$ vertices with edge density at least $\alpha + c$. It is well-known that when $r=2$, every $\alpha \in [0,1)$ is a jump [@erdos1966; @erdos1946]. Moreover, every $\alpha \in \left[0, \frac{r!}{r^r}\right)$ is a jump for $r \geq 3$ [@erdos1971]. In 1984, Frankl and Rödl disproved the jumping constant conjecture when they found the first instance of a nonjump for each $r \geq 3$[@frankl1984]. Since then many infinite sequences of nonjumps have been found, but the smallest known nonjump to date is $\frac{5r!}{2r^r}$ for each $r \geq 3$ determined by Frankl, Peng, Rödl, and Talbot in [@frankl2007]. The only additional jumps that have been found are all $\alpha \in [0.2299,0.2316), \left[0.2871,\frac{8}{27} \right)$ for $r=3$ found by Baber and Talbot in [@baber2011] using Razborov’s flag algebra method [@razborov2007].
Extremal issues of these kinds have also been considered for digraphs and multigraphs (with bounded multiplicity) in [@brown1973] and [@brown1969] and for the more general directed multi-hypergraphs in [@brown1984]. In [@brown1969], Brown and Harary determined the extremal numbers for several types of specific directed graphs including all tournaments - that is, a digraph with one edge in some orientation between every pair of vertices. In [@brown1973], Brown, Erdős, and Simonovits determined the general structure of extremal sequences for every forbidden family of digraphs analogous to the Turán graphs for simple graphs.
In [@brown1984], Brown and Simonovits proved several general extremal results about $r$-uniform directed $q$-hypergraphs. In this model the edges are ordered $r$-tuples of vertices with multiplicity up to $q$ for some fixed positive integer $q$. Among their results on this model are three that will be reproduced in this paper in a more general setting: Supersaturation, Continuity, and Approximation. Roughly speaking supersaturation implies that a large graph with an edge-density more than the Turán density for a particular forbidden family must contain many copies of members of that family. Continuity shows that given an infinite forbidden family, we can get arbitrarily close to its extremal number with a finite subfamily. Approximation is a structural result that shows that given a forbidden family, we can approximate an extremal sequence to an arbitrarily small difference by taking some sequence of graphs that all exclude this family and which all fall into some “nice" form. All of these notions will be made rigorous in the paper.
In [@langlois2010] and [@langlois2009], Langlois, Mubayi, Sloan, and Gy. Turán studied extremal properties of certain small configurations in a directed hypergraph model. This model can be thought of as a $2 \rightarrow 1$ directed hypergraph where each edge has three verticies, two of which are “tails" and the third is a “head." They determined the extremal number for one such subgraph with two edges, and found the extremal number of a second configuration with two edges up to asymptotic equivalence. In [@cameron2015] and [@cameron2015deg], this author followed up this work and found the exact extremal numbers for every $2 \rightarrow 1$ directed hypergraph with exactly two edges. This $2 \rightarrow 1$ model is one instance of the class of models discussed in this paper.
The graph theoretic properties of a more general definition of a directed hypergraph were studied by Gallo, Longo, Pallottino, and Nguyen in [@gallo1993]. There a directed hyperedge was defined to be some subset of vertices with a partition into head vertices and tail vertices. This is a nonuniform version of models considered in this paper.
The totally directed hypergraph model considered in [@brown1984] and the $r \rightarrow 1$ directed hypergraph model resulting from the study of Horn clauses both lead to the natural question of all possible ways to define a directed hypergraph. The definition in this paper of the class of general directed hypergraph models attempts to unify all of the possible “natural" ways one could define a directed hypergraph so that certain extremal questions can be answered about all of them at once. Adding to the motivation of considering more general structures is the recent interest in Razborov’s flag algebra method which applies to all relational theories and not just undirected hypergraphs. The fact that the $d$-simplex model studied by Leader as well as many other somewhat geometric models come out of the class defined in this paper was a very interesting accident.
This paper is organized as follows. In Section 2, we define the class of generalized directed hypergraphs and extend the concepts of Turán density, blowups, and supersaturation to this setting. In Section 3, we define the idea of a jump for a given model of directed hypergraphs and prove several results about these jumps and how the jumps from one instance of the class relate to jumps in another. In Section 4, we adapt a couple of results proved in [@brown1984] for totally directed hypergraphs with multiplicity to any GDH. In Section 5, we ask some questions that arose from studying these structures and discuss alternate definitions that would generalize the concept further.
Basic Definitions and Results
=============================
The following definition for a generalized directed hypergraph is intended to include most uniform models that could reasonably be called uniform directed hypergraphs. This includes models where the edges are $r$-sets each under some partition into $k$ parts of fixed sizes $r_1,\ldots,r_k$ with some linear ordering on the $k$ parts. The definition only includes structures where an $r$-set could include multiple edges up to the number of possible orientations allowed. That is, we do not consider the “oriented" versions of the models where only one edge is allowed per $r$-set. The definition is given in terms of logic and model theory for convenience only. No deep results from those subfields are used. The use of this notation also makes further generalizations like nonuniform directed hypergraphs or oriented directed hypergraphs easy.
\[GDHdef\] Let $\mathcal{L}=\{E\}$, a language with one $r$-ary relation symbol $E$. Let $T$ be an $\mathcal{L}$-theory that consists of a single sentence of the form $$\forall x_1 \cdots x_r E(x_1,\ldots,x_r) \implies \bigwedge_{i \neq j} x_i \neq x_j \land \bigwedge_{\pi \in J_T} E(x_{\pi(1)},\ldots,x_{\pi(r)})$$ for some subgroup of the group of permutations on $r$ elements, $J_T \subseteq S_r$. Call such a theory a *generalized directed hypergraph theory* and any finite model of $T$ is a *generalized directed hypergraph (GDH)*.
Note that this definition includes graphs, hypergraphs, and $r \rightarrow 1$ directed hypergraphs. For example, the theory for a $2 \rightarrow 1$ directed hypergraph is $$T = \{\forall xyz E(x,y,z) \implies x \neq y \land x \neq z \land y \neq z \land E(y,x,z)\}.$$ It is easy to see that when $r=2$ we have only two GDH theories. The theory associated with the group $S_2$ is the theory of graphs, and the theory associated with the trivial group is the theory of directed graphs.
When $r=3$ there are six subgroups of $S_3$. Three of these are all isomorphic to $\mathbb{Z}_2$ with each generated by a permutation that swaps two elements. The corresponding GDH theory for any of these can be thought of as having pointed $3$-sets for edges or as being ($2 \rightarrow 1$)-graphs. Of the other subgroups, $S_3$ itself gives the theory of undirected $3$-uniform hypergraphs, the trivial group gives totally directed $3$-edges, and the subgroup generated by a three-cycle isomorphic to $\mathbb{Z}_3$ yields a GDH theory where the edges can be thought of as $3$-sets that have some kind of cyclic orientation - either clockwise or counter-clockwise. Figure \[lattice\] summarizes the models of GDHs when $r=3$. Note that in general, $S_r$ always corresponds to the normal undirected $r$-graph model and the trivial group always corresponds to totally directed hypergraphs.
at (0,3) [$S_3$]{}; at (-1,2) [$\mathbb{Z}_3 $]{}; at (1,1) [$\mathbb{Z}_2$]{}; at (0,0) [$<i>$]{}; (-1,2) – (0,3); (1,1) – (0,3); (0,0) – (-1,2); (0,0) – (1,1);
at (2.5, 1.5) [$\implies$]{};
(5.5,3.25) circle (1pt); (6,3.25) circle (1pt); (6.5,3.25) circle (1pt); (5.5,3.25) – (6.5,3.25);
(4.25,2.25) circle (1pt); (4.75,1.75) circle (1pt); (3.75,1.75) circle (1pt); (4.25,2.25) – (4.75,1.75); (4.75,1.75) – (3.75,1.75); (3.75,1.75) – (4.25,2.25);
(7.25,1.25) circle (1pt); (7.25,0.75) circle (1pt); (7.75,1) circle (1pt); (7.25,1.25) – (7.25,0.75); (7.25,1) – (7.75,1);
(5.5,-0.25) circle (1pt); (6,-0.25) circle (1pt); (6.5,-0.25) circle (1pt); (5.5,-0.25) – (6.5,-0.25);
(6,0) – (5,2); (6,0) – (7,1); (7,1) – (6,3); (5,2) – (6,3);
A fun thought experiment is to consider the kinds of edges that arise when $r=4$. Many of them are geometric in nature. For instance, the alternating group $A_4$ gives a theory where edges can be thought of tetrahedrons (at least in an abstract sense). In fact, in [@leader2010] Leader and Tan study the “oriented" versions of the models that come from the alternating groups for any $r \geq 3$.
In this paper when the theory is not specified we are simply discussing GDHs that are all models of the same fixed theory. When discussing multiple theories we will often refer to $T$-graphs to mean models of a GDH theory $T$. Throughout the paper, $J_T$ will always stand for the subgroup $J_T \subseteq S_r$ that determines the GDH theory $T$ and $m_T$ will always be the order of this subgroup, $m_T = |J_T|$. Also, $V_G$ and $E_G$ will be used to denote the underlying set of elements of a model $G$ and its relation set respectively.
The following basic propositions are given without proof. The first is a simple consequence thatwe are working in a relational language, and the second results from the fact that $J_T$ is a group.
For any GDH theory $T$ and any nonnegative integer $n$, there exists a GDH $G \models T$ on $n$ elements. Moreover, for any nonnegative integer $k<n$, the substructure of $G$ induced on any $k$-subset of the elements of $G$ is also a $T$-graph.
Given a GDH $G$ with $r$-ary relation set $E_G$, there exists an equivalence relation $\sim$ on $E_G$ defined by $$(a_1,\ldots,a_r) \sim (b_1,\ldots,b_r)$$ if and only if for each $i$, $b_i = a_{\pi(i)}$ for some $\pi \in J_T$.
We can now use these propositions to extend the concepts of extremal graph theory to GDHs in a natural way.
For any GDH $G$, an *edge* of $G$ will always refer to an equivalence class of $[E_G]_{\sim}$.
Given a GDH $G$ on $n$ elements, denote the number of edges of $G$ by $e_T(G)$ and let the *edge density* of $G$ be defined as $$d_T(G) := \frac{e_T(G)}{\frac{r!}{m_T} {n \choose r}}.$$
Note that since $$e_T(G) = \frac{|E_G|}{m_T},$$ then the density is $$d_T(G) = \frac{(n-r)!|E_G|}{n!}$$ and could have been defined this way while mostly avoiding talk of edges as equivalence classes of $E_G$. However, the above definition makes the following extremal concepts reduce to their standard definitions in the undirected case.
Given two GDHs $G$ and $H$ and a function $\psi: V_H \rightarrow V_G$, we say that $\psi$ is a *homomorphism* if for all $(a_1,\ldots,a_r) \in E_H$, $(\psi(a_1),\ldots,\psi(a_r)) \in E_G$.
We say that $G$ contains a copy of $H$ if there exists some injective homomorphism, $\psi:V_H \rightarrow V_G$. Otherwise, we say that $G$ is $H$-free. Similarly, we would say that a GDH $G$ is $\mathcal{F}$-free for some family $\mathcal{F}$ of GDHs if $G$ is $F$-free for all $F \in \mathcal{F}$.
Given a family of GDHs $\mathcal{F}$ and a positive integer $n$, let the $n$th *extremal number*, $\text{ex}_T(n,\mathcal{F})$, be defined as the maximum number of edges over all $\mathcal{F}$-free GDHs on $n$ elements, $$\text{ex}_T(n,\mathcal{F}) := \max_{\mathcal{F}\text{-free } G_n} \{e_T(G_n) \}.$$ The *Turán density* of $\mathcal{F}$ is defined as $$\pi_T(\mathcal{F}) := \lim_{n \rightarrow \infty} \frac{\text{ex}_T(n,\mathcal{F})}{\frac{r!}{m_T} {n \choose r}}.$$
Our first main result is to show that these Turán densities exist for any GDH theory. The proof is the standard averaging argument used to show that these limiting densities exist for families of undirected hypergraphs [@keevash2011].
For any GDH family $\mathcal{F}$ the Turán density exists.
Let $G$ be an $\mathcal{F}$-free GDH on $n$ elements with $\text{ex}_T(n,\mathcal{F})$ edges. For each $i=1,\ldots,n$ let $G^i$ be the subGDH of $G$ induced by removing the $i$th vertex. Each edge of $G$ appears in exactly $n-r$ of these subGDHs. Therefore, $$(n-r)e_T(G) = e_T(G^1)+ \cdots e_T(G^n).$$ Moreover, $e_T(G) = \text{ex}_T(n,\mathcal{F})$ and each $G^i$ is also $\mathcal{F}$-free so $e_T(G^i) \leq \text{ex}_T(n-1,\mathcal{F})$. Therefore, $$\text{ex}_T(n,G) \leq \frac{n}{n-r} \text{ex}_T(n-1,\mathcal{F}).$$ So $$\frac{\text{ex}_T(n,G)}{\frac{r!}{m_T} {n \choose r}} \leq \frac{n}{n-r} \frac{\text{ex}_T(n-1,\mathcal{F})}{\frac{r!}{m_T} {n \choose r}} = \frac{\text{ex}_T(n-1,\mathcal{F})}{\frac{r!}{m_T} {n-1 \choose r}}.$$ Therefore, the sequence of these extremal densities is monotone decreasing as a function of $n$ in the range $[0,1]$. Hence, the limit exists.
Blowups and Blowup Density
--------------------------
We’ll now extend the concept of the blowup of uniform hypergraphs to the more general setting of GDHs and define the corresponding notion of the blowup density. As with hypergraphs, the blowup of a GDH can be thought of as the replacement of each vertex with many copies and taking all of the resulting edges. Formally,
Let $G$ be a GDH with $V_G = \{x_1,\ldots,x_n\}$, and let $t = (t_1,\ldots, t_n)$ be a tuple of positive integers. Define the *$t$-blowup* of $G$ to be the $\mathcal{L}$-structure $G(t)$ where $$V_{G(t)} = \{x_{11},\ldots,x_{1t_1},\ldots,x_{n1},\ldots,x_{nt_n}\}$$ and $$(x_{i_1j_1},\ldots,x_{i_rj_r}) \in E_{G(t)} \iff (x_{i_1},\ldots,x_{i_r}) \in E_G.$$
Let $G$ be a GDH on $n$ vertices, and let $t=(t_1,\ldots,t_n)$ be a tuple of positive integers. Then the $t$-blowup of $G$ is also a GDH.
We need only show that the $\mathcal{L}$-structure $G(t)$ models $T$. So let $$(x_{i_1j_1},\ldots,x_{i_rj_r}) \in E_{G(t)}.$$ Then $(x_{i_1},\ldots,x_{i_r}) \in E_G$. Since $G \models T$ this implies that $i_a \neq i_b$ whenever $a \neq b$. Hence, the elements $x_{i_aj_a} \neq x_{i_bj_b}$ whenever $a \neq b$. It also implies that $(x_{i_{\pi(1)}},\ldots,x_{i_{\pi(r)}}) \in E_G$ for any $\pi \in J_T$. Hence, $$(x_{i_{\pi(1)}j_{\pi(1)}},\ldots,x_{i_{\pi(r)}j_{\pi(r)}}) \in E_{G(t)}$$ for any $\pi \in J_T$. Therefore, $G(t) \models T$.
Next, we consider the edge density of a given blowup by defining the edge polynomial for a GDH.
Let $G$ be a GDH on $n$ vertices. For each $r$-set $R \in {V_G \choose r}$, let $e_R$ be the number of edges of $G$ in $R$. Then let the *edge polynomial* be $$p_G(x) := \sum_{R \in {V_G \choose r}} e_R\prod_{i \in R} x_i.$$
This polynomial is a simple generalization of the standard edge polynomial for undirected hypergraphs. To see this more easily note that for a given GDH $G$, the edges of $G$ are in bijection with the monomials the sum $p_G$ were we to write the sum out with no coefficients greater than one.
From this we see that the edge density of the $(t_1,\ldots,t_n)$-blowup of $G$ is $$\frac{p_G(t_1,\ldots,t_n)}{\frac{r!}{m_T}{t \choose r}} = m_T\frac{p_G(t_1,\ldots,t_n)}{t(t-1)\cdots(t-r+1)}$$ where $t= \sum t_i$. Let $t$ increase to infinity and for each $t$ pick a vector $(t_1,\ldots,t_n)$ that maximizes this edge density. Then this sequence of densities is asymptotically equivalent to the sequence of numbers $$m_T p_G\left(\frac{t_1}{t},\ldots,\frac{t_n}{t}\right).$$ This motivates the following definition.
Let $G$ be a GDH on $n$ vertices. Let $$S^n = \left\{(x_1,\ldots,x_n)|x_i \geq 0 \land \sum_{i=1}^nx_i = 1\right\},$$ the standard $(n-1)$-dimensional simplex. Define the *blowup density* of $G$ as $$b_T(G) = m_T \max_{x \in S^n} \{p_G(x)\}.$$
Since any $x \in S^n$ is the limit of some sequence $\left\{\left(\frac{t_1}{t},\ldots,\frac{t_n}{t}\right)\right\}$ with positive $t_i$ as $t \rightarrow \infty$, then the blowup density of a GDH $G$ is the best limiting density of any sequence of blowups of $G$.
The remaining definition and basic result about blowups given in this subsection will be useful when extending results about jumps and nonjumps from undirected hypergraphs to GDHs generally in Section 3.
Let $T'$ and $T$ be GDH theories such that $J_{T'} \subseteq J_{T} \subseteq S_r$. For a $T$-graph $F$ and a $T'$-graph $F'$ we say that $F$ *contains* $F'$ if $V_F = V_{F'}$ and every edge of $F'$ is contained in some edge of $F$ (where the edges are considered under their equivalence class definition as subsets of $E_F$ and $E_{F'}$). We say that $F$ is the *minimum $T$-container* of $F'$ if $F$ has no edges that do not contain edges of $F'$.
\[containing\] Let $T'$ and $T$ be GDH theories such that $J_{T'} \subseteq J_{T} \subseteq S_r$. Let $F'$ be a $T'$-graph and let $F$ be the minimum $T$-container of $F'$. Then $$\frac{m_{T'}}{m_T}b_T(F) \leq b_{T'}(F') \leq b_T(F)$$ with equality on the left if $F'$ has exactly one edge contained in each edge of $F$ and equality on the right if each edge of $F$ contains all $\frac{m_T}{m_{T'}}$ possible edges of $F'$.
Moreover, if $F'$ has exactly $k$ edges contained in each edge of $F$, then $$b_{T'}(F') =\frac{km_{T'}}{m_T}b_T(F).$$
Let $|V_{F'}| = |V_{F}|=v$, then for any $x \in S^v$, $$p_F(x) \leq p_{F'}(x) \leq \frac{m_T}{m_{T'}}p_F(x)$$ with equality on the left if $F'$ has exactly one edge contained in each edge of $F$ and equality on the right if each edge of $F$ contains all $\frac{m_T}{m_{T'}}$ possible edges of $F'$. Hence, $$\max_{x \in S^v}{p_F(x)} \leq \max_{x \in S^v}{p_{F'}(x)} \leq \max_{x \in S^v}{\frac{m_T}{m_{T'}}p_F(x)}.$$ This implies that $$\frac{m_{T'}}{m_T}b_T(F) \leq b_{T'}(F') \leq b_T(F).$$ In particular, if $F'$ has exactly $k$ edges contained in each edge of $F$, then for any $x \in S^v$, $$p_{F'}(x) = kp_F(x)$$ which implies the result.
Supersaturation and Related Results
-----------------------------------
Supersaturation holds for GDHs as it does for undirected hypergraphs, and the proof of this result is the same as the one for hypergraphs found in [@keevash2011] with only minor differences.
\[supersaturation\] Let $F$ be a GDH on $k$ elements. Let $\epsilon > 0$. For sufficiently large $n \geq n_0(F,\epsilon)$, any GDH $G$ on $n$ elements with density $d(G) \geq \pi_T(F) + \epsilon$ will contain at least $c {n \choose k}$ copies of $F$ for some constant $c = c(F,\epsilon)$.
Fix some positive integer $l$ so that $$\text{ex}_T(l,F) < \left(\pi_T(F) + \frac{\epsilon}{2}\right) \frac{r!}{m_T} {l \choose r}.$$ Let $G$ be a GDH on $n > l$ elements with edge density $d_T(G) \geq \pi(F) + \epsilon$. Then $G$ must contain more than $\frac{\epsilon}{2}{n \choose l}$ $l$-sets with density at least $\pi_T(F) + \frac{\epsilon}{2}$. Otherwise, at most $\frac{\epsilon}{2}{n \choose l}$ $l$-sets contain more than $\left(\pi_T(F) + \frac{\epsilon}{2}\right){l \choose r}$ edges. Therefore, we can count the number of edges in $G$ by $l$-sets and get an upper bound of $${n-r \choose l-r}e_T(G) \leq \frac{\epsilon}{2} {n \choose l} {l \choose r} \frac{r!}{m_T} + \left(1-\frac{\epsilon}{2}\right){n \choose l}\left(\pi_T(F)+\frac{\epsilon}{2}\right){l \choose r}\frac{r!}{m_T}.$$ We can now replace $e_T(G)$ since $$e_T(G) \geq \left(\pi_T(F)+\epsilon\right){n \choose r} \frac{r!}{m_T}.$$ This is enough to get the contradiction.
Since $G$ contains more than $\frac{\epsilon}{2}{n \choose l}$ $l$-sets with density at least $\pi_T(F) + \frac{\epsilon}{2}$, then it contains a copy of $F$ in each. A given copy of $F$ appears in ${n-k \choose l-k}$ $l$-sets of $G$. Therefore, there are more than $$\frac{\epsilon}{2}{n \choose l}{n-k \choose l-k}^{-1} = c {n \choose k}$$ distinct copies of $F$ in $G$ where $$c = \frac{\epsilon}{2}{l \choose k}^{-1}.$$
Similarly, the following theorem is an extension from the same result for undirected hypergraphs, and the proof is an adaptation of the one found in [@keevash2011].
\[blowup\] Let $F$ be a GDH on $k$ vertices and let $t=(t_1,\ldots,t_k)$ be an $k$-tuple of positive integers. Then $\pi_T(F) = \pi_T(F(t))$.
That $\pi_T(F) \leq \pi_T(F(t))$ is trivial since $F(t)$ contains a copy of $F$ so any $F$-free GDH is automatically $F(t)$-free.
Therefore, we only need to show that $\pi_T(F) \geq \pi_T(F(t))$. Suppose not, then for sufficiently large $n$ there exists some $F(t)$-free GDH $G$ on $n$ elements with edge density strictly greater than $\pi_T(F)$. By supersaturation this implies that $G$ contains $c{n \choose k}$ copies of $F$.
Define $G^*$ to be the $k$-uniform hypergraph where $V_{G^*} = V_{G}$ and $\{a_1,\ldots,a_k\} \in E_{G^*}$ iff and only if $\{a_1,\ldots,a_k\}$ contains a copy of $F$ in $G$. Since the edge density of $G^*$ is $c>0$, then for large enough $n$, $G^*$ must contain an arbitrarily large complete $k$-partite subgraph.
For each edge $F$ maps to the vertices in at least one out of $k!$ total possible ways to make an injective homomorphism in $G$. Therefore, by Ramsey Theory, if we take the parts of this complete $k$-partite subgraph large enough and color the edges by the finite number of non-isomorphic ways that $F$ could possibly map to the $k$ vertices, we will get an arbitrarily large monochromatic $k$-partite subgraph where each part has $t$ vertices. This must have been a copy of $F(t)$ in $G$, a contradiction.
The fact that the Turán density of a blowup equals the Turán density of the original GDH leads to the following nice characterization of degenerate families of GDH - those families with Turán density zero.
\[degenerate\] Let $\mathcal{F}$ be some family of GDHs, then $\pi_T(\mathcal{F}) = 0$ if and only if some member $F \in \mathcal{F}$ is a subGDH of the $t$-blowup of a single edge for some vector, $t=(t_1,\ldots,t_r)$, of positive integers. Otherwise, $\pi(\mathcal{F}) \geq \frac{m_T}{r^r}$.
Suppose that no member of $\mathcal{F}$ is such a blowup. Then no member is contained in the $(t,t,\ldots,t)$-blowup of $S$. Let $S(t)$ stand for this blowup, then the sequence of GDHs, $\{S(t)\}_{t=1}^{\infty}$, is an $\mathcal{F}$-free sequence. The density of any such $S(t)$ is $$d_T(S(t)) = \frac{t^r}{\frac{r!}{m_T} {tr \choose r}} = \frac{m_T t^r (tr-r)!}{(tr)!}.$$ These densities tend to $\frac{m_T}{r^r}$ as $t$ increases. Therefore, $$\pi_T(\mathcal{F}) \geq \frac{m_T}{r^r} > 0.$$
Conversely, suppose some $F \in \mathcal{F}$ is a $(t_1,\ldots,t_r)$-blowup of a single edge. By Theorem \[blowup\], $\pi_T(F) = \pi_T(S) = 0$ since $\text{ex}_T(n,S)$=0 for all $n$. Therefore, $\pi_T(\mathcal{F})=0$.
Jumps
=====
Now we turn to the issue of finding jumps and nonjumps for GDH theories. The definition of a jump for undirected hypergraphs extends naturally to this setting as does the important connection between jumps and blowup densities.
Let $T$ be a GDH theory, then $\alpha \in [0,1)$ is a *jump* for $T$ if there exists a $c >0$ such that for any $\epsilon > 0$ and any positive integer $l$, there exists a positive integer $n_0(\alpha, \epsilon, l)$ such that any GDH $G$ on $n \geq n_0$ elements that has at least $(\alpha + \epsilon) \frac{r!}{m_T} {n \choose r}$ edges contains a subGDH on $l$ elements with at least $(\alpha + c) \frac{r!}{m_T} {l \choose r}$ edges.
Note that by Theorem \[degenerate\] every $\alpha \in \left[0,\frac{m_T}{r^r}\right)$ is a jump for any $r$-ary GDH theory $T$. This generalizes the well-known result of Erdős [@erdos1964] that every $\alpha \in [0,\frac{r!}{r^r})$ is a jump for $r$-graphs. The following important theorem on jumps for GDH theories was originally shown by Frankl and Rödl [@frankl1984] for undirected hypergraphs. Their proof works equally well in this setting so the differences here are in name only.
\[mainguy\] The GDH theory $T$ has a jump $\alpha$ if and only if there exists a finite family $\mathcal{F}$ of GDHs such that $\pi_T(\mathcal{F}) \leq \alpha$ and $b_T(F) > \alpha$ for each $F \in \mathcal{F}$.
Let $\alpha$ be a jump and let $c$ be the supremum of all corresponding “lengths" $c$ to the jump. Fix a positive integer $k$ so that $${k \choose r} \left(\alpha + \frac{c}{2}\right) > \alpha \frac{k^r}{r!}.$$ Let $\mathcal{F}$ be the family of all GDHs on $k$ elements with at least $\left(\alpha + \frac{c}{2}\right) {k \choose r} \frac{r!}{m_T}$ edges. Then $\pi_T(\mathcal{F}) \leq \alpha$ since any slightly larger density implies arbitrarily large subsets with density $\alpha + c$. This in turn would imply the existence of a $k$-subset with density at least $\alpha + c$. This $k$-subset would include some member of $\mathcal{F}$. On the other hand, a given $F \in \mathcal{F}$ will have blowup density $$b_T(F) \geq m_T p_F\left(\frac{1}{k},\ldots,\frac{1}{k}\right)>\alpha.$$
Conversely, suppose that such a finite family $\mathcal{F}=\{F_1,\ldots,F_k\}$ exists. Let $\epsilon >0$ and let $\{G_n\}$ be an infinite sequence of GDHs with density that tends to $\alpha + \epsilon$. As in the proof of Theorem \[supersaturation\], for any positive integer $l$, $G_n$ must contain at least $\frac{\epsilon}{2}{n \choose l}$ $l$-subsets with density at least $\alpha + \frac{\epsilon}{2}$.
Let $l$ be large enough so that any GDH on $l$ vertices with density at least $\alpha + \frac{\epsilon}{2}$ contains some $F_i$ from $\mathcal{F}$. Therefore, any $G_n$ with $n > l$ contains $\frac{\epsilon}{2}{n \choose l}$ $l$-sets each with some $F_i$. Since there are only $k$ members of $\mathcal{F}$, then this implies that at least $\frac{\epsilon}{2k}{n \choose l}$ $l$-sets contain the same $F_i$.
Let $|V(F_i)|=v_i$. By the proof of Theorem \[supersaturation\] this implies that there is some positive constant $b$ such that $G_n$ contains at least $b{n \choose v_i}$ distinct copies of $F_i$. By the proof of Theorem \[blowup\] this shows that if $n$ is large enough, then we get a copy of an arbitrarily large $t$-blowup of $F_i$.
Let $c = \min_{F_i \in \mathcal{F}}{b_T(F_i)}$. For some subset $\mathcal{F}' \subseteq \mathcal{F}$, each $F_i \in \mathcal{F}'$ yields an infinite subsequence of $\{G_n\}$ which contains arbitrarily large $t$-blowups of $F_i$. The densities of these blowups all tend to at least $c$. Therefore, for any positive integer $m$, there exists an $m$-set of each $\{G_n\}$ for sufficiently large $n$ with density at least $\alpha + c$. Hence, $\alpha$ is a jump.
The following proposition is needed to compare jumps between different GDH theories.
\[thmB\] The GDH theory $T$ has a jump $\alpha$ if and only if there exists some $c > 0$ such that for all families $\mathcal{F}$ of GDHs, either $\pi_T(\mathcal{F}) \leq \alpha$ or $\pi_T(\mathcal{F}) \geq \alpha + c$.
Let $\alpha$ be a jump for $T$ and let $c>0$ be some corresponding “length" to the jump. Suppose that $\mathcal{F}$ is a finite family of GDHs of type $T$ for which $\alpha < \pi_T(\mathcal{F}) < \alpha + c$. Let $\{G_n\}$ be a sequence of extremal $\mathcal{F}$-free GDHs. For each positive integer $k$ there exists some $G_n$ that contains a $k$-subset with at least $(\alpha + c){k \choose r} \frac{r!}{m_T}$ edges. Take the sequence of these subsets. They are all $\mathcal{F}$-free by assumption, and the limit of their densities is at least $\alpha + c$. Therefore, $\pi_T(\mathcal{F}) \geq \alpha + c$, a contradiction.
Conversely, assume that $\alpha$ is not a jump. Let $c>0$, then for some $0<\epsilon<c$ and some positive integer $l$, there exists an infinite sequence of GDHs, $\{G_n\}$ for which each GDH has density at least $\alpha + \epsilon$ and all $l$-sets have strictly less than $(\epsilon + c){l \choose r}\frac{r!}{m_T}$ edges. Hence, $\{G_n\}$ is $\mathcal{F}$-free where $\mathcal{F}$ is the set of all $l$-GHDs with at least $(\alpha + c){l \choose r} \frac{r!}{m_T}$ edges. So $\pi_T(\mathcal{F}) \geq \alpha + \epsilon$. Since any GDH with density at least $\alpha + c$ must have an $l$-set with density at least $\alpha + c$, then $\pi_T(\mathcal{F}) < \alpha + c$.
We will now look at how jumps are related between two different GDH theories for some fixed edge size $r$. We will see that in general jumps always “pass up" the subgroup lattice. That is, if $J_{T'} \subseteq J_T$ for GDH theories $T'$ and $T$, then a jump for $T'$ is a jump for $T$. The converse is not true in general. In fact, for any GDH theories $T'$ and $T$ with $J_{T'} \subseteq J_T$ such that the order of $J_T$ is at least three times that of $J_{T'}$ we will show that the set of jumps for $T'$ is not equal to the set of jumps for $T$. The case where $m_T = 2m_{T'}$ is open.
Jumps pass up the lattice
-------------------------
First, we will show that for GDH theories $T$ and $T'$ with $J_{T'} \subseteq J_T$ the set of Turán densities of forbidden families of $T$-graphs is a subset of the set of Turán densities for $T'$.
\[thmC\] Let $T$ and $T'$ be two GDH theories such that $J_{T'} \subseteq J_{T}$. Then for any family $\mathcal{F}$ of $T$-graphs there exists a family $\mathcal{F}'$ of $T'$-graphs for which $\pi_{T'}(\mathcal{F}') = \pi_{T}(\mathcal{F})$. Moreover, if $\mathcal{F}$ is a finite family, then $\mathcal{F}'$ is also finite.
For each $F \in \mathcal{F}$ let $F_{T'}$ be the set of all $T'$-graphs that have exactly one edge contained in every edge of $F$. That is, since $J_{T'} \subseteq J_{T}$, then there are $\frac{m_T}{m_{T'}}$ possible $T'$ edges contained within one $T$ edge. So $F_{T'}$ is a finite set with at most $\left(\frac{m_T}{m_{T'}}\right)^{e_T(F)}$ members. Let $$\mathcal{F}' = \bigcup_{F \in \mathcal{F}} F_{T'}.$$ Then $\mathcal{F}'$ is a family of $T'$-graphs. Moreover, $\mathcal{F}'$ is finite if $\mathcal{F}$ is finite. We want to show that $\pi_{T'}(\mathcal{F}') = \pi_{T}(\mathcal{F})$.
First, let $\{G_n'\}$ be an extremal $\mathcal{F}'$-free sequence of $T'$-graphs. For each $G_n'$ let $G_n$ be the $T$-graph constructed by replacing each $T'$-edge of $G_n'$ with its containing $T$-edge (multiple $T'$-edges could correspond to the same $T$-edge but each $T$-edge can only be added once).
The sequence $\{G_n\}$ is $\mathcal{F}$-free since otherwise some $G_n$ contains some $F \in \mathcal{F}$ which means that $G_n'$ must have contained at least one member of $F_{T'}$. Therefore, $$\pi_{T}(\mathcal{F}) \geq \lim_{n \rightarrow \infty} d_T(G_n) \geq \lim_{n \rightarrow \infty} \frac{\frac{m_{T'}}{m_T} e_{T'}(G_n')}{\frac{r!}{m_T}{n \choose r}} = \lim_{n \rightarrow \infty} \frac{ex_{T'}(n,\mathcal{F}')}{\frac{r!}{m_{T'}}{n \choose r}} = \pi_{T'}(\mathcal{F}').$$
Conversely, now let $\{G_n\}$ be an extremal $\mathcal{F}$-free sequence of $T$-graphs. For each $G_n$ construct a $T'$-graph $G_n'$ by replacing each $T$-edge with all $\frac{m_T}{m_{T'}}$ $T'$-edges contained in it. The sequence $\{G_n'\}$ is $\mathcal{F}'$-free with $\frac{m_T}{m_{T'}}\text{ex}_T(n,\mathcal{F})$ edges. Therefore, $$\pi_{T'}(\mathcal{F}') \geq \lim_{n \rightarrow \infty} \frac{\frac{m_T}{m_{T'}}\text{ex}_T(n,\mathcal{F})}{\frac{r!}{m_{T'}} {n \choose r}} = \pi_T(\mathcal{F}).$$ So $\pi_{T'}(\mathcal{F}') = \pi_{T}(\mathcal{F})$.
The converse of Theorem \[thmC\] is false in general. For example, the permutation subgroup for the theory $T'$ of ($2 \rightarrow 1$)-uniform directed hypergraphs is a subgroup of the permutation group for the theory $T$ of undirected $3$-graphs, $S_3$. The extremal number for the directed hypergraph is $F = \{ab \rightarrow c, cd \rightarrow e\}$ (see Figure \[F\]) is $$ex_{T'}(n,F) = \left\lfloor \frac{n}{3} \right\rfloor {\left\lceil \frac{2n}{3} \right\rceil \choose 2}$$ as shown in [@langlois2010]. Therefore, the Turán density is $\pi_{T'}(F) = \frac{4}{27}$. However, it is well-known that no Turán densities exist for $3$-graphs in the interval $\left(0,\frac{6}{27}\right)$.
(-2,2) circle (1pt); (-2,0) circle (1pt); (-2,2) – (-2,0); (0,1) circle (1pt); (-2,1) – (0,1); (0,-1) circle (1pt); (0,1) – (0,-1); (0,0) – (2,0); (2,0) circle (1pt);
\[jumpsgoup\] Let $T$ and $T'$ be two GDH theories such that $J_{T'} \subseteq J_{T}$. If $\alpha$ is a jump for $T'$, then it is also a jump for $T$.
If $\alpha$ is not a jump for $T$, then for any $c > 0$ there exists by Proposition \[thmB\] a family $\mathcal{F}$ such that $\alpha < \pi_T(\mathcal{F}) < \alpha + c$. So by Theorem \[thmC\] there exists a family $\mathcal{F}'$ of $T'$-graphs with $\alpha < \pi_{T'}(\mathcal{F}') < \alpha + c$. So $\alpha$ is not a jump for $T'$.
Corollary \[jumpsgoup\] immediately implies that all nonjumps found for $r$-uniform undirected hypergraphs must also be non-jumps for any GDH with an $r$-ary relation. However, the converse is not true in general.
Jumps do not pass down the lattice
----------------------------------
Roughly speaking, the current best method of demonstrating that a particular $\alpha$ is not a jump for $r$-uniform hypergraphs is to construct a sequence of hypergraphs each with blowup densities that are strictly larger than $\alpha$ but for which any relatively small subgraph has blowup density at most $\alpha$. This method originated in [@frankl1984] and generalizes to GDHs as the following definition and lemma demonstrate.
Let $\alpha \in [0,1)$. Call $\alpha$ a *demonstrated nonjump* for a GDH theory $T$ if there exists an infinite sequence of GDHs, $\{G_n\}$, such that $b_T(G_n) > \alpha$ for each $G_n$ in the sequence and for any positive integer $l$ there exists a positive integer $n_0$ such that whenever $n \geq n_0$ then any subGDH $H \subseteq G_n$ on $l$ or fewer vertices has blowup density $b_T(H) \leq \alpha$.
Every demonstrated nonjump is a nonjump.
Suppose not. Assume that $\alpha$ is a demonstrated nonjump but is a jump. Then there exists a finite family of GDHs $\mathcal{F}$ such that $\pi_T(\mathcal{F}) \leq \alpha$ and $b_T(F) > \alpha$ for each $F \in \mathcal{F}$. Let $l$ be the maximum number of vertices over the members of $\mathcal{F}$. Let $n$ be large enough so that any subGDH on $l$ or fewer vertices has blowup density at most $\alpha$. Then some large enough blowup of $G_n$ contains some $F \in \mathcal{F}$ as a subGDH since the blowup density of each $G_n$ tends to something strictly greater than $\alpha$. Let $H$ be the minimal subGDH of $G_n$ for which the corresponding blowup contains this copy of $F$. Since $H$ has at most $l$ vertices, then it has a blowup density at most $\alpha$. Hence, $$b_T(F) \leq b_T(H(t)) \leq b_T(H) \leq \alpha,$$ a contradiction.
We can now show that a demonstrated nonjump for a GDH theory $T$ yields multiple nonjumps of equal and lesser values down the lattice to GDH theories $T'$ for which $J_{T'} \subseteq J_T$.
Let $T$ and $T'$ be GDH theories such that $J_{T'} \subseteq J_{T}$. Let $\alpha$ be a demonstrated nonjump for $T$. Then $\frac{km_{T'}}{m_T} \alpha$ is a demonstrated nonjump for $T'$ for $k = 1, \ldots, \frac{m_T}{m_{T'}}$.
Let $\alpha$ be a demonstrated nonjump for $T$. Let $\{G_n\}$ be the corresponding infinite sequence of GDHs. Fix some $k \in \{1,\ldots,\frac{m_T}{m_{T'}}\}$. For each $n$ let $G_n'$ be a $T'$-graph constructed from $G_n$ by replacing each $T$-edge with $k$ $T'$-edges in any orientation. Then by Proposition \[containing\] we know that $$b_{T'} (G_n') = \frac{km_{T'}}{m_T}b_T(G_n)$$ and any $H' \subseteq G_n'$ corresponding to $H \subseteq G_n$ also gives: $$b_{T'} (H') = \frac{km_{T'}}{m_T}b_T(H).$$ Therefore, $b_{T'}(G_n') > \frac{km_{T'}}{m_T} \alpha$ for each $n$ and for any positive integer $l$, there exists a $n_0$ such that $b_{T'}(H) \leq \alpha$ for any subGDH $H \subseteq G_n$ for all $n \geq n_0$.
Constructions of sequences of undirected $r$-graphs which show that $\frac{5r!}{2r^r}$ is a demonstrated nonjump for each $r \geq 3$ were given in [@frankl2007]. This gives the following corollary.
Let $T$ be an $r$-ary GDH theory for $r \geq 3$. Then $\frac{5m_Tk}{2r^r}$ is a nonjump for $T$ for $k=1,\ldots,\frac{r!}{m_T}$.
This in turn shows that the set of jumps for a theory $T'$ is a proper subset of the set of jumps for $T$ for any $T$ such that $J_{T'} \subseteq J_T$ and $m_T \geq 3m_{T'}$.
Let $T$ and $T'$ be $r$-ary GDH theories such that $J_{T'} \subseteq J_{T}$ and $m_T \geq 3m_{T'}$. Then there exists an $\alpha$ that is a nonjump for $T'$ and a jump for $T$.
Take $k=1$, then $\frac{5m_{T'}}{2r^r}$ is a nonjump for $T'$. Since $m_T \geq 3m_{T'}$, then $m_T > 2.5 m_{T'}$. So $$\frac{5m_{T'}}{2r^r} < \frac{m_T}{r^r}.$$ Therefore, $\frac{5m_{T'}}{2r^r}$ is a jump for $T$ since every $\alpha \in \left[0,\frac{m_T}{r^r}\right)$ is a jump for $T$.
Continuity and Approximation
============================
The following two results are direct adaptations of two theorems from [@brown1984]. They are both general extremal results related to everything discussed in this paper but did not fit nicely into the other sections. The first result, Continuity, relates extremal numbers of any infinite family of GDHs to the extremal numbers of its finite subfamilies. The second, Approximation, discusses structural aspects of (nearly) extremal sequences for any forbidden family.
Let $\mathcal{F}$ be an infinite family of $T$-graphs. For each $\epsilon > 0$ there exists a finite subfamily $\mathcal{F}_{\epsilon} \subset \mathcal{F}$ such that $$\text{ex}_T\left(n,\mathcal{F}\right) \leq \text{ex}_T\left(n,\mathcal{F}_{\epsilon}\right) < \text{ex}_T\left(n,\mathcal{F}\right) + \epsilon n^r$$ for sufficiently large $n$.
Let $\mathcal{F}$ be the infinite family of GDHs. For each positive integer $k$ let $\mathcal{F}_k$ be the subfamily of $\mathcal{F}$ where each member has at most $k$ vertices. Let $$\gamma_k = \lim_{n \rightarrow \infty} \frac{\text{ex}_T\left(n, \mathcal{F}_k\right)}{\frac{r!}{m_T} {n \choose r}}$$ and let $$\gamma = \lim_{n \rightarrow \infty} \frac{\text{ex}_T\left(n, \mathcal{F}\right)}{\frac{r!}{m_T} {n \choose r}}.$$ Since $\mathcal{F}_k \subset \mathcal{F}$, then $\{\gamma_k\}_{k=1}^{\infty}$ is a monotone decreasing sequence and $\gamma_k \geq \gamma$ for all $k$.
Assume for some $\epsilon > 0$ that $\gamma_k > \gamma + \epsilon$ for all $k$. Note that $$\frac{\text{ex}_T\left(n,\mathcal{F}_k\right)}{\frac{r!}{m_T} {n \choose r}} \geq \gamma_k$$ is true for all $n$. In particular, when $n=k$ there is an $\mathcal{F}_n$-free GDH on $n$ vertices with strictly more than $\left( \gamma + \epsilon \right) \frac{r!}{m_T} {n \choose r}$ edges. Since an $\mathcal{F}_n$-free GDH on $n$ vertices is also necessarily $\mathcal{F}$-free, then this implies that $$\text{ex}_T\left(n, \mathcal{F} \right) > \left(\gamma + \epsilon\right) \frac{r!}{m_T} {n \choose r},$$ a contradiction.
Theorem 6 in [@brown1984] is the Approximation Theorem for totally directed $r$-uniform hypergraphs with bounded multiplicity. We will use the following equivalent statement (in the case of multiplicity one) written in terms of Turán densities as a lemma to prove that this approximation result holds for all GDHs.
\[approx\] Let $\mathcal{F}'$ be a family of forbidden totally directed $r$-graphs ($r$-GDHs under the trivial group), and let $\epsilon>0$. Then there exists some totally directed $r$-graph $G'$ such that every blowup of $G'$ is $\mathcal{F}'$-free and $$\pi(\mathcal{F}') \geq b(G) > \pi(\mathcal{F}') - \epsilon.$$
Let $\mathcal{F}$ be a family of forbidden $T$-graphs, and let $\epsilon > 0$, then there exists some $T$-graph $G$ for which all blowups of $G$ are $\mathcal{F}$-free and $$\pi_T(\mathcal{F}) \geq b_T(G) > \pi_T(\mathcal{F}) - \epsilon.$$
Let $\mathcal{F}'$ be the family of totally directed $r$-graphs as defined in the proof of Theorem \[thmC\]. That is, the family of directed hypergraphs for which we know that $\pi(\mathcal{F}') = \pi_T(\mathcal{F})$. We know from the proof of that theorem that any $T$-graph that is the minimal container for an $\mathcal{F}'$-free graph is $\mathcal{F}$-free. By Lemma \[approx\] there exists some totally directed $\mathcal{F}'$-free $r$-graph , $G'$, such that $$\pi(\mathcal{F}') \geq b(G) > \pi(\mathcal{F}') - \epsilon.$$ By Proposition \[containing\] we know that if $G'$ is the minimal containing $T$-graph of $G$, then $b_T(G) \geq b(G')$. Hence, $$\pi_T(\mathcal{F}) \geq b_T(G) \geq b(G') > \pi(\mathcal{F}') - \epsilon = \pi_T(\mathcal{F}) - \epsilon.$$
Conclusion
==========
Some questions naturally come up in studying GDHs. Most notably it would be nice to show that the set of jumps for some GDH theory $T'$ is a proper subset of the set of jumps of any theory $T'$ up the lattice including those for which $m_T = 2m_{T'}$. Or on the other hand it would be very interesting to learn that this is not true in certain cases for $r \geq 3$!
Let $T'$ and $T$ be $r$-ary GDH theories for $r \geq 3$ such that $J_{T'} \subseteq J_T$ and $m_T = 2m_{T'}$. Then there exists some $\alpha \in [0,1)$ for which $\alpha$ is a jump for $T$ but not for $T'$.
It is known by a result in [@brown1984] that every $\alpha \in [0,1)$ is a jump for digraphs. Therefore, the conjecture is not true when $r=2$. On a related note, is it always true that when $J_{T'} \subset J_T$, there always exists a family $\mathcal{F}'$ of $T'$-graphs such that $\pi_{T'}(\mathcal{F}')$ is not contained in the set of Turán densities for $T$?
Let $T'$ and $T$ be theories such that $J_{T'} \subseteq J_T$. Then there exists some family $\mathcal{F}'$ of $T'$-graphs such that $\pi_{T'}(\mathcal{F}')$ is not contained in the set of Turán densities for $T$.
Finally, it would be nice to generalize the definition of a GDH to include other combinatorial structures. For instance we could easily change the current formulation to include multiple relations in order to capture nonuniform GDHs and those with edges that have bounded multiplicity like the structures studied in [@brown1984]. We could even allow these theories to contain general statements that relate the different relations. An example of this might be the theory of some kind of GDH with an edge-coloring that behaves in a certain way (at least locally). In another direction we could take away the requirement that all vertices of an edge be distinct to allow for kinds of generalized loops or add a condition that the existence of certain edges preclude the existence of others such as in the oriented cases studied in [@leader2010], [@cameron2015], and [@cameron2015deg].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Two newly observed bands built on a two-quasiparticle configuration in $^{130}$Ba have been investigated for the first time with the microscopic projected shell model. The experimental energy spectra and the available electromagnetic transition probabilities are well reproduced. The wobbling character of the higher band is revealed by the angular momentum projected wavefunctions via the *K plot* and the *azimuthal plot*. This provides the first strong microscopic evidence for wobbling motion based on a two-quasiparticle configuration in even-even nuclei.'
address:
- 'State Key Laboratory of Nuclear Physics and Technology, School of Physics, Peking University, Beijing 100871, China'
- 'School of Physical Science and Technology, Northwestern Polytechnical University, Xi’an 710129, China'
author:
- 'Y. K. Wang'
- 'F. Q. Chen'
- 'P. W. Zhao'
title: 'Two quasiparticle wobbling in the even-even nucleus $^{130}$Ba'
---
Nuclear wobbling ,Projected shell model ,*K plot* ,*Azimuthal plot*
\[sec1\] Introduction
=====================
The wobbling motion is one of the most intriguing quantum phenomena of a triaxial rotating nucleus, proposed by Bohr and Mottelson [@Bohr1998]. It is a quantum analog to the motion of a free classical asymmetric top, whose rotation around the principal axis with the largest moment of inertia is usually energy favored and stable. The term “stable” here means that at slightly larger energies, the rotating axis would not be very far away from the space-fixed angular momentum vector, but instead, it executes harmonic precession oscillations about the space-fixed angular momentum. For quantal nuclear systems, these oscillations appear as equidistant excitations and, thus, the energy spectrum is a series of $\Delta I = 2\hbar$ rotational bands corresponding to increasing phonon quanta $n$ and alternating signature quantum numbers. Moreover, the $\Delta I = 1\hbar$ transitions between the bands with $n$ and $n+1$ phonons are collectively enhanced.
A clear evidence for wobbling in this purely collective form, which is seen in all asymmetric top molecules, has not been found so far in nuclear systems. Instead, wobbling evidences have only been reported in odd-$A$ triaxial nuclei, e.g, $^{161}$Lu [@Bringel2005Eur.Phys.J.A167], $^{163}$Lu [@Odegard2001Phys.Rev.Lett.5866; @Jensen2002Phys.Rev.Lett.142503], $^{165}$Lu [@Schoenwaser2003Phys.Lett.B9], $^{167}$Lu [@Amro2003Phys.Lett.B197], $^{167}$Ta [@Hartley2009Phys.Rev.C41304], $^{135}$Pr [@Matta2015Phys.Rev.Lett.82501; @Sensharma2019Phys.Lett.B170], and $^{105}$Pd [@Timar2019Phys.Rev.Lett.62501], where either an odd proton or a neutron occupying a high angular momentum orbital is coupled to the triaxial rotor, and considerably influences the wobbling motion. As a result, the experimentally observed wobbling energies, i.e., energy differences between the wobbling bands, have been found to decrease with increasing spin, contrary to the behavior expected for even-even nuclei [@Bohr1998]. This is interpreted as the so-called “transverse wobbling” [@Frauendorf2014Phys.Rev.C14322], where the odd-nucleon angular momentum alignment is assumed to be frozen and perpendicular to the axis with the maximal moment of inertia. This interpretation stimulates great theoretical interests to clarify the modified wobbling mode in odd-$A$ nuclei using different models [@Chen2014Phys.Rev.C44306; @Tanabe2017Phys.Rev.C64315; @Raduta2017Phys.Rev.C54320; @Shimada2018Phys.Rev.C24318; @Budaca2018Phys.Rev.C24302].
Only a few indications for even-even wobbler have been reported. For instance, in $^{112}$Ru [@Hamilton2010Nucl.Phys.A28], the ground band, and the odd and even spin members of the “$\gamma$-band” are proposed as the zero-, one-, and two-phonon wobbling bands, respectively. However, this wobbling interpretation is not very solid because no electromagnetic transition data were reported. A very interesting example of wobbling in even-even nuclei is the recently reported band structure in $^{130}$Ba [@Petrache2019Phys.Lett.B241], where a pair of bands with even and odd spins, labeled S1 and S1’, were interpreted as the zero- and one-phonon wobbling bands [@Chen2019Phys.Rev.C61301]. It should be mentioned that the configuration of these two bands are built on two aligned protons in the bottom of the $h_{11/2}$ shell. Therefore, the wobbling excitation in $^{130}$Ba is not in a purely collective form, but in the presence of two aligned particles.
On the theoretical side, nuclear wobbling bands have been extensively studied with the triaxial particle-rotor model [@Frauendorf2014Phys.Rev.C14322; @Matta2015Phys.Rev.Lett.82501; @Timar2019Phys.Rev.Lett.62501]. However, such analyses are all phenomenological and are fitted to the data in one way or another. There are also many efforts to extend the microscopic cranking mean-field model to study the wobbling motion. The cranking mean-field model yields only the lowest state for a given configuration and, thus, one has to go beyond the mean-filed level to describe the wobbling excitations. This has been done by incorporating the quantum correlations by means of random phase approximation (RPA) [@Matsuzaki2002Phys.Rev.C41303; @Nakatsukasa2016Phys.Scr.73008] or by the angular momentum projection methods [@Oi2000Phys.Lett.B53; @Shimada2018Phys.Rev.C24318].
The projected shell model (PSM) carries out the shell-model configuration mixing based on Nilsson mean field with the angular momentum projection technique [@Hara1995Int.J.Mod.Phys.E637]. The implementation can also be rooted on the self-consistent relativistic [@Zhao2016Phys.Rev.C41301] and nonrelativistic [@Konieczka2018Phys.Rev.C34310] density functional theories. The PSM was used to understand the wobbling motion in $^{135}$Pr [@Sensharma2019Phys.Lett.B170], where the observed energy spectra and electromagnetic transitions for the wobbling bands are well reproduced. However, an illustration for the underlying wobbling geometry of the angular momentum was missing. The difficulty lies in the fact that the angular momentum geometry is defined in the intrinsic frame, while the angular momentum projected wavefunctions are written in the laboratory frame. In Ref. [@Chen2017Phys.Rev.C51303], focusing on the chiral doublet bands in triaxial nuclei [@Frauendorf1997Nucl.Phys.A131], the *K plot* and the *azimuthal plot* are introduced to illustrate the chiral geometry with the angular momentum projected wavefunctions.
In this work, we report a microscopic investigation on the recently observed two-quasiparticle bands S1 and S1’ of $^{130}$Ba with the PSM. This is the first example of wobbling motion based on a two-quasiparticle configuration. In particular, the influence of the two quasiparticles on the angular momentum geometry of the wobbling bands is illustrated in terms of the *K plot* and the *azimuthal plot*.
\[sec2\] Theoretical Framework
==============================
The framework of the PSM starts from the standard pairing plus quadrupole Hamiltonian [@Ring1980], $$\label{eq1}
\hat{H} = \hat{H}_0 - \frac{\chi}{2}\sum_{\mu}\hat{Q}^\dag_\mu\hat{Q}_\mu - G_M\hat{P}^\dag\hat{P} - G_Q\sum_{\mu}\hat{P}^\dag_\mu\hat{P}_\mu,$$ which includes a spherical single-particle shell model Hamiltonian, a quadrupole-quadrupole interaction, a monopole pairing interaction, and a quadrupole pairing interaction. The intrinsic vacuum state $|\Phi_0\rangle$ can be calculated by the following variational equation, $$\label{eq2}
\delta\langle\Phi_0|\hat{H}-\lambda_n\hat{N}-\lambda_p\hat{Z}|\Phi_0\rangle = 0,$$ with the Lagrange multipliers $\lambda_n$ and $\lambda_p$ determined by the neutron number $N$ and proton number $Z$, respectively.
Based on the intrinsic vacuum state $|\Phi_0\rangle$, the two quasiparticle states $|\Phi_\kappa\rangle$ for even-even nuclei can be constructed with $$\label{eq3}
|\Phi_\kappa\rangle \in \{\hat{\beta}^\dag_{\nu_i}\hat{\beta}^\dag_{\nu_j}|\Phi_0\rangle, \hat{\beta}^\dag_{\pi_i}\hat{\beta}^\dag_{\pi_j}|\Phi_0\rangle\},$$ where $\hat{\beta}^\dag_\nu$ and $\hat{\beta}^\dag_\pi$ are the quasiparticle creation operators for neutron and proton, respectively. The rotational symmetry of the intrinsic states $|\Phi_\kappa\rangle$ can be restored by the projection $\{\hat{P}^I_{MK}|\Phi_\kappa\rangle\}$, in which $\hat{P}^I_{MK}$ denotes the three-dimensional angular momentum projection operator [@Ring1980].
The Hamiltonian (\[eq1\]) is diagonalized in the space consisting of the projected two quasiparticle states and the vacuum, and this leads to the Hill-Wheeler equation, $$\label{eq6}
\sum_{\kappa'K'}\{\langle\Phi_{\kappa}|\hat{H}\hat{P}^I_{KK'}|\Phi_{\kappa'}\rangle-E^{I\sigma}\langle\Phi_{\kappa}|\hat{P}^I_{KK'}|
\Phi_{\kappa'}\rangle\}f^{I\sigma}_{K'\kappa'} = 0,$$ where $\sigma$ labels different eigenstates with the same spin $I$. The norm matrix element $\mathcal{N}_I(K,\kappa;K',\kappa') = \langle\Phi_\kappa|\hat{P}^I_{KK'}|\Phi_{\kappa'}\rangle$ and the energy kernel $\mathcal{H}_I(K,\kappa;K',\kappa') = \langle\Phi_\kappa|\hat{H}\hat{P}^I_{KK'}|\Phi_{\kappa'}\rangle$ can be calculated by using the Pfaffian algorithm [@Bertsch2012Phys.Rev.Lett.42505; @Hu2014Phys.Lett.B162].
By solving the Hill-Wheeler equation (\[eq6\]), one can obtain the eigenvalues $E^{I\sigma}$ and the corresponding eigenfunctions $$\label{eq7}
|\Psi^\sigma_{IM}\rangle = \sum_{K\kappa} f^{I\sigma}_{K\kappa}\hat{P}^I_{MK}|\Phi_\kappa\rangle,$$ with which the electromagnetic transitions can be calculated. It is known that the projected basis $\{\hat{P}^I_{MK}|\Phi_\kappa\rangle\}$ are not orthogonal and therefore, the coefficients $f^{I\sigma}_{K\kappa}$ in Eq. (\[eq7\]) should not be understood as probability amplitudes. However, one can construct the orthogonal and normalized collective wavefunctions [@Ring1980] $$\label{eq8}
g^{I\sigma}(K,\kappa) = \sum_{K'\kappa'}\mathcal{N}_I^{1/2}(K,\kappa; K',\kappa')f^{I\sigma}_{K'\kappa'},$$ which are interpreted as probability amplitudes and are used to construct the *K plot* and the *azimuthal plot* [@Chen2017Phys.Rev.C51303]. The *K plot* is defined as the probability distributions of the components of the angular momentum on the three axes of the intrinsic frame, $$\label{eq9}
p^{I\sigma}(|K|) = \sum_\kappa \left|g^{I\sigma}(K,\kappa)\right|^2 + \left|g^{I\sigma}(-K,\kappa)\right|^2.$$ The *azimuthal plot* is defined as the probability distributions of the polar and the azimuthal angles $(\theta,\phi)$ of the angular momentum in the intrinsic frame, $$\label{eq10}
\mathcal{P}(\theta,\phi) = \sum_\kappa\int_0^{2\pi} d\psi'\left|G^{II}(\psi',\theta,\pi-\phi,\kappa)\right|^2,$$ where $\theta$ denotes the angle between the total angular momentum and the long (*l*) axis, and $\phi$ denotes the angle between the projection of the total angular momentum on the intermediate-short (*i-s*) plane and the *i* axis. The integrand $G^{II}(\psi',\theta,\pi-\phi)$ reads, $$\label{eq11}
G^{II}(\psi',\theta,\pi-\phi,\kappa) = \sqrt{\frac{2I+1}{8\pi^2}}\sum_K g^I(K,\kappa)D^{I\ast}_{IK}(\psi',\theta,\pi-\phi).$$
In the present calculations for the observed bands S1 and S1’ in Ref. [@Petrache2019Phys.Lett.B241], both the proton and neutron orbitals are taken from the $N = 3,\, 4,\, 5$ major shells. The monopole pairing strength, $G_M = 0.15$ MeV for neutron and $G_M = 0.16$ MeV for proton, is determined by the odd-even mass differences. The strength of quadrupole pairing interaction $G_Q$ is assumed to be $0.2\, G_M$; similar to many other calculations [@Sun1994Phys.Rev.Lett.3483; @Sun1996Phys.Rep.375]. The strength of quadrupole force $\chi$ is connected with the quadrupole deformation $(\beta,\gamma)$ by the self-consistent relation as shown in Ref. [@Hara1995Int.J.Mod.Phys.E637]. The deformation parameters $(\beta,\gamma)$ are chosen as $(0.24,30^\circ)$, which is similar to the results given by the calculations of the cranking covariant density functional theory [@Zhao2011Phys.Rev.Lett.122501; @Zhao2011Phys.Lett.B181; @Zhao2015Phys.Rev.Lett.22501] with PC-PK1 [@Zhao2010Phys.Rev.C54319]. Since the configuration of bands S1 and S1’ is assigned as $\pi(h_{11/2})^2$ [@Petrache2019Phys.Lett.B241], only the proton orbitals in the $h_{11/2}$ shell are used to construct the configuration space.
\[sec3\] Results and discussion
===============================
![(Color online) Calculated energy spectra (top) for the zero- ($n = 0$) and one-phonon ($n = 1$) bands as well as the wobbling energy (bottom), in comparison with data [@Petrache2019Phys.Lett.B241].[]{data-label="Energy"}](Fig1.eps){width="40.00000%"}
In the upper panel of Fig. \[Energy\], the calculated excitation energies with respect to the bandhead ($I=10\hbar$) for the zero- ($n=0$) and one-phonon ($n=1$) states are presented in comparison with data [@Petrache2019Phys.Lett.B241]. It is seen that the observed excitation energies are well reproduced, in particular for the lower spin states. For the higher spin states, the calculated energies are slightly higher than data, but this could be improved by including more configurations in the model space. The wobbling energies $E_{\mathrm{wob}}$, defined as $$E_{\mathrm{wob}} (I) = E_{n=1}(I) - \left[ E_{n=0}(I+1)+E_{n=0}(I-1) \right]/2,$$ were calculated from the energy spectra and are shown in the lower panel of Fig. \[Energy\]. The calculated results are in good agreement with data, and in particular, the decreasing tendency of the wobbling energy as a function of angular momentum is presented. This has been suggested as the hallmark of the transverse wobbling in Ref. [@Frauendorf2014Phys.Rev.C14322].
![(Color online) Calculated transition probability ratios $B(E2)_{\rm out}/B(E2)_{\rm in}$ (top) and $B(M1)_{\rm out}/B(E2)_{\rm in}$ (bottom) for the transitions from the one-phonon ($n = 1$) band to the zero-phonon ($n = 0$) band in comparison with data available [@Petrache2019Phys.Lett.B241].[]{data-label="Transition"}](Fig2.eps){width="40.00000%"}
The experimental and theoretical transition probability ratios $B(M1)_{\rm out}/B(E2)_{\rm in}$ and $B(E2)_{\rm out}/B(E2)_{\rm in}$ for the transitions from the one-phonon band to the zero-phonon band are shown in Fig. \[Transition\]. The calculated results agree well with the data. Different from an ideal wobbler, here the values of $B(M1)_{\rm out}/B(E2)_{\rm in}$ are comparable with the $B(E2)_{\rm out}/B(E2)_{\rm in}$ ones. This is attributed to the fact that two quasiparticles in the $h_{11/2}$ shell are involved in the configurations, which enlarges the $M1$ matrix elements. Therefore, the $E2$ component of the $\Delta I=1$ transitions is not expected to be that dominate like in other wobbling bands [@Matta2015Phys.Rev.Lett.82501].
{width="70.00000%"}
In order to examine the orientation of the angular momentum on the intrinsic $(\theta, \phi)$ plane, Fig. \[Aplot\] depicts the *azimuthal plots*, i.e., the profiles $\mathcal{P}(\theta, \phi)$ for the zero- (even spin) and one- (odd spin) phonon bands at several selected angular momenta. In the $\theta$ direction, for all states, the distributions $\mathcal{P}(\theta, \phi)$ are mainly concentrated around $\theta \sim 90^\circ$, and this means that the angular momentum locates mainly in the plane of short and intermediate axes. For both the zero- and one-phonon states, the distributions in the $\theta$ direction are more diffuse at lower spin states. In the $\phi$ direction, however, distinct patterns are found for the zero- and one-phonon states. For the zero-phonon states with $I = 10,\,16,\,22 \hbar$, the distributions $\mathcal{P}(\theta, \phi)$ are mainly concentrated around $\phi\sim90^\circ$, and they become more and more diffuse with the increasing spin. For the one-phonon states with $I = 11,\,17,\,23 \hbar$, the angular momenta orientate equally at two directions with $\phi\sim 120^\circ$ and $\phi\sim 60^\circ$, respectively, and the separation of the two directions become larger with the increasing spin. In contrast to the maximum at $\phi \sim 90^\circ$ for the distribution $\mathcal{P}(\theta, \phi)$ of the zero-phonon states, there is a minimum at $\phi \sim 90^\circ$ in the distribution for the one-phonon states. Therefore, the distribution in the $\phi$ direction is symmetric for the zero-phonon states and is antisymmetric for the one-phonon states. This pattern is consistent with the expectation of a wobbling motion, i.e., the precession of the total angular momentum around the short axis.
{width="70.00000%"}
In Fig. \[Kplot\], the *K plot*, i.e., the $K$ distributions of angular momenta on the three principle axes for the zero- (even spin) and one- (odd spin) phonon bands are depicted at several selected angular momenta. At $I = 10\hbar$, the angular momentum is mainly along the short axis, due to the alignment of the two quasiparticles in the $h_{11/2}$ shell. Therefore, it is seen in Fig. \[Kplot\] that the probability is peaked at very small $K$ values on the *long* and *intermediate* axes, while at $K\sim 10\hbar$ on the *short* axis. At $I = 11\hbar$, the $K$ distributions changes mainly on the *intermediate* axis, i.e., the remarkable probability at $K_i\sim 0$ for $I = 10 \hbar$ vanishes. This indicates that a symmetric wavefunction with respect to $K_i \sim 0$ at $I = 10 \hbar$ changes to an antisymmetric one at $I = 11\hbar$. It is consistent with the pattern shown in the *azimuthal plots* (see Fig. \[Aplot\]), and reflects the fact that the states at $I = 10\hbar$ and $11\hbar$ correspond to the zero- and one-phonon states, respectively.
Increasing the spin to $I = 16\hbar$ and $22\hbar$, the $K$ distributions on the *long* axis barely change, which means that the angular momentum grows very little in the *long*-axis direction. Comparing the $K$ distributions on the other two axes, it is found that the angular momentum is generated mainly in the short axis. Though the two $h_{11/2}$ quasiparticles contribute only about $10\hbar$ in the short axis, the peaks of $K$ distributions on the short axis can reach $16\hbar$ and $22\hbar$ for the $I = 16\hbar$ and $22\hbar$ states, respectively. This indicates that the scenario of the transverse wobbling [@Frauendorf2014Phys.Rev.C14322] is realized in the present microscopic calculations. The $K$ distributions on the *intermediate* axis become very broad for the $I = 16\hbar$ and $22\hbar$ states, while the significant probabilities at $K_i\sim 0$ remain. This feature, together with the vanishing probabilities at $K_i\sim 0$ for the $I = 17 \hbar$ and $23\hbar$ states, reflects again the wobbling character of the odd-spin states.
For the $I = 23 \hbar$ state, it can be seen that the components of the angular momentum on the *intermediate* and *short* axes are comparable. This might indicate that $I = 23 \hbar$ is very close to the so-called critical spin, and above this spin, the transverse wobbling would not be stable and the tilted axis rotation may appear.
\[sec4\] Summary
================
In summary, a pair of newly observed bands built on a two-quasiparticle configuration in $^{130}$Ba have been investigated with the microscopic projected shell model. The experimental energy spectra, energy differences between the two bands, as well as the available electromagnetic transition probabilities are well reproduced. Different from an ideal wobbler, the values of $B(M1)_{\rm out}/B(E2)_{\rm in}$ are comparable with the $B(E2)_{\rm out}/B(E2)_{\rm in}$ ones due to the involvement of the two quasiparticles in the configurations. Nevertheless, the wobbling character of the higher band can be demonstrated in terms of the *K plot* and the *azimuthal plot*. This provides the first strong microscopic evidence of wobbling motion based on a two-quasiparticle configuration in even-even nuclei.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work was partly supported by the National Key R&D Program of China (Contract No. 2018YFA0404400 and No. 2017YFE0116700), the National Natural Science Foundation of China (Grants No. 11621131001, No. 11875075, No. 11935003, and No. 11975031), and the Laboratory Computing Resource Center at Argonne National Laboratory.
[36]{} natexlab\#1[\#1]{}\[1\][`#1`]{} \[2\][\#2]{} \[1\][\#1]{} \[1\][[](http://dx.doi.org/#1)]{} \[1\][[](pmid:#1)]{} \[2\][\#2]{} , , , , . . , , , , , , , , , , , , , , , , , , , , , , , () . . , , , , , , , , , , , , , , , , , , , , () . . , , , , , , , , , , , , , , , , , , , , , , , () . . , , , , , , , , , , , , , , , , , , , , , , , () . . , , , , , , , , , , , , , , , , , , , () . . , , , , , , , , , , , , , , , , , , , , , , , , , , , , () . . , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , () . . , , , , , , , , , , , , , , , , , , , , , , () . . , , , , , , , , , , , , , , , , , , , , , , , , , , , , () . . , , () . . , , , , () . . , , () . . , , , () . . , , , , () . . , () . . , , , , , , , , , , , () . . , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , () . . , , , () . . , , , () . . , , , , () . . , , , , () . . , , () . . , , , () . . , , , () . . , , , , , () . . , , () . . , , , , . , , () . . , , , () . , , , () . . , , () . . , , , , , () . . , , , , , , () . . , , , () . . , , , , () . .
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper attempts to address the question of how best to assure the correctness of saturation-based automated theorem provers using our experience developing the theorem prover [[[Vampire]{}]{}]{}. We describe the techniques we currently employ to ensure that Vampire is correct and use this to motivate future challenges that need to be addressed to make this process more straightforward and to achieve better correctness guarantees.'
author:
- Giles Reger
- Martin Suda
- Andrei Voronkov
bibliography:
- 'bib.bib'
title: 'Testing a Saturation-Based Theorem Prover: Experiences and Challenges (Extended Version) [^1] '
---
[^1]: This work was supported by EPSRC Grant EP/K032674/1. Martin Suda and Andrei Voronkov were partially supported by ERC Starting Grant 2014 SYMCAR 639270. Martin Suda was also partially supported by the Austrian research projects FWF S11403-N23 and S11409-N23. Andrei Voronkov was also partially supported by the Wallenberg Academy Fellowship 2014 - TheProSE.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In intensity-modulated radiation therapy, optimal intensity distributions of incoming beams are decomposed into linear combinations of leaf openings of a multileaf collimator (segments). In order to avoid inefficient dose delivery, the decomposition should satisfy a number of dosimetric constraints due to suboptimal dose characteristics of small segments. However, exact decomposition with dosimetric constraints is only in limited cases possible. The present work introduces new heuristic segmentation algorithms for the following optimization problem: Find a segmentation of an approximated matrix using only allowed fields and minimize the approximation error. Finally, the decomposition algorithms were implemented into an optimization programme in order to examine the assumptions of the algorithms for a clinical example. As a result, identical dose distributions with much fewer segments and a significantly smaller number of monitor units could be achieved using dosimetric constraints. Consequently, the dose delivery is more efficient and less time consuming.'
author:
- Antje Kiesel$^1$ Tobias Gauer$^2$
bibliography:
- 'IMRT.bib'
title: 'Approximated segmentation considering technical and dosimetric constraints in intensity-modulated radiation therapy with electrons'
---
$^1$ Institute for Mathematics, University of Rostock, Rostock, Germany\
E-Mail: [email protected]
$^2$ Department of Radiotherapy and Radio-Oncology, University Medical Center Hamburg-Eppendorf, Hamburg, Germany\
E-Mail: [email protected]
[**Keywords:**]{} IMRT planning, intensity matrix, approximated segmentation, dosimetric and technical constraints, multileaf collimator
[**Mathematics Subject Classification (2000):**]{} MSC 90C90, MSC 92C50, MSC 49M25, MSC 49M27
Introduction
============
In intensity-modulated radiation therapy (IMRT), intensity matrices with nonnegative integer entries are computed for each irradiation field. After discretization of the field into bixels, each entry of the matrix corresponds to the required intensity within this bixel. The segmentation step consists in decomposing the matrix into a linear combination of subfields (segments) shaped by a multileaf collimator (MLC). The first intuition is that a treatment plan is optimal, if the linear combination of the chosen segments equals the matrix. Such a plan consists of various segments possibly including those segments where most of the irradiation field is covered and only few bixels receive radiation.
For dosimetric reasons, however, the model assumption is not given in practice. Irradiation of small photon or electron segments result in a much lower dose output compared to conventional conformal fields. Therefore, the linearity assumption, that irradiating one segment is equivalent to dividing it into two parts and irradiating them separately, only holds, if the two parts are still sufficiently large. In addition, the penetration depth of electrons decreases with decreasing field size and is almost independent of the beam energy for approximately $1$ cm $\times$ $1$ cm fields. However, the energy dependence of the penetration depth is necessary for our new IMRT technique with electron beams to adjust the dose to the target volume by use of various beam energies. Figure \[fig:properties\] shows that electron fields of approximately $3$ cm $\times$ $3$ cm are necessary to keep an output factor of nearly $1$ and an energy-dependent penetration depth.
![Electron dose output at the dose maximum normalized to the dose output of the $10$ cm $\times$ $10$ cm field and electron penetration depth of the 90 % depth-dose as a function of square field size and electron energy (from [@Gau08]). The fields were shaped by an add-on MLC for electrons presented in Figure \[fig:mlc\]. A minimum MLC field size of approximately $3$ cm $\times$ $3$ cm is necessary for decomposing intensity distributions into leaf openings to ensure an output factor of nearly $1$ and an energy-dependent penetration depth.](figure1a.eps "fig:") ![Electron dose output at the dose maximum normalized to the dose output of the $10$ cm $\times$ $10$ cm field and electron penetration depth of the 90 % depth-dose as a function of square field size and electron energy (from [@Gau08]). The fields were shaped by an add-on MLC for electrons presented in Figure \[fig:mlc\]. A minimum MLC field size of approximately $3$ cm $\times$ $3$ cm is necessary for decomposing intensity distributions into leaf openings to ensure an output factor of nearly $1$ and an energy-dependent penetration depth.](figure1b.eps "fig:") \[fig:properties\]
As a consequence, a treatment plan should consist of segment shapes satisfying certain constraints that ensure a minimum field size. For practical purposes it is also necessary that the field openings are connected and do not degenerate into two or more parts. Besides those dosimetric constraints, there are also technical constraints reducing the number of allowed shapes. One is the leaf overtravel constraint that accommodates the fact that the left (respectively right) leaf of the MLC cannot be shifted further than a threshold to the right (respectively left). These constraints have the consequence, that not every intensity matrix is decomposable in segments satisfying the constraints. This leads us to the task to find an approximation matrix and its decomposition into ”good” segments, that differs from the given intensity matrix as little as possible. The aim is to generate equivalent treatment plans with good segments leading to a reduction in the segment number and monitor units, respectively.
The decomposition problem for the exact case without concerning any additional constraints is well studied. Algorithms for the minimization of the beam-on time can be found in [@Ahu05; @Baa05; @Bor94; @Kal05; @Kal05a; @Kam03a]. Approaches for minimizing the number of used segments are given in [@Eng05; @Kal04b; @Nus06]. A variety of technical constraints are considered, see [@Bol04; @Kal08] for the interleaf collision constraint, that prohibits an overlap of adjacent leaf pairs, and [@Kal08; @Kal08a; @Kam04a; @Kam04b; @Que04] for the tongue-and-groove constraint. Kamath et. al. [@Kam03a] also investigate the minimum separation constraint that requires a minimum leaf opening in each row and develop a criterion for a matrix being decomposable under this constraint. Engelbeen and Fiorini [@Engelbeen08] deal with the interleaf distance constraint where the allowed difference between two left (respectively right) leaf positions is bounded by some given threshold.
An approximation problem with the aim of reducing the total beam-on time was first formulated in [@EngKie08] and generalized to approximated decomposition with interleaf-collision constraint in [@Kal08c] and [@KalKie08]. The dependence between field size and output factors, penetration depth and depth-dose fall-off is outlined in [@Gau08]. These considerations lead to the decomposition problem using segments that satisfy some minimum field size constraints. Under these constraints, an exact decomposition of the intensity matrices is, in general, no longer possible (cf. [@Kam03a]) and an approximation problem has to be formulated.
Another algorithmic approach that aims at minimizing the number of segments while keeping the quality of the treatment plan is the direct aperture optimization that combines the choice of beams, apertures and weights without computing a leaf sequencing step. Shepard et al. [@She02] allow only a limited number of apertures for each beam, Bedford and Webb [@Bed06] also integrate constraints on the segment shape and size in the direct aperture optimization approach. Our algorithm is applicable if one uses intensity profile segmentation and wants to compute segmentations satisfying certain field size constraints and reducing the complexity of the plan. Matuszak et al. [@Mat] deal with the minimization of the monitor units by smoothing the intensity profiles.
The paper is organized as follows. Section \[sec:definitions\] gives two definitions of what we call a segment with good dosimetric properties, one basic definition and an extended one including one further constraint. We concretely define the approximation problem and in Section \[sec:algorithm\] propose heuristic algorithms for both definitions, each of them consisting of seven different steps. The different parts of the algorithm and their properties are analyzed in Section \[sec:parts\]. We especially outline, that the solutions of the subproblems in step $1$ and $2$ are indeed optimal. Section \[sec:case\] introduces the clinical case we used for testing the quality of our segmentations. Section \[sec:results\] gives computational results for the test case and detailed numerical results for the segmentation of clinical matrices from different IMRT treatment plans.
Problem formulation and definitions {#sec:definitions}
===================================
Throughout the paper we use the notation $$[k]=\{1,2,\ldots,k\}\quad\text{and}\quad[k,l]=\{k,k+1,\ldots,l\}$$ for integers $k$ and $l$, $k \le l$. Let $A=(a_{ij})$ denote the given fluence matrix of size $m\times n$. Feasible leaf positions of the MLC are modeled as binary matrices $S=(s_{ij})$, called *segments*, that satisfy the consecutive-ones-property in each row. In other words, $S$ is a segment, if there are integral intervals $[\ell_i,r_i]$ for all $i \in [m]$, representing the positions of the left and the right leaf, such that $$\label{eqn:segment}
s_{ij}=\begin{cases}1 \mbox{ if } l_i \le j \le r_i\\ 0 \mbox{ otherwise} \end{cases}\quad ((i,j)\in[m]\times[n]).$$
Furthermore, for each segment $S$, we define $s_{0,j}=s_{m+1,j}=0$ for all $j \in [n]$. For the described reasons, we introduce five parameters $b_l$, $b_r$, $g_1$, $g_2$ and $f$ representing the following constraints:
1. Left Leaf Overtravel Constraint: For all $i \in [m]$, we require $l_i \le b_l$. In each row the left leaf cannot be shifted more to the right than to the bixel with index $b_l$.
2. Right Leaf Overtravel Constraint: For all $i \in [m]$, we require $r_i \ge b_r$. In each row the right leaf cannot be shifted more to the left than to the bixel with index $b_r$.
3. Minimum Separation Constraint and Row Overlap: If a row $i \in [m]$ is not totally covered, we require $r_i-l_i \ge g_1-1$. Similarly, if rows $i$ and $i+1$ are not completely covered, we claim $\min(r_i,r_{i+1})-\max(l_i,l_{i+1}) \ge g_1-1$. At least $g_1$ consecutive bixels in each row receive radiation and the irradiated area of two consecutive rows overlaps in at least $g_1$ bixels.
4. Minimum Vertical Gap: We require a minimum vertical field size and a minimum vertical size of the covered regions, i.e. in each column consecutive ones or zeros should have a minimum number $g_2$. In detail, if $s_{i-1,j}=0$, $s_{ij}=s_{i+1,j}=\dots=s_{k-1,j}=1$ and $s_{kj}=0$ for some column $j$, we have $k-i \ge g_2$. Analogously, we require the same for consecutive zeros framed by ones.
5. Minimum Total Field Height: At least $f$ consecutive rows of the field are not totally covered, i.e. there are at least $f$ consecutive rows with $l \le r$. This ensures, that the total size of the field is reasonably large.
Of course, these parameters only make sense if $1 \le b_l,b_r,g_1 \le n$, $1 \le g_2,f \le m$ as well as $1 \le b_r < b_l \le n$ and $g_2 \le f$.
The case that one row of the field is totally covered and receives no radiation at all, is throughout this paper represented by the leaf positions $l=n+1$ and $r=0$. In practice, one will of course choose leaf positions of the form $l=r+1$ with $l \le b_l$ and $r \ge b_r$ that respect the leaf overtravel constraints (i) and (ii).
The Minimum Vertical Gap can be formulated in terms of the leaf position as follows: If $l_i<l_{i-1}$ for some $i \ge 2$, then we also require $l_{i+1} \le l_i, l_{i+2} \le l_i, \dots, l_{i+g_2-1} \le l_i$. Analogously, if $r_i>r_{i-1}$ for some $i \ge 2$, we also have $r_{i+1} \ge r_i, r_{i+2} \ge r_i, \dots, r_{i+g_2-1} \ge r_i$. This ensures, that in vertical direction, we always have at least $g_2$ bixels open. Therefore, we additionally not allow that $l_i>l_{i-1}$ or $r_i<r_{i-1}$ for $2 \le i \le g_2$ and forbid also $l_i<l_{i-1}$ as well as $r_i>r_{i-1}$ for $m-g_2+1 \le i \le m$. Similarly, we require at least $g_2$ bixels closed in vertical direction, if there are open bixels above and below in this column of the matrix. Thus, we make sure that also thin shapes in vertical direction, having negative dosimetric properties as discussed in the introduction, are forbidden.
\[fig:vertical\_gap\]
A segment is called *connected* if the irradiated area that corresponds to its leaf positions does not resolve into two or more parts, i.e. if the corresponding rectilinear polygon (considered as an open set) is connected.
As the realization of the minimum vertical gap turns out to be the most difficult task, we introduce two different definitions of “good” segments.
\[defn:segment\] Given the parameters $b_l,b_r,g_1,f \in \mathbb{N}$ with $1 \le b_l,b_r,g_1 \le n$, $1 \le f \le m$ as well as $1 \le b_r < b_l \le n$, a *segment with good dosimetric properties* $S$ is a connected segment satisfying the constraints (i), (ii),(iii) and (v).
Given the parameters $b_l,b_r,g_1,g_2,f \in \mathbb{N}$ with $1 \le b_l,b_r,g_1 \le n$, $1 \le g_2,f \le m$ as well as $1 \le b_r < b_l \le n$ and $g_2 \le f$, a *segment with very good dosimetric properties* is a connected segment satisfying the constraints (i)-(v).
For brevity of notation we will call the segments with good dosimetric properties from now on simply segments and the segments with very good dosimetric properties advanced segments.
All in all, we have two optimization problems: Given a matrix $A$ with positive integer entries and suitable parameters $b_l,b_r,g_1,g_2,f$, find a segmentation $$B=\sum\limits_{k=1}^{t} u_k S_k$$ where the $u_k$ are positive integers and
- [**Approximated Segmentation into segments (ASS)**]{}: the $S_k$ are segments
- [**Approximated Segmentation into advanced segments (ASAS)**]{}: the $S_k$ are advanced segments
such that $$\sum\limits_{i=1}^{m} \sum\limits_{j=1}^{n} |a_{ij}-b_{ij}| \rightarrow \min.$$
The value of the objective function of the optimization problem is called *total change*. The delivery time of the segmentation is $\sum_{k=1}^{t} u_k$ and the number of segments is $t$.
Obviously, for the parameter set $b_l=n$, $b_r=1$, $g_1=g_2=f=1$, segments are simply connected segments in the sense of Equation (\[eqn:segment\]), the approximation problem has $0$ as value of the objective function and degenerates to a segmentation problem into connected segments defined by Equation (\[eqn:segment\]).
Approximated segmentation {#sec:algorithm}
=========================
Now we introduce two basic algorithms for [**ASS**]{} and [**ASAS**]{}, each consisting of seven different steps. Two of the steps are identical in both algorithms as they are computed before the segmentation step and therefore do not affect the parameter $g_2$. Five of the steps differ subject to whether the constraint $g_2$ is regarded or not (namely steps $3$-$7$). Steps $1$ and $2$ will be solved exactly, whereas steps $3$-$7$ are heuristic.
The basic structure of our algorithms is as follows:
1. Solve the [**Leaf Overtravel Constraint Problem (LOC)**]{} on $A$: Given $b_l,b_r \in \mathbb{N}$ with $1 \le b_l < b_r \le n$, find an approximation $B$ with nonnegative integer entries that can be decomposed with respect to the Leaf Overtravel Constraint such that the total change is minimal.
$A:=B.$
2. Solve the [**Minimum Separation Constraint Problem (MSC)**]{} on $A$: Given $g_1 \in [n]$, find an approximation $B$ that can be decomposed into segments with $r_i-l_i \ge g_1-1$ for all $i \in [m]$ such that the total change is minimal.
$A:=B.$
3. ASS: Compute an approximated segmentation $\mathcal{S}$ into connected segments satisfying (i)-(iii), but not necessarily (v).
ASAS: Compute an approximated segmentation $\mathcal{S}$ into connected segments satisfying (i)-(iv), but not necessarily (v).
The approximated segmentation may violate (v) and belongs to an approximation matrix $B$ that might have a large total change.
4. [**Combination of Fields**]{}: Combine stepwise two disjoint fields, if this is possible with small total change. For ASAS, make sure that the new field satisfies (iv).
5. [**Make-two-of-two**]{}: For $S,S' \in \mathcal{S}$, where $S$ violates (v) and $S'$ satisfies (v), compute a substitution $S+S'=\hat{S}+\hat{S}'$, such that $\hat{S}$ and $\hat{S}'$ satisfy (v). For ASS, make sure that $\hat{S}$ and $\hat{S}'$ satisfy (i)-(iii). For ASAS, $\hat{S}$ and $\hat{S}'$ must also satisfy (iv).
6. [**Handle Critical Segments**]{}: If there are still segments violating (v), try to combine them with a feasible segment such that the total change of the combination is smaller than omitting the critical segment. Take care, that (i)-(iii) (resp. (i)-(iv)) hold.
7. [**Total Change Improvement**]{}: For all segments and all rows $i$, check whether an increase or decrease of $l_i$ (respectively $r_i$) reduces the total change. For ASS, change $l_i$ or $r_i$ only if (i)-(iii) still hold. For ASAS, change $l_i$ or $r_i$ only if (i)-(iv) still hold. Look at all the segments cyclically until no more changes are possible.
The output of the algorithm is a segmentation of an approximation matrix that consists only of segments for ASS and only of advanced segments for ASAS. The LOC- and the MSC-approximation aim at producing a first approximation that can be better decomposed with the given constraints (i)-(iii) in the segmentation step than the initial matrix. The combination of fields, the make-two-of-two-step and the handling of critical segments try to provide segments satisfying (v) without producing too much total change. Finally, the total change improvement is computed in order to improve the approximation.
A very important feature of our heuristic is that further constraints can easily be taken into consideration. For example, if adjacent left and right leafs may not overlap (interleaf collision constraint) one can allow only leaf positions that respect this constraint in the optimization steps $3$-$7$.
Subproblems {#sec:parts}
===========
Now we describe in detail the subalgorithms and outline some of their basic properties. Throughout the steps that follow the leaf overtravel approximation, we allow only leaf positions that respect the leaf overtravel constraint. For simplicity, we will not mention this basic fact in every step.
Leaf Overtravel Constraint
--------------------------
As the leaf overtravel constraint only affects a single row of the matrix, the problem LOC can be solved for each row independently. Thus, we compute an optimal approximation of a vector ${\boldsymbol{a}}$. Segmentations reduce to sums of intervals $[l,r]$. Segments are simply $0$-$1$-vectors ${\boldsymbol{s}}$ with consecutive ones.
\[lem:loc\] A vector ${\boldsymbol{a}}$ has a segmentation ${\boldsymbol{a}}=\sum\limits_{i=1}^{k} {\boldsymbol{s}}_i$ with corresponding leaf positions $l_i \le b_l$ and $r_i \ge b_r$ iff $a_j \ge a_{j+1}$ for all $j \in [b_l,n-1]$ and $a_j \ge a_{j-1}$ for all $j \in [2,b_r]$.
Let $a_0=a_{n+1}:=0$ and $a_+:=\max\{0,a\}$. On the one hand, the algorithm of Bortfeld (see [@Bor94]) provides a segmentation where the left leaf position is $j$ for exactly $(a_j-a_{j-1})_+$ segments. Analogously, the right leaf position is $j$ for $(a_j-a_{j+1})_+$ segments and no other leaf positions occur. On the other hand, it is obvious that if $a_j>a_{j-1}$ (respectively $a_j>a_{j+1}$) there will be a segment with left (respectively right) leaf position $j$ in every segmentation. This concludes the proof.
Therefore, we have to find an approximation vector, that has no up-steps after index $b_l$ and no down-steps before index $b_r$. As we assume $b_r<b_l$, we can use symmetry to solve the approximation problem for the right leaf positions. Besides, the criterion from Lemma \[lem:loc\] shows, that $b_j=a_j$ for $j \in [b_r+1,b_l-1]$ for each optimal solution of LOC. We simply need to solve the following problem for the subvector $(a_{b_l},\dots,a_n)$ and the left overtravel constraint:
Given a vector ${\boldsymbol{v}}=(v_1,\dots,v_k)$, find an approximation vector ${\boldsymbol{w}}$ with $w_j \ge w_{j+1}$ for $j \in [k-1]$ such that $\parallel {\boldsymbol{v}}-{\boldsymbol{w}} \parallel_1 = \sum\limits_{j=1}^{k} |v_j-w_j| \rightarrow \min$.
The algorithm for solving the problem LOC-left is described in Algorithm \[alg:loc\] in the appendix. It uses a graph theoretical approach and computes a shortest path in a layered digraph, where the $j$-th layer consists of nodes representing the possible entries of the $j$-th component of the approximation vector. The problem LOC-left is similar to the Monotone Discrete Approximation Problem (MDAP) formulated in [@EngKie08] and the algorithm follows the same idea.
Let $min:=\min_{j \in [k]} v_j$ and $max:=\max_{j \in [k]} v_j$ and let $tc_{ij}$ be the objective value of an optimal solution of LOC-left with $w_j=i$. Let $pre_{ij}$ be the corresponding predecessor $w_{j-1}$. With respect to Algorithm \[alg:loc\] (that uses the notation above) we yield the following
\[thm:loc\] Algorithm \[alg:loc\] computes an optimal solution of LOC-left.
The initial values $tc_{i1}$ are trivially correct. Let now $j>1$ and let $(w_1,\dots,w_j)$ be an optimal approximation of $(v_1,\dots,v_j)$ with $w_j=i$. By induction, $tc_{w_{j-1},j-1}$ is computed correctly and thus $$\sum\limits_{l=1}^{j} |v_l-w_l| =tc_{w_{j-1},j-1}+|v_j-i| \ge tc_{ij}.$$ Therefore $tc_{ij}$ is a lower bound for the total change. The choice of $i_{opt}$ makes sure, that the optimal value of $w_k$ is chosen and obviously, the approximation vector from Algorithm \[alg:loc\] realizes the lower bound for the total change of $tc_{i_{opt},k}$.
Minimum separation constraint
-----------------------------
Like the Overtravel Constraint, the Minimum Separation Constraint can be handled independently for each row of the matrix. Thus the task is the following:
Given a vector ${\boldsymbol{v}}=(v_1,\dots,v_n)$ with nonnegative integer entries, find an approximation vector ${\boldsymbol{w}}$ with nonnegative integer entries, such that ${\boldsymbol{w}}$ has a decomposition into intervals of length $\ge g_1$ and $\parallel {\boldsymbol{v}}-{\boldsymbol{w}} \parallel_1 = \sum\limits_{j=1}^{n} |v_j-w_j| \rightarrow \min$.
We know from Kamath et al. (see [@Kam03a]) that a vector ${\boldsymbol{a}}$ can be decomposed without violating the minimum separation constraint, if the optimal decomposition of their algorithm SINGLEPAIR does not violate the minimum separation constraint. For example, the vector ${\boldsymbol{a}}=(1,2,1)$ cannot be decomposed with $g_1=3$, as the optimal decomposition is $(1,2,1)=(1,1,0)+(0,1,1)$ and the used intervals do not have a minimum length of $3$. This motivates the approximation problem defined above.
Obviously, the problem MSC-Row can be formulated as an integer linear programming problem as follows: $$\begin{aligned}
\sum\limits_{j=1}^{k} \sum\limits_{j'=\min(k,j+g_1-1)}^{n} u_{j,j'} - \gamma_k & \le a_k & k \in [n]\\
-\sum\limits_{j=1}^{k} \sum\limits_{j'=\min(k,j+g_1-1)}^{n} u_{j,j'} - \gamma_k & \le -a_k & k \in [n]\\
u_{j,j'} & \ge 0 & j,j' \in [n], j' \ge j+g_1-1 \\
u_{j,j'} & \in \mathbb{Z} & j,j' \in [n], j' \ge j+g_1-1 \\
\gamma_k & \in \mathbb{Z} & k \in [n]\\
\sum\limits_{k=1}^{n} \gamma_k & \rightarrow \min\end{aligned}$$
We solve this integer program for each row of the matrix using SCIP [@Achterberg2007] with SoPlex [@Wunderling] as LP solver. The problem can also be solved by a combinatorial algorithm using a minimum cost flow formulation which is shown in [@Eng_Fio_Kie09].
Segmentation
------------
Let $B$ be the approximation matrix resulting from the MSC-step. The basic idea of the segmentation is to consider the current total change to be the sum of the absolute values of the entries of $B$ and to iteratively compute a segment $S$ whose subtraction reduces the current total change. In each step, the matrix $B$ is updated by setting $B:=B-S$. At the end of the segmentation, a positive entry in $B$ represents a bixel with underdose and a negative entry a bixel with overdose.
### Segmentation for ASS
Let ${\boldsymbol{b}}_i$ denote the $i$-th row of $B$ for all $i \in [m]$. A segmentation consists of segments $S$ each represented by its leaf positions $l_i,r_i$ for $i \in [m]$. The main body for the segmentation step is described in Algorithm \[alg:seg\_ASS\] in the appendix. This algorithm uses the subroutine [**Find interval ASS**]{} that is precisely described in Algorithm \[alg:find\_interval\_ASS\].
The idea behind this heuristic choice of the segment $S$ being subtracted from $B$ in each step is, that we compute the first interval from a sliding window segmentation (see again [@Bor94]), with $l$ as index of the first up-step and $r$ as index of the first down-step in the corresponding row. If these values already satisfy all requirements, we stop. Otherwise, we lengthen the interval by changing $B$, such that the overlap with the previous row increases. In order to keep the total change small, we neglect this approximation and close the row, if there is no overlap at all.
The segmentation resulting from Algorithm \[alg:seg\_ASS\] satisfies the constraints (i)-(iii) and the connectedness, but may contain segments that do not have the minimum total field height $f$.
### Segmentation for ASAS
The segmentation step for ASAS differs a little bit from the ASS segmentation. For ASS, we always find a segment that has its first nonzero row exactly in that row where the current matrix $B$ has its first nonzero row. Going through the rows, we add further ones to the segment if the current sliding window interval overlaps with the previous row.
Computational tests have shown that for ASAS a different technique makes sense because the vertical criterion (iv) plays a role. In detail, whenever we decide for $l_{i+1}<l_i$ (respectively $r_{i+1}>r_i$), this immediately implies $l_{i+k}<l_i$ (respectively $r_{i+k}>r_i$) for $k=2,\dots,g_2$. Additionally, as we need at least $g_1$ consecutive ones in each row, we also know $r_{i+k} \ge \max(l_i-1,l_{i+1}+g_1-1)$ (respectively $l_{i+k} \le \min(r_i+1,r_{i+1}-g_1+1)$). Therefore, we use a matrix $S=(s_{ij})$ to store unavoidable ones, i.e. if $l_{i+1}<l_i$, we put $$\begin{aligned}
\label{eqn:1}
s_{i+k,j}=1 \mbox{ for } 2 \le k \le g_2, \ l_{i+1} \le j \le \max(l_i-1,l_{i+1}+g_1-1)\end{aligned}$$ and if $r_{i+1}>r_i$, we define $$\begin{aligned}
\label{eqn:2}
s_{i+k,j}=1 \mbox{ for } 2 \le k \le g_2, \ \min(r_i+1,r_{i+1}-g_1+1) \le j \le r_{i+1}\end{aligned}$$ As we require consecutive ones in each row, we also put $$\begin{aligned}
\label{eqn:3}
s_{i+k,j}=1 \mbox{ for } 2 \le k \le g_2, \ l_{i+1} \le j \le r_{i+1}\end{aligned}$$ if $l_{i+1}<l_i$ and $r_{i+1}>r_i$ at the same time.
We also have to take care that the covered regions have a vertical minimum size. If $l_{i+1}>l_i$ (respectively $r_{i+1}<r_i$), we analogously put unavoidable zeros into our matrix $S$ using the corresponding rules to (\[eqn:1\]) and (\[eqn:2\]). (\[eqn:3\]) is not necessary here, as zeros do not have to be consecutive in the rows.
Let $g_1=g_2=3$ and let the $(*)$-entries of the matrix above denote the open bixels for row $1$ and $2$. Before choosing the leaf positions for row $3$, there are some unavoidable ones and zeros that have to be respected. $$\begin{pmatrix}
& * & * & * & * & & &\\
* & * & * & * & & & &\\
1 & 1 & 1 & & 0 & & &\\
1 & 1 & 1 & & 0 & & &\\
\end{pmatrix}$$
Thus, the choice of the leaf positions in one row produces unavoidable ones or zeros in other rows. Our algorithm will choose $l_i$ and $r_i$ such that the total change of this row and the corresponding unavoidable ones is minimal. Therefore it might happen, that we do not use the first nonzero row and we also do not use the sliding window technique anymore, as the minimum vertical gap constraint (iv) prohibits so many leaf positions that we can compare the remaining ones with regard to the resulting total change. Algorithm \[alg:seg\_ASAS\] and \[alg:find\_interval\_ASAS\] in the appendix show the corresponding segmentation steps.
The idea behind this algorithm is, that for each segment and in each row we look at all feasible leaf positions. For each pair $(l,r)$, we compute the value $benchmark$, which is the difference between the number of positive entries and the number of nonpositive entries in this row as well as in the corresponding unavoidable ones. The larger this value is, the better the pair $(l,r)$ suits to the segmentation. The unavoidable zeros are not taken into account because it is not necessarily bad if an entry $b_{i,j}>0$ is closed, as this entry can be part of the following segments. We close a row, if the corresponding optimal value of benchmark is zero.
Finally, our procedure leads to a segmentation satisfying (i)-(iv) and the connectedness, but not necessarily satisfies the minimum total field height $f$.
One might argue that it is possible for ASS to compute exactly the segment, that reduces the total change in this step as much as possible. For example, one can consider a layered digraph with $m$ layers of nodes. In layer $i\in [m]$, we have nodes $(i,l,r)$ representing feasible leaf positions and we draw an edge between $(i,l,r)$ and $(i+1,l',r')$, if the combination of these two leaf positions satisfies all constraints. Furthermore, we draw edges from a source to all nodes in the first layer and from all nodes in the last layer to a sink. The edge weights are just the total change reductions caused by the leaf positions of the end node of the edge and zero for all edges whose end node is the sink. For a detailed description of the graph see [@Bol04]. The optimal segment can then be found by shortest path computation in the digraph. But indeed, such a choice is not a good idea because reducing the total change as much as possible leads to badly decomposable residual matrices. For example, for $g_1=2$, ${\boldsymbol{b}}=(1,2,1)$ would be reduced by $(1,1,1)$ and the residuum $(0,1,0)$ is badly decomposable. Our used sliding-window-technique is better and leads to $(1,2,1)=(1,1,0)+(0,1,1)$.
For ASAS the constraints are too complex anyway to compute the optimal segment in one step, that reduces the total change by a maximum value.
Combination of fields
---------------------
Given two segments $S=((l_1,r_1),\dots,(l_m,r_m))$ and $S'=((l'_1,r'_1),\dots,(l'_m,r'_m))$, let us consider the open regions $i_1,\dots,i_2$ and $i'_1,\dots,i'_2$ of $S$ and $S'$, precisely $$\begin{aligned}
\{i_1,\dots,i_2\}=\{i \in [m] \ : \ l_i<n+1\} \label{open1}\\
\{i'_1,\dots,i'_2\}=\{i \in [m] \ : \ l'_i<n+1\} \label{open2}\end{aligned}$$ If $i'_1=i_2+1$ and $\min(r_{i_2},r'_{i'_1})-\max(l_{i_2},l'_{i'_1}) \ge g_1-1$, we merge $S$ and $S'$ and get one new segment $S''$ with $$s''_{ij}=
\begin{cases}
1, & \mbox{ if } s_{ij}=1 \mbox{ or } s'_{ij}=1,\\
0, & \mbox{ otherwise.}
\end{cases}$$
We iterate this step, until no two segments can be merged by this procedure. Obviously, this step does not affect the total change.
Afterwards, we compute a second combination step and merge segments if $i'_1=i_2+2$ and $\min(r_{i_2},r'_{i'_1})-\max(l_{i_2},l'_{i'_1}) \ge g_1-1$. This means, there is only one closed row between the two segments. We compute leaf positions $l \in [\max(l_{i_2},l'_{i'_1}),\min(r_{i_2},r'_{i'_1})-g_1+1]$ and $r \in [l+g_1-1,\min(r_{i_2},r'_{i'_1})]$ such that putting ones to the interval $[l,r]$ in row $i_2+1$ produces the smallest total change with respect to the current approximation matrix.
Again, we drop $S$ and $S'$ out of our segmentation and this time add $S''$ with $$s''_{ij}=
\begin{cases}
1, & \mbox{ if } s_{ij}=1 \mbox{ or } s'_{ij}=1 \mbox{ or } \\
& \quad (i=i_2+1, l \le j \le r),\\
0, & \mbox{ otherwise.}
\end{cases}$$
Again, we iterate this procedure, until no more such merges are possible. Obviously, this second combination step affects the total change, as we increase the approximation matrix by adding ones to the segments. After both of our combination steps, all segments still satisfy (i)-(iii), while even more segments satisfy (v) now. Another positive consequence is a reduction of the total number of (not necessarily pairwise different) segments, called the Delivery Time. The combination step is demonstrated in Figure 3.
For ASAS, we only combine two segments according to one of the two steps described above, if the criterion with the minimum vertical zeros (iv) is not violated after the combination.
\[fig:combination\]
Make-two-of-two
---------------
We now define a substitution step $S+S'=\hat{S}+\hat{S}'$. For ASAS, the same step is computed if $\hat{S}$ and $\hat{S}'$ still satisfy (iv).
Let us again consider a segment $S=((l_1,r_1),\dots,(l_m,r_m))$ and its open region $i_1,\dots,i_2$ defined by (\[open1\]). If $S$ violates (v), we call $S$ a critical segment. Now we check whether we find a segment $S'=((l'_1,r'_1),\dots,(l'_m,r'_m))$ with open region $i'_1,\dots,i'_2$ defined by (\[open2\]) such that $$\begin{aligned}
i_2-i'_1 \ge f-1 \mbox{ and } i'_2-i_1 \ge f-1.\end{aligned}$$ Note that in this case, we have $i'_1<i_1$ and $i'_2>i_2$ due to $i_2-i_1<f-1$. Thus, the set of open rows of $S$ is a subset of the set of open rows of $S'$. If $$\begin{aligned}
\min(r_{i_1},r'_{i_1-1})-\max(l_{i_1}-l'_{i_1-1}) \ge g_1-1, \label{two1}\end{aligned}$$ we substitute $S$ and $S'$ by segments $\hat{S}=((\hat{l}_1,\hat{r}_1),\dots,(\hat{l}_m,\hat{r}_m))$ and $\hat{S}'=((\hat{l}'_1,\hat{r}'_1),\dots,(\hat{l}'_m,\hat{r}'_m))$ defined as follows $$\begin{aligned}
(\hat{l}_i,\hat{r}_i)=
\begin{cases}
(l'_i,r'_i) & \mbox{ if } i < i_1,\\
(l_i,r_i) & \mbox{ if } i \ge i_1,
\end{cases}
\quad
(\hat{l}'_i,\hat{r}'_i)=
\begin{cases}
(l_i,r_i)=(n+1,0) & \mbox{ if } i < i_1,\\
(l'_i,r'_i) & \mbox{ if } i \ge i_1.
\end{cases}\end{aligned}$$
The result is that we add the upper part of segment $S'$ to segment $S$ in order to enlarge $S$, while $S'$ remains sufficiently large. If (\[two1\]) is not satisfied, we can have a second try and check whether $$\begin{aligned}
\min(r_{i_2},r'_{i_2+1})-\max(l_{i_2}-l'_{i_2+1}) \ge g_1-1 \label{two2}.\end{aligned}$$ If this condition is true, we can analogously add the lower part of $S'$ to $S$ and close all rows $\ge i_2+1$ of $S'$. The Make-two-of-two-procedure is illustrated in Figure 4.
\[fig:make\_two\_of\_two\]
The make-two-of-two-step is computed for all critical segments. If we find a partner, we compute the substitution immediately. If no partner is found throughout the segmentation, the segment $S$ is dropped and stored in a new list of critical segments violating (v). Whereas the substitution with a partner does not affect the total change, the elimination of segments from the segmentation that find no partner leads to an increase of the total change.
Handle Critical Segments
------------------------
Let $S=((l_1,r_1),\dots,(l_m,r_m))$ be a critical segment stored in the make-two-of-two-step and let $i_1,\dots,i_2$ be its open region defined by (\[open1\]). We try to combine $S$ with a partner $S'=((l'_1,r'_1),\dots,(l'_m,r'_m))$ from the segmentation with open region $i'_1,\dots,i'_2$ defined by (\[open2\]) if this combination step causes lower total change then simply omit $S$.
More precisely, if $i'_2 \ge i_2$ and $i'_1 \le i_1$, we define a new segment $S''=((l''_1,r''_1),\dots,(l''_m,r''_m))$ with $$(l''_i,r''_i)=
\begin{cases}
(\min(l_i,l'_i),\max(r_i,r'_i)), \mbox{ if } i_1 \le i \le i_2,\\
(l'_i,r'_i), \mbox{ otherwise.}
\end{cases}$$
The resulting segment $S''$ emanates from $S'$ by attaching certain (or all) parts of $S$ and possibly also adding connecting ones. We now drop $S'$ from our segmentation and $S$ from the list of critical segments and add $S''$ to the segmentation, if the total change that is caused by this decision is smaller than simply omitting $S$. If no partner for $S$ is found, we simply delete $S$ from the list of critical segments and accept the corresponding total change.
For ASAS, we execute the step if (iv) is not violated.
\[fig:critical\_segments\]
Total Change Improvement
------------------------
As the previous segmentation steps are indeed not optimal, it may happen, that some leaf positions can be increased (respectively decreased) without violating the constraints and the total change becomes smaller.
Let $B$ be the approximation matrix that corresponds to the current segmentation. The improvement step is computed using Algorithm \[alg:improvement\] which is described in the appendix and adjusts the segmentation such that $B$ gets closer to $A$ with regard to the total change.
Clinical case {#sec:case}
=============
A clinical case was set-up to examine the efficiency of our proposed segmentation algorithms. For a patient with cancer of the right breast, electron irradiation plans using various segmentation settings and different optimization constraints were created with a self-designed IMRT optimization programme based on our previous studies [@Eng05a]. The planning target volume was the right breast, which should receive a total dose of 50.4 Gy (1.8 Gy per fraction). In addition, the target volume should be covered by the $95\%$ isodose line ($95\%$ of the prescribed dose). The ipsilateral lung was considered to be organ at risk.
The optimization programme provides simultaneous optimization of beam orientation, energy and intensity for dose delivery with an add-on MLC for electrons (Euromechanics, Schwarzenbruck, Germany) presented in Figure \[fig:mlc\] and [@Gau06; @Gau08]. Electron dose calculation was performed by Monte Carlo simulations with the treatment planning system *Pinnacle* from Philips (Version 8.1s). Final dose calculation of the treatment plans was conducted using a dose grid size of $3$ mm and a dose calculation uncertainty of $2\%$. The following optimization steps are necessary to generate an electron IMRT plan:
1. Simultaneous optimization of beam orientation, energy and intensity: A set of radiation incidence angles (typically 10–15) is determined given by table and gantry angles [@Eng05a; @Lim07]. For each configuration, the algorithm calculates the optimal fluence distribution, given by a nonnegative integer matrix.
2. The intensity matrices are approximately decomposed into a superposition of allowed segments such that the deviation between desired and actual fluence is minimal. The result is a set of segments, each of them is given by the corresponding MLC leaf positions and its dose weight.
3. The segments from step 2 are treated as candidates for the treatment plan. In a third optimization step, the dose of the candidates are calculated for all beam energies and then optimized for a given weight proportion between best target coverage and minimum dose to critical organs in order to find the final set of segments with optimal beam energies and their corresponding monitor units.
In this paper, we have focused on step 2 and introduced optimization algorithms for an approximate decomposition of intensity matrices.
Until now, the segmentation step consisted in exactly decomposing the intensity matrices using all deliverable segments. In our approach, we admit a decrease in the decomposition accuracy in order to obtain segments which satisfy the dosimetric and technical constraints. Step 3 justifies the approximation approach in Step 2, as a larger approximation error does not necessarily result in a suboptimal treatment plan. Indeed, larger segments produce homogeneous dose distributions and thus, the same final fluence can be generated using fewer larger segments. The acceptability of a treatment plan is decided after step 3 by means of dose volume histograms (see Section \[sec:results\]) and a plan is only presumed if the required dose constraints are not exceeded. Therefore, the danger of cumulative deviation in the approximation step does not really exist, as the computed segments are just candidates for the treatment plan that pass through a further optimization step.
Results {#sec:results}
=======
At first, we compare electron IMRT plans created with different segmentation settings for the clinical case prescribed in Section \[sec:case\]. The comparison was conducted using two different optimization settings: one setting to achieve with a better dose coverage of the breast and the other one to reach a better sparing of the lung. Note, that the segmentation settings refer to parameters in the segmentation step, which is discussed here, whereas the optimization settings play a role in the final optimization step that is not part of this paper. Finally, we give a detailed evaluation for the results of the decomposition step.
A treatment plan with a segmentation setting $xyz$ uses the decomposition algorithm with a minimum total field height of $f=x$, a minimum separation constraint and row overlap of $g_1=y$ and a minimum vertical gap of $g_2=z$. The decomposed matrices vary in their vertical size $m$ and their horizontal size $n$, as they describe only parts of the beam head where the target volume is located. Thus, in practice, the overtravel parameters $b_l$ and $b_r$ will depend on the positioning of the matrix and are put individually for each matrix. Our electron MLC is capable of shifting the leaves edges to $3/4$ of the radiation field.
The plan quality was evaluated by means of dose volume histograms that indicate the amount of dose delivered to a certain volume of the patient (here: the right breast and the right lung). Thus, dose homogeneity in the target volume and dose exposure to the organs at risk can be examined. In Figure \[fig:dvh\], the dose volume histograms for both optimization settings demonstrate that almost identical dose distributions can be achieved using smaller or larger minimum MLC openings (cf. setting $111$ and $441$). In fact, the treatment plan could be slightly improved by use of a minimum vertical gap parameter of $2$ which avoids single leaf openings and closings.
Table \[table1\] and \[table2\] illustrate the main benefit of our approach, as identical results can be achieved with approximately two thirds fewer segments and a significantly smaller number of monitor units by use of dosimetric constraints. As a result, the dose delivery is more efficient and less time consuming. The minimum number of segments is reached for setting $442$ and computational tests have shown, that this configuration produces the optimal results. As the leaf width is $0.7$ cm, fields with a horizontal and vertical height of $4$ bixels have a size of approximately $3$ cm $\times$ $3$ cm and this confirms our dosimetric constraint of $3$ cm $\times$ $3$ cm minimum segment size (cf. Figure \[fig:properties\]). It can be also demonstrated that minimum segment sizes greater than setting $442$ do not necessarily result in fewer segments (cf. Table \[table1\]), although the number of segments in Table \[table2\] is slightly lower for setting $552$. For both optimization settings, the dose volume histograms were considerably better when using minimum segment sizes smaller than setting $552$.
\[fig:dose\]
It is important to underline that the number of segments and the number of monitor units in Table \[table1\] and \[table2\] belong to the final IMRT plan and result from the third optimization step and not from the decomposition step of our algorithms. In fact, the monitor units have another scale here and are not directly comparable with the delivery time from the segmentation. In contrast, the total change information stems from the decomposition step. Note, that the total change of the segmentation itself is not a significant quantity, because if the matrix entries are large, a larger total change is acceptable. Therefore, we compute the total sum of entries for each intensity matrix and then calculate the relative total change which is the ratio between total change and total sum of entries. The smaller the relative total change, the better is the decomposition.
0.2cm
[@l\*[15]{}[@l]{}]{} Setting &Mean Relative Total Change &Number of Segments &Number of Monitor Units\
111 & 0.04 &90 &85586\
221 & 0.16 &79 &65677\
222 & 0.34 &54 &63461\
331 & 0.22 &49 &38598\
332 & 0.36 &32 &21265\
333 & 0.40 &40 &36789\
441 & 0.30 &55 &34262\
442 & 0.41 &26 &12860\
443 & 0.45 &28 &13119\
444 & 0.49 &27 &16899\
551 & 0.35 &40 &19162\
552 & 0.46 &30 &10337\
0.2cm
[@l\*[15]{}[@l]{}]{} Setting &Mean Relative Total Change &Number of Segments &Number of Monitor Units\
111 & 0.04 &78 & 59399\
221 & 0.16 &71 & 46627\
222 & 0.34 &51 & 42411\
331 & 0.22 &52 & 25896\
332 & 0.36 &29 & 16700\
333 & 0.40 &36 & 24957\
441 & 0.30 &45 & 23509\
442 & 0.41 &28 & 10200\
443 & 0.45 &28 & 9995\
444 & 0.49 &27 & 11318\
551 & 0.35 &37 & 13787\
552 & 0.46 &25 & 8832\
For the detailed evaluation of our algorithms, we use a set of $264$ clinical intensity matrices that originate from electron treatment plans for different patients and beam angles. The matrices are produced during the optimization step 1 of the treatment planning that was introduced in Section \[sec:case\] and uses the algorithm from [@Eng05a]. Exemplarily, we compute segmentations for the settings $f=3$, $g_1=3$ and $g_2=1$ as well as $f=4$ and $g_1=g_2=2$. The values of the overtravel parameters are also produced in the pre-segmentation step. The results are shown in Table \[table3\] and \[table4\] and demonstrate how much total change is caused respectively avoided by the steps of the algorithms. The overtravel-approximation and MSC-approximation lead to a certain total change of the matrix that is put into the segmentation step. As an exact decomposition in the segmentation step is impossible, the total change increases here again. Both the combination step and the make-two-of-two step try to eliminate segments not satisfying the parameter $f$ and again cause some total change. Finally, the last two steps of the algorithm improve the performance and reduce the approximation error as much as possible. One can see that the combination step and the make-two-of-two-step are performed more often for ASS, as for ASAS the vertical gap ensures that the fields already have a reasonable size after the segmentation step. Of course, the larger the parameters and thus the minimum field size, the larger becomes the total change.
The first column in Table \[table3\] and \[table4\] gives the average results over the $264$ matrices, while the second (respectively third) column represent the single results for the matrix with the smallest (respectively largest) relative total change. Homogeneous matrices with large nonzero areas can be decomposed quite well, while matrices with only few nonzero entries that do not span connected areas lead to unacceptable results. As a treatment plan is a superposition of several intensity profiles from different beam angles, the approximation errors balance each other and lead to applicable treatment plans as described above. Furthermore, a certain part of the total change is unavoidable if one requires the constraints (i)-(v), e.g. the total change after MSC-approximation is a good lower bound for the achievable total change. All in all, taking the vertical gap $g_2$ into account increases the total change while reducing the number of used segments and the monitor units.
0.2cm
[@l\*[15]{}[@l]{}]{} &Average &Rel. TC min. &Rel. TC max.\
m &17.55 &20 &11\
n &21.44 &19 &13\
Total sum of entries &886.4 &1329 &152\
Total change &82.24 &12 &64\
Delivery time &15.39 &13 &3\
Number of segments &15 &13 &3\
TC after Overtravel-Approximation &26.34 &6 &75\
TC after MSC-Approximation &57.09 &11 &88\
TC after Segmentation &74.33 &13 &91\
TC change after combination &80.96 &28 &91\
TC after make-two-of-two &93.77 &28 &101\
TC after handle critical segments &88.72 &28 &94\
TC after improvement &82.24 &12 &94\
Combinations &9.19 &5 &0\
Successful make-two-of-two &2.78 &0 &0\
Relative total change &0.16 &0.009 &0.62\
0.2cm
[@l\*[15]{}[@l]{}]{} &Average &Rel. TC min. &Rel. TC max.\
m &17.55 &23 &22\
n &21.44 &26 &28\
Total sum of entries &886.4 &997 &384\
Total change &157.7 &46 &269\
Delivery time &7.69 &4 &4\
Number of segments &6.82 &3 &4\
TC after Overtravel-Approximation &26.34 &7 &89\
TC after MSC-Approximation &44.39 &7 &134\
TC after Segmentation &144.8 &46 &267\
TC change after combination &144.8 &46 &267\
TC after make-two-of-two &159.2 &46 &270\
TC after handle critical segments &158.8 &46 &270\
TC after improvement &157.7 &46 &269\
Combinations &0.24 &0 &0\
Successful make-two-of-two &0.57 &0 &0\
Relative total change &0.27 &0.05 &0.70\
Conclusion
==========
In the present study, dosimetric and technical constraints have been taken into consideration in intensity-modulated radiation therapy (IMRT). A set of $5$ parameters has been introduced, two of them for modelling the leaf overtravel constraint, the other three to ensure a minimum field size and to avoid thin field shapes. We proposed algorithms for approximate segmentation of intensity matrices using segments that satisfy the constraints. We basically distinguish between two approximation problems depending on whether the vertical gap parameter is considered or not. The objective function of the optimization is the deviation between the desired and the approximated intensity profile that has to be minimized. The segmentation step is part of an IMRT optimization process which was examined by comparisons of dose volume histograms of treatment plans with small and large segments as well as with and without thin segment shapes. The histograms show that the use of larger segments results in equal IMRT plans with fewer segments and monitor units respectively. Although the approximation error of the segmentations rises with increasing minimum field size, equivalent or even better dose distributions could be achieved. Concluding, this first approach to approximated segmentation in IMRT planning shows the potential of these ideas and there is a need for further research in related approximation problems.
Appendix
========
vector ${\boldsymbol{v}}$ $tc_{i1}=|v_i-i|$ $tc_{ij}=\infty$ for all $i \in [min,max],j>1$ $tc_{i'j}=tc_{i,j-1}+|v_j-i'|$ $pre_{i'j}=i$ $opt=\min_{i \in [min,max]} tc_{i,k}$ Let $i_{opt}$ be one of the indices with $tc_{i_{opt},k}=opt$. $w_j=i_{opt}$ If still $j>1$, then $i_{opt}=pre_{i_{opt},j}$ vector ${\boldsymbol{w}}$
matrix $B$ $i=1;$ $l_i=n+1$; $r_i=0$; $i=i+1$; Choose $l_i$ and $r_i$ for row $i$ such that $r_i-l_i \ge g_1-1$ (see Remark 2). : Choose $l_{i+1}$ and $r_{i+1}$ for row $i+1$ such that the overlap with the previous row is $\ge g_1$ or close row $i+1$. $i=i+1$; Close all remaining rows with $l_i=n+1$ and $r_i=0$. $b_{ij}=b_{ij}-1$; Store $((l_1,r_1),\dots,(l_m,r_m))$ in the segmentation. Segmentation
$i, l_i, r_i$ $finish=0;$ $start=\min \{ j \ | \ b_{i+1,j}>0 \}$; $end=\min \{ j \ | \ b_{i+1,j}>b_{i+1,j+1} \}$; $l=\max(start,l_i)$; $r=\min(end,r_i)$; $l_{i+1}=start$; $r_{i+1}=end$; $finish=1$; $l_{i+1}=n+1$; $r_{i+1}=0$; $finish=1$; $b_{i+1,start-1}=b_{i+1,start-1}+1$; $b_{i+1,end+1}=b_{i+1,end+1}+1$; $l_{i+1}$, $r_{i+1}$
The interval $[l,r]$ for the first open row is computed analogously to the interval for the other rows, only ignoring the overlap constraint and instead requiring $r-l>g_1-1$.
matrix $B$ $i=1$; $s_{ij}=0$ for all $i$ and $j$; $l_i=n+1$; $r_i=0$; $i=i+1$; $start=i$; $closed=0;$ $closed=1$; Choose $l_i$ and $r_i$ for row $i$ such that $r_i-l_i \ge g_1-1$ or close row $start$ (see Remark 3). $closed=1$; : Choose $l_{i+1}$ and $r_{i+1}$ for row $i+1$ such that the overlap with the previous row is larger than $g_1$ or close row $i+1$.\
$i=i+1$; $l_i=$index of the first one in ${\boldsymbol{s_i}}$; $r_i=$index of the last one in ${\boldsymbol{s_i}}$; $l_i=n+1$, $r_i=0$; ${\boldsymbol{b_{start}}}={\boldsymbol{0}}$; $b_{ij}=b_{ij}-1$; Store $((l_1,r_1),\dots,(l_m,r_m))$ in the segmentation. Segmentation
$i+1, l_i, r_i$, current matrix $S$ $v_1=n+1$; $v_2=0$; $v_1=$ index of the first one in ${\boldsymbol{s_i}}$; $v_2=$ index of the last one in ${\boldsymbol{s_i}}$; $t_1=$ index of the first positive entry in ${\boldsymbol{b_i}}$; $t_2=$ index of the last positive entry in ${\boldsymbol{b_i}}$; $opt=0$; $l_{i+1}=n+1$; $r_{i+1}=0$; $benchmark=0$; $benchmark=benchmark-1$; $benchmark=benchmark+1$; $benchmark=benchmark-1$; $benchmark=benchmark+1$; $opt=benchmark$; $l_{i+1}=l$; $r_{i+1}=r$; Put unavoidable ones or zeros corresponding to the decision for $l_{i+1}$ and $r_{i+1}$ into $S$. $l_{i+1}$, $r_{i+1}$, current matrix $S$
The choice of $(l,r)$ in row $start$ is again computed analogously, only ignoring the overlap constraint with the previous row.
Segmentation $l_i=l_i-1;$ $r_i=r_i+1;$ $l_i=l_i+1;$ $r_i=r_i-1;$ Update approximation matrix Segmentation
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
A universal description of the hyperfine splittings (HFS) in bottomonium and the $B_q\; (q=n,s,c)$ mesons is obtained with a universal strong coupling constant $\alpha_s(\mu)=0.305(2)$ in a spin-spin potential. Other characteristics are calculated within the Field Correlator Method, taking the freezing value of the strong coupling independent of $n_f$. The HFS $M(B^*)-
M(B)=45.3(3)$ MeV, $M(B_s^*) - M(B_s)=46.5(3)$ MeV are obtained in full agreement with experiment both for $n_f=3$ and $n_f=4$. In bottomonium, $M(\Upsilon(9460))- M(\eta_b)=70.0(4)$ MeV for $n_f=5$ agrees with the BaBar data, while a smaller HFS, equal to 64(1) MeV, is obtained for $n_f=4$. We predict HFS $M(\Upsilon(2S))-M(\eta_b(2S))=36(1)$ MeV, $M(\Upsilon(3S))-
M(\eta(3S))=27(1)$ MeV, and $M(B_c^*) - M(B_c)= 57.5(10)$ MeV, which gives $M(B_c^*)=6334(1)$ MeV, $M(B_c(2\,{}^1S_0))=6865(5)$ MeV, and $M(B_c^*(2S\,{}^3S_1))=6901(5)$ MeV.
author:
-
-
-
title: 'The Hyperfine Splittings in Bottomonium and the $B_q (q=n,s,c)$ Mesons'
---
Introduction
============
Recently, $\eta_b(1S)$ has been discovered by the BaBar collaboration in the radiative decays $\Upsilon(3S)\rightarrow \gamma \eta_b(1S)$ [@ref.1] and $\Upsilon(2S)\rightarrow \gamma \eta_b(1S)$ [@ref.2], with a mass (averaged over two results) $M(\eta_b)=9391.1\pm 3.1$ MeV. It gives a rather large hyperfine splitting (HFS), $\Delta(b\bar b)=
M(\Upsilon(1S))-M(\eta_b(1S))=69.9\pm~3.1$ MeV. Later this mass was confirmed by the CLEO collaboration also in the radiative $\Upsilon(3S)\rightarrow \gamma\eta_b(1S)$ decay [@ref.3]. This important new information allows to test again our understanding of the hyperfine (HF) interaction in QCD.
Although a spin-spin potential between heavy quarks was used in numerous studies, the parameters defining this potential significantly differ in different models. As a result, theoretical predictions for the mass difference $\Delta(b\bar b)= M(\Upsilon(9460))-M(\eta_b(1S))$ vary in a wide range: $35-90$ MeV [@ref.4]–[@ref.9] and in most cases they are smaller than the experimental number. On a fundamental level the spin-spin potential $V_{ss}({\rm lat})$ has been studied in detail in quenched QCD (see [@ref.10] and references therein). This lattice potential appears to be compatible with zero at distances $r\geq
0.30$ fm and (for unknown reasons) has negative sign at smaller $r$ (with a large magnitude); in any case the lattice potential does not contradict the Fermi-Breit potential with $\delta^3({\mbox{\boldmath${\it r}$}})$ [@ref.11], although the behavior of the spin-spin potential at $r\leq 0.3$ fm remains uncertain.
On the other hand, a detailed phenomenological analysis given in Ref. [@ref.4] has demonstrated the importance of the smearing of the $\delta^3({\mbox{\boldmath${\it r}$}})$-function, from which one may expect that for heavy mesons, containing a $b$-quark, the use of $\delta^3({\mbox{\boldmath${\it r}$}})$ may be a good approximation. For lighter mesons, like $D$, $D_s$, and charmonium, a nonperturbative spin-spin potential may be essential, giving a contribution $\sim 10\%$ [@ref.12]. Here we concentrate on bottomonium and the $B_q~(q=n,s,c)$ mesons, for which nonperturbative contributions are small, and neglect the smearing effect, in this way avoiding to introduce several unknown parameters.
Our main goal here is the extraction of the strong coupling $\alpha_{\rm hf}(\mu)$ from known HFS. In theoretical models two typical choices of $\alpha_{\rm hf}$ are used:
1. In the first one, “a universal" $\alpha_{\rm hf}$ is used. For example, in Ref. [@ref.5] $\alpha_{\rm hf}=0.36$ taken, was obtained from a fit to the mass difference $M(J/\psi)-
M(\eta_c(1S))=117$ MeV, but their HFS, $M(\Upsilon(9460))-
M(\eta_b) = 87$ MeV, is $\sim 25\%$ larger than the experimental number. In Ref. [@ref.9], using a smaller $\alpha_{\rm
hf}=0.339$ a good description of the HFS of the $B$ and $B_s$ mesons was obtained. However, a comparison of their and our results is difficult, because a large string tension, $\sigma=0.257$ GeV$^2$, was taken in [@ref.9], while here and in Ref. [@ref.4] the conventional value $\sigma=0.18$ GeV$^2$ is used.
2. The second choice, with a scale $\mu$ dependent on the quark mass, is mostly used in pQCD, where $\alpha_{\rm hf}(m_b)\sim 0.18$ and $\alpha_{\rm hf}(m_c)\sim 0.26$. Just due to such a small value of $\alpha_{\rm hf}(m_b)$, taken in bottomonium, small HFS were obtained in Ref. [@ref.12], although their wave functions (w.f.) at the origin gave excellent descriptions of the dielectron widths for $\Upsilon(nS)
(n=1,2,3)$ [@ref.13].
Here we use instead of the Fermi-Breit potential a spin-spin potential derived using the Field Correlator Method (FCM) [@ref.14], where relativistic corrections are taken into account and with the mass of a light quark $m_n= 5$ MeV $(n=u,d)$ and $m_s=200$ MeV for an $s$-quark, the $B, B_s$ mesons can be considered on the same footing as the $B_c$ mesons.
It can be shown that the HFS are sensitive to the value of $\Lambda_{\overline{\rm MS}}(n_f)$ taken. Since $\Lambda_{\overline{\rm MS}}$ is known only for $n_f=5$ and $\Lambda_{\overline{\rm MS}}$, used for $n_f=3,4$, varies in a wide range, we make here the assumption, already used in [@ref.4], that the freezing value of the vector coupling constant (denoted as $\alpha_{\rm crit}(n_f)$) is the same for $n_f=3,4,5$.
Then it appears possible to obtain a good description of the HFS for the $B_q$ mesons ($q=n,s,c$) and bottomonium, taking a universal $\alpha_s(\mu)=0.305(2)$ (with small one-loop corections); its value is smaller than in Refs. [@ref.5] and [@ref.9].
We also predict the HFS of the as yet undiscovered $\eta_b(2S)$ and $\eta_b(3S)$, and the mass of $B_c^*$.
The HF potential in the Field Correlator Method
===============================================
The Fermi-Breit potential, $$\hat V_{ss}(r) ={\mbox{\boldmath${\it s}$}}_1 \cdot {\mbox{\boldmath${\it s}$}}_2 \frac{32\pi}{9}\,
\frac{\alpha_{\rm hf} (\mu)}{\tilde m_1 \tilde m_2}\, \delta^3 ({\mbox{\boldmath${\it r}$}}),
\label{eq.01}$$ widely used in heavy quarkonia, contains the constituent quark masses $\tilde m_1$ and $\tilde m_2$, which are very much model-dependent.
In Eq. (\[eq.01\]) the strong coupling constant $\alpha_{\rm hf}(\mu)$ may differ from $\alpha_s(\mu)$ (in the $\overline{\rm MS}$ renormalization scheme) due to higher order perturbative corrections. Here we take the one-loop corrections from [@ref.15], $$\alpha_{\rm hf} (\mu) =\alpha_s (\mu) \left[ 1+ \frac{\alpha_s
(\mu)}{\pi} \rho (n_f)\right],
\label{eq.02}$$ where $$\rho= \frac{5}{12}\beta_0 - \frac{8}{3} - \frac{3}{4}\ln2.
\label{eq.02a}$$
Although higher corrections are small: $\sim 6\%$ for $n_f=3$, $\sim 3\%$ for $n_f=4$, and $\leq 0.1\%$ for $n_f=5$, still they are needed to improve the accuracy and have a more self-consistent picture.
The important role of relativistic corrections, even for the $B_c$ meson, has already been underlined in Refs. [@ref.4] and [@ref.9], and recently in the lattice calculations of the $B_c^*$ mass [@ref.16], [@ref.17]. We take them into account using the spin-spin potential (without smearing), derived in the FCM [@ref.14], [@ref.18]: $$\hat V_{ss}(r) ={\mbox{\boldmath${\it s}$}}_1 \cdot {\mbox{\boldmath${\it s}$}}_2 \frac{32\pi}{9}\,
\frac{\alpha_{\rm hf} (\mu)}{\omega_1\omega_2}\, \delta^3 ({\mbox{\boldmath${\it r}$}}),
\label{eq.03}$$ for which the HFS is $$\Delta_{\rm hf} (nS) =
\frac{8}{9} \frac{\alpha_{\rm hf}(\mu)}{\omega_1\omega_2}|R_n(0)|^2.
\label{eq.04}$$ In Eqs. (\[eq.03\]) and (\[eq.04\]) the variables $\omega_1(nS),
\omega_2(nS)$ are the averaged kinetic energies of a quark 1 and a antiquark 2, which play a role of the dynamical masses: $$\omega_1(nS) = \langle \sqrt{{\mbox{\boldmath${\it p}$}}^2 + m_1^2} \rangle_{nS}, \quad
\omega_2(nS) = \langle \sqrt{{\mbox{\boldmath${\it p}$}}^2 +m_2^2} \rangle_{nS}
\label{eq.05}.$$ The important point is that in Eq. (\[eq.05\]) the masses $m_1$ and $m_2$ are well defined; they are the pole masses of $c$ and $b$ quarks (known now with an accuracy of $\sim 70$ MeV for a $b$ quark and $\sim 100$ MeV for a $c$ quark [@ref.19]). In leading order, the pole mass does not depend on the number of flavors, while to order ($\alpha_s(\bar m_Q)^2$) it slightly depends on $n_f$ . We take here $m_1= m_n=5$ MeV for a light quark $(n=u,d)$; $m_s=200$ MeV for an $s$ quark; the pole mass $m_c=1.41$ GeV; $m_b=4.79$ GeV for $n_f=3$ and $m_b=4.82$ GeV for $n_f=4,5$.
The quantities $\omega_i$ and the w.f. are calculated with the use of the relativistic string Hamiltonian (RSH), also derived in the FCM [@ref.20], $$H_0=\frac{\omega_1}{2} +\frac{\omega_2}{2} +\frac{m^2_1}{2\omega_1}+
\frac{m^2_2}{2\omega_2} +\frac{{\mbox{\boldmath${\it p}$}}^2}{2\omega_{\rm red}} +V_{\rm B}(r),
\label{eq.06}$$ where the variables $\omega_i$ enter as the kinetic energy operators. However, if one uses an einbein approximation [@ref.18], [@ref.21] and considers the spin-dependent potential as a perturbation, then $\omega_i$ should be replaced by its matrix elements (m.e.) (\[eq.05\]).
A simple expression for the spin-averaged mass $M(nS)$ follows from the RSH [@ref.21]: $$M(ns)=\frac{\omega_{1}}{2}+\frac{m_1^2}{2\omega_1}+\frac{\omega_{2}}{2}
+\frac{m_2^2}{2\omega_1} + E_{nS}(\omega_{\rm red}).
\label{eq.07}$$ Here, the excitation energy $E_{nS}(\omega_{\rm red})$ depends on the reduced mass: $\omega_{\rm red}=\frac{\omega_1\omega_2}{\omega_1
+\omega_2}$. The mass formula (\[eq.07\]) does not contain any additive constant in the case of bottomonium, while for the $B$ and $B_s$ mesons a negative (not small) self-energy term, proportional to $(\omega_q)^{-1}$ ($q=n,s$), has to be added to their masses [@ref.22].
Then the variables $\omega_i(nS)$, the excitation energy $E_{nS}(\omega_{\rm red})$, and the w.f. are calculated from the Hamiltonian (\[eq.06\]) and two extremum conditions, $\partial\,M(nS)/\partial\omega_i=0~(i=1,2)$, which are put on the mass $M(nS)$ [@ref.18]: $$\begin{aligned}
\nonumber
&& \left[\frac{\omega_1}{2}+\frac{\omega_2}{2}
+\frac{m_1^2}{2\omega_1} +
\frac{m^2_2}{2\omega_2}+
\frac{\bm{p}^2}{2\omega_{\rm red}}+V_{\rm B}(r)
\right]\varphi_{nS}(r)\\
\label{eq.08}&&= E(nS)\varphi_{nS},\end{aligned}$$ $$\omega_i^2(nS) = m_i^2 -
\frac{\partial E(nS, \mu_{\rm red})}{\partial \omega_i(nS)},\; (i=1,2).
\label{eq.09}$$
In a Hamiltonian approach the choice of the static potential $V_{\rm
B}(r)$ is of great importance; we take it as a sum of a linear confining term and the OGE -type term: this property of additivity is well established now in analytical studies [@ref.14] and on the lattice [@ref.23], [@ref.24]: $$V_{\rm B}(r)=\sigma\ r +\frac{4\alpha_{\rm B}(r)}{3\ r}.
\label{eq.10}$$
For the string tension a conventional value, $\sigma=0.18$ GeV$^2$, is used here for all mesons.
The main uncertainty comes from the vector coupling $\alpha_V(r)$, which is taken here from Refs. [@ref.25], [@ref.26] and denoted as $\alpha_{\rm B}(r)$. Two important conditions have to be put on the vector coupling:
\(i) As in pQCD, it must possess the property of asymptotic freedom (AF); precisely owing to this property the static interaction depends on the number of flavors.
\(ii) The vector coupling freezes at large distances. The property of freezing was widely used in phenomenology [@ref.4]-[@ref.7], [@ref.27] and observed in lattice calculations of the static potential [@ref.23], [@ref.24].\
Unfortunately, one cannot use the static potential and the freezing (critical) constant from lattice studies, where the latter is found to be significantly smaller than in phenomenology and background perturbation theory (BPT). There $\alpha_{\rm
B}(\rm crit)=0.58-0.60$ is used (these numbers are close to the value from [@ref.4]). On the lattice, $\alpha_{\rm crit}({\rm
lat})\sim 0.30$ in full QCD ($n_f=3$) [@ref.24] and $\alpha_{\rm crit}({\rm lat})\sim 0.22$ in quenched calculations [@ref.10], [@ref.23] were obtained.
In Eq. (\[eq.10\]) the vector coupling $\alpha_{\rm B}(r)$ is defined via the vector coupling $\alpha_{\rm B}(q^2)$ in momentum space [@ref.26]: $$\alpha_{\rm B}(r) =
\frac{2}{\pi}\int\limits_0^\infty dq\frac{\sin(qr)}{q}\alpha_{\rm B}(q),
\label{eq.11}$$ where the vector coupling $\alpha_{\rm B}(q^2)$ is taken in two-loop approximation, $$\alpha_{\rm B}(q) =\frac{4\pi}{\beta_0t_{\rm B}}\left(1-\frac{\beta_1}{\beta_0^2}
\frac{\ln t_{\rm B}}{t_{\rm B}}\right)
\label{eq.12}$$ with the logarithm $$t_{\rm B}=\frac{q^2+M_{\rm B}^2}{\Lambda_{\rm B}^2},
\label{eq.13}$$ containing the vector constant $\Lambda_{\rm B}(n_f)$, which differs from the QCD constant $\Lambda_{\overline{\rm MS}}(n_f)$. The relation between them has been established in Ref. [@ref.28]: $$\Lambda_{\rm B}(n_f)=\Lambda_{\overline{\rm MS}}\exp\left(-\frac{a_1}{2\beta_0}\right),
\label{eq.14}$$ with $\beta_0=11 -\frac{2}{3}n_f$ and $a_1=\frac{31}{3}-\frac{10}{9}n_f$. Therefore the constant $\Lambda_{\rm B}$ is always larger than $\Lambda_{\overline{\rm MS}}$, $$\begin{aligned}
\nonumber
\Lambda_{\rm B}^{(5)}&=&1.3656\,\Lambda_{\overline{\rm MS}}^5
(n_f=5),\;\\ \label{eq.15}
\Lambda_{\rm B}^{(4)}&=&1.4238\,\Lambda_{\overline{\rm MS}}^4
(n_f=4),\;\\ \nonumber
\Lambda_{\rm B}^{(3)}&=&1.4753\,\Lambda_{\overline{\rm MS}}^3 (n_f=3).\end{aligned}$$
At present, only the QCD constant $\Lambda_{\overline{\rm MS}}(n_f=5)$ is known with a good accuracy, while for $n_f=3,4$ it is defined with an accuracy $\sim 10\%$ [@ref.19]). For a given $\Lambda_{\overline{\rm
MS}}(n_f=5)$ one can define $\alpha_{\rm crit}(n_f=5)$. Then, assuming that the freezing constant is the same for $n_f=3,4$, the QCD constant, $\Lambda_{\overline{\rm MS}}$ for $n_f=3,4$, can be calculated.
The mass $M_{\rm B}$ under the logarithm $t_{\rm B}$ is proportional to $\sqrt\sigma$ and for $\sigma=0.18$ GeV${}^2$ its value is equal to $M_{\rm B}=1.0\pm 0.05$ GeV [@ref.29].
The critical constants $\alpha_{\rm B}(\rm crit)=\alpha_{\rm B}(q^2=0)$ in momentum space and $\alpha_{\rm B}(r\rightarrow \infty)$ in coordinate space are proved to be equal [@ref.26].
Our calculations give small relativistic corrections for bottomonium: $\omega_b(1S) - m_b \sim 190$ MeV ($\sim 4\%$) and $\sim 7\%$ for the $2S$ and $3S$ states. It is of interest to notice that the relativistic correction to the $b-$quark mass is even smaller in the $B_c$ meson, $\omega_b(1S)
-m_b\sim 83$ MeV ($\sim 2\%$), while for a $c-$quark the difference $\omega_c(1S)- m_c \sim 250$ MeV is already $\sim 20\%$. The values of $\omega_q(1S)-m_q$ are given in Table \[tab.1\] for the $B_q$ mesons ($q=n,s,c$).
Meson $ B$ $B_s$ $B_c$
------------------ ------- ------- -------
$ m_q $ 5 200 1410
$\omega_q -m_q$ 611 489 248
$\omega_b$ 4805 4825 4888
$\omega_b -m_b $ 25 25 83
: \[tab.1\]The kinetic energies $\omega_q(1S)~(q=n,s,c)$ and $\omega_b(1S)$ (in MeV) for the static potential $V_{\rm B}(r)$ (\[eq.10\]) with $n_f=4$ and $\alpha_{\rm crit}=0.58$
It is convenient to introduce the ratio $g_{B_q}$, $$g_{B_q}(nS) = \frac{|R_n(0)|^2}{\omega_1(nS)\omega_2(nS)},
\label{eq.16}$$ which directly enters the HFS (\[eq.04\]) and appears to be weakly dependent on small variations of the masses $m_1$ and $m_2$, which are compatible with a good description of the meson spectrum.
The w.f. at the origin are sensitive to the values of $\Lambda_{\rm B}(n_f)$ (in [@ref.30] only the case with $n_f=3$ was considered). However, if the same freezing value of the coupling constant is taken for $n_f=3,4,5$, then the differences between the w.f. at the origin for $n_f=3$ and $n_f=4$ are $\leq 3\%$.
We use here two values for $\alpha_{\rm crit}$: $\alpha_{\rm crit}=0.580$ and 0.604, for which corresponding values of $\Lambda_{\rm B}(n_f)$, $\Lambda_{\overline{\rm MS}}(n_f)$ are given below in Eq. (\[eq.17\]).
$ B $ $ B_s $ $ B_c$
------------ ----------- ----------- -----------
$|R(0)|^2$ 0.477 0.551 1.669
$g_{B_q}$ 0.161 0.166 0.205
: \[tab.2\] The ratios $g_{B_q}$ (\[eq.09\]) (in GeV) and $|R_1(0)|^2$ (in GeV${}^3$) for the $B_q(1S)$ mesons ($\alpha_{\rm crit}=0.58, n_f=4$)
In bottomonium the difference between $g_b$ for $n_f=4$ and $n_f=5$ appears to be larger, $\sim 10\%$ (see Table \[tab.3\]), where in both cases $\alpha_{\rm crit}=0.604$ is used.
1S 2S 3S
------------------------- ------- --------------- --------------
$|R_n(0)|^2 (n_f=5)$ 6.476 3.398 () 2.682()
$g_b (n_f=5) $ 0.258 0.134 (1) 0.105(1)
$|R_n(0)|^2 (n_f=4)$ 5.668 3.126 2.508
$ g_b (n_f=4) $ 0.230 0.127 0.100
$ \Delta_{hf}(n_f=5) $ 70.3 36.3 28.4
: \[tab.3\] The ratios $g_b(nS)$ (in GeV) and $|R_n(0)|^2$ (in GeV${}^3$) for the $1S$, $2S$, and $3S$ bottomonium states with $\alpha_{\rm crit}=0.604$ and for $n_f=4,5$
For the values of $\alpha_{\rm crit}(n_f)=0.604 (0.58)$ and $\Lambda_{\overline{\rm MS}}(n_f=5)=0.245 (0.236)$ GeV, the following vector constants: $\Lambda_{\rm B}(n_f=3)=0.40 (0.389)$ GeV; $\Lambda_{\rm
B}(n_f=4)=0.372 (0.360)$ GeV, $\Lambda_{\rm B}(n_f=5)=0.335 (0.323)$ GeV are obtained. Then from the relation (\[eq.14\]) we have $$\begin{aligned}
\nonumber
\Lambda_{\overline{\rm MS}}(n_f=3)&=&271 (264)~{\rm MeV},\;\\
\Lambda_{\overline{\rm MS}}(n_f=4)&=&261 (253)~{\rm MeV},\;\\ \nonumber
\lambda_{\overline{\rm MS}}(n_f=5)&=&245 (236)~{\rm MeV}.
\label{eq.17}\end{aligned}$$ For $n_f=5$ it gives $\alpha_s(M_Z)({\rm two-loop})=0.1194$ for $\alpha_{\rm crit}=0.604$ and $\alpha_s(M_Z)=0.1188$ for $\alpha_{\rm
crit}=0.58$; both numbers agree with the world averaged value, $\alpha_s(M_Z)=0.1176\pm 0.0020$ [@ref.19] within its error bar.
Results
=======
The experimental error in the HFS $$\Delta_{\rm hf}(b\bar b) =M(\Upsilon(9460))- M(\eta_b)=69.9\pm 3.1~{\rm MeV}
\label{eq.18}$$ is small, $\pm 3$ MeV, and it is even smaller, $\leq 1$ MeV, for the mass differences $M(B^*)-M(B)$, $M(B_s^*)-M(B_s)$ [@ref.19]. Therefore, we expect that the coupling constant $\alpha_{\rm hf}$ can be extracted with a good accuracy from these data.
Notice that from Eq. (\[eq.02\]) and the value $\alpha_s(\mu)=0.305(2)$ one obtains the following coupling constants with one-loop corrections: $\alpha_{\rm hf}(n_f=5)=\alpha_s(\mu)=0.305(2)$, $\alpha_{\rm hf}(n_f=4)=
0.314(2)$, and $\alpha_{\rm hf}(n_f=3)= 0.323(2)$. Precisely these values are used in our analysis.
For $g_b=0.230$ GeV ($n_f=4$) and $g_b=0.258$ GeV ($n_f=5$) (see Table \[tab.3\]) and taking $\alpha_{\rm
hf}(n_f=5)=\alpha_s(\mu)=0.305(2)$ and $\alpha_{\rm hf}(n_f=4)=0.314(2)$, we obtain $\Delta_{\rm hf}(b\bar b)= 64.2(4)$ MeV ($n_f=4$) and 70.0(4) MeV for $n_f=5$. The difference between them, $\sim 10\%$, is not small and one may conclude that the HFS in bottomonium is in full agreement with the BaBar data [@ref.1], [@ref.2] only for $n_f=5$.
For the $2S$ and $3S$ bottomonium states the difference between the cases with $n_f=4$ and $n_f=5$ is small; their HFS coincide within 2 MeV. Therefore we predict that $\Delta(b\bar b)$ is equal to 36(2) MeV and 27(2) MeV, respectively, for the $3S$ and $2S$ states.
For the $B_q$ mesons, the cases with $n_f=3,4$ have been considered and in both cases agreement with experiment is obtained (see Table \[tab.4\]); it happens because the by $\sim 3\%$ smaller value of $g_{B_q}$ is compensated by the $\sim 3\%$ larger value of $\alpha_{\rm hf}(n_f=3)$. Thus for $B$ and $B_s$ the number of flavors cannot be fixed, if only data on the HFS (and the spectrum) are fitted; therefore some additional information, like decay constants, is neededd to fix $n_f$.
$B$ $B_s$ $B_c(1S)$ $B_c(2S)$
-------------------------- ------------------ ---------------- ----------- -----------
$\Delta_{\rm hf}(n_f=4)$ 45.2 46.5 57.8 38.0
$\Delta_{\rm hf}(n_f=3)$ 45.4 $ 46.1 $ 57.2 37.4
$\Delta_{\rm hf}$(exp) $45 .78\pm 0.35$ $46.5\pm 1.25$ absent absent
: \[tab.4\] The HFS (in MeV) of the $B_q$ mesons with $\alpha_s(\mu)=0.307$ ($\alpha_{\rm hf}(n_f=4)=0.316$, $\alpha_{\rm
hf}(n_f=3)=0.324 )$
From our calculations we predict the following masses of the triplet and singlet $c\bar b(2S)$ states: $M(B_c^*(2\,{}^3S_1))=6901(5)$ MeV and $M(B_c(2\,{}^1S_0))= 6865(5)$ MeV, which are calculated in single-channel approximation.
Notice that the extracted strong coupling constant, $\alpha_s(\mu)=0.305(2)$, is smaller than the one used in Refs. [@ref.5] and [@ref.9]. The renormalization scale, $\mu\sim 1.6-1.65$ GeV, corresponding to this coupling, is rather large but agrees with the existing interpretation of the spin-spin potential as dominantly perturbative, thus partly justifying the use of the $\delta^3({\mbox{\boldmath${\it r}$}})$-function.
Conclusion
==========
Our study of the HFS is performed assuming that the freezing value of the coupling constant is the same for $n_f=3,4,5$ and considering $\alpha_{\rm crit}=0.58$ and 0.60. The calculated HFS of the $B$ and $B_s$ mesons are in good agreement with experiment for both freezing constants. It happens that the HFS for $n_f=3$ and $n_f=4$ coincide with each other (within 0.5 MeV), if the one-loop correction is taken into account in $\alpha_{\rm hf}(n_f)$. The HFS, averaged over two results, are $M(B^*) - M(B)= 45.3(2)$ MeV, $M(B_s^*)-M(B_s)=46.5(3)$ MeV, and $M(B_c^*)-M(B_c)=57.5(10)$ MeV. The latter number gives the mass of the as yet unobserved $B_c^*$ meson, $M(B_c^*)=6.334\pm 1$ MeV. For excited $B_c(2S)$ states we predict the masses: $M(B_c^*(2S)=6901(5)$ MeV and $M(B_c(2S)=6865(5)$ MeV, which are calculated neglecting open channels.
In bottomonium the choice $n_f=5$, where $\Delta_{\rm hf}(b\bar b)=70.0(4)$ MeV, is considered preferable, since for $n_f=4$ the HFS is smaller, equal to 64.2(4) MeV .
For the $2S$ and $3S$ bottomonium states, the calculated HFS are 36(1) MeV and 27(1) MeV, respectively.
The extracted coupling, $\alpha_s(\mu)=0.305(2)$, is smaller than in many other analyses with a universal coupling; it determines the characteristic scale of the spin-spin interaction, being $\mu
\sim 1.6-1.65$ GeV. Knowledge of this scale can help to better understand the behavior of the spin-spin potential at small distances.
This work is supported by the Grant RFFI-09-02-00629 a.
[99]{} B. Aubert et al. (BaBar Collab.), Phys. Rev. Lett. [**103**]{}, 161801 (2009).
B. Aubert et al. (BaBar Collab.), Phys. Rev. Lett. [**101**]{}, 071801 (2008).
G. Bonvicini et al. (CLEO Collab.), arXiv:0909.5474 \[hep-ex\].
S. Godfrey and N. Isgur, Phys. Rev. D [**32**]{}, 189 (1985).
E.J. Eichten and C. Quigg, Phys. Rev. D [**49**]{}, 5845 (1994);\
W. Buchmuller and S.H.H. Tye, Phys. Rev. D [**24**]{}, 132 (1981).
S. Godfrey and J.L. Rosner, Phys. Rev. D [**64**]{}, 074011 (2001); Erratum ibid. [**65**]{}, 039901 (2002).
P. Moxhay and J.L. Rosner, Phys. Rev. D [**28**]{}, 1182 (1983);\
D. Ebert, R.N. Faustov, and V.O. Galkin, Mod. Phys. Lett. A [**18**]{}, 601, 1597 (2003);\
S.M. Ikhdair and S. Sever, Int. J. Phys. A [**21**]{}, 3989 (2006).
B.A. Kniehl et al., Phys. Rev.Lett. [**92**]{}, 242001 (2001); S. Recksiegel and Y. Zumino, Phys. Lett. B [**578**]{}, 369 (2004);\
N. Brambilla and A. Vairo, Acta Phys. Polon. B [**38**]{}, 3429 (2007).
M. DiPierro and E.J. Eichten, Phys. Rev. D [**64**]{}, 114004 (2001).
Y. Koma and M. Koma, Nucl. Phys. B [**769**]{}, 79 (2007).
A. De Rujula, H. Georgi, and S. L. Glashow, Phys. Rev. D [**12**]{}, 147 (1975).
A.M. Badalian and B.L.G. Bakker, Phys. Lett. B [**646**]{}, 29 (2007); Phys. Atom. Nucl.[**70**]{}, 1764 (2007).
A.M. Badalian, A.I. Veselov, and B.L.G. Bakker, J. Phys. G [**31**]{}, 417 (2005);\
A.M. Badalian, B.L.G. Bakker, and I.V. Danilkin, arXiv:0903.3643 \[hep-ph\] (to be published in Yad. Fiz.).
A. DiGiacomo, H.G..Dosch, V.I..Shevchenko, and Yu.A..Simonov, Phys.Rep. [**372**]{}, 319 (2002);\
Yu.A. Simonov, “QCD and topics in hadron physics”, arXiv: hep-ph/9911237;\
H.G. Dosch and Yu.A.Simonov, Phys. Lett. [**B205**]{}, 339 (1988).
J. Pantaleone, S.H.H. Tye, and Y.J. Ng, Phys. Rev. D [**33**]{}, 777 (1986).
E. B. Gregory et al. (HPQCD Collab.), arXiv: 0909.4462 \[hep-lat\].
E. Follana et al. (HPQCD and Fermilab Collab.) Phys. Rev. Lett [**94**]{}, 172001 (2005).
A.M. Badalian, A.V. Nefediev, and Yu.A. Simonov, Phys. Rev. D [**78**]{}, 114020 (2008);\
Yu.A. Simonov, Lectures at 17th Autumn School, Lisbon, 29 Sept.-4 Oct. (1999), arXiv: hep-ph/ 9911237.
C. Amsler et al. (Particle Data Group) Phys. Lett. B [**667**]{}, 1 (2008).
A.Y. Dubin, A.B. Kaidalov, and Yu.A. Simonov, Phys. Atom. Nucl. [**56**]{}, 1745 (1993);\
Phys. Lett. B [**323**]{}, 41 (1994);\
A. V. Dubin and E. L. Gubankova, Phys. Lett. B [**334**]{}, 180 (1994).
Yu.A. Simonov,Z.Phys. C [**53**]{}, 419 (1992);\
Yu.S. Kalashnikova, A.V. Nefediev, and Yu.A. Simonov, Phys. Rev. D [**64**]{}, 014037 (2001);\
A.M. Badalian, B.L. G. Bakker and I.V. Danilkin, Phys. Rev. D [**79**]{}, 037505 (2009);\
A.M. Badalian and I.V. Danilkin, Phys. Atom. Nucl. [**72**]{}, 1206 (2009); arXiv: 0801.1614 \[hep-ph\].
Yu.A. Simonov, Phys. Lett. B [**515**]{}, 137 (2001).
G.S. Bali, Phys. Rep. [**343**]{}, 1 (2001) (and references therein).
C. Bernard et al., (MILC Collab.) Phys. Rev. D [**64**]{}, 054506 (2001).
Yu.A. Simonov, Phys. Atom. Nucl.[ **65**]{}, 135 (2002); ibid. [**58**]{}, 107 (1995);\
A.M. Badalian and Yu.A. Simonov, Phys. Atom. Nucl. [**60**]{}, 630 (1997) \[Yad. Fiz. [**60**]{}, 714 (1997).
A.M. Badalian and D.S. Kuzmenko, Phys. Rev. D [**65**]{}, 016004 (2002);\
A.M. Badalian, Phys. Atom. Nucl. [**63**]{}, 2173 (2000).
J.M. Cornwall, Phys. Rev. D [**26**]{}, 1453 (1982);\
G. Parisi and R. Petronzio, Phys. Lett. B [**94**]{}, 51 (1980);\
A.C. Mattingly and P.M. Stevenson, Phys. Rev. D [**49**]{}, 437 (1994).
Y. Schröder, Phys. Lett. B [**447**]{}, 321 (1999);\
M. Peter, Phys. Rev. Lett. [**78**]{}, 602 (1997); Nucl. Phys. B [**501**]{}, 471 (1997).
A.M. Badalian and A.I. Veselov, Phys. Atom. Nucl. [**68**]{}, 582 (2005).
A. M. Badalian, B. L. G. Bakker, and Yu. A. Simonov, Phys. Rev. D [**75**]{}, 116001 (2007).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We study the two-body anti-triplet charmed baryon decays of ${\bf B}_c\to {\bf B}_n V$, with ${\bf B}_c=(\Xi_c^{0},-\Xi_c^{+},\Lambda_c^+)$ and ${\bf B}_n(V)$ the baryon (vector meson) states. Based on the $SU(3)$ flavor symmetry, we predict that ${\cal B}(\Lambda^{+}_{c}\to \Sigma^{+}\rho^{0},\Lambda^0 \rho^+)
=(0.61\pm 0.46,0.74\pm 0.34)\%$, in agreement with the experimental upper bounds of $(1.7,6)\%$, respectively. We also find ${\cal B}(\Lambda^+_c \to
\Xi^0 K^{*+},\Sigma^0 K^{*+},\Lambda^0 K^{*+})
=(8.7 \pm 2.7,1.2\pm 0.3,2.0\pm 0.5)\times 10^{-3}$ to be compatible with the pseudoscalar counterparts. For the doubly Cabibbo-suppressed decay $\Xi^+_c \to p\phi$, measured for the first time, we predict its branching ratio to be $(1.5\pm 0.7)\times 10^{-4}$, together with ${\cal B}(\Xi^+_c \to p \bar K^{*0},\Sigma^+ \phi)
=(7.8 \pm 2.2,1.9\pm0.9)\times 10^{-3}$. The ${\bf B}_c\to{\bf B}_n V$ decays with ${\cal B}\simeq {\cal O}(10^{-4}-10^{-3})$ are accessible to the BESIII, BELLEII and LHCb experiments.
author:
- 'Y.K. Hsiao'
- Yu Yao
- 'H.J. Zhao'
title: 'Two-body charmed baryon decays involving vector meson with $SU(3)$ flavor symmetry '
---
Introduction
============
The two-body ${\bf B}_c\to{\bf B}_n V$ decays have not been abundantly measured as the ${\bf B}_c\to{\bf B}_n M$ counterparts, where ${\bf B}_c=(\Xi_c^{0},-\Xi_c^{+},\Lambda_c^+)$ are the anti-triplet charmed baryon states, together with ${\bf B}_n$ and $V(M)$ the baryon and vector (pseudo-scalar) meson states, respectively. For example, all Cabibbo-favored (CF) $\Lambda_c^+\to {\bf B}_n M$ decays have been measured [@pdg], including the recent BESIII observation for $\Lambda_c^+\to \Sigma^+\eta'$ [@Ablikim:2018czr], whereas for the CF vector modes only $\Lambda_c^+\to p \bar K^{*0},\Sigma^+\omega,\Sigma^+\phi$ have absolute branching fractions [@pdg]. In addition, the first absolute branching ratio for the $\Xi_c^0$ decays is $\Xi_c^0\to \Xi^-\pi^+$ [@Li:2018qak], instead of any $\Xi_c^0\to {\bf B}_n V$ decays.
Nevertheless, the ${\bf B}_c\to{\bf B}_n V$ decays are not less important than the ${\bf B}_c\to{\bf B}_n M$ counterparts. First, the participations of BESIII, BELLEII and LHCb Collaborations are expected to make more accurate measurements for ${\bf B}_c\to{\bf B}_n V$, such as $\Lambda^{+}_{c}\to \Sigma^{+}\rho^{0},\Lambda^0 \rho^+$, presented as ${\cal B}(\Lambda_c^+\to \Sigma^{+}\rho^{0},\Lambda^0 \rho^+)<(1.7,6)\%$ due to the previous measurements [@pdg]. Second, in the three-body ${\bf B}_c\to {\bf B}_nMM'$ decays, the $MM'$ meson pair is assumed to be mainly in the S-wave state [@Geng:2018upx]. However, the resonant ${\bf B}_c\to{\bf B}_n V, V\to MM'$ decay causes $MM'$ to be in the P-wave state, of which the contribution to the total ${\cal B}({\bf B}_c\to {\bf B}_nMM')$ needs clarification. Note that (S,P) denote $L=(0,1)$ as the quantum numbers for the orbital angular momentum between $M$ and $M'$. Third, the three-body $\Xi_c^+$ decays can be measured as the ratios of ${\cal B}(\Xi_c^+\to{\bf B}_n V)/{\cal B}(\Xi_c^+\to {\bf B}_nMM')$. Particularly, the doubly Cabibbo-suppressed $\Xi_c^+\to p\phi$ decay is observed for the first time, with ${\cal B}(\Xi_c^+\to p\phi)/{\cal B}(\Xi_c^+\to pK^-\pi^+)
=(19.8\pm 0.7\pm 0.9\pm 0.2)\times 10^{-3}$ [@Aaij:2019kss]. The information of ${\cal B}(\Xi_c^+\to{\bf B}_n V)$ is hence helpful to determine ${\cal B}(\Xi_c^+\to {\bf B}_nMM')$.
Since the study of ${\bf B}_c\to{\bf B}_n V$ is necessary, it is important to provide a corresponding theoretical approach. The factorization approach for the heavy hadron decays [@ali; @Geng:2006jt; @Hsiao:2014mua] seems applicable to ${\bf B}_c\to{\bf B}_n V$. Nonetheless, it has been shown that, besides the factorizable effects, there exist significant non-factorizable contributions in ${\bf B}_c\to{\bf B}_nM$ [@Zhao:2018mov], such that the factorization approach fails to explain the data. In contrast, with both factorizable and non-factorizable effects [@He:2000ys; @Fu:2003fy; @Hsiao:2015iiu; @He:2015fwa; @He:2015fsa; @Savage:1989qr; @Savage:1991wu; @h_term; @He:2018joe], the $SU(3)$ flavor symmetry ($SU(3)_f$) approach can accommodate the measurements for ${\bf B}_c\to{\bf B}_nM$ [@Lu:2016ogy; @Wang:2017azm; @Wang:2017gxe; @Geng:2017esc; @Geng:2018bow; @Geng:2018plk; @Geng:2018rse], such as the purely non-factorizable $\Lambda_c^+\to\Xi^0 K^+$ decay [@Ablikim:2018bir]. In addition, the predicted values of ${\cal B}(\Lambda_c^+\to \Sigma^+\eta')$ and ${\cal B}(\Xi_c^0 \to \Xi^-\pi^+)$ are in agreement with the recent observations [@Ablikim:2018czr; @Li:2018qak; @Geng:2017esc; @Geng:2018plk]. Therefore, we propose to extend the $SU(3)_f$ symmetry to ${\bf B}_c\to{\bf B}_nV$, while the existing observations have been sufficient for the numerical analysis. In this report, we will extract the $SU(3)_f$ amplitudes, and predict the to-be-measured ${\bf B}_c\to{\bf B}_nV$ branching fractions.
Formalism
=========
To obtain the amplitudes for the two-body ${\bf B}_c\to {\bf B}_n V$ decays, where ${\bf B}_{c(n)}$ is the singly charmed (charmless) baryon state and $V$ the vector meson, we present the relevant effective Hamiltonian $({\cal H}_{eff})$ for the tree-level $c$ quark decays, given by [@Buras:1998raa] $$\begin{aligned}
\label{Heff}
{\cal H}_{eff}&=&\sum_{i=+,-}\frac{G_F}{\sqrt 2}c_i
\left(V_{cs}V_{ud}O_i+V_{cq}V_{uq} O_i^q+V_{cd}V_{us}O'_i\right),\end{aligned}$$ with $q=d$ or $s$, where $G_F$ is the Fermi constant, $V_{ij}$ are the CKM matrix elements, and $c_{\pm}$ the scale-dependent Wilson coefficients. In Eq. (\[Heff\]), $O_\pm^{(q,\prime)}$ are the four-quark operators: $$\begin{aligned}
\label{O12}
&&
O_\pm={1\over 2}\left[(\bar u d)(\bar s c)\pm (\bar s d)(\bar u c)\right]\,,\;\nonumber\\
&&
O_\pm^q={1\over 2}\left[(\bar u q)(\bar q c)\pm (\bar q q)(\bar u c)\right]\,,\;\nonumber\\
&&
O'_\pm={1\over 2}\left[(\bar u s)(\bar d c)\pm (\bar d s)(\bar u c)\right]\,,\end{aligned}$$ with $(\bar q_1 q_2)\equiv \bar q_1\gamma_\mu(1-\gamma_5)q_2$. By neglecting the Lorentz indices, the operator of $(\bar q_1q_2)(\bar q_3 c)$ transforms as $(\bar q^i q_k \bar q^j)c$ under the $SU(3)_f$ symmetry, where $q_i=(u,d,s)$ represent the triplet of $3$. The operator can be decomposed as irreducible forms, which is accordance with $(\bar 3\times 3\times \bar 3)c=(\bar 3+\bar 3'+6+\overline{15})c$. One hence has [@Savage:1989qr; @Savage:1991wu] $$\begin{aligned}
\label{irreOi}
&&O_{-(+)}\sim {\cal O}_{6(\overline{15})}=
\frac{1}{2}(\bar u d\bar s\mp\bar s d\bar u)c\,,\nonumber\\
%
&&O_{-(+)}^q \sim{\cal O}_{6(\overline{15})}^q=
{1\over 2}(\bar u q\bar q\mp \bar q q\bar u)c\,,\nonumber\\
%
&&O'_{-(+)}\sim {\cal O'}_{6(\overline{15})}=
{1\over 2}(\bar u s\bar d\mp \bar d s\bar u)c\,,\end{aligned}$$ with the subscripts $(6,\overline{15})$ denoting the two irreducible $SU(3)_f$ operators. By substituting ${\cal O}_{6(\overline{15})}^{(q,\prime)}$ for $O_{-(+)}^{(q,\prime)}$, the effective Hamiltonian in Eq. (\[Heff\]) becomes $$\begin{aligned}
\label{Heff2}
{\cal H}_{eff}&=&\frac{G_F}{\sqrt 2}\left[c_- { \epsilon^{ijl} \over 2}H(6)_{lk}+c_+H(\overline{15})_k^{ij}\right]\,,\end{aligned}$$ where the tensor notations of $1/2\epsilon_{ijl}H(6)^{lk}$ and $H(\overline{15})_{ij}^k$ contain ${\cal O}_{6}^{(q,\prime)}$ and ${\cal O}_{\overline{15}}^{(q,\prime)}$, respectively. In terms of $(V_{cs}V_{ud},V_{cd}V_{ud},V_{cs}V_{us},V_{cd}V_{us})=(1,-s_c,s_c,-s_c^2)$ with $s_c\equiv \sin\theta_c$, where $\theta_c$ represents the well-known Cabbibo angle, we have $H_{22}(6)=2$, $H^{23,32}(6)=-2s_c$, $H^2_{12,21}(\overline{15})=-H^3_{13,31}(\overline{15})=s_c$, $H^{33}(6)=2s_c^2$, and $H^3_{12,21}(\overline{15})=-s_c^2$ as the non-zero entries [@Savage:1989qr]. Note that $n=0,1$ and 2 in $s_c^n$ correspond to the Cabibbo-flavored (CF), singly Cabibbo-suppressed (SCS), and doubly Cabibbo-suppressed (DCS) decays, respectively. We also need ${\bf B}_c$ and ${\bf B}_n$ ($V$) to be in the irreducible representation of the $SU(3)_f$ symmetry, given by $$\begin{aligned}
\label{b_octet}
({\bf B}_{c})_i&=&(\Xi_c^0,-\Xi_c^+,\Lambda_c^+)\,,\nonumber\\
%
({\bf B}_n)^i_j&=&\left(\begin{array}{ccc}
\frac{1}{\sqrt{6}}\Lambda^0+\frac{1}{\sqrt{2}}\Sigma^0 & \Sigma^+ & p\\
\Sigma^- &\frac{1}{\sqrt{6}}\Lambda^0 -\frac{1}{\sqrt{2}}\Sigma^0 & n\\
\Xi^- & \Xi^0 &-\sqrt{\frac{2}{3}}\Lambda^0
\end{array}\right)\,,\nonumber\\
%
(V)^i_j&=&\left(\begin{array}{ccc}
\frac{1}{\sqrt{2}}(\rho^0+ \omega) & \rho^- & K^{*-}\\
\rho^+ & \frac{-1}{\sqrt{2}}(\rho^0-\omega) & \bar K^{*0}\\
K^{*+} & K^{*0}& \phi
\end{array}\right)\,.\end{aligned}$$ Subsequently, ${\cal H}_{eff}$ in Eq. (\[Heff2\]) is enabled to be connected to the initial and final states in Eq. (\[b\_octet\]), such that we derive the amplitudes of ${\bf B}_c\to {\bf B}_n V$ as $$\begin{aligned}
&&{\cal A}({\bf B}_c\to {\bf B}_n V)
=\langle {\bf B}_n V|{\cal H}_{eff}|{\bf B}_c\rangle
=\frac{G_F}{\sqrt 2}T({\bf B}_{c}\to {\bf B}_nV)\,,\end{aligned}$$ instead of introducing the details of the QCD calculations for the hadronization. Explicitly, the $T$ amplitudes ($T$-amps) are given by [@Savage:1989qr; @Savage:1991wu] $$\begin{aligned}
\label{Tamp}
T({\bf B}_c\to {\bf B}_n V)&=&T({\cal O}_6)+T({\cal O}_{\overline{15}})\,,\nonumber\\
%
T({\cal O}_6)&=&
\bar a_1 H^{ij}(6)T_{ik}({\bf B}_n)^k_l (V)^l_j+
\bar a_2 H^{ij}(6)T_{ik}(V)^k_l ({\bf B}_n)^l_j\nonumber\\
&+&
\bar a_3 H^{ij}(6)({\bf B}_n)^k_i (V)^l_j T_{kl}+
\bar h H^{ij}(6)T_{ik}({\bf B}_n)^k_j (V)^l_l\,,\nonumber\\
%
T({\cal O}_{\overline{15}})&=&
\bar a_4H^{i}_{jk}(\overline{15}) (V)^j_l ({\bf B}_n)^k_i({\bf B}_{c})^l
+\bar a_5 H(\overline{15})^{i}_{jk} ({\bf B}_{c})^j({\bf B}_n)^k_l (V)^l_i\nonumber\\
&+&
\bar a_6H(\overline{15})^i_{jk}({\bf B}_n)^j_l (V)^k_i ({\bf B}_{c})^l
+\bar a_7 H(\overline{15})^i_{jk}({\bf B}_{c})^j (V)^k_l({\bf B}_n)^l_i \nonumber\\
&+&
\bar h' H^{i}_{jk}(\overline{15})({\bf B}_n)^k_i (V)^l_l ({\bf B}_{c})^j\,,\end{aligned}$$ where $T_{ij} \equiv ({\bf B}_c)_k\epsilon^{ijk}$, and $(c_-,c_+)$ have been absorbed into the $SU(3)$ parameters $(\bar a_i,\bar h^{(\prime)})$.
[|c|l|]{} $\Xi_c^0$&$\;\;\;\;\;\;\;\;$CF $T$-amp\
$\Sigma^{+} K^{*-} $ & $ 2(\bar a_{2}+\frac{\bar a_{4} + \bar a_{7}}{2})$\
$\Sigma^{0}\bar{K}^{*0}$ &$-\sqrt{2}(\bar a_{2}+\bar a_{3}$ $-\frac{\bar a_{6}-\bar a_{7}}{2})$\
$\Xi^{0} \rho^{0} $ & $ -\sqrt{2}(\bar a_{1}-\bar a_{3}$ $-\frac{\bar a_{4}-\bar a_{5}}{2})$\
$ \Xi^{0} \omega $ & $ \sqrt{2}(\bar a_1- \bar a_3 +2\bar h $ $+ \frac{\bar a_4+\bar a_5+2 \bar h'}{2} )$\
$ \Xi^{0} \phi $ &$\bar a_2 +\bar h + \frac{\bar a_7 + \bar h'}{2} $\
$\Xi^{-} \rho^{+} $ & $ 2(\bar a_{1}+\frac{\bar a_{5} + \bar a_{6}}{2})$\
$\Lambda^{0} \bar{K}^{*0} $ & $-\sqrt{\frac{2}{3}}(2\bar a_1-\bar a_2-\bar a_3$\
& $+\frac{2\bar a_5-\bar a_6-\bar a_7}{2})$\
$\Xi_c^+$&$\;\;\;\;\;\;\;\;$CF $T$-amp\
$\Sigma^{+} \bar{K}^{*0} $ &$ -2(\bar a_{3}-\frac{\bar a_{4} + \bar a_{6}}{2})$\
$\Xi^{0} \rho^{+} $ & $2(\bar a_{3}+\frac{\bar a_{4} + \bar a_{6}}{2})$\
$\Lambda_c^+$&$\;\;\;\;\;\;\;\;$CF $T$-amp\
$\Sigma^{0} \rho^{+} $ &$-\sqrt{2}(\bar a_1-\bar a_2-\bar a_3$ $-\frac{\bar a_5-\bar a_7}{2})$\
$\Sigma^{+} \rho^{0} $ & $\sqrt{2}(\bar a_{1}-\bar a_{2}-\bar a_{3}$ $-\frac{\bar a_{5}-\bar a_{7}}{2})$\
$ \Sigma^{+} \omega $ & $ \sqrt{2}(-\bar a_1-\bar a_2+\bar a_3-2\bar h$\
&$+\frac{\bar a_5+\bar a_7+2\bar h'}{2}) $\
$ \Sigma^{+} \phi $ &$ \bar a_4-2\bar h+\bar h' $\
$\Xi^{0} K^{*+} $ & $-2(\bar a_{2}-\frac{\bar a_{4} + \bar a_{7}}{2})$\
$p \bar{K}^{*0} $ & $ -2(\bar a_{1}-\frac{\bar a_{5} + \bar a_{6}}{2})$\
$\Lambda^{0} \rho^{+} $ & $-\sqrt{\frac{2}{3}}(\bar a_1+\bar a_2+\bar a_3$\
&$-\frac{\bar a_5-2\bar a_6+\bar a_7}{2})$\
------------------------------------------------------------------------------------------------------------------
$\Xi_{c}^{0}$ $\;\;\;\;\;\;\;\;\;$SCS $T$-amp
------------------------- ----------------------------------------------------------------------------------------
$\Sigma^{+} \rho^{-} $ $-2(\bar a_{2}+\frac{\bar a_{4} + \bar a_{7}}{2})s_c$
$\Sigma^{-} \rho^{+} $ $-2(\bar a_{1}+\frac{\bar a_{5} + \bar a_{6}}{2})s_c$
$\Sigma^{0} \rho^{0} $ $-(\bar a_{2}+\bar a_{3}$ $-\frac{\bar a_{4}-\bar a_{5}+\bar a_{6}-\bar a_{7}}{2})s_c$
$ \Sigma^{0} \omega $ $[-(\bar a_1+\bar a_2+2\bar h$
$+\frac{\bar a_4+\bar a_5-\bar a_6+\bar a_7+2\bar h'}{2})
]s_c$
$ \Sigma^{0} \phi $ $[\sqrt{2}(\bar a_3-\bar h-\frac{\bar a_6+\bar h'}{2})]s_c$
$\Xi^{-} K^{*+} $ $ 2(\bar a_{1}+\frac{\bar a_{5} + \bar a_{6}}{2})s_c$
$p K^{*-}$ $2(\bar a_{2}+\frac{\bar a_{4} + \bar a_{7}}{2})s_c$
$\Xi^{0} K^{*0} $ $2(\bar a_{1}-\bar a_2-\bar a_{3}$ $+\frac{\bar a_{5}-\bar a_{7}}{2})s_c$
$n \bar K^{*0} $ $-2(\bar a_{1}-\bar a_{2}-\bar a_{3}+\frac{\bar a_{5}-\bar a_{7}}{2})s_c$
$\Lambda^{0} \rho^{0} $ $\sqrt{\frac{1}{3}}(\bar a_1+\bar a_2-2\bar a_3$
$+\frac{\bar a_4-\bar a_5-\bar a_6-\bar a_7}{2})s_c$
$ \Lambda^{0} \omega $ $ \frac{\sqrt{3}}{3}(\bar a_1+\bar a_2-2\bar a_3+6\bar h $
$+\frac{3\bar a_4+\bar a_5+\bar a_6+\bar a_7+6\bar h'}{2}) s_c$
$ \Lambda^{0} \phi $ $\frac{\sqrt{6}}{2}(2\bar a_1+2\bar a_2-\bar a_3+3\bar h $
$+\frac{2\bar a_5-\bar a_6+2\bar a_7+3\bar h'}{2})s_c $
$\Xi_{c}^{+}$ $\;\;\;\;\;\;\;\;\;$SCS $T$-amp
$\Sigma^{0} \rho^{+} $ $\sqrt 2(\bar a_{1}-\bar a_{2}$
$+\frac{\bar a_{4}-\bar a_{5}+\bar a_{6}+\bar a_{7}}{2})s_c$
$\Sigma^{+} \rho^{0} $ $-\sqrt 2(\bar a_{1}-\bar a_{2}$
$-\frac{\bar a_{4}+\bar a_{5}+\bar a_{6}-\bar a_{7}}{2})s_c$
$ \Sigma^{+} \omega $ $\sqrt{2}(\bar a_1+\bar a_2+2\bar h$
$-\frac{\bar a_4+\bar a_5+\bar a_6+\bar a_7-2\bar h'}{2})
s_c$
$ \Sigma^{+} \phi $ $-2(\bar a_3-\bar h-\frac{\bar a_6-\bar h'}{2})s_c$
$\Xi^{0} K^{*+} $ $2(\bar a_2+\bar a_{3}+\frac{\bar a_{6} - \bar a_{7}}{2})s_c$
$p \bar K^{*0} $ $2(\bar a_1-\bar a_{3}+\frac{\bar a_{4} - \bar a_{5}}{2})s_c$
$\Lambda^0\rho^+$ $\sqrt{\frac{2}{3}}(\bar a_1+\bar a_2-2\bar a_3$
$-\frac{3\bar a_4+\bar a_5+\bar a_6+\bar a_7}{2})s_c$
$\Lambda_{c}^{+}$ $\;\;\;\;\;\;\;\;\;$SCS $T$-amp
$\Sigma^{+} K^{*0} $ $-2(\bar a_{1}-\bar a_{3}-\frac{\bar a_{4}-\bar a_{5}}{2})s_c$
$\Sigma^{0} K^{*+} $ $-\sqrt{2}(\bar a_1-\bar a_3-\frac{\bar a_4+\bar a_5}{2})s_c$
$p \rho^{0} $ $ -\sqrt 2(\bar a_{2}+\bar a_3-\frac{\bar a_{6} - \bar a_{7}}{2})s_c$
$ p \omega $ $\sqrt{2}(\bar a_2-\bar a_3+2\bar h$
$+\frac{\bar a_6-\bar a_7-2\bar h'}{2})s_c $
$ p \phi $ $ -2(-\bar a_1-\bar h$
$+\frac{\bar a_4+\bar a_5+\bar a_6+\bar h'}{2})s_c$
$n\rho^+$ $-2(\bar a_{2}+\bar a_3-\frac{\bar a_{4} + \bar a_{7}}{2})s_c$
$\Lambda^{0} K^{*+} $ $-\sqrt{\frac{2}{3}}(\bar a_1-2\bar a_2+\bar a_3$
$-\frac{3\bar a_4-\bar a_5+2\bar a_6+2\bar a_7}{2})s_c$
------------------------------------------------------------------------------------------------------------------
: The $T$-amps for the ${\bf B}_{c}\to {\bf B}_n V$ decays, where CF denotes the Cabibbo-favored processes, while SCS (DCS) the singly (doubly) Cabibbo-suppressed ones.[]{data-label="tab1"}
-----------------------------------------------------------------------------------------------------------
$\Xi_c^0$ $\;\;\;\;\;\;\;\;$DCS $T$-amp
------------------------- ---------------------------------------------------------------------------------
$p\rho^-$ $-2(\bar a_{2}+\frac{\bar a_{4} + \bar a_{7}}{2})s_c^2$
$\Sigma^{-} K^{*+} $ $-2(\bar a_{1}+\frac{\bar a_{5} + \bar a_{6}}{2})s_c^2$
$\Sigma^{0}{K}^{*0}$ $ \sqrt 2(\bar a_{1}+\frac{\bar a_{5} - \bar a_{6}}{2})s_c^2$
$n \rho^{0} $ $\sqrt 2(\bar a_{2}-\frac{\bar a_{4} - \bar a_{7}}{2})s_c^2$
$ n \omega $ $-\sqrt{2}( \bar a_2-2\bar h + \frac{\bar a_4 -\bar a_7 -2\bar h'}{2} )
s_c^2$
$ n \phi $ $-2( \bar a_1 -\bar a_3 +\bar h +\frac{\bar a_5+\bar h'}{2})s_c^2$
$\Lambda^{0} {K}^{*0} $ $-\sqrt{\frac{2}{3}}(\bar a_1-2\bar a_2-2\bar a_3$
$+\frac{\bar a_5+\bar a_6-2\bar a_7}{2})s_c^2$
\[40mm\] $\Xi_c^+$ $\;\;\;\;\;\;\;\;$DCS $T$-amp
$\Sigma^{0} {K}^{*+} $ $ \sqrt 2(\bar a_{1}-\frac{\bar a_{5} - \bar a_{6}}{2})s_c^2$
$\Sigma^{+} {K}^{*0} $ $ 2(\bar a_{1}-\frac{\bar a_{5} + \bar a_{6}}{2})s_c^2$
$p \rho^0 $ $\sqrt 2(\bar a_{2}+\frac{\bar a_{4} - \bar a_{7}}{2})s_c^2$
$ p \omega $ $[\sqrt{2}(-\bar a_2 + 2\bar h+\frac{\bar a_4+\bar a_7 + 2\bar h'}{2})] s_c^2 $
$ p \phi $ $ -2(\bar a_1 - \bar a_3 +h- \frac{\bar a_5 +\bar h'}{2})s_c^2$
$n \rho^{+} $ $2(\bar a_{2}-\frac{\bar a_{4} + \bar a_{7}}{2})s_c^2$
$\Lambda^0 K^{*+}$ $\sqrt{\frac{2}{3}}(\bar a_1-2\bar a_2-2\bar a_3$
$-\frac{\bar a_5+\bar a_6-2\bar a_7}{2})s_c^2$
\[15mm\] $\Lambda_c^+$ $\;\;\;\;\;\;\;\;$DCS $T$-amp
$p {K}^{*0} $ $ 2(\bar a_{3}-\frac{\bar a_{4} + \bar a_{6}}{2})s_c^2$
$nK^{*+}$ $- 2(\bar a_{3}+\frac{\bar a_{4} + \bar a_{6}}{2})s_c^2$
\[40mm\]
-----------------------------------------------------------------------------------------------------------
: The $T$-amps for the ${\bf B}_{c}\to {\bf B}_n V$ decays, where CF denotes the Cabibbo-favored processes, while SCS (DCS) the singly (doubly) Cabibbo-suppressed ones.[]{data-label="tab1"}
With the full expansion of $T$-amps in Table \[tab1\], the two-body ${\bf B}_c\to {\bf B}_nV$ decays are presented with the $SU(3)_f$ parameters. Since $\omega=(u\bar u+d\bar d)/\sqrt 2$ and $\phi=s\bar s$ actually mix with $\omega_1=(u\bar u+d\bar d+s\bar s)/\sqrt 3$ and $\omega_8=(u\bar u+d\bar d-2s\bar s)/\sqrt 6$, the $(\bar h,\bar h')$ terms that are related to $(V)^l_l =\sqrt 2 \omega+\phi=\sqrt 3 \omega_1$ can contribute to the decays with $(\omega,\phi)$ only. In terms of the equation for the two-body decays, given by [@pdg] $$\begin{aligned}
\label{p_space}
&&{\cal B}({\bf B}_c\to {\bf B}_n V)=
\frac{|\vec{p}_{cm}|\tau_{\bf{B}_c}}{8\pi m_{{\bf B}_c}^2 }|{\cal A}({\bf B}_c\to {\bf B}_n V)|^2\,,\nonumber\\
&&|\vec{p}_{cm}|=\frac{
\sqrt{[(m_{{\bf B}_c}^2-(m_{{\bf B}_n}+m_V)^2]
[(m_{{\bf B}_c}^2-(m_{{\bf B}_n}-m_V)^2]}}{2 m_{{\bf B}_c}}\,,\end{aligned}$$ we can compute the branching ratio with the $SU(3)_f$ amplitudes, where $\tau_{\bf{B}_c}$ denotes the $\bf{B}_c$ lifetime. The $SU(3)_f$ amplitudes are accounted to be 9 complex numbers, leading to 17 independent parameters to be extracted, whereas there exist 10 data points for the numerical analysis. To have a practical fit, we follow Refs. [@Lu:2016ogy; @Geng:2017esc; @Geng:2018plk; @Geng:2018rse] to reduce the parameters. In ${\cal H}_{eff}\propto c_-H(6)+c_+H(\overline{15})$, since the QCD calculation at the scale $\mu=1$ GeV leads to $(c_+,c_-)=(0.76,1.78)$ in the naive dimensional regularization (NDR) scheme [@Li:2012cfa; @Fajfer:2002gp], the ratio of $(c_-/c_+)^2\simeq 0.17$ indicates the suppression of $H(\overline{15})$. We hence ignore $(\bar a_{4,5, ...,7},h')$. On the other hand, $(\bar a_{1,2,3},h)$ from $H(6)$ are kept for the fit, represented as $$\begin{aligned}
\label{7p}
\bar a_1, \bar a_2e^{i\delta_{\bar a_2}},\bar a_3e^{i\delta_{\bar a_3}},
\bar h e^{i\delta_{\bar h}}\,,\end{aligned}$$ with the phases $\delta_{\bar a_{2,3},\bar h}$, and $\bar a_1$ set to be relatively real.
Numerical analysis
==================
For the numerical analysis, we collect (the ratios of) the branching fractions for the observed ${\bf B}_c\to{\bf B}V$ decays in Table \[data\], where ${\cal B}(\Xi^+_c \to p \bar K^{*0},\Sigma^+ \phi,\Sigma^+\bar K^{*0})$ are in fact measured to be relative to ${\cal B}(\Xi_c^+\to\Xi^-\pi^+\pi^+)$ [@pdg; @Link:2003cd], recombined as ${\cal R}_{1,2}(\Xi^+_c)$. We obtain ${\cal B}( \Xi^0_c \to \Lambda^0 \phi)$ from ${\cal B}( \Xi^0_c \to \Lambda^0 \phi)/{\cal B}( \Xi^0_c \to \Xi^-\pi^+)$ [@pdg], with the input of ${\cal B}( \Xi^0_c \to \Xi^- \pi^+)$ measured by BELLE [@Li:2018qak]. In addition, the ratio of ${\cal R}(\Lambda_c^+)=
(\Lambda_c^+ \to \Sigma^+ \rho^0)/{\cal B}(\Lambda_c^+\to \Sigma^+ \omega)$ comes from the data events in Ref. [@Kubota:1993pw]. Besides, $s_c=0.22453\pm 0.00044$ [@pdg] is the theoretical input for the CKM matrix elements. By using the equation of [@Zhao:2018mov] $$\begin{aligned}
\label{chi_eq}
\chi^2=
\sum_{i} \bigg(\frac{{\cal B}^i_{th}-{\cal B}^i_{ex}}{\sigma_{ex}^i}\bigg)^2+
\sum_{j}\bigg(\frac{{\cal R}^j_{th}-{\cal R}^j_{ex}}{\sigma_{ex}^j}\bigg)^2\,,\end{aligned}$$ we are able to obtain the minimum $\chi^2$ value, such that the $SU(3)_f$ parameters can be extracted with the best fit. Note that ${\cal B}^i({\cal R}^j)$ represents (the ratio of) the branching fraction, with the subscript $th$ ($ex$) denoting the theoretical (experimental) input, while $\sigma_{ex}^{i(j)}$ stands for the experimental error. As the inputs in Eq. (\[chi\_eq\]), ${\cal B(R)}_{th}$ come from the $T$-amps in Table \[tab1\], while ${\cal B(R)}_{ex}$ and $\sigma_{ex}$ the data points in Table \[data\].
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
(Ratio of) Branching fraction This work Data
--------------------------------------------------------------------------------------------------------------- -------------------- ------------------------------------------
$10^2{\cal B}(\Lambda_c^+ \to p \bar K^{*0})$ $1.9 \pm 0.3$ $1.94\pm 0.27$ [@pdg]
$10^2{\cal B}(\Lambda_c^+ \to \Sigma^+ \omega)$ $1.6 \pm0.7$ $1.69\pm 0.21$ [@pdg]
$10^3{\cal B}(\Lambda_c^+ \to \Sigma^+ \phi)$ $3.9\pm0.6$ $3.8\pm 0.6$ [@pdg]
${\cal R}(\Lambda^+_c )= $0.4\pm 0.3$ $0.3\pm 0.2$ [@Kubota:1993pw]
\frac{{\cal B}( \Lambda_c^+ \to \Sigma^+ \rho^0)}{{\cal B}(\Lambda_c^+ \to \Sigma^+ \omega)}$
$10^3{\cal B}(\Lambda_c^+ \to \Sigma^+ K^{*0})$ $2.3 \pm 0.6$ $3.4\pm 1.0$ [@pdg]
$10^4{\cal B}(\Lambda_c^+ \to p \omega)$ $11.4\pm5.4$ $9.4\pm 3.9$ [@Aaij:2017nsd]
$10^4{\cal B}(\Lambda_c^+ \to p \phi)$ $10.4\pm 2.1$ $10.6\pm 1.4$ [@pdg]
$10^4{\cal B}( \Xi^0_c \to \Lambda^0 \phi)$ $8.4\pm3.9$ $6.1\pm 2.2$ [@pdg; @Li:2018qak]
${\cal R}_1(\Xi^+_c )=\frac{{\cal B}( \Xi^+_c \to p \bar K^{*0})}{{\cal B}(\Xi^+_c \to \Sigma^+\bar K^{*0})}$ $(1.6\pm0.2)s_c^2$ $(2.8\pm1.0)s_c^2$ [@pdg]
${\cal R}_2(\Xi^+_c )=\frac{{\cal B}( \Xi^+_c \to \Sigma^+ \phi)}{{\cal B}(\Xi^+_c \to \Sigma^+\bar K^{*0})}$ $(0.4\pm0.1)s_c^2$ $(1.7\pm 1.2)s_c^2$ [@pdg; @Link:2003cd]
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
: The (ratios of) branching fractions of the ${\bf B}_c\to {\bf B}_n V$ decays. In column 2, the numbers are calculated with the extracted parameters, in comparison with the initial experimental inputs in column 3. []{data-label="data"}
Subsequently, the global fit gives $$\begin{aligned}
\label{su3_fit}
&&(\bar a_1,\bar a_2,\bar a_3,\bar h)=(0.22\pm 0.02,0.23\pm 0.04,
0.39\pm 0.05,0.16\pm0.01)\,\text{GeV}^3\,,\nonumber\\
&&(\delta_{\bar a_2},\delta_{\bar a_3},\delta_{\bar h})=
(-85.5\pm13.0,78.4\pm 8.8,99.3\pm 7.7)^\circ\,,\nonumber\\
&&\chi^2/n.d.f=6.3/3=2.1\,,\end{aligned}$$ where $n.d.f$ represents the number of the degree of freedom. With the fit results in Eq. (\[su3\_fit\]), we calculate the branching ratios, ${\cal R}(\Lambda_c^+)$ and ${\cal R}_{1,2}(\Xi_c^+)$ to be compared to their data inputs in Table \[data\]. Moreover, we predict the branching fractions for the ${\bf B}_c\to{\bf B}_n V$ decays, given in Table \[tab\_result\].
$\Xi_c^0$ our results
------------------------------------------- ----------------------
$10^3{\cal B}_{\Sigma^{+} K^{*-}}$ $9.3 \pm 2.9$
$10^3{\cal B}_{\Sigma^{0} \bar{K}^{*0}}$ $2.7 \pm 2.2$
$10^2{\cal B}_{\Xi^{0} \rho^{0}}$ $1.4 \pm 0.4$
$10^3{\cal B}_{\Xi^{0} \omega}$ $1.0^{+8.6}_{-1.0}$
$10^4{\cal B}_{\Xi^{0} \phi}$ $1.5 ^{+7.1}_{-1.5}$
$10^3{\cal B}_{\Xi^{-} \rho^{+}}$ $8.6 \pm 1.2$
$10^3{\cal B}_{\Lambda^{0} \bar{K}^{*0}}$ $4.6 \pm 2.1$
$\Xi_c^+$ our results
$10^2{\cal B}_{\Sigma^{+} \bar{K}^{*0}}$ $10.1\pm 2.9$
$10^2{\cal B}_{\Xi^{0} \rho^{+}}$ $9.9 \pm2.9 $
$\Lambda_c^+$ our results
$10^3{\cal B}_{\Sigma^{0} \rho^{+}}$ $6.1\pm 4.6$
$10^3{\cal B}_{\Sigma^{+} \rho^{0}}$ $6.1\pm4.6$
$10^3{\cal B}_{\Xi^{0} K^{*+}}$ $8.7 \pm 2.7$
$10^3{\cal B}_{\Lambda^{0} \rho^{+}}$ $7.4\pm 3.4$
: The numerical results of the ${\bf B}_{c}\to {\bf B}_n V$ decays, with ${\cal B}_{{\bf B}_nV}\equiv {\cal B}({\bf B}_c\to {\bf B}_nV)$. []{data-label="tab_result"}
$\Xi_c^0$ our results
--------------------------------------- ---------------------
$10^4{\cal B}_{\Sigma^{+} \rho^{-}}$ $5.6 \pm 1.8$
$10^4{\cal B}_{\Sigma^{-} \rho^{+}}$ $5.3 \pm 0.7$
$10^5{\cal B}_{\Sigma^{0} \rho^{0}}$ $8.2 \pm 6.7$
$10^4{\cal B}_{\Sigma^{0} \omega}$ $1.0 \pm 0.8$
$10^4{\cal B}_{\Sigma^{0} \phi}$ $2.4 \pm1.1$
$10^4{\cal B}_{\Xi^{-} K^{*+}}$ $3.9 \pm 0.5$
$10^4{\cal B}_{\Xi^{0} K^{*0}}$ $6.3 \pm 2.0$
$10^4{\cal B}_{p K^{*-}}$ $3.0\pm 2.2$
$10^4{\cal B}_{n \bar{K}^{*0}}$ $4.5 \pm 3.4$
$10^4{\cal B}_{\Lambda^{0} \rho^{0}}$ $9.2 \pm 2.2$
$10^4{\cal B}_{\Lambda^{0} \omega}$ $0.1^{+2.5}_{-0.1}$
$\Xi_c^+$ our results
$10^3{\cal B}_{\Sigma^{0} \rho^{+}}$ $1.9 \pm 0.6$
$10^3{\cal B}_{\Sigma^{+} \rho^{0}}$ $1.9 \pm 0.6$
$10^4{\cal B}_{\Sigma^{+} \omega}$ $8.2\pm5.9$
$10^3{\cal B}_{\Sigma^{+} \phi}$ $1.9\pm0.9$
$10^4{\cal B}_{\Xi^{0} K^{*+}}$ $9.6 \pm7.9$
$10^3{\cal B}_{p \bar{K}^{*0}}$ $7.8 \pm 2.2$
$10^3{\cal B}_{\Lambda^{0} \rho^{+}}$ $7.1 \pm1.7$
$\Lambda_c^+$ our results
$10^4{\cal B}_{p \rho^{0}}$ $3.5 \pm 2.9$
$10^4{\cal B}_{n \rho^{+}}$ $7.0 \pm 5.8$
$10^3{\cal B}_{\Sigma^{0} K^{*+}}$ $1.2\pm 0.3$
$10^3{\cal B}_{\Lambda^{0} K^{*+}}$ $2.0 \pm 0.5$
: The numerical results of the ${\bf B}_{c}\to {\bf B}_n V$ decays, with ${\cal B}_{{\bf B}_nV}\equiv {\cal B}({\bf B}_c\to {\bf B}_nV)$. []{data-label="tab_result"}
$\Xi_c^0$ our results
------------------------------------- ---------------
$10^5{\cal B}_{p \rho^{-}}$ $3.6 \pm 1.1$
$10^5{\cal B}_{\Sigma^{-} K^{*+}}$ $2.5 \pm 0.3$
$10^5{\cal B}_{\Sigma^{0} K^{*0}}$ $1.3 \pm 0.2$
$10^5{\cal B}_{n \rho^{0}}$ $1.8 \pm 0.6$
$10^5{\cal B}_{n \omega}$ $9.9\pm1.6$
$10^5{\cal B}_{n \phi}$ $3.7\pm1.8$
$10^4{\cal B}_{\Lambda^{0} K^{*0}}$ $8.1\pm 7.2$
$\Xi_c^+$ our results
$10^5{\cal B}_{\Sigma^{0} K^{*+}}$ $5.0\pm 0.7$
$10^5{\cal B}_{\Sigma^{+} K^{*0}}$ $9.9 \pm 1.3$
$10^5{\cal B}_{p \rho^{0}}$ $7.1 \pm 2.2$
$10^4{\cal B}_{p \omega}$ $3.9\pm0.6$
$10^4{\cal B}_{p \phi}$ $1.5\pm0.7$
$10^4{\cal B}_{n \rho^{+}}$ $1.4\pm 0.4$
$10^5{\cal B}_{\Lambda^{0} K^{*+}}$ $3.2 \pm 2.9$
$\Lambda_c^+$ our results
$10^4{\cal B}_{p K^{*0}}$ $1.6 \pm 0.5$
$10^4{\cal B}_{n K^{*+}}$ $1.6 \pm 0.5$
: The numerical results of the ${\bf B}_{c}\to {\bf B}_n V$ decays, with ${\cal B}_{{\bf B}_nV}\equiv {\cal B}({\bf B}_c\to {\bf B}_nV)$. []{data-label="tab_result"}
Discussions and Conclusions
===========================
With $\chi^2/n.d.f\simeq 2$ to present a reasonable fit, the approach based the $SU(3)_f$ symmetry is demonstrated to be reliable for ${\bf B}_c\to {\bf B}_n V$. Besides, our prediction $$\begin{aligned}
\label{1st_pre}
{\cal B}(\Lambda^{+}_{c}\to \Sigma^{+}\rho^{0},\Lambda^0 \rho^+)
&=&(0.61\pm 0.46,0.74\pm 0.34)\%\,,\end{aligned}$$ agrees with the experimental upper bounds of $(1.7,6)\%$, respectively [@pdg]. We also find $$\begin{aligned}
&&{\cal B}(\Lambda^+_c \to
\Xi^0 K^{*+},\Sigma^0 K^{*+},\Lambda^0 K^{*+})
=(8.7 \pm 2.7,1.2\pm 0.3,2.0\pm 0.5)\times 10^{-3}\,,\end{aligned}$$ to be compatible with the pseudo-scalar counterparts. According to Table \[tab1\], we obtain $$\begin{aligned}
&&T(\Lambda_c^+\to\Sigma^{0} \rho^{+})
+T(\Lambda_c^+\to\Sigma^{+} \rho^{0})=0\,,\nonumber\\
%
&&T(\Lambda_c^+\to\Sigma^{+} K^{*0} )
-\sqrt 2 T(\Lambda_c^+\to\Sigma^{0} K^{*+})=-2\bar a_5 s_c\,,\nonumber\\
%
&&T(\Lambda_c^+\to n\rho^+)
-\sqrt 2 T(\Lambda_c^+\to p \rho^{0} )=
-(\frac{\bar a_4+\bar a_6}{2})s_c\,,\nonumber\\
%
&&T(\Lambda_c^+\to nK^{*+})
+T(\Lambda_c^+\to p {K}^{*0})=-2(\bar a_4+\bar a_5)s_c^2\,.
%&&T(\Lambda_c^+\to \Sigma^{+} \phi)=\bar a_4-2\bar h+\bar h'\,.\end{aligned}$$ By ignoring the parameters in $H(\overline{15})$, we obtain $$\begin{aligned}
&&{\cal B}(\Lambda_c^+\to\Sigma^{0} \rho^{+},\Sigma^{+} \rho^{0})
=(6.1\pm4.6)\times 10^{-3}\,,\nonumber\\
&&{\cal B}(\Lambda_c^+\to p \rho^{0} )=\frac{1}{2}{\cal B}(\Lambda_c^+\to n\rho^+)
=(3.5 \pm 2.9)\times 10^{-4}\,,\nonumber\\
&&{\cal B}(\Lambda_c^+\to nK^{*+}, p {K}^{*0})
=(1.6 \pm 0.5)\times 10^{-4}\,,\end{aligned}$$ which respect the isospin symmetry. We also get $$\begin{aligned}
\label{re2}
&&
\frac{1}{\sqrt 2}T(\Lambda_c^+\to p \bar {K}^{*0})-
\frac{1}{s_c}T(\Lambda_c^+\to p \rho^{0})
=T(\Lambda_c^+\to\Sigma^{0} \rho^{+})\,,\nonumber\\
&&
\frac{1}{\sqrt 2}T(\Lambda_c^+\to p \bar {K}^{*0})+
\frac{1}{s_c}T(\Lambda_c^+\to p \rho^{0})
=\sqrt 3T(\Lambda_c^+\to\Lambda^{0} \rho^{+})\,,\end{aligned}$$ which lead to $$\begin{aligned}
\label{prho0}
{\cal B}(\Lambda_c^+\to p \rho^0)&\simeq &\frac{s_c^2}{2}
[3.6{\cal B}(\Lambda_c^+\to \Lambda^0 \rho^+)+1.3{\cal B}(\Lambda_c^+\to \Sigma^0 \rho^+)
-1.1{\cal B}(\Lambda_c^+\to p \bar K^{*0})]\,,\end{aligned}$$ where the pre-factors (3.6,1.3,1.1) have taken into account the differences for $|\vec{p}_{cm}|$ in Eq. (\[p\_space\]). It is interesting to note that the $\Lambda_c^+\to p \pi^0$ decay has a similar relation to that in Eq. (\[prho0\]), where $(\rho,\bar K^{*0})$ are replaced by $(\pi,\bar K^{0})$. However, the relation for $\Lambda_c^+\to p \pi^0$ causes ${\cal B}(\Lambda_c^+\to p \pi^0)\simeq 5\times 10^{-4}$, disapproved by the data [@pdg]. This indicates that, even though the ignoring of $H(\overline{15})$ is viable, the possible interferences between $H(6)$ and $H(\overline{15})$ might give sizeable contributions to some decay modes [@Geng:2018rse]. In this work, since the fit still accommodates the data, it is not clear which of the $\Lambda_c^+\to {\bf B}_n V$ decays receives sizeable interferences between $H(6)$ and $H(\overline{15})$. Like the ${\cal B}(\Lambda_c^+\to p \pi^0)$ case, the precise measurement of ${\cal B}(\Lambda_c^+\to p \rho^0)$ can test the ignoring of $H(\overline{15})$. For the $\Xi_c^+$ decays, we obtain $$\begin{aligned}
&&
{\cal B}(\Xi^+_c \to \Sigma^+\bar K^{*0},\Xi^{0} \rho^{+})
=(10.1\pm 2.9,9.9 \pm2.9)\times 10^{-2}\,,\nonumber\\
&&
{\cal B}(\Xi^+_c \to p \bar K^{*0},\Sigma^+ \phi)
=(7.8 \pm 2.2,1.9\pm0.9)\times 10^{-3}\,.\end{aligned}$$ With $f_{\tau_{{\bf B}_c}}\equiv \tau_{\Xi_c^+}/\tau_{\Lambda_c^+}\simeq 2.2$, ${\cal B}(\Xi^+_c \to \Sigma^+\bar K^{*0},\Xi^{0} \rho^{+})
\simeq (2-4)f_{\tau_{{\bf B}_c}}{\cal B}(\Lambda_c^+ \to p \bar K^{*0})$ is found to be in accordance with $|\bar a_3|^2\simeq (2-4)|\bar a_1|^2$, which can be tested by more accurate measurements.
By means of ${\cal B}({\bf B}_c\to {\bf B}_n V,V\to MM')=
{\cal B}({\bf B}_c\to {\bf B}_n V){\cal B}(V\to MM')$, the resonant contribution to the total ${\cal B}({\bf B}_c\to {\bf B}_n MM')$ can be investigated, where $MM'$ from the vector meson decay are in the P-wave state. On the other hand, the theoretical study of ${\bf B}_c\to{\bf B}_n MM'$ needs $MM'$ to be mainly in the S-wave state [@Geng:2018upx]. Using ${\cal B}(\rho^{0(+)}\to\pi^+\pi^{-(0)})\simeq 100\%$ [@pdg] and the predictions for ${\cal B}(\Lambda_c^+\to \Sigma\rho,\Lambda^0\rho^+)$, we obtain $$\begin{aligned}
{\cal B}(\Lambda_c^+\to \Sigma^+ \rho^0,\rho^0\to\pi^+\pi^-)
&=&(6.1\pm 4.6)\times 10^{-3}\,,\nonumber\\
%
{\cal B}(\Lambda_c^+\to \Sigma^0 \rho^+,\rho^+\to\pi^+\pi^0)
&=&(6.1\pm 4.6)\times 10^{-3}\,,\nonumber\\
%
{\cal B}(\Lambda_c^+\to \Lambda^0\rho^+,\rho^+\to\pi^+\pi^0)
&=&(7.4\pm 3.4)\times 10^{-3}\,,\end{aligned}$$ which are within the total branching ratios of $(4.42\pm 0.28,2.2\pm 0.8,7.0\pm 0.4)\times 10^{-2}$ [@pdg], respectively, showing that the P-wave contributions from $V\to MM'$ are indeed minor to these decays. By putting ${\cal B}( \Xi^+_c \to p\phi)=(1.5\pm 0.7)\times 10^{-4}$ into the measured ratio of ${\cal B}( \Xi^+_c \to p\phi)/{\cal B}(\Xi^+_c \to pK^- \pi^+)
=(19.8\pm 0.7\pm 0.9\pm 0.2)\times 10^{-3}$ [@Aaij:2019kss], we obtain ${\cal B}(\Xi^+_c \to pK^- \pi^+)=(0.8\pm 0.4)\%$, which is a little smaller than the predicted value of $(1.7\pm 0.5)\%$ [@Wang:2019dls].
In sum, within the framework of the $SU(3)_f$ symmetry, we have studied the ${\bf B}_c\to{\bf B}_n V$ decays. We have predicted ${\cal B}(\Lambda^{+}_{c}\to \Sigma^{+}\rho^{0},\Lambda^0 \rho^+)
=(0.61\pm 0.46,0.74\pm 0.34)\%$, in agreement with the experimental upper bounds of $(1.7,6)\%$, respectively. It has also been shown that ${\cal B}(\Lambda^+_c \to
\Xi^0 K^{*+},\Sigma^0 K^{*+},\Lambda^0 K^{*+})
=(8.7 \pm 2.7,1.2\pm 0.3,2.0\pm 0.5)\times 10^{-3}$. For the $\Xi_c^+$ decays, we have obtained ${\cal B}(\Xi^+_c \to \Sigma^+\bar K^{*0},\Xi^{0} \rho^{+})
=(10.1\pm 2.9,9.9 \pm2.9)\times 10^{-2}$, ${\cal B}(\Xi^+_c \to p \bar K^{*0},\Sigma^+ \phi)
=(7.8 \pm 2.2,1.9\pm0.9)\times 10^{-3}$ and ${\cal B}( \Xi^+_c \to p\phi)=(1.5\pm 0.7)\times 10^{-4}$. The predicted ${\cal B}({\bf B}_c\to{\bf B}_n V)$ can be compared to the future measurements by BESIII, BELLEII and LHCb.
ACKNOWLEDGMENTS {#acknowledgments .unnumbered}
===============
This work was supported by National Science Foundation of China (11675030).
[99]{}
M. Tanabashi [*et al.*]{} \[Particle Data Group\], Phys. Rev. D [**98**]{}, 030001 (2018).
M. Ablikim [*et al.*]{} \[BESIII Collaboration\], arXiv:1811.08028 \[hep-ex\].
Y.B. Li [*et al.*]{} \[Belle Collaboration\], arXiv:1811.09738 \[hep-ex\].
C.Q. Geng, Y.K. Hsiao, C.W. Liu and T.H. Tsai, arXiv:1810.01079 \[hep-ph\].
R. Aaij [*et al.*]{} \[LHCb Collaboration\], arXiv:1901.06222 \[hep-ex\].
A. Ali, G. Kramer and C.D. Lu, Phys. Rev. D[**58**]{}, 094009 (1998).
C.Q. Geng, Y.K. Hsiao and J.N. Ng, Phys. Rev. Lett. [**98**]{}, 011801 (2007).
Y.K. Hsiao and C.Q. Geng, Phys. Rev. D [**91**]{}, 116007 (2015).
H.J. Zhao, Y.K. Hsiao and Y. Yao, arXiv:1811.07265 \[hep-ph\].
X.G. He, Y.K. Hsiao, J.Q. Shi, Y.L. Wu and Y.F. Zhou, Phys. Rev. D [**64**]{}, 034002 (2001).
H.K. Fu, X.G. He and Y.K. Hsiao, Phys. Rev. D [**69**]{}, 074002 (2004).
Y.K. Hsiao, C.F. Chang and X.G. He, Phys. Rev. D [**93**]{}, 114002 (2016).
X.G. He and G.N. Li, Phys. Lett. B [**750**]{}, 82 (2015).
M. He, X.G. He and G.N. Li, Phys. Rev. D [**92**]{}, 036010 (2015). M.J. Savage and R.P. Springer, Phys. Rev. D [**42**]{}, 1527 (1990).
M.J. Savage, Phys. Lett. B [**257**]{}, 414 (1991).
G. Altarelli, N. Cabibbo and L. Maiani, Phys. Lett. [**57B**]{}, 277 (1975).
X.G. He, Y.J. Shi and W. Wang, arXiv:1811.03480 \[hep-ph\].
C.D. Lu, W. Wang and F.S. Yu, Phys. Rev. D [**93**]{}, 056008 (2016).
W. Wang, Z.P. Xing and J. Xu, Eur. Phys. J. C [**77**]{}, 800 (2017). D. Wang, P.F. Guo, W.H. Long and F.S. Yu, JHEP [**1803**]{}, 066 (2018). C.Q. Geng, Y.K. Hsiao, Y.H. Lin and L.L. Liu, Phys. Lett. B [**776**]{}, 265 (2018). C.Q. Geng, Y.K. Hsiao, C.W. Liu and T.H. Tsai, Phys. Rev. D [**97**]{}, 073006 (2018). C.Q. Geng, Y.K. Hsiao, C.W. Liu and T.H. Tsai, Eur. Phys. J. C [**78**]{}, 593 (2018). C.Q. Geng, C.W. Liu and T.H. Tsai, Phys. Lett. B [**790**]{}, 225 (2019). A.J. Buras, hep-ph/9806471.
S. Fajfer, P. Singer and J. Zupan, Eur. Phys. J. C [**27**]{}, 201 (2003).
H.n. Li, C.D. Lu and F.S. Yu, Phys. Rev. D [**86**]{}, 036012 (2012).
J.M. Link [*et al.*]{} \[FOCUS Collaboration\], Phys. Lett. B [**571**]{}, 139 (2003). R. Aaij [*et al.*]{} \[LHCb Collaboration\], Phys. Rev. D [**97**]{}, 091101 (2018).
M. Ablikim [*et al.*]{} \[BESIII Collaboration\], Phys. Lett. B [**783**]{}, 200 (2018). Y. Kubota [*et al.*]{} \[CLEO Collaboration\], Phys. Rev. Lett. [**71**]{}, 3255 (1993).
R. Aaij [*et al.*]{} \[LHCb Collaboration\], Phys. Rev. D [**97**]{}, 091101 (2018).
D. Wang, arXiv:1901.01776 \[hep-ph\].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A $k$-tuple total dominating set ($k$TDS) of a graph $G$ is a set $S$ of vertices in which every vertex in $G$ is adjacent to at least $k$ vertices in $S$. The minimum size of a $k$TDS is called the $k$-tuple total dominating number and it is denoted by $\gamma_{\times k,t}(G)$. We give a constructive proof of a general formula for $\gamma_{\times 3, t}(K_n \Box K_m)$.'
address:
- 'Behnaz Pahlavsay, Department of Mathematics, Hokkaido University, Kita 10, Nishi 8, Kita-Ku, Sapporo 060-0810, Japan.'
- 'Elisa Palezzato, Department of Mathematics, Hokkaido University, Kita 10, Nishi 8, Kita-Ku, Sapporo 060-0810, Japan.'
- 'Michele Torielli, Department of Mathematics, GI-CoRE GSB, Hokkaido University, Kita 10, Nishi 8, Kita-Ku, Sapporo 060-0810, Japan.'
author:
- Behnaz Pahlavsay
- Elisa Palezzato
- Michele Torielli
title: '$3$-tuple total domination number of rook’s graphs'
---
Introduction
============
Domination is well-studied in graph theory and the literature on this subject has been surveyed and detailed in the two books by Haynes, Hedetniemi, and Slater [@HHS5; @HHS6]. Among the many variations of domination, the one relevant to this paper is $k$-tuple total domination, which was introduced by Henning and Kazemi [@HK8] as a generalization of [@HH3]. Throughout this paper, we use standard notation for graphs, see for example [@bondy2008graph]. All graphs considered here are finite, undirected, and simple.
For a graph $G=(V_G,E_G)$ and $k \geq 1$, a set $S \subseteq V_G$ is called a $k$-*tuple total dominating set* ($k$TDS) if every vertex $v \in V$ has at least $k$ neighbours in $S$, i.e., $|N_G(v)\cap S| \geq k$. The *$k$-tuple total domination number*, which we denote by $\gamma_{\times k,t}(G)$, is the minimum cardinality of a $k$TDS of $G$. We use min-$k$TDS to refer to $k$TDSs of minimum size.
An immediate necessary condition for a graph to have a $k$-tuple total dominating set is that every vertex must have at least $k$ neighbours. For example, for $k \geq 1$, a $k$-regular graph $G=(V_G,E_G)$ has only one $k$-tuple total dominating set, namely $V_G$ itself.
In the history of domination problems, a lot of work has been done to study the class of cartesian product of graphs and in particular of rook’s graphs. Given two graphs $G$ and $H$, their *Cartesian product* $G \Box H$ is the graph with vertex set $V_G\times V_H$ where two vertices $(u_{1},v_{1})$ and $(u_{2},v_{2})$ are adjacent if and only if either $u_{1}=u_{2}$ and $v_{1}v_{2}\in E_H$ or $v_{1}=v_{2}$ and $u_{1}u_{2}\in E_G$. For more information on the cartesian product of graphs see [@IK]. We will be particularly interested in the case when $K_n \Box K_m$, where $K_n$ is the complete graph on $n$ vertices. Such graph is known as the $n \times m$ *rook’s graph*, as edges represent possible moves by a rook on an $n \times m$ chess board. The $3 \times 4$ rook’s graph is drawn in Figure \[fi:rook34\], along with a min-$3$TDS.
; (11) to (12) to (13) to (14); (11) to\[bend left\] (13); (12) to\[bend left\] (14); (11) to\[bend left\] (14); (21) to (22) to (23) to (24); (21) to\[bend left\] (23); (22) to\[bend left\] (24); (21) to\[bend left\] (24); (31) to (32) to (33) to (34); (31) to\[bend left\] (33); (32) to\[bend left\] (34); (31) to\[bend left\] (34); (11) to (21) to (31) to\[bend left\] (11); (12) to (22) to (32) to\[bend left\] (12); (13) to (23) to (33) to\[bend left\] (13); (14) to (24) to (34) to\[bend left\] (14);
In [@Viz], Vizing studied the *domination number* of graphs, i.e. the minimal cardinality of a dominating set, and made an elegant conjecture that has subsequently become one the most famous open problems in domination theory.
For any graphs $G$ and $H$, $$\gamma(G)\gamma(H) \leq \gamma(G\Box H),$$ where $\gamma(G)$ and $\gamma(H)$ are the domination numbers of the graphs $G$ and $H$, respectively.
Over more than forty years (see [@BDG] and references therein), Vizing’s Conjecture has been shown to hold for certain restricted classes of graphs, and furthermore, upper and lower bounds on the inequality have gradually tightened. Additionally, researcher have explored inequalities (including Vizing-like inequalities) for different variations of domination [@HHS6]. A significant breakthrough occurred when in [@CS1] Clark and Suen proved that $$\gamma (G)\gamma (H)\leq 2\gamma (G\Box H)$$ which led to the discovery of a Vizing-like inequality for total domination [@HR9; @HPT], i.e., $$\label{eq:HRbound}
\gamma_t(G)\gamma_t(H) \leq 2 \gamma_t(G \Box H),$$ as well as for paired [@BHRo; @HJ; @CM], and fractional domination [@FRDM], and the $\{k\}$-domination function (integer domination) [@BM; @HL; @CMH], and total $\{k\}$-domination function [@HL].
Burchett, Lane, and Lachniet [@BLL] and Burchett [@B2011] found bounds and exact formulas for the $k$-tuple domination number and $k$-domination number of the rook’s graph in square cases, i.e., $K_n \Box K_n$ (where $k$-domination is similar to [$k$-tuple total domination]{}, but only vertices outside of the domination set need to be dominated). The $k$-tuple total domination number is known for $K_n \times K_m$ [@HK7] and bounds are given for supergeneralized Petersen graphs [@KP]. In [@KPS], the authors showed that the graph $K_n \Box K_m$ is an extremal case in the study of $k$TDS of cartesian product of graphs, motivating the study of the class of rook’s graphs. Specifically, they showed that $$\gamma_{\times k,t}(K_n \Box K_m) \leq \gamma_{\times k,t}(G \Box H),$$ when $G$ and $H$ are two graphs with $n$ and $m$ vertices, respectively. Moreover, they computed $\gamma_{\times 2, t}(K_n \Box K_m)$ for all $m\ge n$.
This paper is organized as follows. In Section 2, we recall basic properties on $k$TDS. In Section 3, we describe a special class of $3$TDS matrices. In Section 4, we describe several useful inequalities for $\gamma_{\times 3, t}(K_n \Box K_m)$. In Section 5, we compute $\gamma_{\times 3, t}(K_n \Box K_n)$, for any $n\ge3$. In Section 6, we describe our main result: we determine the value of $\gamma_{\times 3, t}(K_n \Box K_m)$ in Theorem \[theo:thegeneralcase\] for all $m\ge n$.
Preliminares
============
We recall some basic properties of $k$TDS and their relations with $(0,1)$-matrices. Assume the vertex set of the complete graph $K_n$ is $[n]:=\{1,\dots, n\}$. Given $D\subseteq V_{K_n}\times V_{K_m}$, we can associate to it a $n \times m$ $(0,1)$-matrix $S=(s_{ij})$ with $s_{ij}=1$ if and only if $(i,j) \in D$. Let $S=(s_{ij})$ be a $n \times m$ $(0,1)$-matrix. Define $$\begin{aligned}
\kappa_S(i,j) & =&\overbrace{\left( \textstyle\sum_{r \in [m]} s_{ir} \right)}^{\text{$i$-th row sum}} + \overbrace{\left( \textstyle\sum_{r \in [n]} s_{rj} \right)}^{\text{$j$-th column sum}} - 2s_{ij}\\
& =&\mathfrak{r}_S(i) + \mathfrak{c}_S(j) - 2s_{ij}.
\end{aligned}$$ If no confusion arises, we will simply write $\mathfrak{r}(i)$, $\mathfrak{c}(j)$ and $\kappa(i,j)$. Notice that $\mathfrak{r}(i)$ is the number of ones in the $i$-th row of $S$ and, similarly, $\mathfrak{c}(j)$ is the number of ones in the $j$-th column of $S$. Moreover, we will denote by $|S|$ the number of ones in $S$.
A $n \times m$ $(0,1)$-matrix $S=(s_{ij})$ corresponds to a $k$TDS $D$ of $K_n \Box K_m$ if and only it satisfies $$\kappa(i,j) \geq k$$ for all $i\in [n]$ and $j \in [m]$, which we call the *$\kappa$-bound*.
;
We call a $n \times m$ $(0,1)$-matrix $S$ a *$k$TDS matrix* if it satisfies the $\kappa$-bound for all $i \in [n]$ and $j \in [m]$. Furthermore, we call $S$ a *min-$k$TDS matrix* if it has exactly $\gamma_{\times k,t}(K_n \Box K_m)$ ones. Note that a $k$TDS matrix (respectively min-$k$TDS matrix) remains a $k$TDS matrix (respectively min-$k$TDS matrix) under permutations of its rows and/or columns.
\[lm:sparserow\] For $n \geq 1$ and $m \geq 1$, a $n \times m$ $k$TDS matrix with an all-$0$ column or an all-$0$ row has at least $kn$ or $km$ ones, respectively.
Let $S$ be a $n\times m$ $k$TDS matrix. Assume there exists $1\le j_0\le m$ such that $\mathfrak{c}(j_0)=0$. Then to achieve $\kappa(i,j_0) \geq k$ for any $i \in [n]$, we need $\mathfrak{r}(i)\ge k$. Since this is true for every row in $S$, we must have at least $kn$ ones. A similar argument works if there exists $1\le i_0\le n$ such that $\mathfrak{r}(i_0)=0$.
There are instances when $kn$ ones is the least number of ones in any $n \times m$ $k$TDS matrix. We establish some cases in the following proposition, see also Theorem 3.3 from [@KPS].
\[prop:manycols\] When $m \geq n \geq 2$ and $m \geq k$, $$\label{eq:Knmsimplebound}
\gamma_{\times k,t}(K_n \Box K_m) \leq kn$$ with equality when $m \geq kn-1$.
If $m \geq n \geq 2$ and $m \geq k$, the $n \times m$ $(0,1)$-matrix with ones in the last $k$ columns and zeros elsewhere is a $k$TDS matrix with $kn$ ones.
Assume $m \geq kn-1$ and let $S$ be a $n \times m$ $k$TDS matrix. If $S$ has a column of zeros, then $|S|\ge kn$ by Lemma \[lm:sparserow\]. If $S$ has no column of zeros but $m\ge kn$, then $|S|\ge kn$. Thus, assume $m=kn-1$ and $\mathfrak{c}(j)\ge1$ for all $1\le j\le m$. If $|S|< kn$, then $\mathfrak{c}(j)=1$ for all $1\le j\le m$. Therefore, if $s_{ij}=1$, then $\mathfrak{r}(i)\ge k+1$ to satisfy $\kappa(i,j) \geq k$. If this is true for every row, then $|S|\ge(k+1)n > kn$. Otherwise, there is a row of zeros, and Lemma \[lm:sparserow\] implies $|S|\ge km \geq kn$.
Motivated by [@BLL], given a $(0,1)$-matrix $S$, we can construct a graph $\Gamma(S)$ with vertices corresponding to the ones in $S$ and edges between $2$ ones belonging to the same row or column, if there are no other ones between them. The following gives one such example.
;
In this way, every $k$TDS matrix $S$ correspond to a graph, which has, in general, several (connected) components. If the set of vertices of a component of $\Gamma(S)$ is $\{(i_1,j_1),\dots (i_p,j_p)\}$, then we define the corresponding component of $S$ as the submatrix of $S$ formed by the intersection of rows $\{R_{i_1},\dots,R_{i_p}\}$ and columns $\{C_{j_1},\dots,C_{j_p}\}$, where $R_d$ and $C_d$ are the $d$-th row and column of $S$, respectively. We shade two components in the example above. In this example, the $5 \times 7$ matrix is the union of two components (a $4 \times 2$ component and a $1 \times 5$ component). A $k$TDS matrix $S$ with a component $H$, up to permutations of the rows and columns of $S$, looks like one of the following $$\begin{array}{|c|c|}
\hline
H & \emptyset \\
\hline
\emptyset & ? \\
\hline
\end{array},
\quad
\begin{array}{|c|}
\hline
H \\
\hline
\emptyset \\
\hline
\end{array},
\quad
\begin{array}{|c|c|}
\hline
H & \emptyset \\
\hline
\end{array},
\text{ or}
\quad
\begin{array}{|c|}
\hline
H \\
\hline
\end{array},$$ where the question mark ($?$) denotes some $(0,1)$-submatrix, and $\emptyset$ denotes an all-$0$ submatrix. Components of $k$TDS matrices have the following properties
- components have no all-$0$ rows and no all-$0$ columns,
- components are $k$TDS matrices in their own right.
\[rem:2onesrowocol\] If $S$ is a $3$TDS matrix with no all-$0$ rows and no all-$0$ columns, then in order to achieve the $3$-bound, it has at least $2$ ones in each row or in each column. Moreover, if $S$ has at least $2$ ones in each row (or column), the same is true for each of its components. Since we are interested in the study of rook’s graphs with $m\ge n$, we will assume that each $3$TDS matrix has at least $2$ ones in each row
In order to describe our main result, we will need the following $3$TDS matrices. For $x \geq 1$ and $y \geq 1$ such that $x+y\ge5$, we define $J(x,y)$ as the $x \times y$ all-$1$ matrix.
For $x\ge5$, let $D(x,3)$ be the $x\times3$ $3$TDS matrix whose first $x-3$ rows coincide with $(0,1,1)$, the $(x-2)$-th and $(x-1)$-th rows coincide with $(1,0,1)$, and the last row coincides with $(1,1,0)$. Depicted below we have the matrix $D(x,3)$ for $x \in \{5, 6, 7\}$
$\begin{array}{ccc}
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=black!60]| & |[fill=white]| & |[fill=black!60]| \\
|[fill=black!60]| & |[fill=white]| & |[fill=black!60]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=white]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=black!60]| & |[fill=white]| & |[fill=black!60]| \\
|[fill=black!60]| & |[fill=white]| & |[fill=black!60]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=white]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=black!60]| & |[fill=white]| & |[fill=black!60]| \\
|[fill=black!60]| & |[fill=white]| & |[fill=black!60]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=white]| \\
};
\end{tikzpicture}
\end{array}$
On the construction of special $3$TDS matrices
==============================================
We describe how to construct a special class of $3$TDS matrices, looking with particular attention at the shape of their components. Moreover, we compute the number of ones in such matrices. Notice that these matrices are exactly the ones appearing in Table \[ta:smallk3\].
\[prop:compclass2\] For any $m\ge n\ge 6$, except $(n,m)=(6,6)$, there exists a $n \times m$ $3$TDS matrix $S$ with no all-$0$ rows and no all-$0$ columns with at least $2$ ones in each row whose components, up to permutations of the rows and columns, are all $J(1,4)$ or $J(3,2)$, except possibly for
- exactly one $J(4,2)$ component;
- exactly one $J(1,y)$ component with $5 \leq y \leq 6$;
- exactly one $D(5,3)$ component but no $J(4,2)$ component;
- exactly one $D(6,3)$ component and no $D(5,3)$ component or $J(4,2)$ component or $J(1,y)$ component with $5 \leq y \leq 6$.
Notice that by Table \[ta:smallk3\], it is enough to show that
- if $S$ is a $n\times m$ $3$TDS matrix with the properties we require, then we have a way to construct $S'$ a $n\times (m+1)$ $3$TDS matrix with the same properties;
- if $S$ is a $n\times (n+1)$ $3$TDS matrix with the properties we require, then we have a way to construct $S'$ a $(n+1)\times (n+1)$ $3$TDS matrix with the same properties.
Let now $S$ be a $n\times m$ $3$TDS matrix with the properties we require. We will apply the following rules to obtain $S'$ a $n\times (m+1)$ $3$TDS matrix with the same properties.
1. If $S$ contains a $J(1,6)$ component and a $D(5,3)$ component, we obtain $S'$ by transforming these two components in two $J(1,4)$ components and a $J(4,2)$ component, and leaving the other components unchanged.
2. If $S$ contains a $J(1,4)$ component and a $D(6,3)$ component, we obtain $S'$ by transforming these two components in one $J(1,4)$ component and two $J(3,2)$ components, and leaving the other components unchanged.
3. If $S$ contains a $J(1,6)$ component, a $J(4,2)$ component and a $J(3,2)$ component, we obtain $S'$ by transforming these three components in two $J(1,4)$ components and a $D(6,3)$ component, and leaving the other components unchanged.
4. If $S$ contains a $J(1,6)$ component, two $J(3,2)$ components and no $J(4,2)$ or $D(5,3)$ components, we obtain $S'$ by transforming these three components in two $J(1,4)$ components and a $D(5,3)$ component, and leaving the other components unchanged.
5. In all other cases, since $S$ always contains at least one $J(1,y)$ component with $4 \leq y \leq 5$, we obtain $S'$ by transforming the $J(1,y)$ component in a $J(1,y+1)$ component, and leaving the other components unchanged.
Notice that if $S$ has only one $J(1,6)$ component, only one $J(4,2)$ component, or only one $J(3,2)$ component, then $n=4,5$. Let now $S$ be a $n\times (n+1)$ $3$TDS matrix with the properties we require. We will apply the following rules to obtain $S'$ a $(n+1)\times (n+1)$ $3$TDS matrix with the same properties.
1. If $S$ contains a $J(1,y)$ component, with $y=5,6$ and a $J(4,2)$ components, we obtain $S'$ by transforming these two components in a $J(1,y-1)$ component and a $D(5,3)$ component, and leaving the other components unchanged.
2. If $S$ contains a $J(1,y)$ component, with $y=5,6$ and a $D(5,3)$ component, we obtain $S'$ by transforming these two components in a $J(1,y-1)$ component and two $J(3,2)$ components, and leaving the other components unchanged.
3. If $S$ contains two $J(1,4)$ components and a $D(6,3)$ components, we obtain $S'$ by transforming these three components in a $J(1,6)$ component, a $J(3,2)$ component and a $D(5,3)$ component, and leaving the other components unchanged.
4. If $S$ contains two $J(1,4)$ components, a $J(4,2)$ component and no $J(1,y)$ components, with $y=5,6$, we obtain $S'$ by transforming these three components in a $J(1,6)$ component and two $J(3,2)$ components, and leaving the other components unchanged.
5. If $S$ contains a $D(5,3)$ components and no $J(1,y)$ components, with $y=5,6$, we obtain $S'$ by transforming this component in a $D(6,3)$ component, and leaving the other components unchanged.
6. In all other cases, since $S$ always contains at least one $J(3,2)$ component, we obtain $S'$ by transforming this component in a $J(4,2)$ component, and leaving the other components unchanged.
Notice that if $S$ has only one $J(1,4)$ component, only one $J(4,2)$ component and no $J(1,y)$ components, with $y=5,6$, then $n=5$, or $n=m=8$ or $n> m$. Similarly, if $S$ has only one $J(1,4)$ component and one $D(6,3)$ component, then $n=m=7$ or $n> m$.
We can now compute the number of ones in a $3$TDS matrix satisfying the requirements of the previous proposition.
\[prop:numberofones\] For any integer $m\ge n \ge 6$, except $(n,m)=(6,6)$, let $2n\equiv 3m+k \pmod {10}$, where $0 \le k \le 9$. Then the number of ones in a matrix of Proposition \[prop:compclass2\] is given by $$\begin{cases}
\lceil \frac{8n+3m}{5} \rceil & \text{ if $k=0,1,2,3,4,7,8,9;$} \\
\lceil \frac{8n+3m}{5} \rceil +1 & \text{ if $k=5,6.$}
\end{cases}$$
Let $S$ be a $n\times m$ $3$TDS matrix with no all-$0$ rows and no all-$0$ columns with at least $2$ ones in each row as described in Proposition \[prop:compclass2\]. Let $a$ be the number of $J(1,4)$ components in $S$ and let $b$ be the number of $J(3,2)$ components in $S$. To prove our statement we have to analyze six cases.
*Case I*: Assume $S$ has only $J(1,4)$ and $J(3,2)$ components. Then $$\begin{aligned}
n &= a+3b, \\
m &= 4a+2b,\end{aligned}$$ and the number of ones in $S$ is $(4a+6b)=(8n+3m)/5$. In this case, we have $2n\equiv 3m \pmod {10}$.
*Case II*: Assume $S$ has $J(1,4)$ components, $J(3,2)$ components and one $J(4,2)$ component. Then $$\begin{aligned}
n &= a+3b+4, \\
m &= 4a+2b+2,\end{aligned}$$ and the number of ones in $S$ is $(4a+6b+8)=(8n+3m+2)/5= \lceil (8n+3m)/5 \rceil$. In this case, we have $2n\equiv 3m+2 \pmod {10}$.
*Case III*: Assume $S$ has $J(1,4)$ components, $J(3,2)$ components and one $J(1,y)$ component with $5 \leq y \leq 6$. We have $$\begin{aligned}
n &= a+3b+1, \\
m &= 4a+2b+y,\end{aligned}$$ and the number of ones in $S$ is $$\begin{aligned}
(4a+6b)+y &= (8n+3m+2y-8)/5 \\
&=
\begin{cases}
(8n+3m+2)/5= \lceil (8n+3m)/5 \rceil & \text{ if $y=5;$} \\
(8n+3m+4)/5= \lceil (8n+3m)/5 \rceil & \text{ if $y=6.$}
\end{cases}\end{aligned}$$ In this case we have $2n \equiv 3m-3y+2 \pmod {10}$, i.e., $2n \equiv 3m+7, 3m+4 \pmod {10}$ when $y=5,6 $, respectively.
*Case IV*: Assume $S$ has $J(1,4)$ components, $J(3,2)$ components, a $J(1,y)$ component with $5 \leq y \leq 6$ and a $J(4,2)$ component. We have $$\begin{aligned}
n &= a+3b+5, \\
m &= 4a+2b+y+2,\end{aligned}$$ and the number of ones in $S$ is $$\begin{aligned}
(4a+6b)+y+8 &= (8n+3m+2y-6)/5 \\
&=
\begin{cases}
(8n+3m+4)/5= \lceil (8n+3m)/5 \rceil & \text{ if $y=5;$} \\
(8n+3m+6)/5= \lceil (8n+3m)/5 \rceil+1 & \text{ if $y=6.$}
\end{cases}\end{aligned}$$ In this case we have $2n \equiv 3m-3y+4 \pmod {10}$, i.e., $2n \equiv 3m+9, 3m+6 \pmod {10}$ when $y=5,6 $, respectively.
*Case V*: Assume $S$ has $J(1,4)$ components, $J(3,2)$ components and one $D(x,3)$ component with $5 \leq x \leq 6$. We have $$\begin{aligned}
n &= a+3b+x, \\
m &= 4a+2b+3,\end{aligned}$$ and the number of ones in $S$ is $$\begin{aligned}
(4a+6b)+2x &= (8n+3m+2x-9)/5 \\
&=
\begin{cases}
(8n+3m+1)/5= \lceil (8n+3m)/5 \rceil & \text{ if $x=5;$} \\
(8n+3m+3)/5= \lceil (8n+3m)/5 \rceil & \text{ if $x=6.$}
\end{cases}\end{aligned}$$ In this case we have $2n \equiv 3m+2x+1 \pmod {10}$, i.e., $2n \equiv 3m+1, 3m+3 \pmod {10}$ when $x=5,6 $, respectively.
*Case VI*: Assume $S$ has $J(1,4)$ components, $J(3,2)$ components, a $J(1,y)$ component with $5 \leq y \leq 6$ and a $D(5,3)$ component. We have $$\begin{aligned}
n &= a+3b+6,\\
m &= 4a+2b+y+3,\end{aligned}$$ and the number of ones in $S$ is $$\begin{aligned}
(4a+6b)+10+y &= (8n+3m+2y-7)/5 \\
&=
\begin{cases}
(8n+3m+3)/5= \lceil (8n+3m)/5 \rceil & \text{ if $y=5;$} \\
(8n+3m+5)/5= \lceil (8n+3m)/5 \rceil+1 & \text{ if $y=6.$}
\end{cases}\end{aligned}$$ In this case we have $2n \equiv 3m-3y+3 \pmod {10}$, i.e., $2n \equiv 3m+8, 3m+5 \pmod {10}$ when $y=5,6 $, respectively.
\[rem:n45onesrow\] A direct computation shows that when $n\in\{4,5\}$ and $(n,m)=(6,6)$, we can compute the number of ones of the matrices in Table \[ta:smallk3\] with no all-$0$ rows and no all-$0$ columns with the formula of Proposition \[prop:numberofones\].
Useful inequalities for min-$3$TDS
==================================
We prove several inequalities for $\gamma_{\times 3,t}(K_n \Box K_m)$. Specifically, we show how $\gamma_{\times 3,t}$ changes when, in a $3$TDS matrix, we increase the number of rows or columns in the general case, or both in the square case. The first lemma describes a lower bound for the number of ones in a $3$TDS matrix.
\[lemma:nnvs2n+1\] Let $m\ge n\ge3$. Then $\gamma_{\times 3,t}(K_n \Box K_m)\ge2n+2$.
Suppose that $\gamma_{\times 3,t}(K_n \Box K_m)\le2n+1$ and let $S$ be a $n\times m$ $3$TDS matrix with $|S|=2n+1$. Since $2n+1<3n$, by Lemma \[lm:sparserow\], $S$ has no all-$0$ rows or all-$0$ columns. Since by Remark \[rem:2onesrowocol\] we can assume that $\mathfrak{r}(i)\ge2$ for all $1\le i\le n$, then $S$ has one row with $3$ ones and $n-1$ rows with $2$ ones. Without loss of generality, we can assume that the first row of $S$ has $3$ ones in the first three entries. As a consequence, $\mathfrak{r}(i)=2$ for all $2\le i\le n$. For all $1\le j\le3$, $\kappa(1,j)=3+\mathfrak{c}(j)-2=\mathfrak{c}(j)+1\ge3$, and hence $\mathfrak{c}(j)\ge 2$, i.e. each of the first three columns of $S$ has at least $2$ ones. Moreover, if $m\ge4$, since $S$ has no all-$0$ rows or all-$0$ columns, for all $4\le j\le m$ there must exists $2\le i\le n$ such that $S$ has a one in position $(i,j)$. Then $\kappa(i,j)=2+\mathfrak{c}(j)-2=\mathfrak{c}(j)\ge3$, i.e. each of the last $m-3$ columns of $S$ has at least $3$ ones.
Assume $n=3$. If $m\ge4$, since $\mathfrak{c}(j)\ge 2$ for all $1\le j\le 3$ and $\mathfrak{c}(j)\ge 3$ for all $4\le j\le m$, then $|S|\ge6+3(m-3)=3m-3>2n-1$. We can then assume that $m=3$ and that the zeros of $S$ are in position $(2,2)$ and $(3,1)$. However, $\kappa(2,1)=2$ and so $S$ is not a $3$TDS matrix.
Assume now $n=4$. Since $\mathfrak{c}(4)\ge3$, the last column of $S$ is $(0,1,1,1)^t$. Hence, we can assume that the remaining ones of $S$ are in position $(2,1)$, $(3,2)$ and $(4,3)$. However, $\kappa(2,1)=2$ and so $S$ is not a $3$TDS matrix.
Assume now $n\ge5$. Counting the ones of $S$ by columns we obtain that $|S|\ge 6+3(m-3)=3m-3$. However, since $m\ge n\ge5$, $3m-3>2n+1$ and so $S$ is not a $3$TDS matrix.
\[rem:smallcasesdiag\] If $3\le n\le 10$, then by Lemma \[lemma:nnvs2n+1\], the $n\times n$ $3$TDS matrix of Table \[ta:smallk3\] are min-$3$TDS and so $\gamma_{\times 3,t}(K_n \Box K_n)=2n+2$.
We are now able to compute $\gamma_{\times 3,t}(K_3 \Box K_m)$ for all $m\ge3$.
\[lemma:n3case3tds\] If $m\ge3$, then $$\gamma_{\times 3,t}(K_3 \Box K_m)=
\begin{cases}
8 & \text{ if $m=3,4;$} \\
9 & \text{ if $m\ge 5.$}
\end{cases}$$
By Lemma \[lemma:nnvs2n+1\], $\gamma_{\times 3,t}(K_3 \Box K_m)\ge8$. Looking at the $3$TDS matrices in Table \[ta:smallk3\], we obtain that $\gamma_{\times 3,t}(K_3 \Box K_3)=\gamma_{\times 3,t}(K_3 \Box K_4)=8$.
Let now $m\ge5$. Suppose that there exists $S$ a $3\times m$ $3$TDS matrix with $|S|=8$. By Remark \[rem:2onesrowocol\], this implies that there exists $1\le i\le 3$ such that $\mathfrak{r}(i)=2$. Without loss of generality we can assume that $i=3$ and that the last row of $S$ coincides with $(0,\dots,0,1,1)$. If $m-1\le j\le m$, then $\kappa(3,j)=2+\mathfrak{c}(j)-2\ge3$. This implies that $\mathfrak{c}(m-1)=\mathfrak{c}(m)=3$. Moreover, by Lemma \[lm:sparserow\], $S$ has no all-$0$ rows or all-$0$ columns, and hence $\mathfrak{c}(j)\ge1$ for all $1\le j\le m-2$. This implies that $|S|=\sum_{j=1}^m\mathfrak{c}(j)\ge (m-2)+6>8$, but this is a contraddiction.
The following result describes the relation between min-$3$TDS matrices that have the same number of rows but whose number of columns differs by one.
\[lemma:3dtsgrowsbyrow\] Let $m\ge n\ge3$. Then $$\gamma_{\times 3,t}(K_n \Box K_m)\le \gamma_{\times 3,t}(K_n \Box K_{m+1})\le \gamma_{\times 3,t}(K_n \Box K_m)+1.$$
Firstly we will prove the first inequality, i.e. we will prove that $\gamma_{\times 3,t}(K_n \Box K_m)\le \gamma_{\times 3,t}(K_n \Box K_{m+1})$.
If $\gamma_{\times 3,t}(K_n \Box K_{m+1})=3n$, the first inequality holds by Proposition \[prop:manycols\]. Let $S$ be a $n\times (m+1)$ $3$TDS matrix with $|S|=\gamma_{\times 3,t}(K_n \Box K_m)-1<3n$. By Lemma \[lm:sparserow\], $S$ has no all-$0$ rows or all-$0$ columns. Furthermore, since $|S|<3n$, then there exists $1\le j\le m+1$ such that $\mathfrak{c}_S(j)\le2$. Hence, without loss of generality, we can assume that $j=m+1$. This fact is crucial for the rest of the proof. If $n=3$, the first inequality holds by Lemma \[lemma:n3case3tds\]. Assume now $m\ge n \ge4$. If $\mathfrak{c}_S(m+1)=1$, then we can assume that the last column of $S$ is $(1,0,\dots,0)^t$. Since $\kappa_S(1,m+1)=\mathfrak{r}_S(1)+1-2\ge3$, then $\mathfrak{r}_S(1)\ge 4$. Consider $S'$ the matrix obtained from $S$ by deleting the last column. Notice that $\mathfrak{r}_{S'}(1)\ge 3$. $S'$ is a $n\times m$ matrix with $|S|-1=\gamma_{\times 3,t}(K_n \Box K_m)-2$ ones, and hence $S'$ is not a $3$TDS matrix, by definition of $\gamma_{\times 3,t}$. However, $\kappa_{S'}(i,j)=\kappa_S(i,j)\ge3$, if $2\le i\le n$ and $1\le j\le m$. Since $S'$ is not a $3$TDS matrix, there exists $1\le j\le m$ such that $\kappa_{S'}(1,j)\le2$. This implies that $\mathfrak{r}_{S'}(1)= 3$ and so that $\mathfrak{r}_S(1)= 4$. Since $m+1\ge5$, then the first row of $S$ has at least one zero in the first $m$ entries. We can construct $S''$ a $n\times m$ matrix obtained from $S$ by deleting the last column and putting exactly $1$ one in one of the zeros of the first row. By construction $S''$ is a $n\times m$ $3$TDS matrix with $|S|=\gamma_{\times 3,t}(K_n \Box K_m)-1$ ones, but this is a contradiction.
If $\mathfrak{c}_S(m+1)=2$, then we can assume that the last column of $S$ is equal to $(1,1,0,\dots,0)^t$. Let $S'$ be the matrix obtained from $S$ by deleting the last column. $S'$ is a $n\times m$ matrix with $|S|-2=\gamma_{\times 3,t}(K_n \Box K_m)-3$ ones, and hence it is not a $3$TDS matrix. However, $\kappa_{S'}(i,j)=\kappa_S(i,j)\ge3$, if $3\le i\le n$ and $1\le j\le m$. This implies that at least one of the first two rows of $S$ have exactly $3$ ones. Assume it is the first one. Since $m+1\ge5$, then the first row of $S$ has at least $2$ zeros in the first $m$ entries. If any of the first $m$ columns of $S$ have $2$ zeros in the first two rows, we can construct $S''$ a $n\times m$ matrix obtained from $S$ by deleting the last column and putting 2 ones in the first two entries of such column. If such column does not exist, we can construct $S''$ a $n\times m$ matrix obtained from $S$ by deleting the last column and putting exactly $1$ one in one zero of the first row and, if the second row has a zero, $1$ one there. By construction $S''$ is a $n\times m$ $3$TDS matrix with at most $|S|=\gamma_{\times 3,t}(K_n \Box K_m)-1$ ones, but this is a contradiction. This proves the first inequality.
We are now ready to prove the second inequality, i.e. to prove that $\gamma_{\times 3,t}(K_n \Box K_{m+1})\le \gamma_{\times 3,t}(K_n \Box K_m)+1$. Let $S$ be a minimum $n\times m$ $3$TDS matrix. By Lemma \[lemma:nnvs2n+1\], we have that $\gamma_{\times 3,t}(K_n \Box K_m)\ge 2n+2$ and hence there exists $1\le i\le n$ such that $\mathfrak{r}_S(i)\ge3$. Without loss of generality we can assume that $i=n$. Consider now $S'$ a $n\times (m+1)$ matrix such that the first $m$ columns coincide with $S$ and the last column is $(0,\dots,0,1)^t$. By construction $S'$ is a $n\times (m+1)$ $3$TDS matrix with $|S|+1=\gamma_{\times 3,t}(K_n \Box K_m)+1$ ones.
The next lemma describes the relation between min-$3$TDS matrix that have the same number of columns but whose number of rows differs by one.
\[lemma:3dtsgrowsbycolumn\] Let $m> n\ge3$, and assume $\gamma_{\times 3,t}(K_n \Box K_m)< 3n$.Then $$\gamma_{\times 3,t}(K_n \Box K_m)\le \gamma_{\times 3,t}(K_{n+1} \Box K_m)\le \gamma_{\times 3,t}(K_n \Box K_m)+2.$$
Firstly we will prove the first inequality, i.e. we will prove that $\gamma_{\times 3,t}(K_n \Box K_m)\le \gamma_{\times 3,t}(K_{n+1} \Box K_m)$.
Let $S$ be a $(n+1)\times m$ $3$TDS matrix with $|S|=\gamma_{\times 3,t}(K_n \Box K_m)-1$. Since $|S|<3n$, by Remark \[rem:2onesrowocol\] there exists $1\le i\le n+1$ such that $\mathfrak{r}_S(i)=2$. Without loss of generality, we can assume that $i=n+1$ and that the last row of $S$ is $(0,\dots,0,1,1)$. Consider $S'$ the matrix obtained from $S$ by deleting the last row. $S'$ is a $n\times m$ matrix with $|S|-2=\gamma_{\times 3,t}(K_n \Box K_m)-3$ ones, and then it is not a $3$TDS matrix. However, $\kappa_{S'}(i,j)=\kappa_S(i,j)\ge3$, if $1\le i\le n$ and $1\le j\le m-2$. This implies that at least one of the last two columns of $S$ have exactly $3$ ones. Assume that this column is the last of $S$. Since $n+1\ge4$, the last column of $S$ has at least one zero in the first $n$ entries. If any of the first $n$ rows of $S$ have $2$ zeros in the last two columns, we can construct $S''$ a $n\times m$ matrix obtained from $S$ by deleting the last row and putting $2$ ones in the last two entries of such row. If such row does not exist but the penultimate column of $S$ has a zero, we can construct $S''$ a $n\times m$ matrix obtained from $S$ by deleting the last row and putting exactly $1$ one in one zero of the penultimate column and exactly $1$ one in one zero of the last column. If the penultimate column has no zero, we can construct $S''$ a $n\times m$ matrix obtained from $S$ by deleting the last row and putting exactly $1$ one in one zero of the last column. By construction $S''$ is a $n\times m$ $3$TDS matrix with at most $|S|=\gamma_{\times 3,t}(K_n \Box K_m)-1$ ones, but this is a contradiction. This proves the first inequality.
We are now ready to prove the second inequality, i.e. to prove that $\gamma_{\times 3,t}(K_{n+1} \Box K_m)\le \gamma_{\times 3,t}(K_n \Box K_m)+2$. Let $S$ be a minimum $n\times m$ $3$TDS matrix. By assumption, we have that $\gamma_{\times 3,t}(K_n \Box K_m)< 3n$ and hence there exists $1\le i\le n$ such that $\mathfrak{r}_S(i)=2$. Without loss of generality we can assume that $i=n$ and that the last row of $S$ coincides with $(0,\dots,0,1,1)$. Consider now $S'$ a $(n+1)\times m$ matrix such that the first $n$ rows coincide with $S$ and the last row is $(0,\dots,0,1,1)$. By construction $S'$ is a $(n+1)\times m$ $3$TDS matrix with $|S|+2=\gamma_{\times 3,t}(K_n \Box K_m)+2$ ones.
We now describes the relation between square min-$3$TDS matrix whose number of rows and columns both differ by one.
\[lemma:3dtsgrowsbydiagonal\] Let $n\ge3$. Then $$\gamma_{\times 3,t}(K_n \Box K_n)+2 \le \gamma_{\times 3,t}(K_{n+1} \Box K_{n+1})\le \gamma_{\times 3,t}(K_n \Box K_n)+3.$$
Firstly, we will prove the first inequality, i.e. we will prove that $\gamma_{\times 3,t}(K_n \Box K_n)+2\le \gamma_{\times 3,t}(K_{n+1} \Box K_{n+1})$.
If $n=3,4$, the first inequality follows from Remark \[rem:smallcasesdiag\]. Assume $n\ge5$. Suppose there exists $S$ a $(n+1)\times (n+1)$ $3$TDS matrix with $|S|=\gamma_{\times 3,t}(K_n \Box K_n)+1$. By Remark \[rem:2onesrowocol\] and $\gamma_{\times 3,t}(K_n \Box K_n)+1<3(n+1)$, there exists $1\le i\le n+1$ such that $\mathfrak{r}_S(i)=2$. Without loss of generality we can assume that $i=n+1$ and that the last row of $S$ coincides with $(0,\dots,0,1,1)$. Let now $S'$ be the $n\times (n+1)$ matrix obtained from $S$ by deleting the last row. $S'$ has $|S|-2=\gamma_{\times 3,t}(K_n \Box K_n)-1$ ones, and hence it is not a $3$TDS by Lemma \[lemma:3dtsgrowsbyrow\]. However, $\kappa_{S'}(i,j)=\kappa_S(i,j)\ge3$ for all $1\le i \le n$ and $1\le j\le n-1$. Since $S$ is a $3$TDS matrix, if $n\le j\le n+1$, then $\kappa_S(n+1,j)=2+\mathfrak{c}_{S}(j)-2=\mathfrak{c}_{S}(j)\ge3$, and hence $\mathfrak{c}_{S'}(j)\ge2$. However, since $S'$ is not a $3$TDS matrix, there exist $1\le i\le n$ and $n\le j\le n+1$ such that $\kappa_{S'}(i,j)=2$, and hence $\mathfrak{c}_{S'}(n)=2$ or $\mathfrak{c}_{S'}(n+1)=2$. Without loss of generality, we can assume that $(i,j)=(1,n+1)$, and hence that the last column of $S'$ is $(1,1,0,\dots,0)^t$ and $\mathfrak{r}_{S'}(1)=2$.
Assume that $\mathfrak{c}_{S'}(n)\ge3$. If in $S'$ there is a column with $2$ zeros in the first two entries, we can construct $S''$ a $n\times n$ matrix obtained from $S'$ by deleting the last column and putting $2$ ones in the first two entries of such column. If such column does not exist, then $\mathfrak{r}_{S'}(2)\ge4$. Furthermore, since $|S|<3(n+1)$, there must exists a column with $2$ zeros, one in the first entry and the second one in the $j$-th position, for some $j\ge3$. We can construct $S''$ a $n\times n$ matrix obtained from $S'$ by deleting the last column and putting $1$ one in the first entry and $1$ one in the $j$-th position of such column. By construction $S''$ is a $n\times n$ $3$TDS matrix with $|S'|=\gamma_{\times 3,t}(K_n \Box K_n)-1$ ones, but this is a contradiction.
Assume now that $\mathfrak{c}_{S'}(n)=2$. Denote by $w$ the penultimate column of $S'$. There are four cases. If $w$ has $2$ zeros in the first two entries, then we can construct $S''$ a $n\times n$ matrix obtained from $S'$ by deleting the last column and putting $2$ ones in the first two entries of $w$. By construction $S''$ is a $n\times n$ $3$TDS matrix with $|S'|=\gamma_{\times 3,t}(K_n \Box K_n)-1$ ones, but this is a contradiction.
If $w=(1,0,\dots)^t$, then the first row of $S'$ is equal to $(0,\dots,0,1,1)$. Since $|S|<3(n+1)$, there must exists $1\le j\le n-1$ such that $\mathfrak{c}_{S'}(j)=\mathfrak{c}_S(j)\ge2$. We can construct $S''$ a $n\times n$ matrix obtained from $S'$ by deleting the last column and putting $1$ one in position $(1,j)$ and $1$ one in position $(2,n)$. By construction $S''$ is a $n\times n$ $3$TDS matrix with $|S'|=\gamma_{\times 3,t}(K_n \Box K_n)-1$ ones, but this is impossible.
Assume $w=(0,1,\dots)^t$. If $\mathfrak{r}_{S'}(2)=2$, then the second row of $S'$ is equal to $(0,\dots,0,1,1)$. Since $|S|<3(n+1)$, there must exists $1\le j\le n-1$ such that $\mathfrak{c}_{S'}(j)\ge2$. We can construct $S''$ a $n\times n$ matrix obtained from $S'$ by deleting the last column and putting $1$ one in position $(1,n)$ and $1$ one in position $(2,j)$. By construction $S''$ is a $n\times n$ $3$TDS matrix with $|S'|=\gamma_{\times 3,t}(K_n \Box K_n)-1$ ones, but this is a contradiction. If $\mathfrak{r}_{S'}(2)\ge3$, but there is at least one zero in the second row of $S'$, we can construct $S''$ a $n\times n$ matrix obtained from $S'$ by deleting the last column and putting $1$ one in position $(1,n)$ and exactly $1$ one in one zero of the second row. By construction $S''$ is a $n\times n$ $3$TDS matrix with $|S'|=\gamma_{\times 3,t}(K_n \Box K_n)-1$ ones, but this is a contradiction. If $\mathfrak{r}_{S'}(2)=n+1$, we can construct $S''$ a $n\times n$ matrix obtained from $S'$ by deleting the last column and putting $1$ one in position $(1,n)$ and exactly $1$ one in one zero of the first row. By construction $S''$ is a $n\times n$ $3$TDS matrix with $|S'|=\gamma_{\times 3,t}(K_n \Box K_n)-1$ ones, and this is a contradiction.
If $w=(1,1,0,\dots,0)^t$, then the first row of $S'$ is equal to $(0,\dots,0,1,1)$. Since $|S|<3(n+1)$, there must exists $1\le j\le n-1$ such that $\mathfrak{c}_{S'}(j)\ge2$, and we can assume that $S'$ has ones in positions $(p,j)$ and $(q,j)$, for some $2\le p< q \le n$. If $\mathfrak{r}_{S'}(2)=2$, then the second row of $S'$ is equal to $(0,\dots,0,1,1)$. This implies that the first two entries of the $j$-th column of $S'$ are zero. We can construct $S''$ a $n\times n$ matrix obtained from $S'$ by deleting the last column and putting $1$ one in position $(1,j)$, $1$ one in position $(2,j)$, $1$ one in position $(p,n)$ and putting $1$ zero in position $(p,j)$. By construction $S''$ is a $n\times n$ $3$TDS matrix with $|S'|=\gamma_{\times 3,t}(K_n \Box K_n)-1$ ones, however this is a contradiction. If $\mathfrak{r}_{S'}(2)=3$, then, without loss of generality, we can assume that the second row of $S'$ is equal to $(0,\dots,0,1,1,1)$. Since $\kappa_{S'}(2,n-1)=\kappa_{S}(2,n-1)\ge3$, this implies that $\mathfrak{c}_{S'}(n-1)\ge2$, and we can assume that $S'$ has ones in positions $(2,n-1)$ and $(i,n-1)$, with $3\le i\le n$. We can construct $S''$ a $n\times n$ matrix obtained from $S'$ by deleting the last column and putting $1$ one in position $(1,n-1)$ and $1$ one in position $(i,n)$. By construction $S''$ is a $n\times n$ $3$TDS matrix with $|S'|=\gamma_{\times 3,t}(K_n \Box K_n)-1$ ones, but this is a contradiction. If $\mathfrak{r}_{S'}(2)\ge4$, then, without loss of generality, we can assume that the second row of $S'$ is equal to $(\dots,1,1,1,1)$. We can construct $S''$ a $n\times n$ matrix obtained from $S'$ by deleting the last column and putting $1$ one in position $(1,n-1)$ and $1$ one in position $(1,n-2)$. By construction $S''$ is a $n\times n$ $3$TDS matrix with $|S'|=\gamma_{\times 3,t}(K_n \Box K_n)-1$ ones, but this is a contradiction.
We are now ready to prove the second inequality, i.e. to prove that $\gamma_{\times 3,t}(K_{n+1} \Box K_{n+1})\le \gamma_{\times 3,t}(K_n \Box K_n)+3$. By Proposition \[prop:numberofones\], Remark \[rem:n45onesrow\] and Lemma \[lemma:n3case3tds\], $\gamma_{\times 3,t}(K_{n} \Box K_{n+1})<3n$. By Lemmas \[lemma:3dtsgrowsbyrow\] and \[lemma:3dtsgrowsbycolumn\], $\gamma_{\times 3,t}(K_{n+1} \Box K_{n+1})\le \gamma_{\times 3,t}(K_{n} \Box K_{n+1})+2 \le \gamma_{\times 3,t}(K_n \Box K_n)+3.$
\[rem:onesbycolumnsvsrowk\] Let $m\ge n\ge3$ and $S$ a $n\times m$ $3$TDS matrix with no all-$0$ columns or all-$0$ row. Let $1\le k\le n-1$ and assume that $S$ has $2n+k$ ones. Counting by column, this implies that $S$ has at most $k+2\lfloor\frac{k}{2}\rfloor$ columns that contain a one belonging to a row with at least $3$ ones. Hence all the other columns contain at least $1$ one belonging to a row with $2$ ones, and so such columns all have at least $3$ ones. This shows that $|S|\ge(k+2\lfloor\frac{k}{2}\rfloor) +3(m-k-2\lfloor\frac{k}{2}\rfloor)=3m-2k-4\lfloor\frac{k}{2}\rfloor$.
The square case
===============
We consider the case when $n=m$ and we give an explicit formula for $\gamma_{\times 3,t}(K_n \Box K_n)$ that is independent from the component structure of square $3$TDS matrices. Notice that our formula coincides with the number of ones of the square matrices appearing in Table \[ta:smallk3\].
\[theo:thesquarecase\] For any integer $n \ge 3$, let $n \equiv r \pmod {10}$, where $0 \le r \le 9$. Then $$\gamma_{\times 3,t}(K_n \Box K_n)=
\begin{cases}
2n+2\lfloor \frac{n}{10} \rfloor+2 & \text{ if $r=4,5,6,7,8,9;$} \\
2n+2\lfloor \frac{n}{10} \rfloor+\lceil \frac{r}{3}\rceil & \text{ if $r=0,1,2,3$ and $n\ne3;$} \\
8 & \text{ if $n=3.$}
\end{cases}$$
If $n=3$, then $\gamma_{\times 3,t}(K_n \Box K_n)=8$ by Lemma \[lemma:n3case3tds\]. Assume $n\ge4$. Since the description of Proposition \[prop:numberofones\] and Remark \[rem:n45onesrow\] coincides with our claim when $n=m$, we clearly have that $$\gamma_{\times 3,t}(K_n \Box K_n)\le
\begin{cases}
2n+2\lfloor \frac{n}{10} \rfloor+2 & \text{ if $r=4,5,6,7,8,9;$} \\
2n+2\lfloor \frac{n}{10} \rfloor+\lceil \frac{r}{3}\rceil & \text{ if $r=0,1,2,3$ and $n\ne3.$}
\end{cases}$$ Notice that when $n \equiv r \pmod {10}$ and $r=1,2$, then $(2n+2\lfloor \frac{n}{10} \rfloor+\lceil \frac{r}{3}\rceil)+2=2(n+1)+2\lfloor \frac{n+1}{10} \rfloor+\lceil \frac{r}{3}\rceil$. This implies that if $\gamma_{\times 3,t}(K_n \Box K_n)=2n+2\lfloor \frac{n}{10} \rfloor+\lceil \frac{r}{3}\rceil$, then by Lemma \[lemma:3dtsgrowsbydiagonal\], $\gamma_{\times 3,t}(K_{n+1} \Box K_{n+1})=2(n+1)+2\lfloor \frac{n+1}{10} \rfloor+\lceil \frac{r}{3}\rceil$. Similarly, when $r=4,5,6,7,8$, then $(2n+2\lfloor \frac{n}{10} \rfloor+2)+2=2(n+1)+2\lfloor \frac{n+1}{10} \rfloor+2$. This implies that if $\gamma_{\times 3,t}(K_n \Box K_n)=2n+2\lfloor \frac{n}{10} \rfloor+2$, then by Lemma \[lemma:3dtsgrowsbydiagonal\], $\gamma_{\times 3,t}(K_{n+1} \Box K_{n+1})=2(n+1)+2\lfloor \frac{n+1}{10} \rfloor+2$. Moreover, when $r=9$, $(2n+2\lfloor \frac{n}{10} \rfloor+2)+2=2(n+1)+2\lfloor \frac{n+1}{10} \rfloor$. This implies that if $\gamma_{\times 3,t}(K_n \Box K_n)=2n+2\lfloor \frac{n}{10} \rfloor+2$, then by Lemma \[lemma:3dtsgrowsbydiagonal\], $\gamma_{\times 3,t}(K_{n+1} \Box K_{n+1})=2(n+1)+2\lfloor \frac{n+1}{10} \rfloor$. However, when $r=0$, then $(2n+2\lfloor \frac{n}{10} \rfloor)+3=2(n+1)+2\lfloor \frac{n+1}{10} \rfloor+1$, and hence in this situation we need to prove that $\gamma_{\times 3,t}(K_{n+1} \Box K_{n+1})=\gamma_{\times 3,t}(K_n \Box K_n)+3$. Similarly when $r=3$, $(2n+2\lfloor \frac{n}{10} \rfloor+1)+3=2(n+1)+2\lfloor \frac{n+1}{10} \rfloor+2$, and hence also in this situation we need to prove that $\gamma_{\times 3,t}(K_{n+1} \Box K_{n+1})=\gamma_{\times 3,t}(K_n \Box K_n)+3$. By Lemma \[lemma:3dtsgrowsbydiagonal\], it is enough to show that if $r=0,3$, then $\gamma_{\times 3,t}(K_{n+1} \Box K_{n+1})>\gamma_{\times 3,t}(K_n \Box K_n)+2$.
Assume $r=0$. Suppose there exists $S$ a $(n+1)\times (n+1)$ $3$TDS matrix with $|S|=\gamma_{\times 3,t}(K_n \Box K_n)+2=2(n+1)+k$, where $k=2\lfloor \frac{n}{10}\rfloor$. By Remark \[rem:onesbycolumnsvsrowk\], $|S|\ge 3(n+1)-2k-4\lfloor\frac{k}{2}\rfloor$. Notice that since $r=0$, then $k=\frac{n}{5}$ and it is an even integer. This implies that $|S|\ge3(n+1)-4k$. However, since $k=\frac{n}{5}$, then $2(n+1)+k<3(n+1)-4k$ and hence $S$ is not a $3$TDS matrix. Assume now $r=3$. Suppose there exists $S$ a $(n+1)\times (n+1)$ $3$TDS matrix with $|S|=\gamma_{\times 3,t}(K_n \Box K_n)+2=2(n+1)+k$, where $k=2\lfloor \frac{n}{10}\rfloor+1$. By Remark \[rem:onesbycolumnsvsrowk\], $|S|\ge 3(n+1)-2k-4\lfloor\frac{k}{2}\rfloor$. Notice that since $r=3$, then $k=\frac{n-3}{5}+1$ and it is an odd integer. This implies that $|S|\ge 3(n+1)-4k+2$. However, since $k=\frac{n-3}{5}+1$, then $2(n+1)+k<3(n+1)-4k+2$ and hence $S$ is not a $3$TDS matrix.
The general case
================
We give a formula for $\gamma_{\times 3,t}(K_n \Box K_m)$ that coincides with the number of ones of the $3$TDS matrices in Table \[ta:smallk3\], but the argument is independent of the shape of the components in a $3$TDS matrix.
\[theo:thegeneralcase\] Let $m\ge n \ge 1$. Assume $(n,m)\notin\{(1,1),(1,2),(1,3),$$(2,2)\}$ and $2n\equiv 3m+k \pmod {10}$, where $0 \le k \le 9$. Then $$\gamma_{\times 3,t}(K_n \Box K_m)=
\begin{cases}
4 & \text{$n=1$ and $m\ge4;$} \\
6 & \text{$n=2$ and $m\ge3;$} \\
8 & \text{$n=3$ and $m=3,4;$} \\
3n & \text{if $m\ge \lfloor \frac{7n-1}{3} \rfloor-1;$} \\
\lceil \frac{8n+3m}{5} \rceil & \text{ if $k=0,1,2,3,4,7,8,9;$} \\
\lceil \frac{8n+3m}{5} \rceil +1 & \text{ if $k=5,6.$}
\end{cases}$$
If $(n,m)\in\{(1,1),(1,2),(1,3),(2,2)\}$, then there are no $n\times m$ $3$TDS matrices. If $n=1$ and $m\ge4$, then any $1\times m$ $(0,1)$-matrix with exactly $4$ ones is a min-$3$TDS matrix. If $n=2$ and $m\ge3$, then any $2\times m$ $(0,1)$-matrix with exactly $3$ columns with $2$ ones is a min-$3$TDS matrix. If $n=3$ and $m=3, 4$, by Lemma \[lemma:n3case3tds\], $\gamma_{\times 3,t}(K_3 \Box K_3)=\gamma_{\times 3,t}(K_3 \Box K_4)=8$.
Assume $n\ge4$. By Propositions \[prop:manycols\] and \[prop:numberofones\], and Remark \[rem:n45onesrow\], we have that $$\gamma_{\times 3,t}(K_n \Box K_m)\le
\begin{cases}
\min\{\lceil \frac{8n+3m}{5} \rceil, 3n\} & \text{ if $k=0,1,2,3,4,7,8,9;$} \\
\min\{\lceil \frac{8n+3m}{5} \rceil +1,3n\} & \text{ if $k=5,6.$}
\end{cases}$$ If $m>\frac{7n-1}{3}-2$ (which occurs when $m\ge \lfloor \frac{7n-1}{3} \rfloor-1$), then $\lceil\frac{8n+3m}{5} \rceil +1\ge 3n$, in which case the previous minimums coincide with $3n$. Assume now $m< \lfloor \frac{7n-1}{3} \rfloor-1$. Since we assume $m\ge n$, we can write $m=n+d$, for some $d\ge0$. When $n=m$, a direct computation shows that our formula coincides with the one of Theorem \[theo:thesquarecase\]. Hence, using induction on $m$, it is enough to show that if $\gamma_{\times 3,t}(K_n \Box K_m)$ coincides with our formula, so does $\gamma_{\times 3,t}(K_n \Box K_{m+1})$. We will prove this with a case by case analysis.
*Case I*: Assume $2n\equiv 3m \pmod {10}$ and $m=n+d$. Then we can write $\lceil \frac{8n+3m}{5} \rceil=2n+\lceil\frac{n+3d}{5}\rceil=2n+k$. Notice that $k=\frac{n+3d}{5}$ and it is an even integer. Since $2n\equiv 3m \pmod {10}$, then $2n\equiv 3(m+1)+7 \pmod {10}$. Moreover, $\lceil \frac{8n+3(m+1)}{5} \rceil=2n+\lceil\frac{n+3d+3}{5}\rceil$ and hence $\lceil \frac{8n+3(m+1)}{5} \rceil=2n+k+1$. By hypothesis, $\gamma_{\times 3,t}(K_n \Box K_m)=2n+k$. By Lemma \[lemma:3dtsgrowsbyrow\], $2n+k\le \gamma_{\times 3,t}(K_n \Box K_{m+1}) \le2n+k+1$. Suppose there exists $S$ a $n\times (m+1)$ $3$TDS matrix with $|S|=2n+k$. By Remark \[rem:onesbycolumnsvsrowk\], $|S|\ge3(m+1)-2k-4\lfloor\frac{k}{2}\rfloor$. In this situation, $3(m+1)-2k-4\lfloor\frac{k}{2}\rfloor=3n+3d+3-4(\frac{n+3d}{5})$ and it is strictly bigger than $2n+k$. This implies that $S$ is not a $3$TDS matrix and hence that $\gamma_{\times 3,t}(K_n \Box K_{m+1}) =2n+k+1=\lceil\frac{8n+3(m+1)}{5}\rceil$.
*Case II*: Assume $2n\equiv 3m+7 \pmod {10}$ and $m=n+d$. Then we can write $\lceil \frac{8n+3m}{5} \rceil=2n+\lceil\frac{n+3d}{5}\rceil=2n+k$. Notice that $k=\frac{n+3d+2}{5}$ and it is an odd integer. Since $2n\equiv 3m+7 \pmod {10}$, then $2n\equiv 3(m+1)+4 \pmod {10}$. Moreover, $\lceil \frac{8n+3(m+1)}{5} \rceil=2n+\lceil\frac{n+3d+3}{5}\rceil$ and hence $\lceil \frac{8n+3(m+1)}{5} \rceil=2n+k+1$. By hypothesis, $\gamma_{\times 3,t}(K_n \Box K_m)=2n+k$. By Lemma \[lemma:3dtsgrowsbyrow\], $2n+k\le \gamma_{\times 3,t}(K_n \Box K_{m+1}) \le2n+k+1$. Suppose there exists $S$ a $n\times (m+1)$ $3$TDS matrix with $|S|=2n+k$. By Remark \[rem:onesbycolumnsvsrowk\], $|S|\ge3(m+1)-2k-4\lfloor\frac{k}{2}\rfloor$. In this situation, $3(m+1)-2k-4\lfloor\frac{k}{2}\rfloor=3n+3d+3-4(\frac{n+3d+2}{5})+2$ and it is strictly bigger than $2n+k$. This implies that $S$ is not a $3$TDS matrix and hence that $\gamma_{\times 3,t}(K_n \Box K_{m+1}) =2n+k+1=\lceil\frac{8n+3(m+1)}{5}\rceil$.
*Case III*: Assume $2n\equiv 3m+4 \pmod {10}$ and $m=n+d$. Then we can write $\lceil \frac{8n+3m}{5} \rceil=2n+\lceil\frac{n+3d}{5}\rceil=2n+k$. Notice that $k=\frac{n+3d+4}{5}$ and it is an even integer. Since $2n\equiv 3m+4 \pmod {10}$, then $2n\equiv 3(m+1)+1 \pmod {10}$. Moreover, $\lceil \frac{8n+3(m+1)}{5} \rceil=2n+\lceil\frac{n+3d+3}{5}\rceil$ and hence $\lceil \frac{8n+3(m+1)}{5} \rceil=2n+k$. By Lemma \[lemma:3dtsgrowsbyrow\], $\gamma_{\times 3,t}(K_n \Box K_{m+1}) =\lceil \frac{8n+3(m+1)}{5} \rceil$.
*Case IV*: Assume $2n\equiv 3m+1 \pmod {10}$ and $m=n+d$. Then we can write $\lceil \frac{8n+3m}{5} \rceil=2n+\lceil\frac{n+3d}{5}\rceil=2n+k$. Notice that $k=\frac{n+3d+1}{5}$ and it is an even integer. Since $2n\equiv 3m+1 \pmod {10}$, then $2n\equiv 3(m+1)+8 \pmod {10}$. Moreover, $\lceil \frac{8n+3(m+1)}{5} \rceil=2n+\lceil\frac{n+3d+3}{5}\rceil$ and hence $\lceil \frac{8n+3(m+1)}{5} \rceil=2n+k+1$. By hypothesis, $\gamma_{\times 3,t}(K_n \Box K_m)=2n+k$. By Lemma \[lemma:3dtsgrowsbyrow\], $2n+k\le \gamma_{\times 3,t}(K_n \Box K_{m+1}) \le2n+k+1$. Suppose there exists $S$ a $n\times (m+1)$ $3$TDS matrix with $|S|=2n+k$. By Remark \[rem:onesbycolumnsvsrowk\], $|S|\ge3(m+1)-2k-4\lfloor\frac{k}{2}\rfloor$. In this situation, $3(m+1)-2k-4\lfloor\frac{k}{2}\rfloor=3n+3d+3-4(\frac{n+3d+1}{5})$ and it is strictly bigger than $2n+k$. This implies that $S$ is not a $3$TDS matrix and hence that $\gamma_{\times 3,t}(K_n \Box K_{m+1}) =2n+k+1=\lceil\frac{8n+3(m+1)}{5}\rceil$.
*Case V*: Assume $2n\equiv 3m+8 \pmod {10}$ and $m=n+d$. Then we can write $\lceil \frac{8n+3m}{5} \rceil=2n+\lceil\frac{n+3d}{5}\rceil=2n+k$. Notice that $k=\frac{n+3d+3}{5}$ and it is an odd integer. Since $2n\equiv 3m+8 \pmod {10}$, then $2n\equiv 3(m+1)+5 \pmod {10}$. Moreover, $\lceil \frac{8n+3(m+1)}{5} \rceil+1=2n+\lceil\frac{n+3d+3}{5}\rceil+1$ and hence $\lceil \frac{8n+3(m+1)}{5} \rceil=2n+k+1$. By hypothesis, $\gamma_{\times 3,t}(K_n \Box K_m)=2n+k$. By Lemma \[lemma:3dtsgrowsbyrow\], $2n+k\le \gamma_{\times 3,t}(K_n \Box K_{m+1}) \le2n+k+1$. Suppose there exists $S$ a $n\times (m+1)$ $3$TDS matrix with $|S|=2n+k$. By Remark \[rem:onesbycolumnsvsrowk\], $|S|\ge3(m+1)-2k-4\lfloor\frac{k}{2}\rfloor$. In this situation, $3(m+1)-2k-4\lfloor\frac{k}{2}\rfloor=3n+3d+3-4(\frac{n+3d+3}{5})+2$ and it is strictly bigger than $2n+k$. This implies that $S$ is not a $3$TDS matrix and hence that $\gamma_{\times 3,t}(K_n \Box K_{m+1}) =2n+k+1=\lceil\frac{8n+3(m+1)}{5}\rceil$.
*Case VI*: Assume $2n\equiv 3m+5 \pmod {10}$ and $m=n+d$. Then we can write $\lceil \frac{8n+3m}{5} \rceil+1=2n+\lceil\frac{n+3d}{5}\rceil+1=2n+k$. Notice that $k=\frac{n+3d}{5}$ and it is an odd integer. Since $2n\equiv 3m+5 \pmod {10}$, then $2n\equiv 3(m+1)+2 \pmod {10}$. Moreover, $\lceil \frac{8n+3(m+1)}{5} \rceil=2n+\lceil\frac{n+3d+3}{5}\rceil$ and hence $\lceil \frac{8n+3(m+1)}{5} \rceil=2n+k$. By Lemma \[lemma:3dtsgrowsbyrow\], $\gamma_{\times 3,t}(K_n \Box K_{m+1}) =\lceil \frac{8n+3(m+1)}{5} \rceil$.
*Case VII*: Assume $2n\equiv 3m+2 \pmod {10}$ and $m=n+d$. Then we can write $\lceil \frac{8n+3m}{5} \rceil=2n+\lceil\frac{n+3d}{5}\rceil=2n+k$. Notice that $k=\frac{n+3d+2}{5}$ and it is an even integer. Since $2n\equiv 3m+2 \pmod {10}$, then $2n\equiv 3(m+1)+9 \pmod {10}$. Moreover, $\lceil \frac{8n+3(m+1)}{5} \rceil=2n+\lceil\frac{n+3d+3}{5}\rceil$ and hence $\lceil \frac{8n+3(m+1)}{5} \rceil=2n+k+1$. By hypothesis, $\gamma_{\times 3,t}(K_n \Box K_m)=2n+k$. By Lemma \[lemma:3dtsgrowsbyrow\], $2n+k\le \gamma_{\times 3,t}(K_n \Box K_{m+1}) \le2n+k+1$. Suppose there exists $S$ a $n\times (m+1)$ $3$TDS matrix with $|S|=2n+k$. By Remark \[rem:onesbycolumnsvsrowk\], $|S|\ge3(m+1)-2k-4\lfloor\frac{k}{2}\rfloor$. In this situation, $3(m+1)-2k-4\lfloor\frac{k}{2}\rfloor=3n+3d+3-4(\frac{n+3d+2}{5})$ and it is strictly bigger than $2n+k$. This implies that $S$ is not a $3$TDS matrix and hence that $\gamma_{\times 3,t}(K_n \Box K_{m+1}) =2n+k+1=\lceil\frac{8n+3(m+1)}{5}\rceil$.
*Case VIII*: Assume $2n\equiv 3m+9 \pmod {10}$ and $m=n+d$. Then we can write $\lceil \frac{8n+3m}{5} \rceil=2n+\lceil\frac{n+3d}{5}\rceil=2n+k$. Notice that $k=\frac{n+3d+4}{5}$ and it is an odd integer. Since $2n\equiv 3m+9 \pmod {10}$, then $2n\equiv 3(m+1)+6 \pmod {10}$. Moreover, $\lceil \frac{8n+3(m+1)}{5} \rceil+1=2n+\lceil\frac{n+3d+3}{5}\rceil+1$ and hence $\lceil \frac{8n+3(m+1)}{5} \rceil=2n+k+1$. By hypothesis, $\gamma_{\times 3,t}(K_n \Box K_m)=2n+k$. By Lemma \[lemma:3dtsgrowsbyrow\], $2n+k\le \gamma_{\times 3,t}(K_n \Box K_{m+1}) \le2n+k+1$. Suppose there exists $S$ a $n\times (m+1)$ $3$TDS matrix with $|S|=2n+k$. By Remark \[rem:onesbycolumnsvsrowk\], $|S|\ge3(m+1)-2k-4\lfloor\frac{k}{2}\rfloor$. In this situation, $3(m+1)-2k-4\lfloor\frac{k}{2}\rfloor=3n+3d+3-4(\frac{n+3d+4}{5})+2$ and it is strictly bigger than $2n+k$. This implies that $S$ is not a $3$TDS matrix and hence that $\gamma_{\times 3,t}(K_n \Box K_{m+1}) =2n+k+1=\lceil\frac{8n+3(m+1)}{5}\rceil$.
*Case IX*: Assume $2n\equiv 3m+6 \pmod {10}$ and $m=n+d$. Then we can write $\lceil \frac{8n+3m}{5} \rceil+1=2n+\lceil\frac{n+3d}{5}\rceil+1=2n+k$. Notice that $k=\frac{n+3d+1}{5}$ and it is an odd integer. Since $2n\equiv 3m+6 \pmod {10}$, then $2n\equiv 3(m+1)+3 \pmod {10}$. Moreover, $\lceil \frac{8n+3(m+1)}{5} \rceil=2n+\lceil\frac{n+3d+3}{5}\rceil$ and hence $\lceil \frac{8n+3(m+1)}{5} \rceil=2n+k$. By Lemma \[lemma:3dtsgrowsbyrow\], $\gamma_{\times 3,t}(K_n \Box K_{m+1}) =\lceil \frac{8n+3(m+1)}{5} \rceil$.
*Case X*: Assume $2n\equiv 3m+3 \pmod {10}$ and $m=n+d$. Then we can write $\lceil \frac{8n+3m}{5} \rceil=2n+\lceil\frac{n+3d}{5}\rceil=2n+k$. Notice that $k=\frac{n+3d+3}{5}$ and it is an even integer. Since $2n\equiv 3m+3 \pmod {10}$, then $2n\equiv 3(m+1) \pmod {10}$. Moreover, $\lceil \frac{8n+3(m+1)}{5} \rceil=2n+\lceil\frac{n+3d+3}{5}\rceil$ and hence $\lceil \frac{8n+3(m+1)}{5} \rceil=2n+k$. By Lemma \[lemma:3dtsgrowsbyrow\], $\gamma_{\times 3,t}(K_n \Box K_{m+1}) =\lceil \frac{8n+3(m+1)}{5} \rceil$.
Directly from the formula of Theorem \[theo:thegeneralcase\], we can generalize the statement of Lemma \[lemma:3dtsgrowsbydiagonal\] and obtain the following.
Let $m\ge n\ge3$. Then $$\gamma_{\times 3,t}(K_n \Box K_m)+2 \le \gamma_{\times 3,t}(K_{n+1} \Box K_{m+1})\le \gamma_{\times 3,t}(K_n \Box K_m)+3.$$
Since, in general, $3n-1>\lfloor \frac{7n-1}{3} \rfloor-1$, we obtain a better bound than the one described in Proposition \[prop:manycols\] for the case when $\gamma_{\times 3,t}(K_n \Box K_m)=3n$.
$\begin{array}{|c|c|c|c|c|c|c|c|c|c|c|c|}
\cline{2-11}
\multicolumn{1}{c|}{} & m=3 & m=4 & m=5 & m=6 & m=7 & m=8 & m=9 & m=10 & m=11 & m=12 \\ \hline
n=2 &
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
};
\end{tikzpicture}
\\
\hline
n=3 &
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| &
|[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| &
|[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| &
|[fill=black!60]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| &
|[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| &
|[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| &
|[fill=black!60]| & |[fill=black!60]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
};
\end{tikzpicture}
\\
\hline
n=4 &
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| &
|[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| &
|[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| &
|[fill=black!60]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| &
|[fill=white]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| &
|[fill=white]| & |[fill=white]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| &
|[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
};
\end{tikzpicture}
\\
\hline
n=5 & & &
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| &
|[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| &
|[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| &
|[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| &
|[fill=black!60]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| &
|[fill=white]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| &
|[fill=white]| & |[fill=white]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| &
|[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| &
|[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| &
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| &
|[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
};
\end{tikzpicture}
\\
\hline
n=6 & & & &
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| &
|[fill=white]| & |[fill=white]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=white]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=white]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=white]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=white]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| &
|[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=white]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=white]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| &
|[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| &
|[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| \\
};
\end{tikzpicture}
\\
\hline
n=7 & & & & &
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=white]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=white]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| & |[fill=white]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| &
|[fill=white]| & |[fill=white]| & |[fill=white]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| &
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| &
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| &|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=white]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=white]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=white]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=white]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| \\
};
\end{tikzpicture}
\\
\hline
n=8 & & & & & &
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| &
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| &
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| &
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=white]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=white]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| \\
};
\end{tikzpicture}
\\
\hline
n=9 & & & & & & &
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=white]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=white]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]|\\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]|& |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]|& |[fill=white]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=white]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=white]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]|\\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]|& |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]|& |[fill=white]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=white]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=white]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]|\\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]|& |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]|& |[fill=white]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| \\
};
\end{tikzpicture}
\\
\hline
n=10 & & & & & & &
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| &
|[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| &
|[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]|\\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| &
|[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]|& |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| &
|[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]|& |[fill=white]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| &
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]|\\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]|& |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]|& |[fill=white]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]|& |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]|& |[fill=white]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| \\
};
\end{tikzpicture}
\\
\hline
n=11 & & & & & & & &
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]|\\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]|& |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]|& |[fill=white]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| \\
};
\end{tikzpicture}
&
\begin{tikzpicture}
\matrix[square matrix]{
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]|\\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]|& |[fill=white]| \\
|[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]|& |[fill=white]| \\
|[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=black!60]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| & |[fill=white]| \\
};
\end{tikzpicture}
\\
\hline
\end{array}$
[20]{}
J. A. Bondy and U. S. R. Murty. [*Graph theory*]{}, [*Graduate texts in mathematics*]{}. Vol. 244, Springer Science and Media, 2008.
B. Brešar, P. Dorbec, W. Goddard, B. Hartnell, M. Henning, S. Klavžar, D. Rall. Vizing’s conjecture: A survey and recent results, [*J. Graph Theory.*]{} [**69**]{} No.1 (2012)
B. Brešar, M. A. Henning, and S. Klavžar. On integer domination in graphs and Vizing-like problems, [*Taiwanese J. Math.*]{} [**10**]{} No.5 (2006) 1317-1328.
B. Brešar, M. A. Henning, D. F. Rall, Paired-domination of Cartesian products of graphs, [*Util. Math.*]{} [**73**]{} (2007) 255-265.
P. A. Burchett, On the border queens problem and $k$-tuple domination on the rook’s graph, [*Congr. Numer.*]{}, [**209**]{} (2011): 179-187.
P. A. Burchett, D. Lane, J. A. Lachniet, $k$-tuple and $k$-domination on the rook’s graph and other results, [*Congr. Numer.*]{}, [**199**]{} (2009): 187-204.
K. Choudhary, S. Margulies, I.V. Hicks, A note on total and paired domination of Cartesian product graphs, [*Electron. J. Combin.*]{} [**20**]{} No.3 (2013).
K. Choudhary, S. Margulies, I. V. Hicks, Integer domination of Cartesian product graphs, [*Discrete Math.*]{} [**338**]{} No.7 (2015).
E. W. Clark, S. Suen, An inequality related to Vizing’s Conjecture, [*Electron. J. Combin.*]{} [**7**]{} Note.4 (2000).
D. C. Fisher, J. Ryan, G. Domke, and A. Majumdar. Fractional domination of strong direct products. [*Discrete Appl. Math.*]{} [**50**]{} No.1 (1994) 89-91.
F. Harary, T. W. Haynes, Double domination in graphs, [*Ars Combin.*]{} [**55**]{} (2000) 201-213.
T. W. Haynes, S. T. Hedetniemi, P. J. Slater, [*Fundamentals of Domination in Graphs*]{}, Marcel Dekker, New York, 1998.
T. W. Haynes, S. T. Hedetniemi, P. J. Slater, [*Domination in Graphs: Advanced Topics*]{}, Marcel Dekker, New York, 1998.
M. A. Henning, A. P. Kazemi, $k$-tuple total domination in cross products of graphs, [*J. Comb. Optim.*]{} [**24**]{} (2011) 339-346.
M. A. Henning, A. P. Kazemi, $k$-tuple total domination in graphs, [*Discrete Appl. Math.*]{} [**158**]{} (2010) 1006-1011.
M. A. Henning, D. F. Rall, On the total domination number of Cartesian products of graphs, [*Graphs Combin.*]{} [**21**]{} (2005) 63-69.
P. T. Ho, A note on the total domination number. [*Util. Math.*]{} [**77**]{} (2008).
X. M. Hou, F. Jiang, Paired domination of Cartesian products of graphs, [*J. Math. Res. Exposition*]{} [**30**]{} No.1 (2010) 181-185.
X. M. Hou, Y. Lu, On the $\{k\}$-domination number of Cartesian products of graphs, [*Discrete Math.*]{} [**309**]{} (2009) 3413-3419.
W. Imrich, S. Klavžar, [*Product Graphs: Structure and Recognition*]{}, John Wiley & Sons, New York, 2000.
A. P. Kazemi, B. Pahlavsay, $k$-tuple total domination in supergeneralized Petersen graphs, [*Comm. Math. App.*]{} [**2**]{} No.1 (2011) 21-30.
A. P. Kazemi, B. Pahlavsay, R. J. Stones, Cartesian product graphs and $k$-tuple total domination, [*Filomat*]{}.
V. G. Vizing, Some unsolved problems in graph theory, [*Usp. Mat. Nauk*]{} [**23**]{} (1968) 117-134.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'M. J. I. Brown'
- 'J. Moustakas'
- 'T. H. Jarrett'
- 'M. Cluver'
title: 'The ultraviolet-infrared color-magnitude relation of star-forming galaxies'
---
Galaxy bimodality, described by the red sequence and blue cloud, has been central to our understanding of galaxy evolution since the turn of this century . Passive galaxies follow a narrow color-magnitude relation while star-forming galaxies in the blue cloud have a broader range of optical colors, resulting from a range of stellar populations, star formation rates and dust obscuration. Although star-forming galaxies are diverse, they do fall along well-established correlations with mass, including the mass-metallicity relation [e.g., @2004ApJ...613..898T], star formation rate versus mass [e.g., @2007ApJ...660L..43N] and declining dust content with decreasing mass [e.g., @2006ApJ...639..157W]. These relations manifest themselves in broadband photometry, albeit outside the optical wavelength range, as illustrated by the dependence of infrared colors on galaxy type [e.g., @2011ApJ...735..112J].
The far-ultraviolet and mid-infrared are both star formation rate tracers, with the former tracing massive stars while the latter traces blackbody emission from warm dust. While far-ultraviolet luminosity is directly proportional to star formation rate, for $\lesssim L^*$ galaxies mid-infrared luminosity is proportional to star formation rate to the power of $\sim 1.3$ [e.g., @2015AA...584A..87C; @2017ApJ...847..136B], which is a consequence of dust content varying with galaxy mass. We thus expect a blue sequence to be present in ultraviolet-infrared color-magnitude diagrams.
To measure the ultraviolet-infrared color-magnitude relation, we use the local galaxy sample of @2014ApJS..212...18B [@2017ApJ...847..136B] and their multiwavelength matched aperture photometry (in AB magnitudes). We limit the sample to galaxies with $m_{W2}-m_{W3}>-0.5$, which excludes passive galaxies from the @2014ApJS..212...18B sample, and we remove active galactic nuclei with the emission line ratio criterion of @2003MNRAS.346.1055K. To correct the GALEX $FUV$ photometry for internal dust obscuration we use $A_{FUV} \propto (M_{FUV} - M_{NUV})$, leaving the constant as a free parameter that we use to minimize the scatter of the color-magnitude relation.
In Figure \[fig:thefigure\] we present the ultraviolet-infrared color-magnitude plot of $z\sim 0$ star-forming galaxies, using the $FUV$, $NUV$, and WISE $W3$ photometry. We find the best relation is produced when $A_{FUV} = 2.6 (M_{FUV} - M_{NUV})$, which is shallower than the dust extinction relation of @2011ApJ...741..124H, where $A_{FUV} = (3.83\pm 0.48) [M_{FUV} - M_{NUV} - (0.022\pm 0.024)]$. As a cross check of our results, in Figure \[fig:thefigure\] we also plot photometry of $z<0.05$ GAMA galaxies [@2016MNRAS.460..765W] with WISE $m_{W2}-m_{W3}>-0.5$, and we find good agreement although GAMA spans a smaller range of $M_{W3}$ than @2014ApJS..212...18B [@2017ApJ...847..136B]. We also observe similar relations when we replace WISE $W3$ with WISE $W4$ or [*Spitzer*]{} $24~{\rm \mu m}$, albeit with more scatter.
![The ultraviolet-infrared (left) and $g-H$ (right) color-magnitude relations for star-forming galaxies. Some of the outliers in the $g-H$ are labelled, and these are often merging galaxies rather than spirals. As the @2014ApJS..212...18B [@2017ApJ...847..136B] sample deliberately selected galaxies to span parameter space, it shows more scatter than the magnitude limited GAMA sample. \[fig:thefigure\]](blueseqplot.pdf "fig:"){width="49.00000%"} ![The ultraviolet-infrared (left) and $g-H$ (right) color-magnitude relations for star-forming galaxies. Some of the outliers in the $g-H$ are labelled, and these are often merging galaxies rather than spirals. As the @2014ApJS..212...18B [@2017ApJ...847..136B] sample deliberately selected galaxies to span parameter space, it shows more scatter than the magnitude limited GAMA sample. \[fig:thefigure\]](gH.pdf "fig:"){width="49.00000%"}
The best-fit color-magnitude relation is given by $$M_{W3} = -14.8 - 2.1 \times \left[ 2.6 (M_{FUV} - M_{NUV}) - M_{W3} \right].$$ Using the @2014ApJS..212...18B [@2017ApJ...847..136B] sample, we find the $1\sigma$ scatter of $M_{W3}$ about the relation is $\sigma_{W3} =1.6~{\rm mag}$. If ultraviolet - infrared color was used as a distance indicator then the 68% scatter of the distance would be a factor of $\sim 2$.
We note color-magnitude relations for blue galaxies have been identified previously, including the median optical color of blue galaxies varying with magnitude [@2004ApJ...600..681B]. Furthermore, @1982ApJ...257..527T identified a tight $B-H$ color-magnitude relation for spiral galaxies, and in the right panel of Figure \[fig:thefigure\] we reproduce this relation for star-forming galaxies using SDSS $g$ and 2MASS $H$-band photometry. @1982ApJ...257..527T attributed this relation to specific star formation rate, chemical abundances and/or initial mass function varying with mass. Interestingly, we do see some outliers in the $g-H$ versus $M_H$ diagram, including merging starbursts. These outliers are not unexpected, given $g$ and $H$ trace different galaxy properties, whereas the ultraviolet and mid-infrared are both (primarily) star formation rate tracers.
In this note we have identified and characterized the ultraviolet-infrared color-magnitude relation of star-forming galaxies. The ultraviolet to mid-infrared flux ratios of star-forming galaxies span over two orders of magnitude and show a clear dependence on absolute magnitude from $M_{W3}\sim -13$ to $M_{W3}\sim -25$, which may present problems for models of galaxy spectral energy distributions that have been largely verified on $\sim L^*$ galaxies. The color-magnitude relation of star-forming galaxies illustrates the (broadband) spectral diversity of star-forming galaxies that results from established correlations between the physical properties and mass, including the mass-metallicity relation.
, I. K., [Glazebrook]{}, K., [Brinkmann]{}, J., et al. , 600 681
, E. F., [Wolf]{}, C., [Meisenheimer]{}, K., et al. 2004, , 608, 752
Blanton, M. R., & Moustakas, J. 2009, , 47, 159
, M. J. I., [Moustakas]{}, J., [Smith]{}, J.-D. T., et al. 2014, , 212, 18
, M. J. I., [Moustakas]{}, J., [Kennicutt]{}, R. C., et al. 2017, , 847, 136
, C., [Gil de Paz]{}, A., [Castillo-Morales]{}, A., et al. 2015, , 584, 87
, A., [Boissier]{}, S., [Madore]{}, B. F., et al. 2007, , 173, 185
, C.-N., [Kennicutt]{}, R. C., [Johnson]{}, B. D., et al. 2011, , 741, 124
, T. H., [Cohen]{}, M., [Masci]{}, F., et al. 2011, , 735, 112
, G., [Heckman]{}, T. M., [Tremonti]{}, C., et al. 2003, , 2003, 346, 1055
Noeske, K. G., Weiner, B. J., Faber, S. M., et al. 2007, , 660, L43
, C. A., [Heckman]{}, T. M., [Kauffmann]{}, G., et al. 2004, , 2004, 613
, R. B., [Mould]{}, J. R. & [Aaronson]{}, M., et al. 1982, , 257, 527
, A. H., [Robotham]{}, A. S. G., [Bourne]{}, N., et al. 2016, , 460, 765
, Y., [Charmandaris]{}, V., [Hao]{}, L., et al. 2006, , 639, 157
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We give a simplification of Belkale’s geometric proof of the Horn conjecture. Our approach uses the geometry of two-step flag manifolds to explain the occurrence of the Horn inequalities in a very straightforward way. The arguments for both necessity and sufficiency of the Horn inequalities are fairly conceptual when viewed in this framework. We provide examples to illustrate the method of proof.'
author:
- |
Kevin Purbhoo\
University of British Columbia
title: 'Two-step flag manifolds and the Horn conjecture'
---
Introduction
============
General approach
----------------
Horn’s conjecture [@H] was originally formulated as a recursive method for solving a problem concerning the eigenvalues of Hermitian matrices. However, as a consequence of work of Klyachko [@Kly], Horn’s conjecture can be reformulated as saying that the non-vanishing Schubert intersection numbers for Grassmannians satisfy a certain recursion. This recursion was first proved by Knutson and Tao [@KT], using combinatorial methods. Later, a geometric proof was given by Belkale [@B], which was the inspiration for this paper.
For our purposes, a Schubert problem will refer to a collection of Schubert varieties $\bar{\Omega}_{\sigma_1}, \ldots,
\bar{\Omega}_{\sigma_s}$, on some partial flag variety $G/P$. A Schubert problem is non-vanishing if the product of the corresponding cohomology classes is non-zero. Equivalently, a Schubert problem is non-vanishing if and only if the general translates $\bar{\Omega}^{F_1}_{\sigma_1}, \ldots,
\bar{\Omega}^{F_s}_{\sigma_s}$ of these Schubert varieties have a non-empty generically transverse intersection. This observation allows one to study the vanishing question inside the tangent space to the $G/P$.
In an effort to better understand the geometry behind Horn’s conjecture, we study the tangent spaces of Schubert varieties of two-step flag manifolds, and show how these are related to the problem. The two-step flag manifold $$Fl(d,r,{\mathbb{C}}^n) = \{ (S,V)\ |\ S \subset V \subset {\mathbb{C}}^n,
\ \dim S = d,\ dim V = r\}$$ fibres over the Grassmannian $Gr(r,{\mathbb{C}}^n)$, with fibre $Gr(d,V)$ at the a point $V \in Gr(r,{\mathbb{C}}^n)$. A Schubert problem on $Gr(r,{\mathbb{C}}^n)$ can be “lifted” in a number of different ways to a Schubert problem on $Fl(d,r,{\mathbb{C}}^n)$, in such a way that a non-empty transverse intersection on $Gr(r,{\mathbb{C}}^n)$ lifts to a non-empty transverse intersection on $Fl(d,r,{\mathbb{C}}^n)$. Each way of lifting corresponds to a non-vanishing Schubert problem inside the fibre $Gr(d,V)$. However, it is sometimes possible to see that the intersection on $Fl(d,r,{\mathbb{C}}^n)$ is non-transverse for some trivial reasons. These trivial conditions are seen to be the Horn inequalities.
The difficult part of the Horn conjecture is to show the sufficiency of the Horn inequalities. In our approach this amounts to showing that a non-transverse intersection on $Gr(r,{\mathbb{C}}^n)$ lifts to something upstairs which is not only non-transverse, but non-transverse for the aforementioned trivial reasons. However, once we have all the appropriate machinery in place, this turns out to be almost as straightforward as the “easy” direction of Horn’s conjecture.
The two-step flag manifolds are not a necessary part of the argument. In principle the entire argument could be formulated inside the tangent space of the Grassmannian. This would probably even lead to a shorter proof. However, it is our opinion that the geometry of the two-step flag manifold plays a fundamental role in this picture, and to undermine its role would be remiss.
Although there are a number of new ideas in this paper, we do not claim to be presenting an original or independent proof of the Horn conjecture. Our objective in writing this was to better understand the argument in [@B], to simplify it in places, and to show how it relates to some of our own previous work [@P1]. A few of the results have been taken directly from [@B], while some of the others merely contain old ideas which have been dressed up in a new context. We will try to indicate whenever possible which ideas and results have been borrowed.
Partial flag manifolds
----------------------
### Schubert varieties and Schubert positions
Let $G = GL(n)$, and let $P \subset G$ be a parabolic subgroup, the stabiliser subgroup of some $k$-step flag $$V^0 = {\{0\}}=V^0_0 \subsetneq V^0_1 \subsetneq \cdots
\subsetneq V^0_k \subsetneq V^0_{k+1} = {\mathbb{C}}^n.$$ We will assume for later convenience that the $V^0_i$ are coordinate subspaces. Let $d_j = \dim V_j$. Then $G/P$ is a partial flag manifold (of type $0<d_1<\cdots<d_k<n$).
Let $$F = {\{0\}}=F_0 \subsetneq F_1 \subsetneq \cdots \subsetneq F_{n-1}
\subsetneq F_n = {\mathbb{C}}^n.$$ be a full flag on $C^n$. Let $B(F) \subset G$ denote its stabiliser.
Let $\sigma = \sigma_1 \ldots \sigma_n$ be a string of length $n$, where $\sigma_l \in \{0, \ldots, k\}$, and the number of $j$’s in this string is $d_{j+1}-d_j$. It is a basic fact that $G/P$ has finitely many $B(F)$-orbits, and that these orbits are indexed by the set of possible $\sigma$ as follows.
Define $$\Omega^F_\sigma = \{V \in G/P\ \big|\
\sum_{j=1}^k \dim V_j \cap F_l - \dim V_j \cap F_{l-1} = \sigma_l\}.$$ $\Omega^F_\sigma$ is a $B(F)$-orbit, called the [**Schubert cell**]{} of $\sigma$ with respect to the flag $F$. When the flag $F$ is understood, or irrelevant, we shall also denote this $\Omega_\sigma$. The [**Schubert variety**]{} of $\sigma$, with respect to the flag $F$, is its closure $\bar{\Omega}^F_\sigma$, which represents the Schubert class $S^\sigma \in H^*(G/P)$.
If the $\sigma_l$ are weakly decreasing, then $\Omega_\sigma$ consists of a single point which is a subflag of $F$. If the $\sigma_i$ are weakly increasing, then $\Omega_\sigma$ is a dense open subvariety of $G/P$, and $S^\sigma = 1 \in H^*(G/P)$.
In general we can easily calculate the dimension of this orbit: $$\dim \Omega_\sigma = \#\{l<l'\ |\ \sigma_l < \sigma_{l'}\}.$$
If $V \in G/P$, then the [**Schubert position**]{} of $V$ with respect to the flag $F$ is the unique $\sigma$ such that $V \in \Omega^F_\sigma$. When we have multiple flags $F^1, \ldots, F^s$ on ${\mathbb{C}}^n$, the Schubert position of $V$ will be the $s$-tuple $(\sigma^1, \ldots \sigma^s)$, where $\sigma^i$ is the Schubert position of $V$ with respect to $F^i$.
### $01$-strings from $\sigma$
From $\sigma$ we construct [*$01$-strings*]{} in two different ways. First, for each pair $(u,v)$, with $0 \leq u < v \leq k$, we define a string $\sigma(uv)$. This is constructed as follows: we delete from $\sigma = \sigma_1 \ldots \sigma_n$ every $\sigma_l \notin \{u,v\}$. What remains will be a string consisting only of ‘$u$’s and ‘$v$’s. To convert this into a $01$-string, we change every ‘$u$’ to a [[‘$0$’]{} ]{}, and every ‘$v$’ to a [[‘$1$’]{} ]{}.
If $\sigma=01312230132$, then $\sigma(13)$ is produced as follows: $$\sigma = 0\mathbf{131}22\mathbf{3}0\mathbf{13}2
\mapsto \mathbf{131313} \mapsto 010101 = \sigma(13).$$
Each $\sigma(uv)$ corresponds to a Schubert cell on a different Grassmannian. In section \[sec:twostep\] we investigate the significance of these $\sigma(uv)$ on a two-step flag manifold.
In a completely different spirit, we define strings $\sigma[j] = \sigma[j]_1 \ldots \sigma[j]_n$, for $1 \leq j \leq k$, defined as follows. $$\sigma[j]_l =
\begin{cases}
0,& \text{if $\sigma_l \leq k-j$} \\
1,& \text{if $\sigma_l > k-j$.}
\end{cases}$$
These have a very natural geometric significance. The $k$-step flag variety has $k$ different projections onto Grassmannians $Gr(d_j, {\mathbb{C}}^n)$. The image of the Schubert cell $\Omega_\sigma$ is projected onto $Gr(d_j, {\mathbb{C}}^n)$ is the Schubert cell $\Omega_{\sigma[j]}$.
If $\sigma=2103210$, then $$\begin{aligned}
\sigma[1]&=0001000 \\
\sigma[2]&=1001100 \\
\sigma[3]&=1101110\end{aligned}$$
### Notation for Grassmannians
When $P$ is a maximal parabolic, $G/P \cong Gr(r,n)$ is a Grassmannian, and we can equivalently index the Schubert varieties and Schubert classes by partitions. Let $\Lambda(r,n-r)$ denote the set of partitions with $r$-parts, whose largest part is at most $n-r$, i.e. $$\Lambda(r,n-r) =
\{ \lambda = \mathbf{0} \leq \lambda_1 \leq \cdots \leq \lambda_r
\leq \mathbf{n-r} \}.$$ There is a simple correspondence between these partitions and $01$-strings: given a partition $\lambda \in \Lambda(r,n-r)$, one can construct a string as follows.
> 0\. Begin with the empty string and $l=1$.\
> 1. Append $\lambda_l - \lambda_{l-1}$ [[‘$0$’]{}s ]{}.\
> 2. If $l=n-r$ stop.\
> 3. Append a [[‘$1$’]{} ]{}.\
> 4. Increment $l$ and repeat steps 1-4.\
$\lambda = \mathbf{0} \leq 0 \leq 1 \leq 3 \leq 3 \leq \mathbf{5}$, corresponds to the string $101001100$.
In the case of Grassmannians, we will sometimes find it more convenient to use $\Lambda(r,n-r)$ to index our Schubert classes and Schubert varieties. We therefore denote these $S^\lambda$ and $\bar{\Omega}_\lambda$ respectively. It is worth noting at this time that $$\dim \Omega_\lambda = \sum_{l=1}^{n-r} \lambda_l =: |\lambda|.$$
### Induced flags
Whenever we have a full flag $F$ on a vector space $W$, and $V \subset W$ is a subspace, we get induced flags $F_V$ and $F_{W/V}$ on $V$ and $W/V$ respectively. These can be thought of as follows.
Consider the chain of subspaces $$F_V = \bigg(
{\{0\}}=F_0 \cap V \subset F_1 \cap V \subset \cdots \subset F_{n-1} \cap V
\subset F_n \cap V = V \bigg).$$ Since $\dim V$ may be less than or equal to $\dim W$, this will no longer be a flag, as some of the $F_{i-1} \cap V \subset F_i \cap V$ may become equalities. However by eliminating all repeated elements, one gets a full flag on $V$. It is easy to check that the elements we keep correspond precisely to the [[‘$1$’]{}s ]{}in the Schubert position of $V$ with respect to $F$.
Similarly we can construct a full flag on $W/V$. Let $\Pi:W \to W/V$ denote the quotient map. Then $$F_{W/V}=
\bigg(
{\{0\}}=\Pi(F_0) \subset \Pi(F_1) \subset \cdots \subset \Pi(F_{n-1})
\subset \Pi(F_n) = W/V \bigg).$$ Again by eliminating repeated elements we obtain a full flag. In this case, we keep the elements corresponding to the [[‘$0$’]{}s ]{}in the Schubert position of $V$.
The generalised Horn conjecture
-------------------------------
We can now state Horn’s conjecture in the language of Schubert calculus. Actually we will give the slightly more general statement which appears in [@B].
To simplify the notation in the statement a little, any time we write $S^\lambda \in H^*(Gr(a,b))$ (or otherwise assume that $\lambda$ indexes such a class), then we implicitly assume $$\lambda = \mathbf{0} \leq \lambda_1 \leq \cdots \leq \lambda_a \leq
\mathbf{b-a} \in \Lambda(a,b-a).$$
\[def:hornineq\] Let $\lambda^1, \ldots, \lambda^s \in \Lambda(r,n-r)$ Suppose that for every $d$, $1 \leq d \leq r$, and every non-zero product $S^{\mu^1} \cdots S^{\mu^s} \in H^*(Gr(d,r))$, the inequality $$\label{eqn:hornineq}
\sum_{i=1}^s \sum_{k=1}^d \lambda^i_{\mu^i_k +k} \geq (s-1)d(n-r)$$ holds. In this case we say that the [**Horn inequalities**]{} hold for $\lambda^1, \ldots, \lambda^s$.
The product $S^{\lambda^1} \cdots S^{\lambda^s} \in H^*(Gr(r,n))$ is non-zero if and only if the Horn inequalities hold for $\lambda^1, \ldots, \lambda^s$.
The standard form of Horn’s conjecture assumes that the product of $S^{\lambda^1} \cdots S^{\lambda^s}$ is of top degree, whereas this formulation (which appears in [@B]) does not. To accommodate this, definition \[def:hornineq\] allows an inequality for all non-zero products in $H^*(Gr(d,r))$ rather than those just those of top degree. This generalisation is easily shown to imply the standard statement.
Definition \[def:hornineq\] differs slightly from the usual definition of the Horn inequalities in another small way. Normally one does not include an inequality for the case $d=r$; one simply uses $d$ for which $1 \leq d \leq r-1$. The $d=r$ inequality simply amounts to saying that $$\sum_{j=1}^s {\mathop{\mathrm{codim}}}\Omega_{\lambda^i} \leq \dim Gr(r,n),$$ i.e. this covers the case where the product vanishes for dimensional reasons. If we assume that $S^{\lambda^1} \cdots S^{\lambda^s}$ is of top degree then this inequality is always satisfied.
Schubert calculus in the tangent space
======================================
General statements for $G/P$
----------------------------
Let $F^1, \ldots, F^s$ be flags on ${\mathbb{C}}^n$. To determine whether a product $S^{\sigma^1} \cdots S^{\sigma^s} \in H^*(Gr(r,n))$ vanishes, it is sufficient to consider whether the Schubert varieties $\bar{\Omega}^{F^i}_{\sigma}$ have a point of intersection when the flags $F^i$ are sufficiently generic. This is due to the Kleiman-Bertini theorem [@Kl], which tells us that if $F^i$ are generic, these Schubert varieties will intersect transversely (in positively oriented points). Moreover, this point of intersection can be assumed to be inside the open cell $\Omega^{F^i}_{\sigma}$
Thus we can take the following approach. Choose the flags $F^i$ such that $V \in G/P$ is an intersection point of the Schubert varieties $\Omega^{F^i}_{\sigma}$. Subject to this restriction, the flags $F^i$ should be as generic as possible. We call generic flags with this restriction [**almost generic**]{} for $V$. The question then becomes whether or not the Schubert varieties with respect to almost generic flags intersect transversely. If they intersect transversely, then Schubert varieties with respect to fully generic flags must have a point of intersection; however if they do not, then the point of intersection is artificial, and fully generic flags will not give any point of intersection.
More importantly, this calculation can be done on the level of tangent spaces. We have the following basic result which appears in $\cite{B, P1}$.
\[lem:tangent1\] $S^{\sigma^1} \cdots S^{\sigma^s} = 0 \in H^*(G/P)$ if and only if the intersection of subspaces $$\bigcap_{i=1}^s T_V \Omega^{F^i}_{\sigma^i} \subset T_V G/P$$ is non-transverse for $F^i$ almost generic.
Our $V$ can be any point of $G/P$, so let us take $V=V^0$. Then $P$ acts transitively on the flags $F^i$ such that $V^0 \in \Omega^{F^i}_{\sigma^i}$. Thus to calculate $T_{V^0} \Omega^{F^i}_{\sigma^i}$, for generic $F^i$, it suffices to compute it for a special $F^i$ and consider the action of a generic element of $P$ on the tangent space $T_{V^0} G/P = {\mathfrak{g}}/{\mathfrak{p}}$.
There is always at least one coordinate flag $\hat{F}(\sigma^i)$ such that $V^0 \in \Omega^{\hat{F}(\sigma^i)}_{\sigma^i}$; we will use one of these. The standard flag on ${\mathbb{C}}^n$ is $${F^{\mathrm{std}}}= {\{0\}}\subsetneq \langle x_1 \rangle
\subsetneq \langle x_1, x_2 \rangle \subsetneq \cdots
\subsetneq \langle x_1, x_2, \ldots, x_{n-1} \rangle$$ The Weyl group $S_n$ acts transitively on the set of coordinate flags of ${\mathbb{C}}^n$, thus $\hat{F}^i = w_i \cdot {F^{\mathrm{std}}}$ for some $w_i \in S_n$. Of all possible choices for the $\hat{F}(\sigma^i)$, we will always choose $\hat{F}(\sigma^i) = w_i \cdot {F^{\mathrm{std}}}$ with $w_i$ minimal in the Bruhat order.
Denote the tangent space to the Schubert variety $\Omega^{\hat{F}(\sigma)}_{\sigma}$ at $V_0$ by $$\hat{Z}_\sigma = T_{V^0} \Omega^{\hat{F}(\sigma)}_{\sigma}
\subset {\mathfrak{g}}/{\mathfrak{p}}.$$ Let $Z_\sigma$ denote a generic $P$-translate of $\hat{Z}_\sigma$. This notation will be convenient, as we will often need to consider intesections $\bigcap_{i=1}^s p_i \hat{Z}_{\sigma_i}$ where $p_i \in P$ are generic; we can now write this simply as $\bigcap_{i=1}^s Z_{\sigma_i}$. For the special cases of Grassmannians and two-step flag manifolds, we will use the letters $X$ or $Y$ respectively instead of $Z$.
Lemma \[lem:tangent1\] can be stated equivalently as follows:
\[lem:tangent2\] $S^{\sigma^1} \cdots S^{\sigma^s} = 0 \in H^*(G/P)$ if and only if the intersection of subspaces $$\bigcap_{i=1}^s Z_{\sigma^i} \subset {\mathfrak{g}}/{\mathfrak{p}}$$ is non-transverse.
Although we have eliminated the flags $F^i$ from this statement, we will sometimes wish to think of $Z_{\sigma^i}$ as being determined by generic flags, rather than as generic translates of $\hat{Z}_{\sigma^i}$.
Whenever we use the notation $Z_{\sigma^i}$, we tacitly assume (unless otherwise specified) that underlying flags $F^i$ are almost generic. If the $F^i$ are almost generic, we say that $Z_{\sigma^1}, \ldots, Z_{\sigma^s}$ are in [**general position**]{}. This is, of course, equivalent to the statement that $Z_{\sigma^i}$ is a *generic* $P$-translate of $\hat{Z}_{\sigma^i}$.
To calculate a particular $\hat{Z}_{\sigma}$, we use the following fact.
\[prop:calcZ\] $\hat{Z}_{\sigma}$ is the image of ${\mathfrak{b}}(\hat{F}(\sigma))$ under ${\mathfrak{g}}\to {\mathfrak{g}}/{\mathfrak{p}}$.
As $\Omega^{\hat{F}(\sigma)}_{\sigma}$ is a $B(\hat{F}(\sigma))$ orbit, the tangent space is generated by ${\mathfrak{b}}(\hat{F}(\sigma))$.
Grassmannians {#sec:grassmannian}
-------------
### The tangent space to a Grassmannian Schubert variety
To distinguish the Grassmannian as a special case, we use the notation $\hat{X}_{\lambda}$ for $\hat{Z}_\sigma$, and $X_\lambda$ for the generic translate $X_\sigma$, where $\lambda$ is the partition corresponding to the $01$-string $\sigma$. We will take $V^0$ to be the coordinate subspace $\langle x_{n-r+1}, \ldots, x_n \rangle \in Gr(r,n)$.
We will now identify the subspace $\hat{X}_\lambda$. Put $V = V^0$, and $Q = {\mathbb{C}}^n/V$. Now ${\mathfrak{g}}/{\mathfrak{p}}$ can be naturally identified with ${\mathrm{Hom}}(V, Q)$, so we can view $\hat{X}_\lambda$ as a set of homomorphisms $\phi:V \to Q$. Both $V$ and $Q$ inherit full flags from the standard flag. $V$ inherits the flag $${F^{\mathrm{std}}}_V = {\{0\}}\subsetneq \langle x_{n-r+1} \rangle
\subsetneq \langle x_{n-r+1}, x_{n-r+2} \rangle \subsetneq \cdots
\subsetneq \langle x_{n-r+1}, \ldots, x_{n-1} \rangle \subsetneq V$$ and $Q$ inherits the image of $${F^{\mathrm{std}}}_Q = V \subsetneq V + \langle x_{1} \rangle
\subsetneq V + \langle x_{1}, x_{2} \rangle \subsetneq \cdots
\subsetneq V + \langle x_{1}, \ldots, x_{n-r-1} \rangle \subsetneq {\mathbb{C}}^n$$ under the quotient map ${\mathbb{C}}^n \to Q$.
Under these identifications, $$\hat{X}_\lambda =
\{\phi \in {\mathrm{Hom}}(V,Q)\ |\ \phi(({F^{\mathrm{std}}}_V)_l) \subset
({F^{\mathrm{std}}}_Q)_{\lambda_l}\ l=1, \ldots, r\}.$$
Let $\sigma$ denote the $01$-string corresponding to $\lambda$, and put $$\alpha_k =
\begin{cases}
\sum_{l=1}^k 1-\sigma_l, &\text{if $\sigma_k = 0$} \\
n-r+\sum_{l=1}^k \sigma_l, &\text{ if $\sigma_k = 1$.}
\end{cases}$$ So $\alpha_k \leq n-r$ if $\sigma_k=0$, and $\alpha_k >n-r$ if $\sigma_k=1$. Then the flag $\hat{F}(\sigma)$ is $$\hat{F}(\sigma)
= {\{0\}}\subsetneq \langle x_{\alpha_1} \rangle
\subsetneq \langle x_{\alpha_1}, x_{\alpha_2}\rangle \subsetneq \cdots
\subsetneq \langle x_{\alpha_1}, \ldots, x_{\alpha_{r-1}} \rangle
\subsetneq {\mathbb{C}}^n.$$ To see this, note that $V \cap \hat{F}(\sigma)_l$ jumps in dimension exactly when $\sigma_l=1$, and $\alpha$ is the smallest permutation that accomplishes this.
By proposition \[prop:calcZ\], $\hat{X}_\sigma$ is is identified with ${\mathfrak{b}}(\hat{F}(\sigma))/{\mathfrak{p}}$. Let ${\mathfrak{n}}$ denote the orthogonal complement to ${\mathfrak{p}}$. Since $V$ and $\hat{F}(\sigma)$ are coordinate flags, we have ${\mathfrak{b}}(\hat{F}(\sigma))/{\mathfrak{p}}= {\mathfrak{b}}(\hat{F}(\sigma)) \cap {\mathfrak{n}}/{\mathfrak{p}}$. Thus it suffices to determine ${\mathfrak{b}}(\hat{F}(\sigma)) \cap {\mathfrak{n}}$. We see that an element $\phi \in {\mathfrak{n}}$ preserves $\hat{F}(\sigma)$ if and only if for $j \in \{n-r+1, \ldots, n\}$, $$\phi(x_j) \subset \langle x_{\alpha_1}, x_{\alpha_2}, \ldots,
x_{\alpha_j} \rangle$$ But since ${\mathop{\mathrm{Image}}}(\phi) \subset \langle x_1, \ldots, x_{n-r} \rangle$, this is equivalent to
$$\phi(x_j) \subset V + \langle x_1, \ldots, x_l \rangle
= ({F^{\mathrm{std}}}_Q)_l$$ where $l$ is the number of [[‘$0$’]{}s ]{}before the $j^{\rm th}$ [[‘$1$’]{} ]{}. Thus $l = \lambda_j$.
In $Gr(4, {\mathbb{C}}^9)$, if $\lambda = \mathbf{0} \leq 0 \leq 1 \leq 3 \leq 3 \leq \mathbf{5}$, then with respect to the standard coordinate bases on $V$ and $Q$, $$\hat{X}_\lambda = \
\left (
\begin{matrix}
0 & * & * & * \\
0 & 0 & * & * \\
0 & 0 & * & * \\
0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0
\end{matrix}
\right)$$ In general, if we write $\hat{X}_\lambda$ in this form, the number of $*$’s in column $j$ will be $\lambda_j$.
The action of $P$ on ${\mathrm{Hom}}(V,Q)$ is given by $$\left(
\begin{matrix}
A & 0 \\
C & B
\end{matrix}
\right)
\cdot \phi
= A \phi B^{-1}.$$ To compute the generic intersection of two subspaces $X_\lambda \cap X_\mu$, it suffices to take $X_\lambda = \hat{X}_\lambda$, and $X_\mu = p \cdot \hat{X}_\mu$, where $$p =
\left(
\begin{matrix}
& &1& & & \\
& {\reflectbox{$\ddots$}}& & &\mbox{\Huge $0$}& \\
1& & & & & \\
& & & & &1 \\
&\mbox{\Huge $0$}& & &{\reflectbox{$\ddots$}}& \\
& & &1& & \\
\end{matrix}
\right)$$ This is the tangent space analogue of intersecting a Schubert variety with an opposite Schubert variety.
\[ex:oppositeintersection\] If $\lambda = \mathbf{0} \leq 0 \leq 1 \leq 3 \leq 3 \leq \mathbf{5}$, $\mu = \mathbf{0} \leq 3 \leq 3 \leq 3 \leq 5 \leq \mathbf{5}$, then $$\begin{aligned}
X_\lambda \cap X_\mu &=
\left (
\begin{matrix}
0 & * & * & * \\
0 & 0 & * & * \\
0 & 0 & * & * \\
0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0
\end{matrix}
\right)
\cap
\left (
\begin{matrix}
* & 0 & 0 & 0 \\
* & 0 & 0 & 0 \\
* & * & * & * \\
* & * & * & * \\
* & * & * & *
\end{matrix}
\right) \\
&=
\left (
\begin{matrix}
0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 \\
0 & 0 & * & * \\
0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0
\end{matrix}
\right)\end{aligned}$$ The intersection is codimension $18$; however the expected codimension is ${\mathop{\mathrm{codim}}}X_\lambda + {\mathop{\mathrm{codim}}}X_\mu = 13+6 =19$. Thus $S^\lambda S^\mu = 0 \in H^*(Gr(r,n))$.
### $X_\lambda$ as a set of homomorphisms {#sec:grassmannianhoms}
Let $\phi \in X_\lambda \subset {\mathrm{Hom}}(V,Q)$. This represents a direction in which we can perturb $V$ and remain in the tangent space to the Schubert variety. The kernel, $\ker \phi$ represents the maximal subspace of $V$ which is preserved by this perturbation. One of the key ideas which we take from Belkale’s proof is to examine $\ker \phi$ for a generic $\phi \in \bigcap_{i=1}^s X_{\lambda^i}$. For now, however, let us restrict our attention to a single $X_\lambda$.
Let us recall that the construction of $X_\lambda$ was actually a tangent space of a Schubert variety relative to some generic flag $F$ on ${\mathbb{C}}^n$. Thus $V$ carries an induced flag $F_V$, and $Q$ carries an induced flag $F_Q$. The action of $GL(V) \times GL(Q)$ on ${\mathrm{Hom}}(V,Q)$ corresponds to changing these induced flags. $F_V$ and $F_Q$ carry all the relevant information (for our purposes) about the original flag $F$, and moreover, the action of $GL(V) \times GL(Q)$ on $Flags(V) \times Flags(Q)$ is transitive. Thus we shall stop thinking about $F_V$ and $F_Q$ as the induced flags of $F$, and instead think of them as an independent pair of generic flags on $V$ and $Q$ respectively.
\[rmk:genericinducedflags\] We have in fact just observed that the induced flags on $V$ and ${\mathbb{C}}^n/V$ are generic, if $V \in Gr(r,n)$ is a generic point of intersection of Schubert varieties in general position.
Let $S$ be a subspace of $V$ whose Schubert position relative to the flag $F_V$ is $\rho$. Put $d = \dim S$. If $\phi \in {\mathrm{Hom}}(V,Q)$ and $\ker \phi
\supset S$, then $\phi$ descends to a map $\tilde{\phi} \in {\mathrm{Hom}}(V/S,Q)$. We now consider the space $$X_\lambda{\big /}S = \{\tilde\phi \in {\mathrm{Hom}}(V/S,Q)\ |\ \phi \in X_\lambda,
\ \ker\phi \supset S\} = X_\lambda \cap {\mathrm{Hom}}(V/S,Q).$$
$X_\lambda{\big /}S = X_{\lambda'}$, where $\lambda' \in \Lambda(r-d, n-r)$ is the subpartition $$\lambda' = \mathbf{0} \leq \lambda_{k_1} \leq \cdots \leq
\lambda_{k_{r-d}} \leq \mathbf{n-r}$$ where $k_j$ are the positions of the [[‘$0$’]{}s ]{}of $\rho$. The flag on $V/S$ associated to $X_{\lambda'}$ is the induced flag $(F_V)_{V/S}=:F_{V/S}$.
The action of $P$ simply gives a change of basis on $V$ and $Q$. Thus it suffices to prove this in the case where $F = \hat{F}(\lambda)$, in which case $X_\lambda = \hat{X_\lambda}$ and $F_V = {F^{\mathrm{std}}}_V$.
We must verify that for every $\phi \in X_\lambda{\big /}S$, that $\phi ((F_{V/S})_j) \subset (F_Q)_{\lambda'_j}$. However, this is straightforward, as $(F_{V/S})_j$ is the image under the quotient map of $(F_V)_{k_j}$ which maps to $(F_Q)_{\lambda_{k_j}} =(F_Q)_{\lambda'_j}$. Thus $X_\lambda{\big /}S \subset
X_{\lambda'}$.
Moreover, every element of $X_{\lambda'}$ can be seen to have a unique lifting to $\{\phi \in X_\lambda\ |\ S \supset \ker \phi\}$. To see this, note that by a change of coordinates which preserves $F_V$ we can make $S$ is a coordinate subspace, in which case, this is obvious. Thus $X_\lambda{\big /}S = X_{\lambda'}$.
### Genericity of $S$
The one remaining fact we will need is that when $S$ is the special subspace $S = \ker \phi$, $\phi \in \bigcap_{i=1}^s X_{\lambda^i}$, then the flags $F_S$ and $F_{V/S}$ are actually [*generic*]{} flags. It is a priori conceivable that by choosing this particular $S$ (its definition involves the flags $F^i$), we could have undone all the genericity we had before, thereby ending up in a position where all the flags $F^i_{V/S}$ are not generic with respect to each other (for example, they could all be equal). This would be most unfortunate; however luckily it does not happen.
Let $\rho^i$ denote the Schubert position of $S \subset V$ with respect to the flag $F^i_V$. Thus $S \in \bigcap_{i=1}^s \Omega^{F^i_V}_{\rho^i}$. We must show that $S$ is in fact a generic point of this intersection of Schubert varieties. This is sufficient, since the induced flags at a generic point of an intersection of Schubert varieties are generic (c.f. remark \[rmk:genericinducedflags\]).
The idea is to fix generic flags $F^i_V$, while the flags $F^i_Q$ vary. We show that there cannot exist a subvariety $T \subset \bigcap_{i=1}^s \Omega^{F^i_V}_{\rho^i}$ such that $\ker \phi \in T$, for a generic choice of $\phi \in \bigcap_{i=1}^s X_{\lambda^i}$, and generic induced flags $F^i_Q$. This is a consequence of a generalisation of the Kleiman moving lemma, due to Belkale.
\[lem:moving\] Let $H$ be an algebraic group acting on a variety $X$. Suppose $\pi: X \to Y$ is an $H$-invariant fibration, such that $H$ acts transitively on the fibres. Let $Z_i \subset X$, and $Y_i=\pi(Z_i) \subset Y$ $i \in \{1, \ldots, s\}$, be subvarieties, such that $\pi|_{Z_i}: Z_i \to Y_i$ is a fibration. Put $Y_0 = \bigcap_{i=1}^s Y_i$, and $X_0 = \pi^{-1}(Y_0)$. Let $T \subset Y_0$, be a subvariety. Then for generic $h_i \in H$, the intersection $$\pi^{-1}(T) \cap \left ( \bigcap_{i=1}^s h_i \cdot Z_i \right )$$ has the expected dimension as an intersection inside $X_0$. That is, $$\label{eqn:moving}
{\mathop{\mathrm{codim}}}\pi^{-1}(T) \cap \left ( \bigcap_{i=1}^s h_i \cdot Z_i \right )
= {\mathop{\mathrm{codim}}}\pi^{-1}(T) +
\sum_{i=1}^s {\mathop{\mathrm{codim}}}(Z_i \cap X_0)$$ (here ${\mathop{\mathrm{codim}}}$ means “codimension inside $X_0$”).
In this example, $$X = {\mathrm{Hom}}_d(V,Q)
= \{\psi \in {\mathrm{Hom}}(V,Q)\ |\ \dim \ker \psi = d\},$$ $$Y = Gr(n-d,V), \qquad \pi(\psi) = \ker(\psi)$$ The group $H$ is $GL(Q)$, which we note preserves the kernel of $\psi \in Z_i$. The action of $GL(Q)$ acts transitively on the induced flags $F^i(Q)$.
The subvariety $Z_i \subset X_{\lambda^i} \subset X$ consists of those elements of $X_{\lambda^i}$ whose kernel is in Schubert position $\rho^i$. Thus a generic translate $h_i \cdot Z_i$ is simply $X_{\lambda^i}$ for a different generic choice of flags $F^i_Q$. Thus the image $\pi(\psi)$ of a generic point in the intersection $\psi \in \bigcap_{i=1}^s h_i \cdot Z_i$, is just the kernel of a generic element of $\bigcap_{i=1}^s X_{\lambda^i}$.
The image of $Z_i$ under $\pi$ is the open Schubert cell $Y_i = \Omega^{F^i_V}_{\rho^i}$. We wish to show that there is no subvariety $T \subset \bigcap Y_i$ such that $\ker \psi \in T$ for all choices above. In other words, for any subspace $T$ we can choose $\psi$ and generic flags $F^i_Q$ such that $\psi \in \bigcap_{i=1}^s X_{\lambda^i}$ and $\ker \psi \notin T$. Equivalently, we must show there exists $\psi \in \bigcap h_i \cdot Z_i$, with $\psi \notin \pi^{-1}(T)$. But this is clear from lemma \[lem:moving\], since otherwise the codimensions in equation (\[eqn:moving\]) would not add.
\[cor:genericdescent\] Let $S = \ker \phi$ for a generic $\phi \in \bigcap_{i=1}^s
X_{\lambda^i}$. Then the $X_{\lambda^i}{\big /}S = X_{{\lambda^i}'}$ are in general position. Moreover there is an element $\tilde \phi \in \bigcap_{i=1}^s X_{{\lambda^i}'}
\subset {\mathrm{Hom}}(V/S,Q)$ such that $\ker \tilde \phi = {\{0\}}$.
The fact that the $X_{\lambda^i}{\big /}S$ are generic simply means that the induced flags $F^i_{V/S}$ and $F^i_{Q}$ are generic, which is what we have just shown. For the second statement, since $\ker \phi = S$, $\phi$ descends to well defined map $\tilde \phi:V/S \to Q$ with $\tilde \phi \in \bigcap_{i=1}^s X_{\lambda^i}{\big /}S$ and $\ker \tilde \phi = {\{0\}}$.
\[cor:transversereduction\] The intersection $\bigcap_{i=1}^s X_{\lambda^i}{\big /}S$ is transverse.
The argument can be seen as special case of [@B]\[lemma 2.18\].
Since the $X_{\lambda^i}{\big /}S = X_{{\lambda^i}'}$ are linear subspaces of ${\mathrm{Hom}}(V/S,Q)$ their intersection is necessarily equidimensional. Thus it suffices to show that the intersection is transverse on an open subset which contains a point of intersection.
Consider the space ${\mathrm{Hom}}_0(V/S,Q) = \{\psi \in {\mathrm{Hom}}(V/S,Q)\ |\ \ker \psi = {\{0\}}\}.$ This is a homogeneous space under the action of $GL(V/S) \times GL(Q)$, and it is a Zariski open subset of ${\mathrm{Hom}}(V/S,Q)$. By corollary \[cor:genericdescent\] the $X_{{\lambda^i}'}$ are generic translates of $\hat X_{{\lambda^i}'}$ by elements of $GL(V/S) \times GL(Q)$. So by Kleiman-Bertini, the intersection $$\bigcap_{i=1}^s \left({\mathrm{Hom}}_0(V/S,Q) \cap X_{\lambda^i}\right)$$ is a transverse intersection. Moreover, it contains the point $\tilde \phi$ which shows that the intersection $\bigcap_{i=1}^s X_{{\lambda^i}'}$ is transverse.
Two-step flag manifolds {#sec:twostep}
-----------------------
### The tangent space to a Schubert variety in a two-step flag manifold
Let $0<d<r<n$. We consider the two-step flag manifold $Fl(d,r,{\mathbb{C}}^n)$. For two-step flag manifolds, we use the notation $\hat{Y}_{\sigma}$ for $\hat{Z}_\sigma$, and $Y_\sigma$ for the generic translate $Z_\sigma$. Here $\sigma$ is a $012$-string, with $d$ [[‘$2$’]{}s ]{}, $r-d$ [[‘$1$’]{}s ]{}, and $n-r$ [[‘$0$’]{}s ]{}. Let $\eta_1 < \eta_2 < \cdots <\eta_{n-r}$ denote the positions of the [[‘$0$’]{}s ]{}(i.e $\sigma_{\eta_m}=0$, for $m \leq n-r$). Let $\eta_{n-r+1} < \cdots < \eta_{n-d}$ denote the positions of the [[‘$1$’]{}s ]{}, and $\eta_{n-d+1} < \cdots < \eta_{n}$ denote the positions of the [[‘$2$’]{}s ]{}.
Our first objective is to describe the space $\hat{Y}_{\sigma}$. Let $V$ denote the coordinate subspace $V = \langle x_{n-r+1}, \ldots, x_n \rangle$, and $S$ denote the coordinate subspace $S = \langle x_{n-d+1}, \ldots, x_n \rangle$. (Eventually $d$ and $S$ will play the same role as they did in section \[sec:grassmannian\].) We take our base flag $V^0$ to be the coordinate two-step flag $$V^0 = {\{0\}}\subset S \subset V \subset {\mathbb{C}}^n.$$
Now ${\mathfrak{g}}/{\mathfrak{p}}$ has a basis descending from the standard basis $$\big\{ E_{jk}
\ \big|\ \mbox{$j \leq n-r$, $k>n-d$, and ($k>n-d$ or $j \leq n-r$)}
\big\}$$ where $E_{jk}$ is the image under ${\mathfrak{g}}\to {\mathfrak{g}}/{\mathfrak{p}}$ of the $n \times n$ matrix whose only non-zero entry is a [[‘$1$’]{} ]{}in the $(j,k)$-position. These basis vectors naturally partition ${\mathfrak{g}}/{\mathfrak{p}}$ into three blocks: the upper left block is spanned by those $E_{jk}$ such that $1 \leq j \leq n-r$, $n-r < k \leq n-d$; the lower right block is spanned by those $E_{jk}$ such that $n-r < j \leq n-d$, $n-d < k \leq n$, and the upper right block is spanned by those $E_{jk}$ such that $1 \leq j \leq n-r$, $n-d < k \leq n$. The first two are naturally viewed as subspaces, while the last is more naturally viewed as a quotient space.
Since $S$ and $V$ are coordinate subspaces, $\hat{Y}_{\sigma}$ is spanned by some subset of the $E_{jk}$.
$\hat{Y}_{\sigma} = \{E_{jk}\ |\ \eta_j < \eta_k\}$.
$\eta = \eta_1 \ldots \eta_n$ is the element of the $S_n$ such that $\eta^{-1} \cdot {F^{\mathrm{std}}}= \hat{F}(\sigma)$. Thus ${\mathfrak{b}}(\hat{F}(\sigma)) = \eta^{-1} \cdot {\mathfrak{b}}({F^{\mathrm{std}}})$. One can easily check that the image in ${\mathfrak{g}}/{\mathfrak{p}}$ is $\{E_{jk}\ |\ \eta_j < \eta_k\}$.
\[ex:Y\] Let $d=2$, $r=5$, $n=9$. Then ${\mathfrak{g}}/{\mathfrak{p}}$ looks like $$\left(
\begin{matrix}
\cdot & \cdot & \cdot & \cdot & * & * & * & * & * \\
\cdot & \cdot & \cdot & \cdot & * & * & * & * & * \\
\cdot & \cdot & \cdot & \cdot & * & * & * & * & * \\
\cdot & \cdot & \cdot & \cdot & * & * & * & * & * \\
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & * & * \\
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & * & * \\
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & * & * \\
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot
\end{matrix}
\right)$$ Let $\sigma = 021010201$. Then $\eta = 146835927$, and $$\hat{Y}_{\sigma} =
\left(
\begin{matrix}
\cdot & \cdot & \cdot & \cdot & * & * & * & * & * \\
\cdot & \cdot & \cdot & \cdot & 0 & * & * & 0 & * \\
\cdot & \cdot & \cdot & \cdot & 0 & 0 & * & 0 & * \\
\cdot & \cdot & \cdot & \cdot & 0 & 0 & * & 0 & 0 \\
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & 0 & * \\
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & 0 & * \\
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & 0 & 0 \\
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot
\end{matrix}
\right )$$ Note that the above pictures of ${\mathfrak{g}}/{\mathfrak{p}}$ and $\hat{Y}_\sigma$ break up into three blocks, $$\left(
\begin{matrix}
* & * & * \\
0 & * & * \\
0 & 0 & * \\
0 & 0 & *
\end{matrix}
\right)
\qquad
\left(
\begin{matrix}
* & * \\
0 & * \\
0 & * \\
0 & 0
\end{matrix}
\right)
\qquad
\left(
\begin{matrix}
0 & * \\
0 & * \\
0 & 0
\end{matrix}
\right )$$ and each block contains a Young-diagram shaped picture. These pictures are $\hat{X}_{\sigma(01)}$, $\hat{X}_{\sigma(02)}$ and $\hat{X}_{\sigma(12)}$.
### Grassmannian problems in the two-step flag manifold
In example \[ex:Y\] one can see that ${\mathfrak{g}}/{\mathfrak{p}}$ consists of three rectangular blocks, and the diagram representing $\hat{Y}_\sigma$ is shaped like a Young diagram when restricted to each of these blocks. Thus these are actually diagrams for some $\hat{X}_\lambda$.
Inside ${\mathfrak{g}}/{\mathfrak{p}}$ there are two natural subspaces, corresponding to the fibres of the forgetful maps $G/P \to Gr(r,{\mathbb{C}}^n)$ and $G/P \to Gr(d,{\mathbb{C}}^n)$. The subspaces are ${\mathrm{Hom}}(S,V)$, and ${\mathrm{Hom}}(V/S, Q)$ (where $Q = {\mathbb{C}}^n/V$). In terms of our basis for ${\mathfrak{g}}/{\mathfrak{p}}$, these correspond to the lower right and upper left blocks respectively. Note that these subspaces are in fact $P$-invariant. Thus the quotient map $$\Pi_{{\mathrm{Hom}}(S,Q)} : {\mathfrak{g}}/{\mathfrak{p}}\to {\mathrm{Hom}}(S, Q) =
({\mathfrak{g}}/{\mathfrak{p}})\big/ \big({\mathrm{Hom}}(S,V) \oplus {\mathrm{Hom}}(V/S,Q) \big )$$ is $P$-equivariant.
By simply noting that the dimensions are correct we see that is it possible to view $X_{\sigma(01)}$ as a subspace of ${\mathrm{Hom}}(V/S,Q)$, $X_{\sigma(12)} \subset {\mathrm{Hom}}(S,V)$, and $X_{\sigma(02)} \subset {\mathrm{Hom}}(S,Q)$. However, better than this, we have the following result.
We have the following identifications.
1. $Y_\sigma \cap {\mathrm{Hom}}(V/S,Q) = X_{\sigma(01)}$.
2. $Y_\sigma \cap {\mathrm{Hom}}(S,V) = X_{\sigma(12)}$.
3. $\Pi_{{\mathrm{Hom}}(S,Q)}(Y_\sigma) = X_{\sigma(02)}$.
Moreover, if $Y_{\sigma^i}$, $i \in \{1, \ldots, s\}$ are in general position, then so are $X_{\sigma^i(01)}$, $X_{\sigma^i(12)}$, and $X_{\sigma^i(02)}$.
Since the spaces ${\mathrm{Hom}}(V/S,Q), {\mathrm{Hom}}(S,V) \subset {\mathfrak{g}}/{\mathfrak{p}}$ are $P$-invariant, it suffices to check statements 1–3 for $\hat{Y}_\sigma$.
For statement 1, let $\zeta_1< \cdots < \zeta_{n-d}$, and $\zeta_{n-d+1} < \cdots <\zeta_{n-r}$ denote the positions of the [[‘$0$’]{}s ]{}and [[‘$1$’]{}s ]{}respectively for $\sigma(01)$, and let $\lambda$ denote the partition corresponding to $\sigma(01)$. Note that $j < \lambda_k$ if and only if $\zeta_j < \zeta_k$ if and only if $\eta_j < \eta_k$. Thus we have $$\hat{Y}_\sigma \cap {\mathrm{Hom}}(V/S,Q) =
\langle E_{jk}\ |\
\mbox{$1 \leq j \leq n-r$, $n-r < k< n-d$, and $\eta_j < \eta_k$}
\rangle$$ On the other hand, $$\begin{aligned}
\hat{X}_{\sigma(01)} &=
\langle E_{jk}\ |\ E_{jk}(({F^{\mathrm{std}}}_{V/S})_l) \subset
({F^{\mathrm{std}}}_Q)_{\lambda_l},\ \forall l \rangle \\
&=
\langle E_{jk}\ |\ E_{jk}(({F^{\mathrm{std}}}_{V/S})_k) \subset
({F^{\mathrm{std}}}_Q)_{\lambda_k} \rangle \\
&=
\langle E_{jk}\ |\ ({F^{\mathrm{std}}}_Q)_j \subset
({F^{\mathrm{std}}}_Q)_{\lambda_k} \rangle \\
&=
\langle E_{jk}\ |\ j < \lambda_k \rangle \\
&=
\langle E_{jk}\ |\ \zeta_j < \zeta_k \rangle \end{aligned}$$ To see the genericity, note that $P$ acts on ${\mathrm{Hom}}(V/S,Q)$, and contains the subgroup $GL(V/S) \times GL(Q)$ (acting in the usual way). Thus if $Y_{\sigma^i}$ are generic, so are $X_{\sigma^i(01)}$.
A similar argument holds for $\sigma(12)$ and $\sigma(02)$.
Even more importantly, we can sometimes use the transversality (or lack thereof) for these Grassmannian Schubert problems to deduce transversality in the two-step flag manifold.
\[lem:splitting\] Let $\sigma^1, \ldots, \sigma^s$ be $012$-strings, giving rise to Schubert classes on the two-step flag manifold $Fl(d,r,{\mathbb{C}}^n)$.
1. If $\bigcap_{i=1}^s X_{\sigma^i(02)}$ is non-transverse, then $\bigcap_{i=1}^s Y_{\sigma^i}$ is non-transverse.
2. If $\bigcap_{i=1}^s X_{\sigma^i(01)}$, $\bigcap_{i=1}^s X_{\sigma^i(12)}$, and $\bigcap_{i=1}^s X_{\sigma^i(02)}$ are all transverse, then $\bigcap_{i=1}^s Y_{\sigma^i}$ is transverse.
These follow from elementary facts about intersections in quotient spaces and subspaces.
One important special case of this theorem is the following, which is a variant on the vanishing criterion in [@P1].
\[cor:almosthorn\] If $\sum_{i=1}^s \dim X_{\sigma^i(02)} < (s-1)d(n-r)$ then $\bigcap_{i=1}^s Y_{\sigma^i}$ non-transverse.
If $\sum_{i=1}^s \dim X_{\sigma^i(02)} < (s-1)d(n-r)$ then $\bigcap_{i=1}^s X_{\sigma^i(02)}$ is non-transverse for dimensional reasons.
### The strings $\sigma[1]$ and $\sigma[2]$
We can view ${\mathrm{Hom}}(V,Q)$ and ${\mathrm{Hom}}(S, {\mathbb{C}}^n/S)$ as quotient spaces of ${\mathfrak{g}}/{\mathfrak{p}}$ as well: ${\mathrm{Hom}}(V,Q) = ({\mathfrak{g}}/{\mathfrak{p}})\big/{\mathrm{Hom}}(S,V)$ and ${\mathrm{Hom}}(S, {\mathbb{C}}^n/S) = ({\mathfrak{g}}/{\mathfrak{p}}) \big/{\mathrm{Hom}}(V/S,Q)$. Let $\Pi_{{\mathrm{Hom}}(V,Q)}$ and $\Pi_{{\mathrm{Hom}}(S, {\mathbb{C}}^n/S)}$ denote the projection maps. Counting [[‘$1$’]{}s ]{}and [[‘$0$’]{}s ]{}, we see $X_{\sigma[2]}$ and $X_{\sigma[1]}$ are of the correct dimensions to view $X_{\sigma[2]} \subset {\mathrm{Hom}}(V,Q)$ and $X_{\sigma[1]} \subset {\mathrm{Hom}}(S, {\mathbb{C}}^n/S)$.
$\Pi_{{\mathrm{Hom}}(V,Q)}(Y_\sigma)$ is an $X_{\sigma[2]}$, and $\Pi_{{\mathrm{Hom}}(S,{\mathbb{C}}^n)}(Y_\sigma)$ is an $X_{\sigma[1]}$; however, they are not in general position.
Let $F$ be a flag such that $(S,V) \in \Omega^F_{\sigma}$. Then $V \in \Omega^F_{\sigma[2]}$, and $S \in \Omega^F_{\sigma[1]}$. It follows that $Y_\sigma = {\mathfrak{b}}(F)/{\mathfrak{p}}$ maps surjectively onto $X_{\sigma[2]} = {\mathfrak{b}}(F)/{\mathop{\mathrm{Stab}}}(V)$ and $X_{\sigma[1]} = {\mathfrak{b}}(F)/{\mathop{\mathrm{Stab}}}(S)$. The quotient maps are precisely $\Pi_{{\mathrm{Hom}}(V,Q)}$ and $\Pi_{{\mathrm{Hom}}(S, {\mathbb{C}}^n/S)}$.
They are not in general position, since the flag $F$ is almost generic for the two step flag ${\{0\}}\subset S \subset V \subset {\mathbb{C}}^n$, but not necessarily almost generic for $V$ or $S$.
We can now tie this to the material in section \[sec:grassmannianhoms\].
Fix a flag $F$ such that $(S,V) \in \Omega^{F}_{\sigma}$. View $X_{\sigma[2]} \subset {\mathrm{Hom}}(V,Q)$ and $X_{\sigma(01)} \subset {\mathrm{Hom}}(V/S,Q)$, as being tangent spaces to Schubert varieties relative the flag $F$ (or its induced flags). Then $X_{\sigma[2]}{\big /}S = X_{\sigma(01)}$.
Relative to the flag $F$, $X_{\sigma[2]}{\big /}S = \Pi_{{\mathrm{Hom}}(V,Q)}(Y_\sigma) \cap {\mathrm{Hom}}(V/S,Q)$ and $X_{\sigma(01)} = Y_\sigma \cap {\mathrm{Hom}}(V/S,Q)$. These are the same as $\Pi_{{\mathrm{Hom}}(V,Q)}$ restricted to ${\mathrm{Hom}}(V/S,Q)$ is the identity map.
Proof of Horn’s conjecture {#sec:proof}
==========================
The two-step flag manifold as a fibration
-----------------------------------------
Consider the map $q: Fl(d,r,{\mathbb{C}}^n) \to Gr(r,{\mathbb{C}}^n)$ defined by $q(S,V) = V$. This is a fibration, and the fibre over $V$ is simply the Grassmannian $Gr(d,V)$.
Schubert varieties behave well under this fibration. Let $F$ be a flag on ${\mathbb{C}}^n$, and $\sigma$ a $012$-string representing a Schubert class on $Fl(d,r, {\mathbb{C}})$. Then the restriction of $q$ to the Schubert cell $\Omega^F_\sigma$ is also a fibration. We have $q(\Omega^F_\sigma) = \Omega^F_{\sigma[2]}$ and the fibre over $V$ is the Schubert cell $\Omega^{F_V}_{\sigma(12)}$. Thus from this picture, we see that intersection of Schubert varieties in $Fl(d,r, {\mathbb{C}}^n)$ is related to two Schubert intersection problems in Grassmannians: one on the base space $Gr(r,{\mathbb{C}}^n)$ and one on the fibre $Gr(d,V)$. The precise relationship is as follows:
Let $F^1, \ldots, F^s$ be generic flags on ${\mathbb{C}}^n$. Assume that $S^{\sigma^1(12)} \cdots S^{\sigma^s(12)} \neq 0 \in H^*(Gr(d,r))$. Then $\bigcap_{i=1}^s \Omega^{F^i}_{\sigma^i}$ has a point of intersection if and only if $\bigcap_{i=1}^s \Omega^{F^i}_{\sigma^i[2]}$ has a point of intersection.
The “only if” direction is clear: if $\bigcap_{i=1}^s \Omega^{F^i}_{\sigma^i}$ is non-empty, then so is its image under $q$.
Thus suppose $I = \bigcap_{i=1}^s \Omega^{F^i}_{\sigma^i[2]}$ is non-empty. $S^{\sigma^1(12)} \cdots S^{\sigma^s(12)} \neq 0 \in H^*(Gr(d,r))$ is equivalent to the Schubert problem in the fibre $\bigcap_{i=1}^s \Omega^{F_V^i}_{\sigma^i(12)}$ having a point of intersection for $V$ which induce generic flags $F^i_V$. But by remark \[rmk:genericinducedflags\], a generic $V \in I$ has this property. Thus there is a point in $\bigcap_{i=1}^s \Omega^{F^i}_{\sigma^i}$ over a generic point in $I$. In particular this intersection is non-empty.
This simple lemma gives us a way (in fact many different ways) of turning a Grassmannian Schubert intersection problem into a Schubert intersection problem on a two-step flag manifold. Note that given any $01$-strings $\tau$ and $\mu$ with the correct number of [[‘$0$’]{}s ]{}and [[‘$1$’]{}s ]{}, we can always construct a $012$-string $\sigma$ such that $\sigma[2]=\tau$ and $\sigma(12) = \rho$.
We call $(\sigma^1, \ldots, \sigma^s)$ a [**lifting**]{} of $({\tau^1}, \ldots, {\tau^s})$, if $\sigma^i[2] = \tau^i$, $\sigma^i(12) = \rho^i$ and $S^{\rho^1} \cdots S^{\rho^s} \neq 0 \in H^*(Gr(d,r))$.
Thus we see that there is a lifting corresponding to each $s$-tuple $(\rho_1, \ldots, \rho_s)$ such that $S^{\rho^1} \cdots S^{\rho^s} \neq 0 \in H^*(Gr(d,r))$. If we allow $d$ to vary between $1$ and $r$, we also have one Horn inequality for each such triple. This is no coincidence: we are about to see that each Horn inequality arises from applying corollary \[cor:almosthorn\] to a lifting.
Necessity of the Horn inequalities
----------------------------------
We show that each lifting of a Grassmannian Schubert problem gives rise to a Horn inequality. This argument will establish the necessity of the Horn inequalities.
Suppose $\lambda^1, \ldots, \lambda^s \in \Lambda(r,n-r)$, with $$S^{\lambda^1} \cdots S^{\lambda^s} \neq 0 \in H^*(Gr(r,{\mathbb{C}}^n))$$ Then $\bigcap_{i=1}^s \Omega^{F^i}_{\lambda^i}$ contains a point of intersection. Thus so does any lifting of this Schubert problem.
Let $(\sigma^1, \ldots, \sigma^s)$ be such a lifting, say, corresponding to an $s$-tuple of $01$-strings $(\rho^1, \ldots, \rho^s)$. Let $\mu^i$ denote the partition corresponding to $\rho^i$. Since $\bigcap_{i=1}^s \Omega^{F^i}_{\sigma^i}$ is non-empty, for generic flags, the intersection of tangent spaces $\bigcap_{i=1}^s Y_\sigma^i$ must be transverse. By corollary \[cor:almosthorn\] this means that $$\label{eqn:althorn}
\sum_{i=1}^s \dim X_{\sigma^i(02)} \geq (s-1)d(n-r)$$ is a necessary inequality.
We now compute $\dim X_{\sigma(02)}$ (for ease of notation we are fixing $i$ and omitting the superscript $i$ from $\sigma, \rho, \mu, \lambda$). Let $z(j)$ denote the number of [[‘$0$’]{}s ]{}in the list $\sigma_1, \cdots, \sigma_{j-1}$. $$\begin{aligned}
\dim X_{\sigma(02)} &= \#\{(j',j)\ |\ j'<j,\ \sigma_{j'}=0,\ \sigma_j=2\} \\
&= \sum_{j\ |\ \sigma_j =2} \lambda_{j - z(j)} \\
&= \sum_{k=1}^d \lambda_{\text{position of $k^{\rm th}$ {{\rm `$1$'} }in $\rho$}} \\
&= \sum_{k=1}^d \lambda_{\mu_k+k}\end{aligned}$$
Rewriting (\[eqn:althorn\]) based on this calculation, we obtain $$\sum_{i=1}^s \sum_{k=1}^d \lambda^i_{\mu^i_k +k} \geq (s-1)d(n-r).$$ which is exactly the inequality (\[eqn:hornineq\]). Thus we see that the inequality (\[eqn:althorn\]) is precisely the Horn inequality corresponding to $(\mu_1, \ldots, \mu_s)$.
Sufficiency of the Horn inequalities
------------------------------------
To prove sufficiency of the Horn inequalities, we must show that whenever $S^{\lambda^1} \cdots \lambda^s = 0 \in H^*(Gr(r,{\mathbb{C}}^n))$, there is a Horn inequality violated by the $\lambda^i$. We have already seen that a violation of the Horn inequalities gives rise to a non-transverse intersection $\bigcap_{i=1}^s X_{\sigma^i(02)}$ for a lifting $(\sigma^1, \ldots, \sigma^s)$ of $(\lambda^1, \ldots, \lambda^s)$. We would now like to prove the reverse: a nontransverse intersection $\bigcap_{i=1}^s X_{\sigma^i(02)}$ leads to a violation of a Horn inequality. This requires an inductive argument.
### The inductive step
\[lem:induction\] Let $\lambda^1, \ldots, \lambda^s \in \Lambda(r,n-r)$, and let $(\sigma^1, \ldots, \sigma^s)$ be a lifting of $(\lambda^1, \ldots, \lambda^s)$ to Schubert varieties in $Fl(d,r,{\mathbb{C}}^n)$. Let $\mu^i$ be the partition corresponding to $\sigma^i(02)$. If $\lambda^1, \ldots, \lambda^s$ satisfies all of its the Horn inequalities, then $\mu^1, \ldots, \mu^s$ must satisfy all of its Horn inequalities.
This lemma allows us to argue by induction. Suppose the Horn conditions are sufficient for all integers $d$, with $d<r$. Suppose $S^{\lambda^1} \cdots S^{\lambda^s} = 0 \in H^*(Gr(r,{\mathbb{C}}^n))$. We show that one of the two things must be true.
1. The product is zero in cohomology for dimensional reasons.
*or*
2. There is a lifting of $(\lambda^1, \ldots, \lambda^s)$ to $(\sigma^1, \ldots, \sigma^s)$ such that the intersection $\bigcap_{i=1}^s X_{\sigma^i(02)}$ is non-transverse.
In the first case, the Horn inequality for $d=r$ is violated. In the second case, our inductive hypothesis tells us that some Horn inequality is violated by $\sigma^1(02), \ldots, \sigma^s(02)$. So by lemma \[lem:induction\], there must be some Horn inequality violated by $\lambda^1, \ldots, \lambda^s$.
Let $V \subset {\mathbb{C}}^n$, with $\dim V = r$ and fix flags $F^1, \ldots, F^s$ on ${\mathbb{C}}^n$ which are almost generic for $V$. Let $S \subset V$, with $\dim S = d$, and $S$ in Schubert position $\sigma^i(12)$ with respect to the induced flag $F^i_V$.
We work inside the two-step flag manifold $Fl(d',d,V)$. let $(\chi^1, \ldots, \chi^s)$ be a lifting of $(\sigma^1(12), \ldots, \sigma^s(12))$. Of course, since this is a lifting $$S^{\chi^1(12)} \cdots S^{\chi^s(12)} \neq 0 \in H^*(Gr(d',S)).$$ Also, the intersection of Schubert varieties $\bigcap_{i=1}^s \Omega^{F^i_V}_{\sigma^i(12)}$ is non-empty, since it contains the point $S$. Thus $\bigcap_{i=1}^s \Omega^{F^i_V}_{\chi^i}$ is non-empty.
Now we consider the fibration $p:Fl(d',d,V) \to Gr(d',V)$. As in the case with $q$, the fibration $p$ maps Schubert varieties map to Schubert varieties. The image is $p(\Omega^{F}_\chi) = \Omega^F_{\chi[1]}$. Thus we see that $$S^{\chi^1[1]} \cdots S^{\chi^s[1])} \neq 0 \in H^*(Gr(d',V)).$$
We now check that the Horn inequality for $\mu^1, \ldots, \mu^s$ corresponding to $(\chi^1(12), \ldots, \chi^s(12))$ is identical to the Horn inequality for $\lambda^1, \ldots, \lambda^s$, corresponding to $(\chi^1[1], \ldots, \chi^s[1])$. The two inequalities are $$\label{eqn:lambdaeqn}
\sum_{i=1}^s \sum_{k=1}^d
\lambda^i_{\text{position of the $k^{\rm th}$ {{\rm `$1$'} }in $\chi^i[1]$}}
\geq (s-1)d'(n-r)$$ and $$\label{eqn:mueqn}
\sum_{i=1}^s \sum_{k=1}^d
\mu^i_{\text{position of the $k^{\rm th}$ {{\rm `$1$'} }in $\chi^i(12)$}}
\geq (s-1)d'(n-r).$$ Now, $\lambda^i_l = \mu^i_{l'}$ where $l-l'$ is the number of [[‘$0$’]{}s ]{}in the list $\chi^i_1, \cdots, \chi^i_{l-1}$, thus $l'$. But $\chi^i[1]_l = 1$ if and only if $\chi^i_l = 2$ if and only if $\chi^i(12)_{l'} = 1$. Thus we see that $$\lambda^i_{\text{position of the $k^{\rm th}$ {{\rm `$1$'} }in $\chi^i[1]$}}
=\mu^i_{\text{position of the $k^{\rm th}$ {{\rm `$1$'} }in $\chi^i(12)$}}$$ and the two Horn inequalities (\[eqn:lambdaeqn\]) and (\[eqn:mueqn\]) are the same.
Since every Horn inequality for $\mu^1, \ldots, \mu^s$ arises in this way, if $(\lambda^1, \ldots, \lambda^s)$ satisfies all its Horn inequalities, then so does $(\mu^1, \ldots, \mu^s)$.
It is perhaps most natural to view this lemma as a statement about the three-step flag manifold $Fl(d',d,r,{\mathbb{C}}^n)$. We take a Schubert problem on $Gr(r,{\mathbb{C}}^n)$ and lift it to $Fl(d,r,{\mathbb{C}}^n)$. We then lift this again, to a problem on $Fl(d',d,r,{\mathbb{C}}^n)$, given by $0123$-strings $\omega^1, \ldots, \omega^s$. We can then interpret both inequalities (\[eqn:lambdaeqn\]) and (\[eqn:mueqn\]), as the statement $\sum_{i=1}^s \dim X_{\sigma^i(03)} \geq (s-1)d(n-r)$, which can be viewed as a Horn inequality for both $\lambda^i$ and $\mu^i$.
### Proof of sufficiency
We now have all the ingredients in place to prove the sufficiency of the Horn conditions.
(Horn’s conjecture) As always, let $V \subset {\mathbb{C}}^n$, with $\dim V = r$, and fix almost generic flags $F^i$. Consider a generic $\phi \in \bigcap_{i=1}^s X_\lambda^i$. Let $S$ be the kernel of $\phi$ in Schubert position $\rho^i$ with respect to generic flags $F^i_V$ on $V$. Let $(\sigma^1, \ldots, \sigma^s)$ be the lifting of $(\lambda^1, \ldots, \lambda^s)$ by $(\rho^1, \ldots, \rho^s)$.
Assume that $S^{\lambda^1} \cdots S^{\lambda^s} = 0 \in H^*(Gr(r,{\mathbb{C}}^n))$. Thus $\bigcap_{i=1}^s Y_{\sigma^i}$ is non-transverse.
There are two possibilities. Either $S=V$ or $\dim S < \dim V$. In the first case, we must have $\phi=0$, which means that $\bigcap_{i=1}^s X_\lambda^i = {\{0\}}$ (otherwise, $\phi=0$ would not be a generic choice). If $\sum_{i=1}^s {\mathop{\mathrm{codim}}}X_\lambda^i = r(n-r)$ then this is a transverse intersection, contradicting $S^{\lambda^1} \cdots S^{\lambda^s} = 0$. Thus $\sum_{i=1}^s {\mathop{\mathrm{codim}}}X_\lambda^i > r(n-r)$ which violates the Horn inequality for $d=r$.
If $\dim S< \dim V$, we note the following facts: $$\bigcap_{i=1}^s X_{\sigma^i(12)} =
\bigcap_{i=1}^s X_{\rho^i}$$ is a transverse intersection, since the Schubert varieties $\Omega^{F^i_V}_{\rho_i}$ are in general position, and contain $S$ as a point of intersection. Also $$\bigcap_{i=1}^s X_{\sigma^i(01)} =
\bigcap_{i=1}^s X_{\sigma^i[1]}{\big /}S
= \bigcap_{i=1}^s X_{\lambda^i}{\big /}S$$ is a transverse intersection, by corollary \[cor:transversereduction\].
If $\bigcap_{i=1}^s X_{\sigma^i(02)}$ is also transverse, then by lemma \[lem:splitting\], $\bigcap_{i=1}^s Y_{\sigma^i}$ would be transverse. Thus it is not a transverse intersection, and so by the inductive argument, some Horn inequality is violated.
Examples
--------
The proof of sufficiency describes a procedure for finding a Horn inequality which is violated, when $S^{\lambda^1} \cdots S^{\lambda^s} = 0.$ We now give some examples to show what happens in this process, in the simplest case which is $s=2$. (The sufficiency of the Horn inequalities is an easy fact when $s=2$; nevertheless, it illustrates the method of the proof fairly adequately.)
\[ex:simple\] Let $$\lambda^1
= {\bf 0} \leq 0 \leq 3 \leq 3 \leq {\bf 4}$$ and $$\lambda^2
= {\bf 0} \leq 1 \leq 3 \leq 3 \leq {\bf 4}$$ As in example \[ex:oppositeintersection\], we can illustrate $X_{\lambda^1}$ and $X_{\lambda^2}$ in general position as $$X_{\lambda^1} =
\left (
\begin{matrix}
0 & * & * \\
0 & * & * \\
0 & * & * \\
0 & 0 & 0
\end{matrix}
\right )
\qquad
X_{\lambda^2} =
\left (
\begin{matrix}
0 & 0 & 0 \\
+ & + & 0 \\
+ & + & 0 \\
+ & + & +
\end{matrix}
\right )$$ We’ll write these both in a single diagram as $$\begin{tabular}{ccc} \hline
{\multicolumn{1}{|c}{}}&{\multicolumn{1}{|c}{$^*$}}& {\multicolumn{1}{|c|}{$^*$}}\\\hline
{\multicolumn{1}{|c}{$_+$}}& {\multicolumn{1}{|c}{$^*_+$}}& {\multicolumn{1}{|c|}{$^*$}}\\\hline
{\multicolumn{1}{|c}{$_+$}}& {\multicolumn{1}{|c}{$^*_+$}}& {\multicolumn{1}{|c|}{$^*$}}\\\hline
{\multicolumn{1}{|c}{$_+$}}& {\multicolumn{1}{|c}{$_+$}}& {\multicolumn{1}{|c|}{$_+$}}\\\hline
\end{tabular}$$ Take a generic point $\phi \in X_{\lambda^1} \cap X_{\lambda^2}$, e.g. $$\phi =
\left (
\begin{matrix}
0 & 0 & 0 \\
0 & 3 & 0 \\
0 & 8 & 0 \\
0 & 0 & 0 \\
\end{matrix}
\right )$$ The kernel of $\phi$ has Schubert position $(101, 101)$. We use this Schubert position to lift $(\lambda^1, \lambda^2)$: $$\begin{tabular}{ccc} \hline
{\multicolumn{1}{|c}{$^*$}}& {\multicolumn{1}{|c}{}}& {\multicolumn{1}{|c|}{$^*$}}\\\hline
{\multicolumn{1}{|c}{$^*_+$}}& {\multicolumn{1}{|c}{$_+$}}& {\multicolumn{1}{|c|}{$^*$}}\\\hline
{\multicolumn{1}{|c}{$^*_+$}}& {\multicolumn{1}{|c}{$_+$}}& {\multicolumn{1}{|c|}{$^*$}}\\\hline
{\multicolumn{1}{|c}{$_+$}}& {\multicolumn{1}{|c}{$_+$}}& {\multicolumn{1}{|c|}{$_+$}}\\\hline
& {\multicolumn{1}{|c}{$_+$}}& {\multicolumn{1}{|c|}{$^*$}}\\\cline{2-3}
\end{tabular}$$ The upper right block is non-transverse for dimensional reasons. Thus we are led to consider the Horn inequality corresponding to $(101,101)$, i.e. $$\lambda^1_1 + \lambda^1_3 + \lambda^2_1 + \lambda^2_3 \geq 8.$$ The positions of the [[‘$1$’]{}s ]{}in $(101,101)$ are the indices which appear to the indices which appear on the left hand side. We see that this inequality is violated by $(\lambda^1, \lambda^2)$, as $$\lambda^1_1 + \lambda^1_3 + \lambda^2_1 + \lambda^2_3 \geq
0+3+1+3 <8.$$
Let $$\lambda^1
= {\bf 0} \leq 0 \leq 2 \leq 3 \leq 3 \leq 3 \leq 4 \leq {\bf 4}$$ and $$\lambda^2
= {\bf 0} \leq 1 \leq 1 \leq 3 \leq 3 \leq 3 \leq 3 \leq {\bf 4}.$$ We illustrate $X_{\lambda^1}$ and $X_{\lambda^2}$ as: $$\begin{tabular}{cccccc} \hline
{\multicolumn{1}{|c}{}}& {\multicolumn{1}{|c}{$^*$}}& {\multicolumn{1}{|c}{$^*$}}& {\multicolumn{1}{|c}{$^*$}}& {\multicolumn{1}{|c}{$^*$}}& {\multicolumn{1}{|c|}{$^*$}}\\\hline
{\multicolumn{1}{|c}{$_+$}}& {\multicolumn{1}{|c}{$^*_+$}}& {\multicolumn{1}{|c}{$^*_+$}}& {\multicolumn{1}{|c}{$^*_+$}}& {\multicolumn{1}{|c}{$^*$}}& {\multicolumn{1}{|c|}{$^*$}}\\\hline
{\multicolumn{1}{|c}{$_+$}}& {\multicolumn{1}{|c}{$_+$}}& {\multicolumn{1}{|c}{$^*_+$}}& {\multicolumn{1}{|c}{$^*_+$}}& {\multicolumn{1}{|c}{$^*$}}& {\multicolumn{1}{|c|}{$^*$}}\\\hline
{\multicolumn{1}{|c}{$_+$}}& {\multicolumn{1}{|c}{$_+$}}& {\multicolumn{1}{|c}{$_+$}}& {\multicolumn{1}{|c}{$_+$}}& {\multicolumn{1}{|c}{$_+$}}& {\multicolumn{1}{|c|}{$^*_+$}}\\\hline
\end{tabular}$$ Let $\phi \in X_{\lambda^1} \cap X_{\lambda^2}$ be a generic element, e.g. $$\phi =
\left (
\begin{matrix}
0 & 0 & 0 & 0 & 0 & 0 \\
0 & 5 & 6 & 7 & 0 & 0 \\
0 & 0 & 8 & 9 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 1
\end{matrix}
\right )$$ The kernel of $\phi$ has Schubert position $(100110,010011)$. We use this Schubert position to lift $(\lambda^1, \lambda^2)$: $$\begin{tabular}{cccccc} \hline
{\multicolumn{1}{|c}{$^*$}}& {\multicolumn{1}{|c}{$^*$}}& {\multicolumn{1}{|c}{$^*$}}& {\multicolumn{1}{|c}{}}&{\multicolumn{1}{|c}{$^*$}}& {\multicolumn{1}{|c|}{$^*$}}\\\hline
{\multicolumn{1}{|c}{$^*_+$}}& {\multicolumn{1}{|c}{$^*_+$}}& {\multicolumn{1}{|c}{$^*$}}& {\multicolumn{1}{|c}{$_+$}}&{\multicolumn{1}{|c}{$^*_+$}}& {\multicolumn{1}{|c|}{$^*$}}\\\hline
{\multicolumn{1}{|c}{$_+$}}& {\multicolumn{1}{|c}{$^*_+$}}& {\multicolumn{1}{|c}{$^*$}}& {\multicolumn{1}{|c}{$_+$}}&{\multicolumn{1}{|c}{$^*_+$}}& {\multicolumn{1}{|c|}{$^*$}}\\\hline
{\multicolumn{1}{|c}{$_+$}}& {\multicolumn{1}{|c}{$_+$}}& {\multicolumn{1}{|c}{$^*_+$}}& {\multicolumn{1}{|c}{$_+$}}&{\multicolumn{1}{|c}{$^*_+$}}& {\multicolumn{1}{|c|}{$_+$}}\\\hline
& & & {\multicolumn{1}{|c}{$_+$}}& {\multicolumn{1}{|c}{$^*_+$}}& {\multicolumn{1}{|c|}{$^*$}}\\\cline{4-6}
& & & {\multicolumn{1}{|c}{$_+$}}& {\multicolumn{1}{|c}{$^*_+$}}& {\multicolumn{1}{|c|}{$^*$}}\\\cline{4-6}
& & & {\multicolumn{1}{|c}{$_+$}}& {\multicolumn{1}{|c}{$_+$}}& {\multicolumn{1}{|c|}{$_+$}}\\\cline{4-6}
\end{tabular}$$ The upper left, and lower right blocks have a transverse intersection. However, the upper right block $$\begin{tabular}{ccc} \hline
{\multicolumn{1}{|c}{}}&{\multicolumn{1}{|c}{$^*$}}& {\multicolumn{1}{|c|}{$^*$}}\\\hline
{\multicolumn{1}{|c}{$_+$}}& {\multicolumn{1}{|c}{$^*_+$}}& {\multicolumn{1}{|c|}{$^*$}}\\\hline
{\multicolumn{1}{|c}{$_+$}}& {\multicolumn{1}{|c}{$^*_+$}}& {\multicolumn{1}{|c|}{$^*$}}\\\hline
{\multicolumn{1}{|c}{$_+$}}& {\multicolumn{1}{|c}{$_+$}}& {\multicolumn{1}{|c|}{$_+$}}\\\hline
\end{tabular}$$ does not. This block represents $X_{{\lambda^1}'} \cap X_{{\lambda^2}'}$ for $${\lambda^1}'
= {\bf 0} \leq 0 \leq 3 \leq 3 \leq {\bf 4}$$ $${\lambda^2}'
= {\bf 0} \leq 1 \leq 3 \leq 3 \leq {\bf 4}$$ In example \[ex:simple\] we found that the Horn inequality ${\lambda^1}'_1 + {\lambda^1}'_3 + {\lambda^2}'_1 + {\lambda^2}'_3 \geq 8$, which corresponds to $(101,101)$ is violated. To find a corresponding Horn inequality which is violated by $(\lambda^1, \lambda^2)$ we lift the Schubert position of $\ker \phi$ by $(101, 101)$ to get $(200120, 020012)$. The positions of the [[‘$2$’]{}s ]{}give the indices which appear in the relevant inequality. In this case we find that the Horn inequality $$\lambda^1_1 + \lambda^1_5 + \lambda^2_2 + \lambda^2_6 \geq 8$$ is violated. Indeed $$\lambda^1_1 + \lambda^1_5 + \lambda^2_2 + \lambda^2_6 =
0+3+1+3 <8.$$
[99]{}
P. Belkale, *Geometric Proofs of Horn and Saturation Conjectures*, J. Algebraic Geometry **15** (2006), no. 1, 133-173.
W. Fulton, *Eigenvalues of sums of Hermitian matrices.* Séminaire Bourbaki. Vol. 1997/98. Astérisque No. 252 (1998), Exp. No. 845, 5, 255–269.
W. Fulton, *Young Tableaux with Applications to Representation Theory and Geometry,* Cambridge U.P., New York, 1997.
W. Fulton, *Eigenvalues, invariant factors, highest weights, and Schubert calculus*, Bull. Amer. Math. Soc. 37 (2000), pp 209-250.
A. Horn, *Eigenvalues of sums of of Hermitian matrices*, Pacific J,. Math 12 (1962), pp. 225-241.
A. Klyachko, *Stable bundles, representation theory, and Hermitian operators*, Institute Mittag-Leffler Preprint 1996-7; earlier version, University of Marne-la-Valée, 1994.
S. Kleiman, *The transversality of a general translate*, Compositio Math. 28 (1974), 287-297.
A. Knutson, T. Tao, *The honeycomb model of $GL_n({\mathbb{C}})$ tensor products I: proof of the saturation conjecture.* J. Amer. Math. Soc. 12 (1999), no. 4, 1055–1090.
K. Purbhoo, *Vanishing and non-vanishing criteria in Schubert calculus*, to appear in Int. Math. Res. Not.
K. Purbhoo, *Root games on Grassmannians*, preprint [math.CO/0310103]{}.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In $s$-wave superconductors the Cooper pair wave function is isotropic in momentum space. This property may also be expected for Cooper pairs entering a normal metal from a superconductor due to the proximity effect. We show, however, that such a deduction is incorrect and the pairing function in a normal metal is surprisingly anisotropic because of quasiparticle interference. We calculate angle resolved quasiparticle density of states in NS bilayers which reflects such anisotropic shape of the pairing function. We also propose a magneto-tunneling spectroscopy experiment which could confirm our predictions.'
author:
- 'Yukio Tanaka$^{1,2}$ Yasuhiro Asano$^{3}$, and Alexander A. Golubov$^{4}$'
title: 'Cooper Pair Shape in Normal-metal/Superconductor Junctions'
---
It is well known that Cooper pairs consisting of two electrons are characterized by electric charge $2e$, macroscopic phase, internal spin, and by time and orbital structures [@deGennes]. The charge $2e$ manifests itself in various experiments, like Shapiro steps, flux quantization and excess current due to the Andreev reflection. The macroscopic phase generates the Josephson current [@deGennes]. The internal spin structure is classified into spin-triplet and spin-singlet states which can be identified by nuclear magnetic resonance. Further, based on a symmetry with respect to the internal time, superconducting state can belong to the even-frequency or the odd-frequency symmetry class [@Berezinskii]. The orbital degree of freedom is described by an angular momentum quantum number $l$. In $s$-wave superconductors with $l=0$, the Cooper pair wave function is spherically symmetric on the Fermi surface, i.e. an angular structure of a Cooper pair is isotropic in momentum space. The shape of Cooper pairs in $p$-wave ($l=1$) and $d$-wave ($l=2$) superconductors is characterized by two-fold and four-fold symmetries, respectively [@Sigrist]. The well established properties listed above hold in bulk superconductors. The presence of perturbations like spin-flip or interface scattering may change the symmetry of Cooper pairs. For instance, unusual odd-frequency property of Cooper pairs in proximity structures was predicted in recent studies [@Bergeret; @Golubov2007]. The shape of Cooper pair wave function in non-uniform systems like superconducting junctions is not necessarily the same as that in the bulk state. Despite the extensive study of the proximity effect during several past decades, rather little attention has been paid to the problem of Cooper pair shape in non-uniform superconducting systems [@Belzig; @GK]. This issue is quite important in view of current interest to the physics of superconducting nanostructures.
The aim of the present Letter is to clarify the consequences of breakdown of translational symmetry in superconductors on the Cooper pair shape. For this purpose, we study the proximity effect in quasi two-dimensional normal metal / superconductor (N/S) junctions by solving the Eilenberger equation, treating self-consistently the spatial variation of the superconducting pair potential. We analyze the pairing function and the local density of states (LDOS) in N/S junctions with spin-singlet $s$-wave and spin-triplet $p_{x}$-wave superconductors. Surprisingly, the pairing function in a normal metal turns out to be strongly anisotropic even in junctions with $s$-wave superconductors. To detect the complex Cooper pair shape, we propose to use scanning tunneling spectroscopy in rotating magnetic field. We show that the calculated tunneling conductance exhibits complex patterns even in the $s$-wave case.
Let us consider a quasi-two dimensional N/S junction as shown in Fig. \[fig:1\] which is the simplest example of non-uniform superconducting system, where the S region is semi-infinite and the normal metal has finite length $L$.
In the case of $p_x$-wave superconductor, we assume for simplicity that triplet Cooper pairs consist of two electrons with opposite spin projections on $z$-axis, i.e. $S_{Z}=0$. These assumptions do not limit the generality of the discussion below. We consider a perfect N/S interface with full transmissivity, while it can be shown that characteristic behavior of Cooper pairs remains qualitatively unchanged even in the presence of a potential barrier at the N/S interface.
The quasiclassical Green’s functions [@Quasi] in a normal metal (N) and a superconductor (S) are parameterized as $$\hat{g}_{\pm }^{(i)}=f_{1\pm }^{(i)} \hat{\tau}_{1}+f_{2\pm }^{(i)}\hat{\tau}%
_{2} +g_{\pm }^{(i)}\hat{\tau}_{3},\ \ (\hat{g}_{\pm }^{(i)})^{2}=\hat{1},$$where a superscript $i (= N, S)$ refers to N and S, $\hat{\tau}_{j}$ ($j=1-3$) are the Pauli matrices, and $\hat{1}$ is a unit matrix. The subscript $+ (-) $ denotes a moving direction of a quasiparticle in the $x$ direction [@Quasi], and $\bar{\Delta}_{+}(x)$ ($\bar{\Delta}_{-}(x)$) is the pair potential for a left (right) going quasiparticle. In a normal metal, $\bar{\Delta}_{\pm}(x)$ is set to zero because the pairing interaction is absent there. The Green’s functions can be expressed in terms of the Ricatti parameters [@Ricatti], $$\begin{aligned}
f_{1\pm }^{(i)} &= \mp \nu_i[\Gamma_{\pm }^{(i)}(x)+\zeta_{\pm }^{(i)}(x)]
/[1+\Gamma_{\pm }^{(i)}(x)\zeta_{\pm }^{(i)}(x)], \\
%f_{1\pm}^{(N)}&= \pm[\Gamma_{\pm}^{(N)}(x) + \zeta_{\pm }^{(N)}(x)]
%/[1 + \Gamma_{\pm }^{(N)}(x)\zeta_{\pm }^{(N)}(x)], \notag \\
f_{2\pm}^{(i)}&=i [\Gamma_{\pm }^{(i)}(x)-\zeta_{\pm }^{(i)}(x)]
/[1 + \Gamma_{\pm }^{(i)}(x)\zeta_{\pm }^{(i)}(x)], \\
g_{\pm}^{(i)}&=[1 - \Gamma_{\pm }^{(i)}(x)\zeta_{\pm }^{(i)}(x)]
/[1 + \Gamma_{\pm }^{(i)}(x)\zeta_{\pm }^{(i)}(x)],\end{aligned}$$ with $\nu_i=1$ for $i=S$ and $\nu_i=-1$ for $i=N$. The parameters $\Gamma_{\pm }^{(i)}(x)$ and $\zeta_{\pm }^{(i)}(x)$ obey the Eilenberger equation of the Ricatti type [@Ricatti], $$\begin{aligned}
i v_{Fx}\partial_{x}\Gamma_{\pm }^{(i)}(x)
& =-\bar{\Delta}_{\pm }(x)
[1+(\Gamma_{\pm}^{(i)}(x))^{2}] +2\varepsilon \nu_i \Gamma_{\pm }^{(i)}
(x), \nonumber
\\
i v_{Fx}\partial _{x}\zeta_{\pm }^{(i)}(x) &
=-\bar{\Delta}_{\pm }(x)
[1 + (\zeta_{\pm}^{(i)}(x))^{2}] -2\varepsilon \nu_i \zeta_{\pm }^{(i)}(x),
\nonumber\end{aligned}$$ with $\bar{\varepsilon}=\varepsilon+i\delta_0$ where $\varepsilon$ is the energy of a quasiparticle measured from the Fermi level and $\delta_0$ is the inverse of the mean free time due to impurity scattering. In the clean limit, we consider $\delta_0 \ll \Delta_0$. Boundary condition at $x=-L$ is given by $\zeta_{\pm}^{(N)}(-L)=-\Gamma_{\mp}^{(N)}(-L)$. Boundary condition at the N/S interface becomes $\zeta_{\pm}^{(S)}(0)=-\Gamma_{\pm}^{(N)}(0)$ and $\zeta_{\pm}^{(N)}(0)=-\Gamma_{\pm}^{(S)}(0)$. The pair potential $\bar{\Delta}_{\pm}(x)$ is expressed by $\bar{\Delta}_{\pm}(x)=\Delta (x)\Phi _{\pm }(\theta )\Theta(x)$, where a form factor $\Phi_{\pm }(\theta )$ is given by $\Phi_{\pm
}(\theta )=1$ for $s$-wave symmetry and $\pm \cos \theta $ for $p_{x}$-wave one with $\theta$ being an incident angle of a quasiparticle measured from the $x$ direction. Bulk pair potential is $\Delta(\infty)=\Delta_{0}$, and we determine the spatial dependence $\Delta (x)$ in a self-consistent way.
For $x \gg L_0$, the angular structure of $f_{2\pm}^{(i)}$ follows that of the pair potential, whereas $f_{1\pm}^{(i)}$ is zero with $L_{0}=v_{F}/2\pi T_{C}$ being a coherence length and $T_{C}$ being the transition temperature. The pairing function $f_{1\pm }^{(i)}$ is generated by inhomogeneity in a system and thus has a finite value only near the interface and in a normal metal. Recent study [@OddNS] showed that $f_{1\pm }^{(i)}$ has an odd-frequency symmetry. Generally speaking, functions $f_{1\pm}^{(i)}$ and $f_{2\pm }^{(i)}$ have opposite parities. If a superconductor has $s$- ($p_{x}$)-wave symmetry, the induced odd-frequency component has the odd (even) parity, respectively. Pairing function $f_{1}$ is defined in the angular domain of $-\pi/2 \le \theta < 3\pi/2$. We denote $f_{1}(\theta)$ by $f_{1+}(\theta)$ in the angle range $-\pi /2 \le
\theta <\pi /2$ and $f_{1}(\theta)=f_{1-}(\pi-\theta)$ for $ \pi/2
\le \theta <3\pi /2$. The angular structure of functions $f_2$ and $g$ is defined in the same manner. LDOS is given by the relation $\rho_{L}(\theta) = {\rm Real}[g(\theta)]$. In what follows, we fix temperature $T=0.05T_{C}$, the length of the normal region $L=5L_{0}$, and $\delta_{0} = 0.01\Delta_{0}$.
In Figs. \[fig:2\] and \[fig:3\] we show polar plots of $f_{1}$, $f_{2}$, and $\rho_{L}$ in a $s$-wave and $p_{x}$-wave junctions for several choices of $\varepsilon$ and $x$. Black, red and blue lines represent, respectively, the results for $x=\infty$ (superconductor), $x=0$ (interface), and $x=-L/2$ (normal metal). The odd-frequency component is always absent for $x=\infty$. Since $\rho_{L}$ is independent of $x$ in N, the resulting value of $\rho_{L}$ at $x=0$ is equal to that at $x=-L/2$.
First, we focus on the $s$-wave case (Fig. \[fig:2\]). At $\varepsilon=0$ in Fig. \[fig:2\](a), the even-frequency component $f_2$ has a circular shape reflecting $s$-wave symmetry for $x=\infty$ and 0. However, at $x=-L/2$, the shape is no longer a simple circular but a double distorted circles. The shape of $f_2$ in superconductor always has the circular shape independent of $\varepsilon$ as shown with black broken lines in Fig.2(a), (d), and (g). At $\varepsilon=0.1\Delta_0$ in Fig.2(d), $f_2$ at the interface (red line) slightly deviates from the circular shape, while the shape in N drastically changes. The tendency is more remarkable at $\varepsilon=0.5\Delta_0$ in Fig.2(g). The butterfly-like pattern seen in $f_2$ at the interface (red line in Fig.2(g)) is completely different from the original circular shape in a superconductor. At $\varepsilon=0$, $f_{1}$ at the interface becomes ellipsoidal as shown in a red line in Fig.2(b). The shape of $f_{1}$ at $x=0$ and $x=-L/2$ shows the butterfly-like pattern (Fig. 2(e)). For $\varepsilon=0.5\Delta_{0}$, the line shape of $f_{1}$ has many spikes as shown in Fig. 2(h). Such anisotropic property of $f_{1}$ and $f_2$ affects the LDOS as shown in Fig.2(c), (f), and (i). In particular, LDOS at the interface for $\varepsilon=0.5\Delta_0$ (red line in (i)) strongly deviates from the circular shape. The LDOS in a superconductor vanishes at $\varepsilon <\Delta_0$ (black lines in Fig.2(f) and (i)). At the interface, the shape of $\rho_{L}$ in Fig.2(c) is quite similar to that of $f_{1}$ in Fig.2(b).
These profiles can be qualitatively understood as follows. At $x=0$, the following relations hold $$f_{1\pm}^{(N)}=\pm \Gamma \frac{1 - \alpha^{2} }{\Xi},\
f_{2\pm}^{(N)}= i \Gamma
\frac{1 + \alpha^{2} }{\Xi}, \
g_{\pm}^N=
\frac{1 + \alpha^{2}\Gamma^{2} }
{\Xi},
\label{eq.g}$$ with $\Xi=1 - \Gamma^{2}\alpha^{2}$, $\Gamma=\Gamma_{\pm}^{(S)}(0)$ and $\alpha=\exp[2i \varepsilon L/(v_{F}\cos\theta)]$. For $\varepsilon \ll \Delta_{0}$, $\Gamma \sim 1/i$ and $f_{1\pm} \sim i
g_{\pm}$ are satisfied. Thus shape of function $f_{1}$ is similar to that of $\rho_{L}$. This argument seems to be valid even for $\varepsilon=0.1\Delta_0$ in Fig.2(e) and (f), and for $\varepsilon=0.5\Delta_0$ in Fig.2(h) and (i). The oscillating behavior in $f_1$, $f_2$, and $\rho_L$ is more remarkable at $\varepsilon =0.5\Delta_{0}$. Although we do not present calculated results of $f_{1}$ and $f_{2}$ at $x=-L/2$ for $\varepsilon=0.5\Delta_0$, the butterfly-like pattern with many spikes in the pairing functions can be seen also in a normal metal. The directions of the spin projections in LDOS are characterized by small value of $\Xi$, which has close relation to the formation of the bound states [@Rowell]. For $\theta \sim \pm \pi/2$, $\alpha$ oscillates rapidly with small variation of $\theta$, which explains the fine structures in LDOS around $\theta=\pm\pi/2$. The quasiparticle interference effect is a source of bound state formation in a normal metal. As a result, the circular shape of Cooper pairs in $s$-wave superconductor is modified into the butterfly-like pattern in a normal metal.
Next, we discuss the results for $p_{x}$-wave junctions shown in Fig. \[fig:3\]. In a superconductor ($x=\infty$), functions $f_{2}$ and $g$ are given by $\Delta_{0} \cos\theta /
\sqrt{\varepsilon^{2} - \Delta_{0}^{2}\cos^{2}\theta}$ and $\varepsilon / \sqrt{\varepsilon^{2} -
\Delta_{0}^{2}\cos^{2}\theta}$, respectively. As shown by black lines in Fig.3(d), (f), (g) and (i), the amplitudes of $f_{2}$ and $g$ become large along the directions $\theta
=\cos^{-1}(\varepsilon/\Delta_0)$. At $\varepsilon=0$ and $x=0$, formation of a mid-gap Andreev resonant state [@TK95; @ZES] significantly enhances the amplitudes of $f_{1}$ and $\rho_{L}$ compared to that of $f_{2}$ as shown by red lines in Fig.3(a), (b), and (c). For $\varepsilon=0$ and $\varepsilon=0.1\Delta_0$, the shapes of $f_{1}$ and $\rho_{L}$ at the N/S interface are similar to those in $s$-wave superconductor junctions (red lines in Fig.3(b), (c), (e) and (f)\]. The shape of $f_{1}$ and $f_{2}$ in the N region is rather complex too. At $\varepsilon=0.5\Delta_{0}$, $f_{1}$, $f_{2}$ and $\rho_{L}$ also exhibit the butterfly-like patterns as shown by red lines in Fig.3(g), (h), and (i). Similar to $s$-wave case, functions $f_{1}$ and $f_{2}$ in N have a complex line shapes with many spikes.
The butterfly-like pattern in angle resolved local density of states could be observed through the LDOS in magnetic fields.
Here, we propose an experimental setup of scanning tunneling spectroscopy (STS) in the presence of the magnetic field. As shown in Fig. \[fig:1\], magnetic field is applied parallel to the N/S plane. Tunneling current at a fixed bias voltage is measured as a function of the angle $\phi$ between the $x$ axis and the direction of magnetic field. The vector potential in this configuration is given by $(A_{x},A_{y})=-\lambda H \exp(-z/\lambda)(\sin \phi, \cos
\phi)$ [@Doppler; @Magneto]. We assume that thickness of a quasi two-dimensional superconductor is sufficiently small compared to the magnetic field penetration depth $\lambda$. Magnetic field shifts the quasiparticle energy $\varepsilon$ to $\varepsilon - H
\Delta_{0} \sin(\phi-\theta)/B_{0}$ where $B_{0}= h/(2e\pi^{2}\xi\lambda)$ and $\xi=\hbar v_F/\pi \Delta_0$. For typical values of $\xi \sim \lambda \sim 100$nm, the magnitude of $B_{0}$ is of the order of 0.02Tesla. Local density of state observed in STS experiments is given by, $ \rho(\phi)= \int^{3\pi/2}_{-\pi/2} \rho_{L}(\theta,\phi) d\theta$.
In Fig. \[fig:4\], $\rho$ is plotted as a function of $\phi$. For $\varepsilon=0$, $\Delta_{0} H/B_{0}$ gives the effective energy of a quasiparticle. In an $s$-wave superconductor, it is evident that $\rho$ is independent of $\phi$ and the amplitude of LDOS is vanishingly small, see curve A in Fig. \[fig:4\](a). In a $p_{x}$-wave superconductor, as shown by curve A in Fig. \[fig:4\](b), $\rho$ depends slightly on $\phi$ due to quasiparticle excitations near the nodal points at the Fermi surface [@Vekhter]. At the N/S interface, $\rho$ in the $s$-wave case has minima at $\phi=\pi/2$ and $3 \pi/2$ and maxima at $\phi=0$ and $\pi$, as shown by curve B in Fig. \[fig:4\](a). On the other hand, in the $p_{x}$-wave case, $\rho$ has maxima at $\phi=\pi/2$ and $3 \pi/2$, see curve B in Fig. \[fig:4\](b). This behavior can be explained in the following way. Eq. (\[eq.g\]) yields approximate expression for $\rho$ at $H \ll B_{0}$ $$\rho \sim \int^{\pi/2}_{-\pi/2} \frac{ 2(1 - a^{2})\; d\theta}
{1 + a^{2} \pm 2a \cos [C \sin(\theta-\phi)/\cos \theta]},$$ where $a=\exp(-4L \delta_{0}/\hbar |{v_{Fx}}| )$, $C=4LH/\hbar
|v_{Fx} |$, and the sign in the denominator is $+$ ($-$) for $s$-wave ($p_{x}$-wave) junctions. We note that $a$ is a positive number almost independent of $\theta$ and $C$ is a small positive number. For $\phi=n\pi + \pi/2$ with integer $n$, the magnitude of the argument of cosine function in the denominator becomes small and the denominator is reduced to $(1 \pm a)^{2}$ for $s$-wave ($p_{x}$-wave) case. When $\phi$ deviates from $n\pi + \pi/2$, the cosine function decreases and therefore $\rho$ has a dip (peak) at $\phi=\pi/2$ and $3\pi/2$ for $s$-wave ($p_{x}$-wave) cases as shown by curve B in Fig. \[fig:4\](a) ((b)). Similar argument also explains the maximum (minimum) of $\rho$ at $\phi=n\pi$ in $s$-wave ($p_{x}$-wave) junctions. When the field $H$ is increases further (curves C in Figs.4(a) and (b)), the shape of the $\rho$ exhibits more complex behavior reflecting the butterfly-like pattern of $\rho_{L}$ shown in Figs. \[fig:2\] and \[fig:3\], because the increase of $H$ has qualitatively similar effect as the increase of $\varepsilon$.
The complicated oscillating features are seen much more clear at $\varepsilon=0.5\Delta_{0}$. As shown by curves $C$ in Figs. 4(c) and (d), the period of oscillations becomes shorter than that in Figs. \[fig:4\](a) and (b), reflecting the butterfly-like patterns in Figs. \[fig:2\](i) and Figs. \[fig:3\](i). At the same time, the magnitude of the oscillations at $\varepsilon=0.5\Delta_{0}$ becomes smaller than that at $\varepsilon=0$, since the integration with respect to $\theta$ averages the butterfly-like pattern in $\rho_L$. The discussed features in LDOS at the N/S interface differ strongly from those in a bulk superconductor. We conclude that STS experiments in magnetic field should resolve the remarkable deformation of Cooper pairs.
In summary, we have studied the Cooper pair shape in normal-metal/superconductor (N/S) junctions by using the quasiclassical Green’s function formalism. The quasiparticle interference leads to striking deformations in the shape of a Cooper pair wave function in a normal metal. As a consequence, the angle resolved local density of states exhibits the butterfly-like patterns. We also show that the anisotropic shape of Cooper pairs could be resolved by scanning tunneling spectroscopy experiments in magnetic field. The Cooper pair deformation is a common feature of non-uniform superconducting systems in the clean limit. This provides a key concept to explore new quantum interference phenomena in superconducting nanostructures.
[99]{} de Gennes [*Superconductivity of Metal and Alloys*]{}, W.A. Benjamin, Inc. (1966).
M. Sigrist and K. Ueda, Rev. Mod. Phys. **63** 239 (1991).
V. L. Berezinskii, JETP Lett. **20**, 287 (1974).
F. S. Bergeret, A. F. Volkov, and K. B. Efetov, Rev. Mod. Phys. **77** 1321 (2005).
Y. Tanaka and A.A. Golubov, Phys. Rev. Lett. **98**, 037003 (2007); Y. Tanaka, A.A. Golubov, S. Kashiwaya and M. Ueda, Phys. Rev. Lett. [**9**9]{} 037005 (2007); M. Eschrig, T. Lofwander, Th. Champel, J.C. Cuevas and G. Schon, J. Low Temp. Phys. **147** 457 (2007).
W. Belzig, C. Bruder, and A. L. Fauchere, Phys. Rev. B **58** 14531 (1998).
A.A.Golubov and M.Yu.Kupriyanov, JETP Lett **67**, 501 (1998).
J.W. Serene and D. Rainer, Phys. Rep. **101** 221 (1983).
M. Eschrig, Phys. Rev. B **61** 9061 (2000), A.Shelankov and M. Ozana, Phys. Rev. B **61**, 7077 (2000); N. Schopohl and K. Maki, Phys. Rev. B **52**, 490 (1995).
Y. Tanaka, Y. Tanuma and A.A. Golubov, Phys. Rev. B [**7**6]{} 054522 (2007).
J. M. Rowell and W. L. McMillan, Phys. Rev. Lett. **16**, 453 (1966); J. M. Rowell, Phys. Rev. Lett. **30**, 167 (1973). Y. Tanaka and S. Kashiwaya, Phys. Rev. Lett. **74**, 3451 (1995); S. Kashiwaya and Y. Tanaka, Rep. Prog. Phys. **63**, 1641 (2000), T. Löfwander, V. S. Shumeiko, and G. Wendin, Supercond. Sci. Technol. **14**, R53 (2001).
L.J. Buchholtz and G. Zwicknagl, Phys. Rev. B **23**, 5788 (1981); J. Hara and K. Nagai, Prog. Theor. Phys. **74**, 1237 (1986); C.R. Hu, Phys. Rev. Lett. **72**, 1526 (1994); C. Bruder, Phys. Rev. B **41**, 4017 (1990).
M. Fogelström, D. Rainer and J. A. Sauls, Phys. Rev. Lett. 79 281 (1997).
Y. Tanaka, Y. Tanuma, K. Kuroki and S. Kashiwaya, J. Phys. Soc. Jpn. **71**, 2102 (2002). I. Vekhter, P. J. Hirschfeld, J. P. Carbotte, and E. J. Nicol, Phys. Rev. B 59 (1999) R9023.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We demonstrate a precision magnetic microscope based on direct imaging of the Larmor precession of a $^{87}$Rb spinor Bose-Einstein condensate. This magnetometer attains a field sensitivity of 8.3 pT/Hz$^{1/2}$ over a measurement area of 120 $\mu$m$^2$, an improvement over the low-frequency field sensitivity of modern SQUID magnetometers. The corresponding atom shot-noise limited sensitivity is estimated to be 0.15 pT/Hz$^{1/2}$ for unity duty cycle measurement. The achieved phase sensitivity is close to the atom shot-noise limit suggesting possibilities of spatially resolved spin-squeezed magnetometry. This magnetometer marks a significant application of degenerate atomic gases to metrology.'
author:
- 'M. Vengalattore$^{1}$'
- 'J. M. Higbie$^{1*}$'
- 'S. R. Leslie$^1$'
- 'J. Guzman$^1$'
- 'L. E. Sadler$^1$'
- 'D. M. Stamper-Kurn$^{1,2}$'
title: 'High-Resolution Magnetometry with a Spinor Bose-Einstein Condensate'
---
[^1]
Precision magnetometers that map magnetic fields with high spatial resolution have been applied to studies of condensed matter systems [@CupSC], biomagnetic imaging [@BioMag1] and tests of fundamental symmetries [@EDM2]. Many of these applications require the measurement of magnetic fields at low ($<$ 10 Hz) frequencies. Current technologies capable of micron-scale magnetic microscopy include superconducting quantum interference devices (SQUIDs), scanning Hall probe microscopes, magnetic force microscopes and magneto-optical imaging techniques [@Bending]. Of these, SQUIDs offer the highest sensitivity, demonstrated at 30 pT/Hz$^{1/2}$ over a measurement area of around 100 $\mu$m$^2$ [@SQUID1]. The low-frequency sensitivity of these devices is limited by $(1/f)$ flicker noise of unknown origins [@Koch]. Magnetic fields may also be sensed by detecting the Larmor precession of spin polarized atomic gases. To date, atomic magnetometers have achieved field sensitivities of 0.5 fT/Hz$^{1/2}$ over measurement volumes of 0.3 cm$^3$ [@Romalis1]. However, attaining high spatial resolution with a hot-vapor medium is precluded by rapid thermal diffusion of the atoms, restricting the minimum resolved length scale of these magnetometers to around 1 mm.
Trapped ultracold gases present an attractive medium for a variety of precision measurements due to their negligible Doppler broadening and long coherence times [@Treutlein; @JMH; @Jo]. Spinor Bose gases, comprised of atoms with non-zero spin, the orientation of which is free to vary, are particularly well suited to magnetic microscopy. In contrast with hot-vapor atomic magnetometers, the suppression of thermal diffusion in a gas through Bose-Einstein condensation enables precise measurements at high spatial resolution. Also, density-dependent mean field shifts, which deleteriously affect other types of precision measurements using dense ultracold gases, do not affect Larmor precession due to the rotational invariance of interparticle interactions in a spinor gas [@Ho; @Ohmi].
Here, we perform precise magnetic microscopy with high two-dimensional spatial resolution using a $^{87}$Rb $F=1$ spinor Bose-Einstein condensate (BEC). In our magnetometer, longitudinally spin-polarized spinor condensates are prepared in an optical trap. Larmor precession is induced by tipping the magnetization perpendicular to a bias field imposed along the axis of the condensate. The spins in each region of the condensate then precess at a rate that is proportional to the [*local*]{} magnetic field. After a variable integration time, the condensate is probed using magnetization-sensitive imaging to extract the local Larmor phase. The [*difference*]{} in this phase between various regions of the condensate reveals the spatial variations of the magnetic field.
The determination of the accrued Larmor phase of a coherent spin state, such as the transversely magnetized condensate, is subject to an uncertainty in the initial phase of $\delta \phi_a = 1/\sqrt{N}$ due to projection noise of measuring $N$ atoms. This noise limits the field sensitivity over a measurement area $A$ to $\delta B = \frac{\hbar}{g \mu_B} \frac{1}{\sqrt{\tau D T}} \frac{1}{\sqrt{\tilde{n} A}}$ where $\tau$ is the Zeeman coherence time and $\tilde{n}$ the local column density of the gas. We assume the measurement is repeated over a total measurement time $T$ at a duty cycle $D$. The $A^{-1/2}$ scaling of field sensitivity with the measurement area $A$ for the atomic magnetometer may be compared with the area scaling for SQUID magnetometers. This scaling ranges between $A^{-3/4}$, for a fixed SQUID sensor coupled optimally to a variable pickup loop, and $A^{-5/8}$, for direct sensing with a SQUID optimized to operate at the quantum limit for the noise energy [@Koch2]. For either scaling, the atomic magnetometer outperforms SQUID magnetometers at small measurement areas (Fig. \[fig:MagComp\]).
![Field sensitivity for repeated measurements using the spinor BEC magnetometer. Curves marked ASN (PSN) represent atom (photon) shot-noise limited sensitivities, assuming $\tau = 250$ ms, $D = 1$, and the atomic column density and probe light levels for our experiment (see text). Diffusion of magnetization limits the sensitivity for a given length scale by imposing a limit on $\tau$ (short dashed line, assuming $D = 1$). The gray line indicates the measured spatial root Allan variance; the sensitivity demonstrated in measurements of an applied localized magnetic field, assuming duty cycles of $D = 0.003$ ($\bullet$) or $D =1$ ($\circ$), is also shown. Results are compared both to the ideal sensitivity of a quantum-limited SQUID magnetometer (dot-dashed line) and to demonstrated low-frequency sensitivities [@SQUID1; @SQUID6] ($\bigtriangleup$).[]{data-label="fig:MagComp"}](arXfig1){width="43.00000%"}
Optical detection of Larmor precession is limited also by photon shot noise. In this work, the Larmor precession phase is measured by repeated phase-contrast imaging of the condensate using circular polarized light [@JMH]. For our probe detuning of $\delta = 2
\pi \times 500$ MHz below the $F=1 \rightarrow F'=2$ (D1) transition of $^{87}$Rb, the phase contrast signal can be written as $s \simeq
1 + 2 \tilde{n} \sigma_0 (\gamma/2 \delta) [c_0 + c_1 \langle F_y
\rangle]$ where $\sigma_0 = 3 \lambda^2/2 \pi$ is the resonant cross-section, $\gamma$ is the natural linewidth and $F_y$ is the projection of the local atomic spin on the imaging axis $\hat{y}$, which is made perpendicular to the field axis. The detuning-dependent constants $c_0 = 0.118$ and $c_1 = 0.274$ describe the isotropic polarizability and optical activity, respectively. We neglect the effects of linear birefringence ($\propto \! \langle
F_y^2\rangle$). An estimate of the Larmor precession phase is obtained by tracking the sinusoidal oscillation of the phase contrast signal across the sequence of phase contrast images. The photon shot-noise limited sensitivity of this estimate is then $\delta \phi_\gamma \approx \sqrt{\frac{2}{\eta
N_p}}{\frac{\sqrt{1+\tilde{n} \sigma_0 (\gamma/2 \delta)
c_0}}{\tilde{n} \sigma_0 (\gamma/2 \delta) c_1}}$, giving a field sensitivity limit of $\delta B = \frac{\hbar}{g \mu_B} \frac{\delta
\phi_\gamma}{\sqrt{\tau D T}}$. Here, $\eta$ is the detection quantum efficiency and $N_p$ is the total photon fluence, integrated across the multi-pulse imaging sequence, within the region of interest.
For our demonstration, spin-polarized $^{87}$Rb condensates of up to $1.4 \times 10^6$ atoms were confined in a single-beam optical dipole trap characterized by trap frequencies $(\omega_x,
\omega_y, \omega_z) = 2 \pi (165, 440, 4.4)$ s$^{-1}$ [@JMH]. The tight confinement along the imaging axis (condensate radius $r_y
= 2.0\, \mu$m) ensured that the condensate is effectively two-dimensional with respect to spin dynamics. Larmor precession of the condensate was induced in the presence a bias field of 165(7) mG aligned along the long axis of the condensate. A measurement integration time of 250 ms was chosen; at longer times, measurements were hampered by uncontrolled motion of the condensate along the weakly confining dimension (see below).
We operated our ultracold-atom magnetometer under two testing conditions. In one, the magnetometer was used to sense the long length-scale inhomogeneous background magnetic field in our apparatus. While these background fields were partially cancelled by suitable electromagnets, the remaining background typically varied from shot to shot. To account for this fluctuating background, the field profile along the long axis of the condensate was determined by a third-order polynomial fit to the magnetometer measurements [@backgroundnote]. The residuals from this fit were then analyzed to characterize noise limits to our magnetometer.
In the second testing condition, we used the magnetometer to measure a deliberately applied, localized magnetic field. This field was simulated using a circularly polarized laser beam at a wavelength of 790 nm. The choice of wavelength and polarization ensured that this beam imposed a local optically-induced Zeeman shift [@Dalibard] on the trapped atoms (Fig. \[fig:LP\]a). The beam was aligned at an angle $\theta \sim 60^\circ$ to the direction of the bias field, incident and focussed in the plane perpendicular to the imaging axis. The magnetic background for each run of the magnetometer was again determined by third-order polynomial fits, but using measurements from regions far from the localized field. The Zeeman shift due to the localized field was extracted from the residual of this fit.
Measurements of this localized field were affected by small center-of-mass oscillations of the condensate along its long axis. An oscillation with amplitude $\delta z$ blurs the magnetic landscape and washes out features comparable to or smaller than $\delta z$. Unable to eliminate such oscillations completely, we monitored the condensate motion for each run of the magnetometer by a sequence of four images spaced at a quarter period of the axial trap frequency, taken prior to the Larmor imaging sequence. We discarded measurements for which an excursion comparable to the width of the localized field was indicated. Two-dimensional maps of the magnetic field were obtained from a pixel-by-pixel analysis of the Larmor precession phase at each coordinate within the profile of the condensate (Fig.\[fig:LP\]b). The frame-to-frame variation of this signal showed the characteristic oscillations due to Larmor precession as well as an overall decay of the condensate number due to off-resonant scattering of probe light. This decay was taken into account in obtaining an unbiased estimate of the local Larmor phase. Our 2D approach was found to be susceptible to imaging aberrations, primarily in the narrow ($\hat{x}$) dimension of the gas.
Alternately, more robust measurements were obtained by integrating the field measurement over the $\hat{x}$ direction to reduce measurements to a single resolved direction along the long axis ($\hat{z}$) of the gas. For this, the aberrated signal profile in the $\hat{x}$ direction was determined at each $z$ coordinate in the images from averages over the multiple frames. The phase-contrast signal height in each image frame and at each $z$ coordinate was then determined. 1D phase profiles were obtained as before (Fig. \[fig:LP\]c).
![(a) Sequence of phase contrast images of a condensate under the influence of a local optically-induced Zeeman shift. (b) The resulting 2-D map of the magnetic field obtained by a pixel-by-pixel estimation of the Larmor precession phase. (c) 1-D phase profile showing the applied field inhomogeneity. The peak strength of the field as determined by this single-shot measurement is $166.2 \pm 1.2$ pT. []{data-label="fig:LP"}](arXfig2){width="44.00000%"}
The demonstrated sensitivity of our magnetometer is shown in Fig. \[fig:MagComp\]. The spatial root Allan variance from the 1D data was determined for each of 15 runs of the magnetometer under the first testing conditions (magnetic background only) and then averaged. Here, the measurement *area* is determined by accounting for the effective $5.3\, \mu$m length over which the aberrated signals are averaged in the $\hat{x}$ direction. This noise level matches closely with photon shot-noise estimates and is $\sim$3 times that due to atomic shot-noise given the number of atoms in the corresponding areas. Excess noise for areas larger than about 20 $\mu$m$^2$ was found to correlate with the local intensity of the probe light, an effect we attribute to probe-light induced shifts of the Larmor precession frequency during imaging. This noise can be reduced by using a linearly polarized probe with a more homogenous intensity profile and by carefully aligning the magnetic bias field to be perpendicular to the imaging axis.
Results from measurements under the second testing condition (background plus localized field) are shown in Fig. \[fig:DR\]. Here, the strength of the applied field (peak value of Gaussian fits) was measured repeatedly at several powers of the field-inducing laser beam. From all these measurements, a calibration between the laser power and the localized field strength was obtained. From the residual scatter in the measurements, we determine the (rms) sensitivity of our Larmor precession phase measurements as $1.0 \times 10^{-2}$ rad, corresponding to a single-shot field sensitivity of 0.9 pT over the 120 $\mu\mbox{m}^2$ area under the Gaussian field profile. This sensitivity was demonstrated for field strengths up to 60 pT. A marginally larger variance for higher fields points to the existence of small systematic effects, e.g. residual motion of the condensate or variations in the localized field strength. Averaging over the entire range of measurements shown in Fig. \[fig:DR\] yields a phase sensitivity of $1.2 \times 10^{-2}$ rad.
![Single-shot measurements of the localized magnetic field imposed by a laser beam focused to $\sigma=24\, \mu$m rms width. Residuals from a linear fit, calibrating the field magnitude [*vs*]{} laser power, are shown. Error bars indicate standard deviations for 10 measurements per setting. The measurement time was 250 ms.[]{data-label="fig:DR"}](arXfig3){width="37.00000%"}
Under repeated operation, our magnetometer, with a low duty cycle of just $D = 3 \times 10^{-3}$, attains a field sensitivity of 8.3 pT/Hz$^{1/2}$, an improvement over that demonstrated for low-frequency ($<$ 10 Hz) field measurements with modern SQUID magnetometers [@SQUID1; @SQUID6]. Plausible extensions of current cold-atom experimental methods should enable duty cycles of order unity. At full duty cycle, our demonstrated single-shot sensitivity would correspond to a field sensitivity of 0.5 pT/Hz$^{1/2}$.
As discussed above, in the photon shot-noise limit, the sensitivity of an atomic magnetometer increases with increasing probe fluence. While calculations based on linear Raman scattering rates indicated that reliable phase measurements could be obtained even at a fluence of 3400 photons/$\mu$m$^2$ per frame, it was found that the light induced losses of the condensate far exceeded those predicted by the calculations. The discrepancy was attributed to superradiant Raman scattering of atoms into the $F=2$ hyperfine states, in which atoms would no longer be observed by our probe. To counter this problem, the superradiant gain was reduced by lowering the probe intensity. Each frame in the imaging sequence was obtained by integrating the light from four pulses of light, each of duration 2.2 $\mu$s and spaced by the Larmor period of $\sim$10 $\mu$s. We also increased the motional decoherence of the $F=2$ atoms produced during superradiance by their preferential scattering of incident light resonant with the $F=2 \rightarrow F'=3$ (D2) transition. Together, these strategies enabled a probe fluence of 750 photons/$\mu$m$^2$ per frame.
Our magnetometry medium, though Bose condensed, is still a gas in which atoms are free to move. Thus, in determining the phase shift accrued due to a local magnetic field, one must consider atomic motion due to both quantum-mechanical and classical effects. For instance, imposing a weak inhomogeneous field of characteristic length $\sigma$ leads to quantum diffusive motion of the fluid. For times $\tau > \tau_Q = m \sigma^2 /
\hbar$, with $m$ the atomic mass, the motion of the spinor gas will greatly reduce the phase accrued due to Larmor precession. This evolution can be considered to be the quantum limit of thermal diffusion observed in NMR studies [@carr54]. In our experiment, the condition $\tau = 250 \, \mbox{ms} > \tau_Q$ is reached at length scales below 10 $\mu$m (Fig. \[fig:MagComp\]). For integration times $\tau \ll \tau_Q$, effects of quantum diffusion require that phase measurements $\phi$ be corrected by an amount $\delta\phi/\phi \sim (\tau/\tau_Q)^4$.
An inhomogeneous magnetic field also exerts forces on magnetic dipoles. In the extreme case of static inhomogenous fields with a Zeeman energy comparable to the chemical potential ($\sim 10^5$ times larger than those studied in this work), these classical forces can modify the density distribution of the condensate [@lowresmag]. In our case, these forces result in small corrections and limits to the accrued Larmor phase. The field strength $B$ and a characteristic length $\sigma$ for its variation define a classical time scale $\tau_C = \sqrt{m
\sigma^2/\mu_B B}$, the time taken by an atom to move $\sigma$ when accelerated by this field. For an integration time $\tau$, this classical motion imposes a limit on the maximum detectable phase shift (when $\tau = \tau_C$) of $\phi_m \simeq \tau_Q/\tau$. It should be noted that neither the diffusion of an imprinted phase nor the limitation on the dynamic range are fundamental; both can be eliminated by constraining atomic motion, e.g. by imposing an optical lattice potential.
To observe the dilution of magnetization due to atomic motion, we imposed a light-induced magnetic field localized to a beam waist of 5.4 $\mu$m onto the transversely-magnetized spinor condensate. Following a 5 ms exposure to the field-inducing laser beam, the condensate magnetization was allowed to evolve freely for variable time before being probed. During this evolution, the imprinted Larmor phase diminished in peak height and grew in extent, matching well with calculations based on a non-interacting spinor gas in a localized field (Fig. \[fig:PD\]).
![Quantum evolution of an imprinted phase. (a) Cross sections of the phase imprint after a free evolution time $t = 0, 40$ and $90$ ms. Traces are offset for clarity. (b) The $1/e$ width of the imprinted phase. The dashed line corresponds to numerical simulations based on a non-interacting spinor gas. (c) The peak phase $vs$ the free evolution time. The dashed line corresponds to the expected scaling imposed by normalization of magnetization.[]{data-label="fig:PD"}](arXfig4){width="46.00000%"}
In conclusion, we have demonstrated a spinor-BEC magnetometer, a powerful application of ultracold atoms to precision measurement of scientific and technological significance. Inasmuch as the Larmor precession phase represents the phase relations among BECs in several Zeeman states, this magnetometer can be regarded as a condensate interferometer with high temporal and spatial resolution. The single-shot phase sensitivity and shot-to-shot variations of 10 mrad achieved here represent an order of magnitude improvement over the performance of current BEC interferometers.
The demonstrated phase sensitivity is close to the atom shot-noise limit. This augurs spin-squeezed magnetometry via continuous quantum non-demolition measurements of the condensate [@spinsqueezing] and novel spatially and temporally resolved studies of spin squeezed ensembles. Finally, the high spatial resolution and sensitivity of the spinor magnetometer render it a powerful detector of quantum spins in optical lattices and may facilitate time-resolved, nondestructive studies of frustration, dipolar interactions and disorder in such systems.
We thank D. Budker and J. Clarke for helpful discussions. This work was supported by the NSF and the David and Lucile Packard Foundation. S. R. L. acknowledges support from the NSERC.
C. C. Tsuei and J. R. Kirtley, Rev. Mod. Phys. [**72**]{}, 969 (2000). K. Kobayashi and Y. Uchikawa, IEEE Trans. Mag. [**39**]{}, 3378 (2003). L. R. Hunter, Science [**252**]{}, 73 (1991). S. J. Bending, Adv. Phys. [**48**]{}, 449 (1999). J. R. Kirtley [*et al*]{}, Appl. Phys. Lett. [**66**]{}, 1138 (1995). R. H. Koch [*et al*]{}, J. Low Temp. Phys. [**51**]{}, 207 (1983). I. K. Kominis [*et al*]{}, Nature [**422**]{}, 596 (2003). P. Treutlein [*et al*]{}, Phys. Rev. Lett. [**92**]{}, 203005 (2004). J. M. Higbie [*et al*]{}, Phys. Rev. Lett. [**95**]{}, 050401 (2005). G. B. Jo [*et al*]{}, cond-mat/0608585 (2006). T. L. Ho, Phys. Rev. Lett. [**81**]{}, 742 (1998). T. Ohmi and K. Machida, J. Phys. Soc. Jpn. [**67**]{}, 1822 (1998). R. H. Koch, D. J. Van Harlingen and J. Clarke, Appl. Phys. Lett. [**38**]{}, 380 (1981). T. S. Lee, E. Dantsker and J. Clarke, Rev. Sci. Inst. [**67**]{}, 4208 (1996). Based on the geometry of the apparatus and the location of magnetic field sources, we estimate field variations across the condensate due to fourth order and higher terms to be $\ll$ 1 fT, negligible even at the high precision of the spinor magnetometer. J. Dalibard and C. Cohen-Tannoudji, J. Opt. Soc. Am. B [**6**]{}, 2023 (1989). H. Carr and E. Purcell, Phys. Rev. [**94**]{}, 630 (1954). S. Wildermuth [*et al*]{}, Nature [**435**]{}, 440 (2005). A. Kuzmich [*et al*]{}, Europhys. Lett. [**42**]{}, 481 (1998).
[^1]: These authors contributed equally to this work.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Julien Baglio,'
- 'Francisco Campanario,'
- 'Seraina Glaus,'
- 'Margarete Mühlleitner,'
- 'Jonathan Ronca,'
- Michael Spira
- and Juraj Streicher
bibliography:
- 'paper.bib'
title: 'Higgs-Pair Production via Gluon Fusion at Hadron Colliders: NLO QCD Corrections'
---
Introduction
============
Since the discovery of a scalar resonance [@Aad:2012tfa; @Chatrchyan:2012xdj] with a mass of $125.09\pm0.24$ GeV [@Khachatryan:2016vau] that is compatible with the Standard Model (SM) Higgs boson [@Higgs:1964ia; @Higgs:1964pj; @Higgs:1966ev; @Englert:1964et; @Guralnik:1964eu; @Kibble:1967sv], the detailed study of the properties of this particle has been a high priority of the analyses at the Large Hadron Collider (LHC). Theoretical uncertainties are a limiting factor for the accuracies reachable at the LHC. This restriction can partly be compensated by increasing the diversity of processes involving the Higgs boson and a broader spectrum of Higgs couplings probed at the LHC. In order to test the nature of the Higgs boson, its self-interactions are of particular interest. It will be the first step towards an experimental reconstruction of the Higgs potential. This plays a crucial role as the origin of electroweak symmetry breaking within the SM. The initial processes that provide a direct sensitivity to the Higgs self-couplings are Higgs-pair production processes. They involve the trilinear Higgs coupling at leading order (LO) [@Glover:1987nx; @Plehn:1996wb; @Dawson:1998py; @Djouadi:1999rca; @Baglio:2012np]. These processes are complementary to indirect effects induced by the Higgs self-interactions in radiative corrections to electroweak observables and single-Higgs processes [@Degrassi:2016wml; @Degrassi:2017ucl] that are plagued by unknown interference effects with other kinds of New Physics.
The Higgs self-interactions are uniquely described by the SM Higgs potential $$V = \frac{\lambda}{2} \left( \phi^\dagger \phi -
\frac{v^2}{2}\right)^2\, ,$$ where $\lambda$ defines the self-interaction strength of the SM Higgs field. In unitary gauge, the Higgs doublet $\phi$ is given by $$\phi = \left( \begin{array}{c} \displaystyle 0 \\ \displaystyle
\frac{v+H}{\sqrt{2}} \end{array} \right)$$ with $v\approx 246$ GeV denoting the vacuum expectation value (vev) and $H$ is the physical Higgs field. In the SM, the self-interaction strength is given in terms of the Higgs mass $M_H$ by $\lambda =
M_H^2/v^2$. Expanding the Higgs field around its vev, the Higgs self-interactions, including the corresponding permutations, are uniquely determined as $$\lambda_{H^3} = 3 \frac{M_H^2}{v}, \qquad \lambda_{H^4} = 3
\frac{M_H^2}{v^2}\, ,$$ where $\lambda_{H^3}$ ($\lambda_{H^4}$) denotes the trilinear (quartic) Higgs self-coupling.
While the quartic Higgs coupling $\lambda_{H^4}$ cannot be probed directly at the LHC, due to the tiny size of the triple-Higgs production cross section [@Plehn:2005nk; @Binoth:2006ym; @Fuks:2015hna; @deFlorian:2016sit; @deFlorian:2019app][^1], the trilinear Higgs coupling can be accessed directly in Higgs-pair production. Higgs-boson pairs are dominantly produced in the loop-induced gluon-fusion mechanism $gg\to HH$ that is mediated by top-quark loops supplemented by a per-cent-level contribution of bottom-quark loops, see Fig. \[fg:hhdia\]. There are destructively interfering box and triangle diagrams at LO with the latter involving the trilinear Higgs coupling [@Glover:1987nx; @Plehn:1996wb]. The box diagrams provide the dominant contributions to the cross section. A rough estimate of the dependence of the cross section on the size of the trilinear coupling is given by the approximate relation $\Delta\sigma/\sigma \sim -\Delta\lambda_{H^3}/\lambda_{H^3}$ in the vicinity of the SM value of $\lambda_{H^3}$. Therefore, in order to determine the trilinear coupling, the theoretical uncertainties of the corresponding cross section need to be small. Thus, the inclusion of higher-order corrections is mandatory. The QCD corrections are fully known up to next-to-leading order (NLO) [@Borowka:2016ehy; @Borowka:2016ypz; @Baglio:2018lrj] and at next-to-next-to-leading order (NNLO) in the limit of heavy top quarks [@deFlorian:2013uza; @deFlorian:2013jea; @Grigo:2014jma]. While the NLO corrections are large, the NNLO contributions are of more moderate size. Very recently, the next-to-next-to-next-to-leading order (N$^3$LO) QCD corrections have been computed in the limit of heavy top quarks resulting in a small further modification of the cross section [@Banerjee:2018lfq; @Chen:2019lzz; @Chen:2019fhs]. This calculation uses the N$^3$LO corrections to the effective Higgs and Higgs-pair couplings to gluons in the heavy-top limit (HTL) [@Spira:2016zna]. The higher-order QCD corrections increase the total LO cross section by about a factor of two. Recently, the full NLO results have been matched to parton showers [@Heinrich:2017kxx; @Jones:2017giv] and the full NNLO results in the limit of heavy top quarks have been merged with the NLO mass effects and supplemented by the additional top-mass effects in the double-real corrections [@Grazzini:2018bsd].
![\[fg:hhdia\] *Generic diagrams contributing to Higgs-boson pair production via gluon fusion. The contribution of the trilinear Higgs coupling is marked in red.*](./plots/dia_gg2hh.pdf){width="110.00000%"}
The goal of this paper is to present in detail the calculation of Ref. [@Baglio:2018lrj] of the full NLO corrections to Higgs pair production in gluon fusion. We rely on a direct numerical integration of the Feynman diagrams, without any tensor reduction. We extend the results presented in Ref. [@Baglio:2018lrj] and study not only the LHC at center-of-mass energies of 13 and 14 TeV, but also present numbers for a potential high-energy upgrade of the LHC (HE-LHC) at 27 TeV [@Abada:2019ono] and for a provisional 100 TeV proton collider within the Future-Circular-Collider (FCC) project [@Abada:2019lih; @Benedikt:2018csr]. Special emphasis will be given to the study of the theoretical uncertainties affecting the results and in particular the scale and scheme uncertainty related to the top-quark mass. We will also study the variation of the trilinear Higgs coupling and show that the NLO mass effects shift the minimum of the total cross section as a function of $\lambda_{H^3}$. They vary substantially over the range of $\lambda_{H^3}$ values.
The paper is organized as follows. We present the notation of our calculation in Section \[sc:lo\] and discuss the results at LO. In Section \[sc:nlo\] we move to the NLO QCD corrections. We discuss the details of the calculation of the virtual corrections in Section \[sc:virtuals\]. We describe the derivation of the real corrections in Section \[sc:reals\]. Our numerical analysis is performed in Section \[sc:results\]. Finally, the conclusions are given in Section \[sc:conclusions\].
Leading-order cross section \[sc:lo\]
=====================================
At LO, Higgs-boson pair production via gluon fusion is mediated by the generic diagrams of Fig. \[fg:hhdia\], including all permutations of the external lines. There are triangle and box diagrams with the former involving the trilinear Higgs coupling through an $s$-channel Higgs exchange. The LO matrix element of $g(q_1) g(q_2) \to H(p_1) H(p_2)$ can be cast into the form $$\begin{aligned}
{\cal M}(g^a g^b \to HH) & = & -i\,\frac{G_F\alpha_s(\mu_R) Q^2}{2\sqrt{2}\pi}
{\cal A}^{\mu\nu} \epsilon_{1\mu} \epsilon_{2\nu} \delta_{ab}
\nonumber \\[0.3cm]
\mbox{with} \qquad {\cal A}^{\mu\nu} & = & F_1 T_1^{\mu\nu} + F_2 T_2^{\mu\nu}
\, , \nonumber \\[0.3cm]
F_1 & = & C_\triangle F_\triangle + F_\Box \, , \qquad
\qquad \qquad
F_2 = G_\Box \, , \nonumber \\[0.3cm]
C_\triangle & = & \frac{\lambda_{H^3} v}{Q^2 - M_H^2 +
iM_H\Gamma_H} \nonumber \\[0.3cm]
\mbox{and} \qquad Q^2 & = & (p_1+p_2)^2 = m_{HH}^2
\label{eq:lomat}\end{aligned}$$ with $Q=m_{HH}$ denoting the invariant Higgs-pair mass. Here $a,b$ denote the color indices of the initial gluons, $\epsilon_{1/2}$ their polarization vectors, $\Gamma_H$ the total Higgs width[^2], $G_F$ the Fermi constant and $\alpha_s(\mu_R)$ the strong coupling at the renormalization scale $\mu_R$. Since in this work we neglect the small bottom-quark contribution, the LO function of the triangle-diagram contribution is given by the top-quark contribution, $$F_\triangle (\tau_t) = \tau_t \Big[ 1 + (1-\tau_t) f(\tau_t) \Big ]
\label{eq:ftriangle}$$ with $\tau_t = 4m_t^2/Q^2$ and the basic function $$\begin{aligned}
f(\tau) & = & \left\{ \begin{array}{ll} \displaystyle \arcsin^2
\frac{1}{\sqrt{\tau}} & \tau \ge 1 \\ \displaystyle - \frac{1}{4} \left[
\log \frac{1+\sqrt{1-\tau}} {1-\sqrt{1-\tau}} - i\pi \right]^2 & \tau <
1 \end{array} \right. \, ,
\label{eq:ftau}\end{aligned}$$ where $m_t$ denotes the top mass, while the more involved analytical expressions for $F_\Box$ and $G_\Box$ can be found in Ref. [@Plehn:1996wb]. In the HTL, the LO form factors approach the values $$F_\triangle \to \frac{2}{3}, \qquad F_\Box \to -\frac{2}{3}, \qquad
G_\Box \to 0 \, .
\label{eq:ffhtl}$$ There are two tensor structures contributing which correspond to the total angular-momentum states with $S_z=0$ and $2$, $$\begin{aligned}
T_1^{\mu\nu} & = & g^{\mu\nu}-\frac{q_1^\nu q_2^\mu}{(q_1q_2)}\, , \nonumber \\
T_2^{\mu\nu} & = & g^{\mu\nu}+\frac{M_H^2 q_1^\nu q_2^\mu}{p_T^2 (q_1q_2)}
-2\frac{(q_2 p_1) q_1^\nu p_1^\mu}{p_T^2 (q_1q_2)}
-2\frac{(q_1 p_1) p_1^\nu q_2^\mu}{p_T^2 (q_1q_2)}
+2\frac{p_1^\nu p_1^\mu}{p_T^2} \nonumber \\
\mbox{with} \quad p_T^2 & = & 2 \frac{(q_1 p_1)(q_2 p_1)}{(q_1 q_2)} -
M_H^2 \, ,\end{aligned}$$ where $p_T$ is the transverse momentum of each of the final-state Higgs bosons. Working in $n=4-2\epsilon$ dimensions, the following projectors on the two form factors can be constructed, $$P_1^{\mu\nu} = \frac{(1-\epsilon) T_1^{\mu\nu} + \epsilon
T_2^{\mu\nu}}{2(1-2\epsilon)}\, , \qquad \qquad
P_2^{\mu\nu} = \frac{\epsilon T_1^{\mu\nu} + (1-\epsilon)
T_2^{\mu\nu}}{2(1-2\epsilon)} \, ,$$ such that $$P_1^{\mu\nu} {\cal A}_{\mu\nu} = F_1 \, , \qquad \qquad
P_2^{\mu\nu} {\cal A}_{\mu\nu} = F_2 \, .$$ Using these projectors, the explicit results of the two form factors $F_{1,2}$ can be obtained in a straightforward manner. The analytical expressions can be found in Refs. [@Glover:1987nx; @Plehn:1996wb]. Working out the polarization and color sums of the matrix element of Eq. (\[eq:lomat\]), the LO partonic cross section $\hat\sigma_{LO}$ is given by $$\hat\sigma_{LO} = \frac{G_F^2\alpha_s^2(\mu_R)}{512 (2\pi)^3}
\int_{\hat t_-}^{\hat t_+} d\hat t \Big[ | F_1 |^2 + |F_2|^2
\Big]$$ with the integration boundaries $$\hat t_\pm = -\frac{1}{2} \left[ Q^2 - 2M_H^2 \mp Q^2
\sqrt{1-4\frac{M_H^2}{Q^2}} \right] \, ,
\label{eq:tbound}$$ where the symmetry factor 1/2 for the identical Higgs bosons in the final state is taken into account. The LO hadronic cross section $\sigma_{LO}$ can then be derived by a convolution with the parton densities $$\sigma_{LO} = \int_{\tau_0}^1 d\tau \frac{d{\cal L}^{gg}}{d\tau}
\hat\sigma_{LO}(Q^2 = \tau s)$$ with the gluon luminosity, given in terms of the gluon densities $g(x,\mu_F)$, $$\frac{d{\cal L}^{gg}}{d\tau} = \int_\tau^1 \frac{dx}{x} g(x,\mu_F)
g\left(\frac{\tau}{x},\mu_F\right)
\label{eq:lgg}$$ at the factorization scale $\mu_F$ and the integration boundary $\tau_0=4M_H^2/s$, where $s$ denotes the hadronic center-of-mass (c.m.) energy squared. The differential cross section with respect to the invariant squared Higgs-pair mass $Q^2$ can be obtained as $$\frac{d\sigma_{LO}}{dQ^2} = \left. \frac{d{\cal L}^{gg}}{d\tau}~
\frac{\hat\sigma_{LO}(Q^2)}{s} \right|_{\tau = \frac{Q^2}{s}} \, .
\label{eq:lodiff}$$ As can be expected from single Higgs-boson production via gluon fusion (see [@Graudenz:1992pv; @Spira:1995rr; @Harlander:2005rq; @Anastasiou:2009kn; @Aglietti:2006tp]), the NLO QCD corrections to these LO expressions will be large.
Next-to-leading-order corrections \[sc:nlo\]
============================================
The NLO QCD corrections to Higgs-pair production via gluon fusion have been computed in the HTL, a long time ago [@Dawson:1998py]. The NLO result for the gluon-fusion cross section can be generically expressed as [@Dawson:1998py] $$\begin{aligned}
\sigma_{NLO}(pp \rightarrow H H + X) & = &
\sigma_{LO} + \Delta
\sigma_{virt} + \Delta\sigma_{gg} + \Delta\sigma_{gq} +
\Delta\sigma_{q\bar{q}} \, , \nonumber $$ $$\begin{aligned}
\sigma_{LO} & = & \int_{\tau_0}^1 d\tau~\frac{d{\cal
L}^{gg}}{d\tau}~\hat\sigma_{LO}(Q^2 = \tau s) \, , \nonumber \\
\Delta \sigma_{virt} & = & \frac{\alpha_s(\mu_R)}
{\pi}\int_{\tau_0}^1 d\tau~\frac{d{\cal L}^{gg}}{d\tau}~\hat
\sigma_{LO}(Q^2=\tau s)~C_{virt}(Q^2) \, , \nonumber \\
\Delta \sigma_{ij} & = & \frac{\alpha_{s}(\mu_R)} {\pi} \int_{\tau_0}^1
d\tau~ \frac{d{\cal L}^{ij}}{d\tau} \int_{\tau_0/\tau}^1 \frac{dz}{z}~
\hat\sigma_{LO}(Q^2 = z \tau s)\, C_{ij}(Q^2,z) \qquad (ij=gg,gq,q\bar q)
\, ,\nonumber \\
C_{gg}(Q^2,z) & = & - z P_{gg} (z) \log \frac{\mu_F^{2}}{\tau s}
+ 6 [1+z^4+(1-z)^4] \left(\frac{\log (1-z)}{1-z} \right)_+
+ d_{gg}(Q^2,z) \, , \nonumber \\
C_{gq}(Q^2,z) & = & -\frac{z}{2} P_{gq}(z)
\log\frac{\mu_F^{2}}{\tau s(1-z)^2} + d_{gq}(Q^2,z) \, , \nonumber \\
C_{q\bar q}(Q^2,z) & = & d_{q\bar q}(Q^2,z)
\label{eq:nlocxn}\end{aligned}$$ with $\hat\sigma_{LO}(Q^2)$ denoting the partonic cross section at LO and the strong coupling $\alpha_s(\mu_R)$ is evaluated at the renormalization scale $\mu_R$. The objects $d{\cal
L}^{ij}/d\tau~(i,j=g,q,\bar q)$ denote the parton-parton luminosities, defined analogously to $d{\cal L}^{gg}/d\tau$ of Eq. (\[eq:lgg\]), using the quark densities $q(x,\mu_F)$, $$\begin{aligned}
\frac{d{\cal L}^{gq}}{d\tau} & = & \sum_{q,\bar q} \int_\tau^1 \frac{dx}{x}
\Big[ g(x,\mu_F) q\left(\frac{\tau}{x},\mu_F\right)
+ q(x,\mu_F) g\left(\frac{\tau}{x},\mu_F\right) \Big] \, , \nonumber \\
\frac{d{\cal L}^{q\bar q}}{d\tau} & = & \sum_q \int_\tau^1 \frac{dx}{x}
\Big[ q(x,\mu_F) \bar q\left(\frac{\tau}{x},\mu_F\right)
+ \bar q(x,\mu_F) q\left(\frac{\tau}{x},\mu_F\right) \Big]\end{aligned}$$ at the factorization scale $\mu_F$ and $P_{ij}(z)~(i,j=g,q,\bar q)$ are the specific Altarelli–Parisi splitting functions [@Altarelli:1977zs].
The quark-mass dependence is in general encoded in the LO cross section $\hat\sigma_{LO}(Q^2)$ and the terms $C_{virt}(Q^2)$, $d_{ij}(Q^2,z)$ for the virtual and real corrections, respectively. These expressions can easily be converted into the differential cross section with respect to $Q^2$, $$\begin{aligned}
\frac{d\Delta\sigma_{virt}}{dQ^2} & = & \left.
\frac{\alpha_s\left(\mu_R\right)}{\pi}~\frac{d{\cal L}^{gg}}{d\tau}~
\frac{\hat{\sigma}_{LO}\left(Q^2 \right)}{s}~C_{virt}
\left(Q^2\right) \right|_{\tau = \frac{Q^2}{s}}, \nonumber \\[0.3cm]
\frac{d\Delta\sigma_{ij}}{dQ^2} & = & \left.
\frac{\alpha_s\left(\mu_R\right)}{\pi}\int_{\frac{Q^2}{s}}^1
\frac{dz}{z^2}~\frac{d{\cal L}^{ij}}{d\tau}~
\frac{\hat{\sigma}_{LO}\left(Q^2\right)}{s}~C_{ij}(Q^2,z) \right|_{\tau
= \frac{Q^2}{zs}} \, ,
\label{eq:nlodiff}\end{aligned}$$ while the differential cross section at LO is given in Eq. (\[eq:lodiff\]).
![\[fg:gghhvirt\] *Typical two-loop triangle (left), one-particle reducible (middle) and box (right) diagrams contributing to Higgs-pair production via gluon fusion at NLO.*](./plots/dia_virt.pdf){width="95.00000%"}
Within the HTL, the Higgs coupling to gluons can be described by an effective Lagrangian [@Ellis:1975ap; @Shifman:1979eb; @Inami:1982xt; @Spira:1995rr; @Kniehl:1995tn] $${\cal L}_\mathrm{eff} = \frac{\alpha_s}{12\pi} G^{a\mu\nu} G^a _{\mu\nu}
\left(C_1 \frac{H}{v} - C_2 \frac{H^2}{2v^2} \right)
\label{eq:leff}$$ involving the Wilson coefficients ($L_t = \log \mu_R^2/m_t^2$) [@Chetyrkin:1997iv; @Kramer:1996iq; @Dawson:1998py; @Schroder:2005hy; @Baikov:2016tgj; @Grigo:2014jma; @Spira:2016zna; @Gerlach:2018hen] $$\begin{aligned}
C_1 & = & 1 + \frac{11}{4} \frac{\alpha_s}{\pi} + \left\{
\frac{2777}{288} +
\frac{19}{16} L_t + N_F \left(\frac{L_t}{3}-\frac{67}{96} \right)
\right\} \left(\frac{\alpha_s}{\pi} \right)^2 + {\cal O}(\alpha_s^3)
\, , \nonumber \\
C_2 & = & C_1 + \left( \frac{35}{24} + \frac{2}{3} N_F \right)
\left(\frac{\alpha_s}{\pi} \right)^2 + {\cal O}(\alpha_s^3)
\label{eq:leffcoeff}\end{aligned}$$ that are known up to N$^4$LO [@Schroder:2005hy; @Baikov:2016tgj; @Spira:2016zna]. Since the top quark is integrated out, the number of active flavours has been chosen as $N_F=5$. If these effective Higgs couplings to gluons in the calculation of the NLO QCD corrections are used, the calculation of these is simplified to a one-loop calculation for the virtual corrections and a tree-level one for the matrix elements of the real corrections. The terms $C_{virt}(Q^2)$ and $d_{ij}(Q^2,z)$, for the virtual and real corrections, approach in the HTL the simple expressions $$\begin{aligned}
C_{virt}(Q^2) & \to & \frac{11}{2} + \pi^2 +
C^\infty_{\triangle\triangle} +
\frac{33-2N_F}{6} \log\frac{\mu_R^2}{Q^2}, \nonumber \\
C_{\triangle\triangle} & = &
\Re e~\frac{\int_{\hat t_-}^{\hat t_+} d\hat t \left\{ c_1 \left[
(C_\triangle F_\triangle + F_\Box) + \frac{p_T^2}{\hat t}
G_\Box \right]^* + (\hat t \leftrightarrow \hat u) \right\}}
{\int_{\hat t_-}^{\hat t_+} d\hat t \left\{ |C_\triangle F_\triangle +
F_\Box |^2 + |G_\Box|^2 \right\}}, \nonumber \\[0.3cm]
C^\infty_{\triangle\triangle} & = & \left. C_{\triangle\triangle}
\right|_{c_1 = 2/9}, \nonumber \\[0.3cm]
d_{gg}(Q^2,z) \!\!\! & \to & \!\!\! - \frac{11}{2} (1-z)^3 \, , \
d_{gq}(Q^2,z) \to \frac{2}{3} z^2 - (1-z)^2 \, , \
d_{q\bar q}(Q^2,z) \to \frac{32}{27} (1-z)^3 \, ,
\label{eq:coeffvirt}\end{aligned}$$ where $\hat s,\hat t, \hat u$ ($\hat s=Q^2$ at LO and for the virtual corrections) denote the partonic Mandelstam variables and $C_{\triangle\triangle}$ is the contribution of the one-particle reducible diagrams, see Fig. \[fg:gghhvirt\].
At NLO QCD, the full mass dependence of the LO partonic cross section has been taken into account, while keeping the virtual corrections $C_{virt}$ and the real corrections $d_{ij}$ in the HTL (“Born-improved” approach) [@Dawson:1998py]. This yields a reasonable approximation for smaller invariant Higgs-pair masses and approximates the full NLO result of the total cross section within about 15% [@Borowka:2016ehy; @Borowka:2016ypz; @Baglio:2018lrj]. The NLO QCD corrections in the HTL increase the cross section by $80-90\%$ [@Dawson:1998py]. Within the Born-improved HTL, the NNLO QCD corrections have been obtained in Refs. [@deFlorian:2013uza; @deFlorian:2013jea; @Grigo:2014jma] increasing the total cross section by a moderate amount of $20-30\%$ [@deFlorian:2013jea]. Beyond these NNLO QCD corrections, the soft-gluon resummation (threshold resummation) has been performed at next-to-next-to-leading logarithmic (NNLL) accuracy for the total cross section and invariant mass distribution, modifying the total cross section further by a small amount if the central scales are chosen as $\mu_R=\mu_F=Q/2$ [@Shao:2013bz; @deFlorian:2015moa]. Very recently, the N$^3$LO QCD corrections have been computed in the Born-improved HTL resulting in a small modification of the cross section beyond NNLO[@Banerjee:2018lfq; @Chen:2019lzz; @Chen:2019fhs; @Spira:2016zna]. These N$^3$LO QCD corrections in the HTL have been merged with the full top-mass effects of the NLO calculation [@Chen:2019fhs]. The calculations in the HTL have been improved by several steps including mass effects partially at NLO. The full mass effects in the real correction terms $d_{ij}$ have been included by means of the full one-loop real matrix elements for $gg\to HHg, gq \to HHq, q\bar q\to
HHg$. This improvement reduces the Born-improved HTL prediction for the total cross section by about 10% [@Frederix:2014hta; @Maltoni:2014eza] and is called the “FTapprox” approximation. The calculation of the full real matrix elements has been performed by using the [MG5\_aMC@NLO]{} framework [@Alwall:2014hca; @Hirschi:2015iia]. Another improvement has been achieved by an asymptotic large-top-mass expansion of the full NLO corrections at the level of the integral [@Grigo:2013rya] and the integrand [@Grigo:2015dia]. This indicated sizable mass effects in the virtual two-loop corrections alone. In addition, the large top-mass expansion has been extended to the virtual NNLO QCD corrections resulting in 5% mass effects estimated on top of the NLO result [@Grigo:2015dia]. The large-top-mass expansion of the NLO QCD corrections has been used to perform a conformal mapping of the expansion parameter and to apply Padé approximants. In this way, an approximation of the full calculation has been achieved for $Q$ values up to about 700 GeV [@Grober:2017uho]. Another approximation builds on an expansion in terms of a variable that dominantly corresponds to the transverse momentum of the Higgs bosons. The results of this approach show good agreement with the full calculation for $Q$ values up to about 900 GeV [@Bonciani:2018omm]. Analytical results are also available in the large-$Q$ limit [@Davies:2018qvx]. The latter have recently been combined with the numerical results of Refs. [@Borowka:2016ehy; @Borowka:2016ypz] for the full QCD corrections [@Davies:2019dfy]. In the following, we will discuss the details of our NLO calculation.
Virtual corrections \[sc:virtuals\]
-----------------------------------
Typical diagrams of the two-loop virtual corrections are shown in Fig. \[fg:gghhvirt\]. They can be arranged in three different classes: (a) triangle, (b) one-particle-reducible and (c) box diagrams[^3]. They contribute to the coefficient $C_{virt}(Q^2)$ of Eq. (\[eq:nlocxn\]), $$C_{virt}(Q^2) = 2 \Re e~\frac{\int_{\hat t_-}^{\hat t_+} d\hat t \left\{
(C_\triangle F_\triangle + F_\Box)^* [C_\triangle (\Delta
F_\triangle) + \Delta F_\Box] + G_\Box^* (\Delta G_\Box) \right\} }
{\int_{\hat t_-}^{\hat t_+} d\hat t \left\{ |C_\triangle F_\triangle +
F_\Box |^2 + |G_\Box|^2 \right\} },$$ where $\Delta F_\triangle, \Delta F_\Box$ and $\Delta G_\Box$ denote the virtual corrections to the corresponding LO form factors. While $\Delta
F_\triangle$ involves only virtual corrections to the triangle diagram, $\Delta F_\Box$ and $\Delta G_\Box$ acquire contributions from the one-particle-reducible and box diagrams.
### Triangle diagrams
![\[fg:triadia\] *Two-loop triangle diagrams contributing to Higgs-pair production via gluon fusion.*](./plots/triangles.pdf){width="90.00000%"}
The generic 2-loop triangle diagrams contributing to the virtual coefficient $C_{virt}(Q^2)$ are shown in Fig. \[fg:triadia\]. They only contribute to the spin-0 form factor $F_1$ of Eq. (\[eq:lomat\]) and can be parametrized as the correction $\Delta F_\triangle$ to the form factor $F_\triangle$, $$\Delta F_\triangle = \frac{\alpha_s}{\pi}~{\cal
C}_{virt}(Q^2)~F_\triangle \, ,$$ where ${\cal C}_{virt}(Q^2)$ denotes the [*complex*]{} virtual coefficient relative to the LO form factor $F_\triangle$ of the amplitude. This virtual coefficient is related to the single-Higgs case so that the relative QCD corrections can be simply obtained from the known (complex) virtual coefficient ${\cal
C}^H_{virt}(M_H^2)$ of single Higgs production [@Spira:1995rr; @Graudenz:1992pv; @Harlander:2005rq; @Anastasiou:2009kn; @Aglietti:2006tp][^4], $${\cal C}_{virt}(Q^2) = \left. {\cal C}^H_{virt}(M_H^2)\right|_{M_H^2 \to
Q^2} \, .$$ In the HTL, this virtual coefficient (before renormalization) approaches the expression $${\cal C}_{virt} (Q^2) \to \frac{\Gamma(1-\epsilon)}{\Gamma(1-2\epsilon)}
\left(\frac{4\pi\mu_0^2 (1-i\bar\epsilon)}{-Q^2}\right)^\epsilon \left\{
-\frac{3}{2\epsilon^2} + \frac{3}{4} - \frac{\pi^2}{4} \right\}$$ with the ’t Hooft scale $\mu_0$, where the (infinitesimal) regulator $\bar\epsilon$ defines the proper analytical continuation of this expression. This result has to be followed by the renormalization of the strong coupling $\alpha_s$ and the top mass $m_t$ that will be discussed in Section \[sc:renorm\]. In addition, we have subtracted the HTL to obtain the pure top-mass effects at NLO (relative to the massive LO expression $F_\triangle$) to ensure that in the end the results of the program [Hpair]{} [@hpair] can be added back. This last step will be discussed in Section \[sc:renorm\], too.
### One-particle-reducible diagrams
The one-particle-reducible contribution is depicted in Fig. \[fg:gghhvirt\] (middle diagram), where a second diagram with the initial gluons interchanged has to be added. These will constitute the $\hat t$- and $\hat u$-channel parts where the second is related to the first just by the interchange $\hat t\leftrightarrow \hat
u$ \[see $C_{\triangle\triangle}$ of Eq. (\[eq:coeffvirt\])\]. The analytical expression of the coefficient $c_1$ can be related to the top contribution of the process $H\to Z\gamma$ [@Cahn:1978nz; @Bergstrom:1985hp]. The basic building block will be the one-loop contribution of the Higgs coupling to an on-shell and an off-shell gluon that is described, after translating all couplings and masses, by the “effective” Feynman rule,\
(100,100)(20,0) (-120,-560)[{width="100.00000%"}]{} (240,58)[$\displaystyle -i \frac{\alpha_s}{\pi v}
\Big[I_1(\tau,\lambda)-I_2(\tau,\lambda)\Big]
\Big[ q_2^\mu q_1^\nu - (q_1 q_2) g^{\mu\nu} \Big] \delta_{ab}\, ,$]{}
\
where the functions $I_{1,2}$ are defined as [@Gunion:1989we] $$\begin{aligned}
I_1(\tau,\lambda) & = & \frac{\tau\lambda}{2(\tau-\lambda)} +
\frac{\tau^2\lambda^2}{2(\tau-\lambda)^2} \left[ f(\tau) - f(\lambda)
\right]
+ \frac{\tau^2\lambda}{(\tau-\lambda)^2} \left[ g(\tau)
- g(\lambda) \right], \nonumber \\
I_2(\tau,\lambda) & = & - \frac{\tau\lambda}{2(\tau-\lambda)}\left[
f(\tau) - f(\lambda) \right],\end{aligned}$$ with $\tau = 4m_t^2/m_H^2$, $\lambda = 4 m_t^2/q_2^2$ and the basic functions $$\begin{aligned}
g(\tau) & = & \left\{ \begin{array}{ll} \displaystyle \sqrt{\tau-1}
\arcsin \frac{1}{\sqrt{\tau}} & \tau \ge 1 \\ \displaystyle
\frac{\sqrt{1-\tau}}{2} \left[ \log \frac{1+\sqrt{1-\tau}}
{1-\sqrt{1-\tau}} - i\pi \right] & \tau < 1 \end{array} \right.\end{aligned}$$ and $f(\tau)$ defined in Eq. (\[eq:ftau\]). Implementing this building block for the two top loops of the one-particle-reducible diagrams, one arrives at the final coefficient $c_1$ of Eq. (\[eq:coeffvirt\]), $$\begin{aligned}
c_1 & = & 2 \Big[ I_1(\tau,\lambda_{\hat t}) -I_2(\tau,\lambda_{\hat
t}) \Big]^2
\label{eq:1prc1}\end{aligned}$$ with $\lambda_{\hat t} = 4 m_t^2/\hat t$ (and $\lambda_{\hat u} = 4
m_t^2/\hat u$ for the $\hat t\leftrightarrow \hat u$ interchanged contribution accordingly). This expression, inserted in the coefficient $C_{\triangle\triangle}$ of Eq. (\[eq:coeffvirt\]), determines the contribution of the one-particle-reducible diagrams analytically and agrees with the previous calculation of Ref. [@Degrassi:2016vss]. In the HTL, this coefficient approaches the value $c_1\to 2/9$ in accordance with Eq. (\[eq:coeffvirt\]). We have subtracted the HTL with $c_1=2/9$ from the coefficient $C_{\triangle\triangle}$ in order to account for the NLO top-mass effects only so that eventually the results of the program [Hpair]{} [@hpair] can be added back. While the total effect of the one-particle-reducible contributions on the total cross section ranges below the per-cent level, the finite mass effects at NLO contribute less than one per mille.
![\[fg:1prcomp\] *Comparison of the approximation of Ref. [@deFlorian:2017qfk] (blue) for the one-particle-reducible contributions and the HTL (red), both normalized to the full analytical expression. The singularity at about 720 GeV is due to a sign change of the exact expression.*](./plots/1pr_comp.pdf){width="80.00000%"}
Reference [@deFlorian:2017qfk] has proposed an approximation of this one-particle-reducible contribution in terms of the triangle form factor of two on-shell external gluons, $$\begin{aligned}
C_{\triangle\triangle} & = &
\Re e~\frac{\int_{\hat t_-}^{\hat t_+} d\hat t \left[
(C_\triangle F_\triangle + F_\Box)^* V_{eff}^2\right]}
{\int_{\hat t_-}^{\hat t_+} d\hat t \left\{ |C_\triangle F_\triangle +
F_\Box |^2 + |G_\Box|^2 \right\}}, \nonumber \\[0.3cm]
V_{eff} & = & F_\triangle (\bar\tau_t)\end{aligned}$$ with $\bar\tau_t = 16 m_t^2/Q^2$ \[i.e. $\tau_t$ of Eq. (\[eq:ftriangle\]) evaluated at half the invariant Higgs-pair mass $Q/2$ instead of $Q$\], where the function $F_\triangle$ can be found in Eq. (\[eq:ftriangle\]). Since Ref. [@deFlorian:2017qfk] works in the HTL, the contribution of the second form factor $F_2$ vanishes, i.e. $G_\Box\to 0$, and the approximation $V_{eff}^2/2$ is in fact treated as an approximation for the coefficient $c_1$ of the exact expression of $C_{\triangle\triangle}$ as given in Eq. (\[eq:coeffvirt\])[^5]. Thus, the approximate expression involving the coefficient $c_1$ has to be compared to the corresponding expression involving the exact coefficient $c_1$ of Eq. (\[eq:1prc1\]). This comparison is presented, normalized to the exact expression, in Fig. \[fg:1prcomp\] and shows that the approximation of Ref. [@deFlorian:2017qfk] is not better than the HTL.
### Box diagrams
The third class of two-loop contributions to the virtual corrections is given by the box diagrams. The generic box diagrams are shown in Figs. \[fg:boxdia1\]–\[fg:boxdia4\] in the Appendix. The simultaneous exchange of the gluons and Higgs bosons has to be added to complete the set of diagrams. The only exception is diagram 44 that is already totally symmetric so that in the final end there are 93 two-loop box diagrams. The generic 47 diagrams are grouped into 6 topology classes. The first 5 topologies contain only a virtual threshold for $Q^2 > 4m_t^2$. The diagrams of topology 6 on the other hand develop a second threshold for $Q^2>0$, because two virtual gluon lines next to the external gluons can be cut. This implies that the form factors are complex in the entire $Q^2$ range. Therefore, a dedicated treatment of this last topology in terms of a suitably constructed infrared subtraction term to isolate the associated infrared singularities is required.
In the following, we will exemplify our method for the boxes 39 of topology 5 and 45 of topology 6. The diagrams of topologies 1–5 are treated analogously to box 39 and those of topology 6 analogously to box 45. The algebraic manipulation of the traces and projections onto the form factors have been performed with the help of the symbolic tools [FORM]{} [@Vermaseren:2000nd; @Kuipers:2012rf], [Reduce]{} [@Hearn:1971zza], and [Mathematica]{} [@Mathematica]. Our method of Feynman parametrization and end-point subtraction to isolate the ultraviolet singularities for the numerical integration has first been applied to the NLO two-loop QCD corrections to $H\to\gamma\gamma, Z\gamma$ in Refs. [@Djouadi:1990aj; @Spira:1991tj] and later to the squark-loop contributions to $h,H \leftrightarrow
gg,\gamma\gamma$ within the minimal supersymmetric extension of the SM [@Muhlleitner:2006wx]. The method of the infrared subtraction as applied to topology 6 originates from numerical cross checks of the full NLO QCD corrections to single Higgs production in Refs. [@Spira:1995es; @Spira:1995rr; @Graudenz:1992pv; @Muhlleitner:2006wx]. The stabilization of virtual thresholds by integration by parts of the integrand has first been applied to the SUSY–QCD corrections to single Higgs production in Refs. [@Muhlleitner:2010nm; @Muhlleitner:2010zz]. The basic idea behind the integration by parts is to reduce the power of the threshold-singular denominator and in this way to stabilize the numerical integration. The treatment of the thresholds in our approach is performed by replacing the squared top mass $m_t^2$ by a complex counter part $$m_t^2 \to m_t^2 (1-i\bar\epsilon)
\label{eq:imaginary}$$ with a positive regulator $\bar\epsilon > 0$ to ensure proper micro-causality. This defines the analytical continuation of our two-loop box integrals. In the following, the parameter $\bar\epsilon$ will be kept finite in our numerical analysis, while the narrow-width limit $\bar\epsilon\to 0$ is achieved by a Richardson extrapolation [@Richardson]. This will be discussed in more detail in the following paragraphs.
#### Box 39
\
![\[fg:box39\] *Explicit definitions of the virtual momenta in box 39.*](./plots/dia_box39.pdf){width="115.00000%"}
Using the definition of real and virtual momenta as in Fig. \[fg:box39\], the contribution to the tensor $A^{\mu\nu}$ \[see Eq. (\[eq:lomat\])\] of the virtual two-loop corrections is given by $$\begin{aligned}
A_{39}^{\mu\nu} & = & \frac{3}{16}\,
\frac{\alpha_s}{\pi}\, (4\pi)^4\, B_{39}^{\mu\nu} \, , \nonumber \\
B_{39}^{\mu\nu} \!\! & \!\!\!\! = \!\!\!\! & \!\! \int \frac{d^n k d^n
q}{(2\pi)^{2n}} \frac{Tr\Big\{ (\,{\!\!\not\! {k}}+{\!\!\not\! {q}}-{\!\!\not\! {p}}_1+m_t)
(\,{\!\!\not\! {k}}+{\!\!\not\! {q}}+m_t) \gamma^\sigma (\,{\!\!\not\! {k}}+m_t)
({\!\!\not\! {k}}+{\!\!\not\! {p}}_2+m_t) \gamma^\nu
(\,{\!\!\not\! {k}}+{\!\!\not\! {q}}_1-{\!\!\not\! {p}}_1+m_t) \gamma^\rho \Big\}}
{[(k+q)^2-m_t^2] [(k+q-p_1)^2-m_t^2] [(k+p_2)^2-m_t^2]
[(k+q_1-p_1)^2-m_t^2]} \nonumber \\[0.3cm]
& & \qquad \times~\frac{g_{\rho\sigma} (2q-q_1)^\mu - g^\mu_\rho
(q-2q_1)_\sigma - g^\mu_\sigma (q+q_1)_\rho}{(k^2-m_t^2)(q-q_1)^2 q^2}
\, ,
\label{eq:mat39}\end{aligned}$$ where $k,q$ are the loop momenta that are integrated over. The Feynman parametrization is first performed for the integration over $k$. We provide Feynman parameters $x_1,\ldots,x_4$ for the first four propagators in the denominator and $1-\sum_i x_i$ for the last one ($k^2-m_t^2$). Performing the substitutions $$x_1 = (1-x)(1-y)\, , \quad x_2 = (1-x)y\, , \quad x_3 = xzr \, ,
\quad x_4 = xz(1-r) \, ,$$ we arrive at a four-dimensional integral over $x,y,z,r$ with integration boundaries from 0 to 1. To symmetrize the $n$-dimensional $k$-integration, we have to perform the shift $$\begin{aligned}
k & \to & k - Q_1 \, , \nonumber \\
Q_1 & = & (1-x)q + xzq_1 + xzr q_2 - [(1-x)y+xz] p_1 \, ,\end{aligned}$$ in both the numerator and denominator. The residual (properly normalized) denominator after the $k$-integration is treated as a propagator for the second loop integration over $q$. We attribute additional Feynman parameters $x_5, x_6$ to this residual propagator and the next one \[$(q-q_1)^2$\] and $1-x_5-x_6$ for the last one ($q^2$) in Eq. (\[eq:mat39\]). Performing the substitution[^6] $$x_5 = s\, , \quad x_6 = (1-s)t\, ,$$ we again arrive at integrals over $s,t$ from 0 to 1. This latter parametrization requires the shift $$\begin{aligned}
q & \to & q - Q_2 \, , \nonumber \\
Q_2 & = & - [zs + (1-s)t] q_1 - zrs q_2 - (y-z)s p_1\end{aligned}$$ in the numerator and denominator to be able to perform the loop integration over $q$ symmetrically. After projecting on the two form factors, we finally arrive at integrals of the type $$\Delta F_i = \frac{\alpha_s}{\pi}~\Gamma(1+2\epsilon)
\left(\frac{4\pi\mu_0^2}{m_t^2}\right)^{2\epsilon} \int_0^1 d^6 x~
\frac{x^\epsilon(1-x)^\epsilon s^{-1-\epsilon} H_i(\vec
x)}{N^{3+2\epsilon}(\vec x)}
\label{eq:fi39}$$ with $\vec x = (x,y,z,r,s,t)$ and $d^6x = dx\,dy\,dz\,dr\,ds\,dt$. $H_i(\vec x)$ denotes the full numerator, including regular factors of the Jacobians due to the Feynman parametrization and substitutions, and singular as well as higher powers of the dimensional regulator $\epsilon$, and $N(\vec x)$ the final denominator, $$\begin{aligned}
N(\vec x) & = & 1 + \rho_s xzr \Big\{ xz + (1-x) [zs+(1-s)t] \Big\}
\nonumber \\
& & \quad - \rho_t x \Big\{ z(1-y-r) + (y-z)[z+(1-x)(1-s)(t-z)] \Big\}
\nonumber \\
& & \quad + \rho_u xzr \Big\{ xz + (1-x) [zs+(1-s)y] \Big\} \nonumber \\
& & \quad - \rho_H \Big\{ [xz+(1-x)y][1-xz-(1-x)y] - x(1-x)s(y-z)^2
\Big\} \, ,\end{aligned}$$ where we define $\rho_s = \hat s/m_t^2 = Q^2/m_t^2$, $\rho_t = (\hat
t-M_H^2)/m_t^2$, $\rho_u = (\hat u-M_H^2)/m_t^2$ and $\rho_H =
M_H^2/m_t^2$. The singular powers in $\epsilon$ of $H_i(\vec x)$ arise from powers of $k^2$ and $q^2$ in the numerators of the final integrations of the loop momenta $k$ and $q$. It is important that the final denominator develops the form of $1+O(1/m_t^2)$ to ensure that no further ultraviolet nor infrared singularities arise from this part of the integrand.
The integral for $\Delta F_i$ of Eq. (\[eq:fi39\]) is singular for $s\to 0$. To separate this singularity from the integral, we perform an endpoint subtraction, $$\begin{aligned}
\Delta F_i & = & \frac{\alpha_s}{\pi}~\Gamma(1+\epsilon)
\frac{\Gamma(1-\epsilon)}{\Gamma(1-2\epsilon)}
\left(\frac{4\pi\mu_0^2}{m_t^2}\right)^{2\epsilon} \left[ \Delta
F_{i,1} + \Delta F_{i,2} \right] \, , \nonumber \\
\Delta F_{i,1} & = & \int_0^1 \frac{d^6 x}{s} \left\{ \frac{H_i(\vec
x)}{N^3(\vec x)} (1+\epsilon L) - \left. \frac{H_i(\vec x)}{N^3(\vec x)}
\right|_{s=0} (1+\epsilon L_0) \right\} \, , \nonumber \\
\Delta F_{i,2} & = & -\frac{1}{\epsilon} \int_0^1 d^5 x \left.
\frac{H_i(\vec x)}{N^3(\vec x)} \right|_{s=0} \left[1+\epsilon L_1 +
\epsilon^2 \left( \frac{L_1^2}{2} + 3\zeta_2 \right) \right] \nonumber \\
\mbox{with} \qquad L & = & \log\frac{x(1-x)}{s} - 2 \log N(\vec x) \, ,
\nonumber \\
L_0 & = & \log\frac{x(1-x)}{s} - 2 \log N(\vec x)|_{s=0} \, , \nonumber \\
L_1 & = & \log [x(1-x)] - 2 \log N(\vec x)|_{s=0} \, ,\end{aligned}$$ where in the second term $\Delta F_{i,2}$ the integration over $s$ has been performed analytically and the integration measure is given by $d^5
x = dx\,dy\,dz\,dr\,dt$. It should be noted that in the terms $L,
L_0, L_1$ the logarithms of the denominator $N$ need to be linear in $N$ to be consistent with the analytical continuation along the proper Riemann sheet. We have checked numerically that the first (subtracted) part $\Delta F_{i,1}$ is finite for each order in the dimensional regulator $\epsilon$ by introducing cuts in the integration boundaries, i.e. integrating from $\tilde\epsilon$ to $1-\tilde\epsilon$, varying $\tilde\epsilon$ down to $10^{-10}$ and checking that the integrals become independent of $\tilde\epsilon$.
These integrals are numerically stable below the virtual $t\bar
t$-threshold, i.e. for $Q^2 < 4 m_t^2$ or $\rho_s < 4$. However, above this threshold, the integrals have to be stabilized. We have achieved this stabilization by means of integration by parts with respect to the Feynman parameter $z$. The denominator is a quadratic polynomial in $z$, $$\begin{aligned}
N(\vec x) & = & a z^2 + b z + c \nonumber \\
\mbox{with} \qquad a & = & x [\rho_s r + \rho_t + \rho_u r + \rho_H]
[1 - (1-x)(1-s)] \, , \nonumber \\
b & = & \rho_s x(1-x)r(1-s)t - \rho_t x [1-r - (1-x)(1-s)(y+t)]
\nonumber \\
& + & \rho_u x(1-x)y r (1-s) - \rho_H x [1-2(1-x)y(1-s)]
\, , \nonumber \\
c & = & 1 - \rho_t x(1-x)y(1-s)t - \rho_H (1-x) y [1-y + xy(1-s)] \, .\end{aligned}$$ To simplify the integration by parts, we insert a unit factor $\Delta/\Delta$ with $\Delta=4ac-b^2$ in the integrand and replace $\Delta$ in the numerator by the expression $$\Delta = 4 a N - (\partial_z N)^2 = 4 a N - (2az+b)^2 \, .$$ Then the following manipulation can be performed, $$\begin{aligned}
\int_0^1 dz~\frac{H_i(\vec x)}{N^3} & = & \frac{1}{\Delta} \left\{
\left. \left[\frac{2a+b}{2N^2} H_i(\vec x) + \frac{\partial_z H_i(\vec
x)}{2N}\right]\right|_{z=1} - \left. \left[\frac{b}{2N^2} H_i(\vec x) +
\frac{\partial_z H_i(\vec x)}{2N}\right]\right|_{z=0} \right. \nonumber \\
& & \left. + \int_0^1 dz~\left[ \frac{3a}{N^2} H_i(\vec x) -
\frac{\partial_z^2 H_i(\vec x)}{2N} \right] \right\}\end{aligned}$$ and analogously for integrals involving additional powers of $\log N$ factors in the numerator of the integrand. The progress achieved with these integrations by parts is that the maximal power of the denominator in the new integral is reduced by one compared to the original integral. One could perform additional integrations by parts with respect to another Feynman parameter. However, we did not investigate this further, since the stability we achieved at this point has been sufficient for the numerical integrations for the top loops[^7].
After performing the integrations by parts, the integral is stable for regulators $\bar \epsilon$ \[see Eq. (\[eq:imaginary\])\] down to $0.05$ for the relevant Higgs mass, top mass and $Q^2$ range. Since this is still apart from the plateau of the narrow-width limit, we performed a Richardson extrapolation [@Richardson] from finite values of $\bar
\epsilon$ down to zero. Richardson extrapolation is possible since the $\bar \epsilon$-dependence of the integral is polynomial for small values of $\bar \epsilon$. The basic principle behind this extrapolation method is very simple: let a function $f(\bar\epsilon)$ behave for small $\bar\epsilon$ as $$f(\bar\epsilon) = f(0) + {\cal O}(\bar\epsilon^n) \, .$$ If we know $f(\bar\epsilon)$ for two different values $\bar\epsilon$ and $t\bar\epsilon$, we can construct the new function $$R_1(\bar\epsilon,t) = \frac{t^n f(\bar\epsilon)-f(t
\bar\epsilon)}{t^n-1} \, .
\label{eq:richardson}$$ This function shows a better convergence towards the value at $\bar\epsilon=0$, $$R_1(\bar\epsilon,t) = f(0) + {\cal O}(\bar\epsilon^{n+1}) \, .$$ Our integrals $I(\bar\epsilon)$ behave for small values of $\bar\epsilon$ as $$I(\bar\epsilon) = I(0) + {\cal O}(\bar\epsilon)$$ so that the first new extrapolation function in our case is given by $$R_1(\bar\epsilon,t) = \frac{t I(\bar\epsilon)-I(t \bar\epsilon)}{t-1}
= I(0) + {\cal O}(\bar\epsilon^2) \, .$$ Using an additional value of $\bar\epsilon$, this method can be repeated iteratively for the new function obtained by applying Eq. (\[eq:richardson\]), $$R_2(\bar\epsilon,t) = \frac{t^2 R_1(\bar\epsilon)-R_1(t
\bar\epsilon)}{t^2-1} = I(0) + {\cal O}(\bar\epsilon^3) \, .$$ In this way, the estimated error is reduced by each additional iteration. We have used this method for a set of $\bar\epsilon$ separated by factors of $t=2$. Then, we obtain the following extrapolation polynomials, $$\begin{aligned}
R_1(\bar\epsilon) & = & 2 I(\bar\epsilon) -I(2\bar\epsilon) = I(0) +
{\cal O}(\bar\epsilon^2) \, , \nonumber \\
R_2(\bar\epsilon) & = & \frac{1}{3} \Big[ 8 I(\bar\epsilon) - 6
I(2\bar\epsilon) + I(4\bar\epsilon) \Big] = I(0) + {\cal
O}(\bar\epsilon^3) \, , \nonumber \\
R_3(\bar\epsilon) & = & \frac{1}{21} \Big[ 64 I(\bar\epsilon) - 56
I(2\bar\epsilon) + 14 I(4\bar\epsilon) - I(8\bar\epsilon) \Big] = I(0)
+ {\cal O}(\bar\epsilon^4) \, , \nonumber \\
R_4(\bar\epsilon) & = & \frac{1}{315} \Big[ 1024 I(\bar\epsilon) - 960
I(2\bar\epsilon) + 280 I(4\bar\epsilon) - 30 I(8\bar\epsilon) +
I(16\bar\epsilon) \Big] = I(0) + {\cal O}(\bar\epsilon^5)\end{aligned}$$ and so on. We have used extrapolation polynomials up to $R_9(\bar\epsilon)$. To determine the extrapolation error, we have chosen different sets of $\bar\epsilon$ values and derived the spread of the extrapolated values appropriately (see Section \[sc:results\] for more details).
#### Box 45
\
![\[fg:box45\] *Explicit definitions of the virtual momenta in box 45.*](./plots/dia_box45.pdf){width="115.00000%"}
Based on the distribution of the loop and external momenta of Fig. \[fg:box45\], the contribution to the two-loop matrix element is given by $$\begin{aligned}
A_{45}^{\mu\nu} & = & \frac{3}{8}\, \frac{\alpha_s}{\pi}\, (4\pi)^4\,
B_{45}^{\mu\nu} \, , \nonumber \\
B_{45}^{\mu\nu} & \!\!\! = \!\!\! & \int \frac{d^n k d^n q}{(2\pi)^{2n}}
\frac{Tr\Big\{ (\,{\!\!\not\! {k}}-{\!\!\not\! {q}}_1+m_t)
(\,{\!\!\not\! {k}}-{\!\!\not\! {q}}_1+{\!\!\not\! {p}}_1+m_t) (\,{\!\!\not\! {k}}+{\!\!\not\! {q}}_2+m_t)
\gamma^\sigma (\,{\!\!\not\! {k}}+{\!\!\not\! {q}}+m_t) \gamma^\rho \Big\}}
{[(k+q)^2-m_t^2] [(k+q_2)^2-m_t^2] [(k+p_1-q_1)^2-m_t^2]
[(k-q_1)^2-m_t^2]} \nonumber \\[0.3cm]
& & \qquad \times~\frac{\Big\{ g_{\rho\tau} (2q+q_1)^\mu - g^\mu_\rho
(q+2q_1)_\tau - g^\mu_\tau (q-q_1)_\rho \Big\}}{(q+q_1)^2 (q-q_2)^2 q^2}
\nonumber \\[0.3cm]
& & \qquad \times~\Big\{ g^{\nu\tau} (q+q_2)_\sigma + g^\nu_\sigma
(q-2q_2)^\tau -g_\sigma^\tau (2q-q_2)^\nu \Big\} \, .
\label{eq:mat45}\end{aligned}$$ Following the same procedure as for box 39 for the Feynman parametrization, we have first performed the parametrization of the $k$-integration following the ordering of the denominator of Eq. (\[eq:mat45\]). The shift in the loop momentum $k$ and the corresponding substitutions of the Feynman parameters are given by $$\begin{aligned}
k & \to & k - Q_1 \, , \nonumber \\
Q_1 & = & (1-x)q - xyq_1 + x(1-y) q_2 + xyz p_1 \, , \nonumber \\
x_1 & = & (1-x)\, , \quad x_2 = x(1-y)\, , \quad x_3 = xyz \, .\end{aligned}$$ Performing the second loop integration over $q$ with the residual (normalized) denominator of the $k$ integration as the first propagator of the $q$ integration, attributing the additional Feynman parameters $x_4, x_5, x_6$ to the remaining propagators in Eq. (\[eq:mat45\]) and applying the substitutions[^8] $$x_4 = rs\, , \quad x_5 = 1-s\, , \quad x_6 = (1-r)st \, ,$$ we arrive at the final expressions for the shift of $q$ and the denominator that contribute to the two form factors, $$\begin{aligned}
q & \to & q - Q_2 \, , \nonumber \\
Q_2 & = & [yrs + 1-s] q_1 - [(1-y)rs+(1-r)st] q_2 - yzrs p_1 \, , \nonumber \\
N(\vec x) & = & r - \rho_s x \Big\{ xy(1-y)r + (1-x)
[1-s+yrs][(1-r)t+(1-y)r] \Big\} \nonumber \\
& & \quad - \rho_t xyzr \Big\{ 1-xy - (1-x)[yrs+1-s] \Big\}
- \rho_H xyzr \Big\{ 1-xyz-(1-x)yzrs \Big\}
\nonumber \\
& & \quad - \rho_u xyzr \Big\{ x(1-y) + (1-x)s [(1-r)t+(1-y)r] \Big\}\end{aligned}$$ and the final integrals of the two form factors ($i=1,2$) can be cast into the form $$\Delta F_i = \Gamma(1+2\epsilon)
\left(\frac{4\pi\mu_0^2}{m_t^2}\right)^{2\epsilon} \int_0^1 d^6 x~
\frac{x^{1+\epsilon} (1-x)^\epsilon r^{1+\epsilon} s^{-\epsilon}
H_i(\vec x)}{N^{3+2\epsilon}(\vec x)} \, ,$$ where $H_i(\vec x)$ contains all additional regular Feynman-parameter factors from Jacobians and the normalization of the denominator of the first loop-integration over $k$. It develops a singular Laurent-expansion in $\epsilon$. The final denominator exhibits the basic form of $r+O(1/m_t^2)$, so that the additional singular behavior is entirely controlled by the limit of small $r$. Since the denominator is of the form $$\begin{aligned}
N(\vec x) & = & a r^2 + b r + c \, , \nonumber \\
\mbox{where} \qquad a & = & x(1-x)ys \Big[-\rho_s (1-y-t) + \rho_t yz -
\rho_u z(1-y-t) + \rho_H y z^2\Big] \, , \nonumber \\
b & = & 1 - \rho_s x \Big\{ xy(1-y) +(1-x)[(1-s)(1-y-t)+yst] \Big\}
- \rho_H xyz (1-xyz) \nonumber \\
& & -\rho_t xyz [1-xy - (1-x)(1-s)] - \rho_u xyz [x(1-y)+(1-x)st]
\, , \nonumber \\
c & = & - \rho_s x(1-x)(1-s)t
\label{eq:abc}\end{aligned}$$ with $a,c = {\cal O}(1/m_t^2)$ and $b=1+{\cal O}(1/m_t^2)$ and the infrared singularities are universal (relative to the LO expressions) the coefficient $a$ does not contribute to the infrared singularity structure, because $a$ is subleading relative to $b$ in the limit $r\to
0$. Thus, we can construct infrared subtraction terms that turn the contributions to the form factors into $$\begin{aligned}
\Delta F_i & = & \frac{\alpha_s}{\pi}~\Gamma(1+2\epsilon)
\left(\frac{4\pi\mu_0^2}{m_t^2}\right)^{2\epsilon} (G_1 + G_2) \, , \nonumber \\
G_1 & = & \int_0^1 d^6 x~x^{1+\epsilon} (1-x)^\epsilon r^{1+\epsilon}
s^{-\epsilon} \left\{ \frac{H_i(\vec x)}{N^{3+2\epsilon}(\vec x)}
-\frac{H_i(\vec x)|_{r=0}}{N_0^{3+2\epsilon}(\vec x)} \right\} \, , \nonumber \\
G_2 & = & \int_0^1 d^6 x~x^{1+\epsilon} (1-x)^\epsilon r^{1+\epsilon}
s^{-\epsilon} \frac{H_i(\vec x)|_{r=0}}{N_0^{3+2\epsilon}(\vec
x)} \nonumber \\
\mbox{with} \qquad N_0(\vec x) & = & br + c \, .\end{aligned}$$ Numerically, we have tested that the subtracted integral $G_1$ (after expansion in the dimensional regulator $\epsilon$) is finite for each coefficient of the expansion in $\epsilon$ individually by integrating the Feynman-parameter integrals from $\tilde\epsilon$ to $1-\tilde\epsilon$ with $\tilde\epsilon$ varied down to $10^{-10}$. The second integral $G_2$ can be integrated over the Feynman parameter $r$ analytically giving rise to hypergeometric functions, $$G_2 = \frac{1}{2+\epsilon} \int_0^1 d^5x~\frac{x^{1+\epsilon}
(1-x)^\epsilon s^{-\epsilon}}{c^{3+2\epsilon}}
~_2F_1\left(3+2\epsilon, 2+\epsilon; 3+\epsilon; -\frac{b}{c}\right)
\left. H_i(\vec x) \right|_{r=0}$$ with $d^5x = dx\, dy\, dz\, ds\, dt$. Since this integral is singular for $c\to 0$, we have to invert the last argument of the hypergeometric function. Using the transformation relation $$\begin{aligned}
~_2F_1 (a,b;c;z) & = &
\frac{\Gamma(c)\Gamma(b-a)}{\Gamma(b)\Gamma(c-a)}(-z)^{-a}~_2F_1\left(
a,1-c+a; 1-b+a; \frac{1}{z}\right) \nonumber \\
& & \qquad + \frac{\Gamma(c)\Gamma(a-b)}{\Gamma(a)\Gamma(c-b)}
(-z)^{-b}~_2F_1\left(b,1-c+b; 1-a+b; \frac{1}{z}\right) \, ,\end{aligned}$$ the special property $$~_2F_1 (a,0;c;z) = 1$$ and suitable end-point subtractions of the residual singular integrals analogous to box 39, we arrive at the final decomposition of the initial Feynman-parameter integral $$\begin{aligned}
\Delta F_i & = & \frac{\alpha_s}{\pi}~\Gamma(1+\epsilon)
\frac{\Gamma(1-\epsilon)}{\Gamma(1-2\epsilon)}
\left(\frac{4\pi\mu_0^2}{m_t^2}\right)^{2\epsilon} \sum_{j=1}^6 S_j
\, , \nonumber \\[0.0cm]
S_1 & = & \int_0^1 d^6x~xr \left\{ \frac{H_i(\vec x)}{N^3(\vec x)} \left[
1+\epsilon L + \epsilon^2 \left( \frac{L^2}{2} + 3\zeta_2\right) \right]
\right. \nonumber \\
& & \qquad\quad - \left. \frac{H_i(\vec x)|_{r=0}}{(c+br)^3} \left[
1+\epsilon L_0 + \epsilon^2 \left( \frac{L_0^2}{2} + 3\zeta_2\right)
\right] \right\} \, , \nonumber \\[0.0cm]
S_2 & = & -\int_0^1 d^6x~x \frac{H_i(\vec x)|_{r=0}}{(b+cr)^3} \left\{
1+\epsilon L_1 + \epsilon^2 \left( \frac{L_1^2}{2} + 3\zeta_2\right)
+ \epsilon^3 \left( \frac{L_1^3}{6} + 3\zeta_2 L_1 \right) \right\}
\, , \nonumber \\[0.0cm]
S_3 & = & -\int_0^1 \frac{d^5x}{2\rho_s (1-x)(1-s)t} \left\{
\frac{H_i(\vec x)|_{r=0}}{b^2} \left[
1-\epsilon (L_2+2) + \epsilon^2 \left( \frac{L_2^2}{2} + 2L_2 + 2\zeta_2
+ 4 \right) \right] \right. \nonumber \\
& & \qquad\quad + \frac{H_i(\vec x)|_{r,t=0,s=1}}{b_0^2} \left[
1-\epsilon (L_3+2) + \epsilon^2 \left( \frac{L_3^2}{2} + 2L_3 + 2\zeta_2
+ 4 \right) \right] \nonumber \\
& & \qquad\quad - \frac{H_i(\vec x)|_{r=0,s=1}}{b_1^2} \left[
1-\epsilon (L_4+2) + \epsilon^2 \left( \frac{L_4^2}{2} + 2L_4 + 2\zeta_2
+ 4 \right) \right] \nonumber \\
& & \qquad\quad \left. - \frac{H_i(\vec x)|_{r,t=0}}{b_2^2} \left[
1-\epsilon (L_5+2) + \epsilon^2 \left( \frac{L_5^2}{2} + 2L_5 + 2\zeta_2
+ 4 \right) \right] \right\} \, , \nonumber \\[0.0cm]
S_4 & = & -\int_0^1 \frac{dx\, dy\, dz\, ds}{2\rho_s (1-x)(1-s)} \left\{
\frac{H_i(\vec x)|_{r,t=0}}{b_2^2} \left[
-\frac{1}{\epsilon}+ L_6+2 - \epsilon \left( \frac{L_6^2}{2} + 2L_6 + 2\zeta_2
+ 4 \right) \right. \right. \nonumber \\
& & \qquad\qquad\qquad\qquad\qquad \left. + \epsilon^2 \left( \frac{L_6^3}{6}
+ L_6^2 + 2(\zeta_2+2) L_6 - 2\zeta_3+4\zeta_2+8 \right) \right] \nonumber \\
& & \qquad\quad - \frac{H_i(\vec x)|_{r,t=0, s=1}}{b_0^2} \left[
-\frac{1}{\epsilon}+ L_7+2 - \epsilon \left( \frac{L_7^2}{2} + 2L_7 + 2\zeta_2
+ 4 \right) \right. \nonumber \\
& & \qquad\qquad\qquad\qquad\qquad \left. \left. + \epsilon^2 \left(
\frac{L_7^3}{6} + L_7^2 + 2(\zeta_2+2) L_7 - 2\zeta_3+4\zeta_2+8 \right)
\right] \right\} \, , \nonumber \\[0.0cm]
S_5 & = & -\int_0^1 \frac{dx\, dy\, dz\, dt}{2\rho_s (1-x)t} \left\{
\frac{H_i(\vec x)|_{r=0,s=1}}{b_1^2} \left[
-\frac{1}{\epsilon}+ L_8+2 - \epsilon \left( \frac{L_8^2}{2} + 2L_8 + \zeta_2
+ 4 \right) \right. \right. \nonumber \\
& & \qquad\qquad\qquad\qquad\qquad \left. + \epsilon^2 \left( \frac{L_8^3}{6}
+ L_8^2 + (\zeta_2+4) L_8 +2\zeta_2+8 \right) \right] \nonumber \\
& & \qquad\quad - \frac{H_i(\vec x)|_{r,t=0, s=1}}{b_0^2} \left[
-\frac{1}{\epsilon}+ L_9+2 - \epsilon \left( \frac{L_9^2}{2} + 2L_9 + \zeta_2
+ 4 \right) \right. \nonumber \\
& & \qquad\qquad\qquad\qquad\qquad \left. \left. + \epsilon^2 \left(
\frac{L_9^3}{6} + L_9^2 + (\zeta_2+4) L_9 +2\zeta_2+8 \right)
\right] \right\} \, , \nonumber \\[0.0cm]
S_6 & = & -\int_0^1 dx\, dy\, dz~\frac{H_i(\vec
x)|_{r,t=0,s=1}}{2\rho_s (1-x)b_0^2} \left\{ \frac{1}{\epsilon^2} -
\frac{1}{\epsilon} (L_{10}+2) + \frac{L_{10}^2}{2} + 2L_{10} + \zeta_2 +
4 \right. \nonumber \\
& & \qquad\qquad\qquad\qquad\qquad \left. - \epsilon \left( \frac{L_{10}^3}{6}
+ L_{10}^2 + (\zeta_2+4) L_{10} +2\zeta_2+8 \right) \right\} \, .\end{aligned}$$ The logarithms used in the expressions above are defined as $$\begin{aligned}
L & = & \log\left(\frac{x(1-x)r}{s}\right) - 2 \log N \, , \qquad\qquad\qquad\,
L_0 = \log\left(\frac{x(1-x)r}{s}\right) - 2 \log (c+br) \, , \nonumber \\
L_1 & = & \log\left(\frac{x(1-x)r}{s}\right) - 2 \log (b+cr) \, , \qquad\qquad
L_2 = \log\left[-\rho_s s(1-s)t\right] + \log b \, , \nonumber \\
L_3 & = & \log\left[-\rho_s s(1-s)t\right] + \log b_0 \, , \qquad\qquad\qquad\ \ \,
L_4 = \log\left[-\rho_s s(1-s)t\right] + \log b_1 \, , \nonumber \\
L_5 & = & \log\left[-\rho_s s(1-s)t\right] + \log b_2 \, , \qquad\qquad\qquad\ \ \,
L_6 = \log\left[-\rho_s s(1-s)\right] + \log b_2 \, , \nonumber \\
L_7 & = & \log\left[-\rho_s s(1-s)\right] + \log b_0 \, , \qquad\qquad\qquad\quad\,
L_8 = \log\left(-\rho_s t\right) + \log b_1 \, , \nonumber \\
L_9 & = & \log\left(-\rho_s t\right) + \log b_0
\, , \qquad\qquad\qquad\qquad\qquad\!\!
L_{10} = \log\left(-\rho_s\right) + \log b_0\end{aligned}$$ and the remaining objects $b_0, b_1, b_2$ as $$\begin{aligned}
b_0 & = & b|_{t=0,s=1} \, , \qquad\qquad b_1 = b|_{s=1} \, , \qquad\qquad
b_2 = b|_{t=0}\end{aligned}$$ with $b$ from Eq. (\[eq:abc\]).
Box 45 contains a second threshold for $Q^2>0$ so that even below the $t\bar t$-threshold, integrations by parts are required to stabilize the integrand numerically. These integrations by parts are performed for the Feynman parameter $r$ in the contributions $S_{1,2}$ along the same lines as for box 39, while the integrals $S_{3-6}$ are stable without integrations by parts.
### Renormalization \[sc:renorm\]
The strong coupling $\alpha_s$ has been renormalized in the $\overline{\rm MS}$ scheme with the top quark decoupled, i.e. the renormalization constant is given by $$\begin{aligned}
\alpha_{s,0} & = & \alpha_s(\mu_R) + \delta\alpha_s \, , \nonumber \\
\frac{\delta\alpha_s}{\alpha_s} & = & \frac{\alpha_s}{\pi}
\Gamma(1+\epsilon) \left(\frac{4\pi\mu_0^2}{\mu_R^2}\right)^\epsilon
\left\{ -\frac{33-2(N_F+1)}{12\epsilon} + \frac{1}{6} \log
\frac{\mu_R^2}{m_t^2} \right\}\end{aligned}$$ with $N_F=5$. This choice ensures that there are no artificial large logarithms of the top mass for the available energy range of the LHC in the final result, since we do not introduce top densities inside the proton, i.e. work in a five-flavour scheme. The additional logarithm of the top mass cancels against the diagrams with a top loop within the external gluon lines, see Fig. \[fg:toploop\]. This leads to the total contribution related to the renormalization of the strong coupling $$\delta_{\alpha_s} F_i = \frac{\alpha_s}{\pi} \Gamma(1+\epsilon)
\left(\frac{4\pi\mu_0^2}{\mu_R^2}\right)^\epsilon \left\{
- \frac{33-2N_F}{12\epsilon} \right\} F_{i,LO} \, ,$$ where the LO form factors $F_i$ have to be used in $n$ dimensions, i.e. including higher orders in the dimensional regulator $\epsilon$.
![\[fg:toploop\] *Typical diagrams with external top loops.*](./plots/dia_toploop.pdf){width="115.00000%"}
For our default prediction, we have renormalized the top mass on-shell so that the renormalization constant is given by $$\begin{aligned}
m_{t,0} & = & m_t - \delta m_t \, , \nonumber \\
\frac{\delta m_t}{m_t} & = & \frac{\alpha_s}{\pi}
\Gamma(1+\epsilon) \left(\frac{4\pi\mu_0^2}{m_t^2}\right)^\epsilon
\left\{ \frac{1}{\epsilon} + \frac{4}{3} \right\} \, .\end{aligned}$$ The explicit contribution of the mass counterterm can either be obtained by calculating the corresponding counterterm diagrams or, in much more elegant manner, by differentiating the LO form factors with respect to the top mass, $$\delta_{m_t} F_i = - \delta m_t \frac{\partial F_{i,LO}}{\partial m_t}
\, ,$$ where we followed the second option. For the renormalization of the top mass in terms of the $\overline{\rm MS}$ mass, a counterterm $$\begin{aligned}
m_{t,0} & = & \overline{m}_t(\mu_t) - \delta \overline{m}_t \, , \nonumber \\
\frac{\delta \overline{m}_t}{\overline{m}_t(\mu_t)} & = & \frac{\alpha_s}{\pi}
\Gamma(1+\epsilon) \left(\frac{4\pi\mu_0^2}{\mu_t^2}\right)^\epsilon
\frac{1}{\epsilon}\end{aligned}$$ has to be used with the LO and NLO expressions of the form factors expressed in terms of the $\overline{\rm MS}$ top mass $\overline{m}_t(\mu_t)$. For the evaluation of the $\overline{\rm MS}$ top mass, we use the N$^3$LO relation between the pole and $\overline{\rm
MS}$ mass [@Gray:1990yh; @Chetyrkin:1999ys; @Chetyrkin:1999qi; @Melnikov:2000qh], $$\begin{aligned}
{\overline{m}}_{t}(m_{t}) & = & \frac{m_{t}}{\displaystyle 1+\frac{4}{3}
\frac{\alpha_{s}(m_t)}{\pi} + K_2
\left(\frac{\alpha_s(m_t)}{\pi}\right)^2 + K_3
\left(\frac{\alpha_s(m_t)}{\pi}\right)^3}
\label{eq:mspole}\end{aligned}$$ with $K_2\approx 10.9$ and $K_3 \approx 107.11$. The scale dependence of the $\overline{\rm MS}$ mass is treated at N$^3$LL, $$\begin{aligned}
{\overline{m}}_{t}\,(\mu_t)&=&{\overline{m}}_{t}\,(m_{t})
\,\frac{c\,[\alpha_{s}\,(\mu_t)/\pi ]}{c\, [\alpha_{s}\,(m_{t})/\pi ]}
\label{eq:msbarev}\end{aligned}$$ with the coefficient function [@Tarasov:1982gk; @Chetyrkin:1997dh] $$\begin{aligned}
c(x)=\left(\frac{7}{2}\,x\right)^{\frac{4}{7}} \, [1+1.398x+1.793\,x^{2}
- 0.6834\, x^3]\, .\end{aligned}$$ Since we are interested in the finite top-mass effects on top of the LO ones, we have subtracted in addition the Born-improved HTL of the virtual corrections involving the full top-mass dependence at LO [@Dawson:1998py]. This yields the additional subtraction term $$\delta_{HTL} F_i = \frac{\alpha_s}{\pi}
\frac{\Gamma(1-\epsilon)}{\Gamma(1-2\epsilon)}
\left(\frac{4\pi\mu_0^2}{-m_t^2 \rho_s}\right)^\epsilon \left\{
\frac{3}{2\epsilon^2} + \frac{33-2N_F}{12\epsilon}
\left(\frac{\mu_R^2}{-m_t^2 \rho_s}\right)^{-\epsilon}
- \frac{11}{4} + \frac{\pi^2}{4} \right\} F_{i,LO} \, .$$ After adding this subtraction term, the result of [Hpair]{} can simply be added back to the NLO top-mass effects obtained in this way for the virtual corrections. Thus, the total counterterm plus HTL-subtraction is given by $$\delta F_i = \delta_{\alpha_s} F_i + \delta_{m_t} F_i + \delta_{HTL} F_i
\, .$$ The addition of this term results in an infrared and ultraviolet finite result for the virtual corrections as we have explicitly checked numerically. It should be noted that we have defined this total subtraction term with the imaginary part $\bar\epsilon$ for the top mass to be consistent with our treatment of the two-loop diagrams. For the two-loop triangle diagrams, this total subtraction term is included in the narrow-width approximation according to the known result for the single-Higgs case.
### Differential cross section
The final numerical integrations have been performed by [Vegas]{} [@Lepage:1980dq] for the differential cross sections $d\sigma/dQ^2$ of Eq. (\[eq:nlodiff\]), i.e. the integration over $\hat t$ is included. Each individual box diagram is divergent in $\hat t$ at the lower and upper bound of the $\hat t$-integration in general. To stabilize the $\hat t$-integration, we have performed a suitable substitution to smoothen the integrand, $$\hat t_1 = m_t^2 e^y + t_{1-}$$ with $\hat t_1 = \hat t-M_H^2, \hat u_1 = \hat u-M_H^2$ and $\hat
t_{1\pm} = \hat t_{\pm}-M_H^2$, where the integration boundaries $\hat
t_{\pm}$ are given in Eq. (\[eq:tbound\]). By means of this substitution, we can rewrite the integration over $\hat t_1$ generically as[^9] $$\int_{\hat t_{1-}}^{\hat t_{1+}} \frac{d\hat t_1}{\hat t_1\hat u_1 -
\hat s M_H^2} f(\hat t_1, \hat u_1) = \int_{y_-}^{y_+}
\frac{dy}{t_+-t_-} \Big[ f(\hat t_1,\hat u_1) + f(\hat u_1,\hat t_1)
\Big] \, ,$$ where $f(\hat t_1, \hat u_1)$ denotes the corresponding virtual matrix element with the (singular) denominator $\hat t_1\hat u_1 - \hat s
M_H^2$ extracted and the integration boundaries read $$\begin{aligned}
y_+ & = & \log \frac{(t_+-t_-)(1-\tilde\epsilon)}{m_t^2} \, , \nonumber \\
y_- & = & \log \frac{(t_+-t_-)\tilde\epsilon}{m_t^2} \, ,\end{aligned}$$ where we have introduced a cut $\tilde\epsilon$ for the upper and lower bound of the $\hat t_1$-integration (after rewriting this into an integral from 0 to 1 and replacing these integration boundaries by $\tilde\epsilon$ and $1-\tilde\epsilon$). We have checked that the total sum of all box diagrams becomes independent of this cut by varying $\tilde\epsilon$ down to $10^{-10}$, i.e. that the total sum is again finite[^10].
Real corrections \[sc:reals\]
-----------------------------
We are left with the evaluation of the real contributions to complete the picture of the NLO QCD corrections. As we are interested in the calculation of the top-mass effects on top of the HTL calculation that is provided by [Hpair]{}, we use the universality of the infrared divergent pieces to subtract the Born-improved HTL contributions $d\sigma_{ij}^{\text{HTL}}$ in such a way that our integration of the real contributions $d\Delta\sigma_{ij}^{\text{mass}} = d\sigma_{ij} -
d\sigma_{ij}^{\text{HTL}}$ is finite. We construct a local subtraction term for the partonic channels $d\hat{\sigma}_{ij}$, $$d\Delta\hat{\sigma}^{\text{mass}}_{ij}(p_k) =
d\hat{\sigma}_{ij}(p_k) - d\hat{\sigma}_{\text{LO}}(\tilde{p}_k)
\frac{d\hat{\sigma}_{ij}^{\text{HTL}}(p_k)}{d\hat{\sigma}_{\text{LO}}^{\text{HTL}}(\tilde{p}_k)},$$ where $p_k$ denote the four-momenta from the full $2\to 3$ phase-space and $\tilde{p}_k$ stand for the mapping of the momenta $p_k$ on a $2\to 2$ sub-phase-space. As the results in the HTL limit are given in the Born-improved approximation in which the pure HTL is rescaled with the full LO matrix elements, we need to map the full $2\to 3$ phase-space onto a projected $2\to 2$ phase-space to construct the subtraction term involving this rescaling to the full LO contribution $ d\hat{\sigma}_{\text{LO}}$.
The mapping is done by using the transformation formulae for initial-state emitter and initial-state spectator in the construction of dipole subtraction terms, i.e. using Eqs. (5.137-5.139) of Ref. [@Catani:1996vz]. The (mapped) momenta of the initial-state partons are $p_{1/2}$ ($\tilde{p}_{1/2}$), the (mapped) momenta of the final-state Higgs bosons are $p_{3/4}$ ($\tilde{p}_{3/4}$), and the momentum of the radiated parton is $p_5$. For the initial-state partons, we use the following mapping, $$\tilde{p}_1 = p_1,\qquad \tilde{p}_2 = p_2 \left(1-\frac{(p_5
p_1) + (p_5 p_2)}{(p_1 p_2)}\right).$$ In order to transform the Higgs momenta, we introduce the variables $K$ and $\tilde{K}$, $$K = p_1+p_2-p_5,\qquad \tilde{K} = \tilde{p}_1+\tilde{p}_2$$ allowing us to define $$\begin{aligned}
\tilde{p}_3
& = p_3 - 2\,\frac{p_3 (K+\tilde{K})}{(K+\tilde{K})^2}
\left(K+\tilde{K}\right) + 2\,\frac{(p_3 K)}{K^2}
\tilde{K},\nonumber\\
\tilde{p}_4
& = p_4 - 2\,\frac{p_4 (K+\tilde{K})}{(K+\tilde{K})^2}
\left(K+\tilde{K}\right) + 2\,\frac{(p_4 K)}{K^2} \tilde{K}.\end{aligned}$$ The HTL matrix elements are calculated analytically. We introduce the partonic center-of-mass energy $\hat{s}$, and the Mandelstam variables $\hat{t}=(p_1-p_5)^2$ and $\hat{u}=(p_2-p_5)^2$. The invariant squared Higgs-pair mass is $Q^2=\hat{s}+\hat{t}+\hat{u}$. The real spin- and colour-averaged matrix elements are $$\begin{aligned}
\overline{\Big|\mathcal{M}_{gg\to HHg}^{\text{HTL}}\Big|^2}
& = \frac{\alpha_s^3(\mu_R)G_F^2}{12\pi}\,
\frac{\hat{s}^4+\hat{t}^4+\hat{u}^4+Q^8}{\hat{s}\hat{t}\hat{u}}\left(1-\frac{3
M_H^2}{Q^2-M_H^2}\right)^2,\nonumber\\
\overline{\Big|\mathcal{M}_{qg\to HHq}^{\text{HTL}}\Big|^2}
& = \frac{\alpha_s^3(\mu_R)G_F^2}{27\pi}\,
\frac{\hat{s}^2+\hat{u}^2}{-\hat{t}}\left(1-\frac{3
M_H^2}{Q^2-M_H^2}\right)^2,\nonumber\\
\overline{\Big|\mathcal{M}_{q\bar{q}\to HHg}^{\text{HTL}}\Big|^2}
& = \frac{8\alpha_s^3(\mu_R) G_F^2}{81\pi}\,
\frac{\hat{t}^2+\hat{u}^2}{\hat{s}}\left(1-\frac{3
M_H^2}{Q^2-M_H^2}\right)^2,\end{aligned}$$ and the LO matrix element in the HTL reads $$\overline{\Big|\mathcal{M}_{\text{LO}}^{\text{HTL}}\Big|^2} =
\frac{\alpha_s^2(\mu_R) G_F^2}{288\pi^2}\,
Q^4 \left(1-\frac{3 M_H^2}{Q^2-M_H^2}\right)^2.$$ The full one-loop matrix elements have been generated with [FeynArts]{} [@Hahn:2000kx] and [FormCalc]{} [@Hahn:1998yk]. They contain triangle, box, and pentagons diagrams. Generic diagrams for the contribution $g g\to H H g$ are given in Fig. \[fg:gghhreal\], generic diagrams for the contributions $q g\to H H q$ and $q\bar{q}\to H H g$ are displayed in Fig. \[fg:gghhreal2\]. The numerical evaluation of the scalar integrals [@tHooft:1978jhc] as well as the tensor reduction has been performed using the techniques developed in Refs. [@vanOldenborgh:1990yc; @Denner:2002ii; @Denner:2005nn; @Denner:2010tr] and implemented in the library [Collier 1.2]{} [@Denner:2016kdg]. The latter has been interfaced to the analytic expressions generated by [FormCalc]{} with an in-house routine. In order to improve our numerical stability, we have implemented a technical collinear cut in the phase-space parametrization. The integration of the scattering angle $\theta$ of the radiated parton in the c.m. system is restricted to the range $|\!\cos\theta|< 1-\delta$ with $\delta=10^{-4}$. We have checked that our results are stable against a variation of $\delta$ from $10^{-4}$ to $10^{-6}$ and therefore they are not affected by our choice for this technical cut.
![\[fg:gghhreal\] *Typical one-loop triangle (upper row), box (middle row), and pentagon (lower row) diagrams for the partonic channel $g g\to H H g$ contributing to the real corrections of Higgs-pair production via gluon fusion at NLO in QCD.*](./plots/dia_real_ggHHg.pdf){width="75.00000%"}
![\[fg:gghhreal2\] *Typical one-loop triangle and box diagrams for the partonic channels $q g\to H H q$ (upper row) and $q\bar{q}\to H H g$ (lower row), contributing to the real corrections of Higgs-pair production via gluon fusion at NLO in QCD.*](./plots/dia_real_qgHHq.pdf "fig:"){width="75.00000%"} ![\[fg:gghhreal2\] *Typical one-loop triangle and box diagrams for the partonic channels $q g\to H H q$ (upper row) and $q\bar{q}\to H H g$ (lower row), contributing to the real corrections of Higgs-pair production via gluon fusion at NLO in QCD.*](./plots/dia_real_qqHHg.pdf "fig:"){width="75.00000%"}
We have cross-checked the final mass-effects of the real corrections against the results presented in the literature [@Frederix:2014hta; @Maltoni:2014eza; @Borowka:2016ehy; @Borowka:2016ypz] and we have obtained agreement.
Results \[sc:results\]
======================
Our numerical results will be presented for the invariant Higgs-pair-mass distributions for different c.m. energies, i.e. 14 TeV for the LHC, 27 TeV for a potential high-energy LHC (HE-LHC) and 100 TeV for a provisional proton collider within the Future-Circular-Collider (FCC) project. The Higgs mass has been chosen as $M_H=125$ GeV and the top pole mass as $m_t=172.5$ GeV. The results for the full NLO cross sections have been obtained with two different PDF sets, [MMHT2014]{} [@Harland-Lang:2014zoa] and [PDF4LHC15]{} [@Butterworth:2015oua], that are taken from the [LHAPDF-6]{} library [@Buckley:2014ana]. The central scale choices for the renormalization and factorization scales are $\mu_F=\mu_R=Q/2$ and the input value $\alpha_s(M_Z)$ is chosen according to the PDF set used. Since [MMHT2014]{} contains a LO set, these PDFs are used for the evaluation of the consistent K-factors with the NLO (LO) cross section calculated with NLO (LO) $\alpha_s$ and PDFs. The whole calculation of the virtual and real corrections has been performed at least twice independently adopting also different Feynman parametrizations of the virtual two-loop diagrams. The real corrections have been derived with different parametrizations of the real phase-space. Both calculations agree within the numerical errors. We work in the narrow-width approximation of the top quark so that the Richardson extrapolation has to be applied to reach this limit for the two-loop box diagrams.[^11].
Differential cross section\[sec:diffxs\]
----------------------------------------
For the differential cross section, we have computed a grid of $Q$-values from 250 GeV to 1.5 TeV. In order to get a reliable result for the total cross section later on, we have used steps of 5 GeV between $Q=250$ GeV and $Q=300$ GeV, steps of 25 GeV between $Q=300$ GeV and $Q=700$ GeV, and steps of 50 GeV for $Q>700$ GeV. After applying the integrations by parts to each individual virtual diagram, we reached reliable results of our numerical integrations for $\bar\epsilon$ values \[see Eq. (\[eq:imaginary\])\] down to about 0.05. In order to obtain the result in the narrow-width approximation ($\bar\epsilon \to 0$), we have performed a Richardson extrapolation applied to the results for different values[^12] of $\bar\epsilon$. We adopt $\bar\epsilon$ values $\bar\epsilon_n= 0.025\times 2^n$ ($n=0\dots 10$). For bins close to threshold, $Q=300,325,350$ GeV, we use the set $n=0\dots 8$. For $Q \in [375,475]$ GeV, we use $n=1\dots 9$ while we use $n=2\dots 10$ for Q values in the range $Q \in [500,700]$ GeV. For $Q$ values starting at 750 GeV, we restrict the extrapolation to $n=2\dots 6$. In this way, we obtain a series of extrapolated results up to the ninth order in the dominant region and up to the fifth order in the tails for large $Q$. We define an estimate of the theoretical error due to the Richardson extrapolation as the difference of the extrapolated results at fifth and fourth order. In addition, we multiply this error by a factor of two close to the virtual $t\bar t$ threshold in order to be conservative. The total estimated Richardson-extrapolation error ranges below the per-cent level and is added in quadrature to the statistical integration error.
Since we have subtracted the (Born-improved) HTL consistently from the virtual and real corrections, we are left with the pure top-mass effects at NLO that are infrared and ultraviolet finite individually after renormalization. This part has then been added to the results of [Hpair]{} [@hpair] to derive the full NLO cross section. The final invariant Higgs-pair-mass distributions are displayed in Figs. \[fig:distrib\_14\]–\[fig:distrib\_100\] for the three c.m. energies, 14, 27, 100 TeV. The blue curves show the Born-improved result in the HTL of Ref. [@Dawson:1998py] as implemented in [Hpair]{} [@hpair], the yellow ones the Born-improved HTL result plus the mass effects of the real corrections, the green curves the Born-improved HTL result plus the mass effects of the virtual corrections and the red curves the full NLO results. The plots on the left side of each figure have been obtained by using [MMHT2014]{} PDFs [@Harland-Lang:2014zoa] and the ones on the right with [PDF4LHC]{} PDFs [@Butterworth:2015oua]. The lower panel on the left shows the consistently defined K-factors $K=d\sigma_{NLO}/d\sigma_{LO}$. The lower panel on the right shows the ratio of the differential NLO cross section to the one obtained in the Born-improved HTL.
While the Born-improved HTL provides a reasonable approximation for $Q$-values close to threshold, the real corrections add a negative mass effect of about $-10\%$ for $\sqrt{s}=14$ TeV (yellow curves) that is approximately uniform in the entire $Q$ range. The (negative) mass effects of the virtual corrections (green curves), however, become large at large values of $Q$ reaching a level of more than 20% for $Q$ beyond about 1 TeV. While the relative mass effects of the virtual corrections at NLO are independent of the collider energy (see the right plots showing the ratios to the HTL in the lower panels) in agreement with Eq. (\[eq:nlodiff\]), the NLO mass effects of the real corrections become larger with rising collider energy, reaching a level of $-20\%$ for $\sqrt{s}=100$ TeV. Both mass effects of the virtual and real corrections add up in the same direction and result in a total modification of the differential cross section of up to $-40\%$ compared to the Born-improved HTL at large $Q$ values for $\sqrt{s}=100$ TeV. While (as for the ratios) the full NLO K-factors shown in the left plots are close to the Born-improved HTL (blue curves) at $Q$ values close to the production threshold, they deviate significantly at larger values of $Q$ due to the additional NLO top-mass effects that decrease the total size of the NLO QCD corrections compared to the HTL as expected from unitarity arguments.
To estimate the theoretical uncertainties, we have varied the renormalization and factorization scales for each bin in $Q$ by a factor of 2 up and down around the central scale $\mu_R=\mu_F=Q/2$ and derived the envelope of a 7-point variation, i.e. excluding points where the renormalization and factorization scales differ by more than a factor of two. The residual uncertainties are shown by the red band around the full NLO results (red curves) in Figs. \[fig:distrib\_14\]–\[fig:distrib\_100\]. They range at the level of 10–15% in total as can be inferred from the explicit numbers for $\sqrt{s}=14$ TeV (using [PDF4LHC]{} PDFs), $$\begin{aligned}
\frac{d\sigma_{NLO}}{dQ}\Big|_{Q=300~{\rm GeV}} & = &
0.02978(7)^{+15.3\%}_{-13.0\%}\
{\rm fb/GeV},\nonumber\\
\frac{d\sigma_{NLO}}{dQ}\Big|_{Q=400~{\rm GeV}} & = &
0.1609(4)^{+14.4\%}_{-12.8\%}\
{\rm fb/GeV},\nonumber\\
\frac{d\sigma_{NLO}}{dQ}\Big|_{Q=600~{\rm GeV}} & = &
0.03204(9)^{+10.9\%}_{-11.5\%}\
{\rm fb/GeV},\nonumber\\
\frac{d\sigma_{NLO}}{dQ}\Big|_{Q=1200~{\rm GeV}} & = &
0.000435(4)^{+7.1\%}_{-10.6\%}\
{\rm fb/GeV} \, .\end{aligned}$$
![Invariant Higgs-pair-mass distributions for Higgs boson pair production via gluon fusion at the 14 TeV LHC as a function of $Q=m_{HH}$. LO results (in black), HTL results (in blue), HTL results including the full real corrections (in yellow), HTL results including the full virtual corrections (in green, including the numerical errors), and the full NLO QCD results (in red, including the numerical errors). Left: Results with the [MMHT2014]{} PDF set, the panel below displays the K-factors for the different results. Right: Results with the [PDF4LHC15]{} PDF set, the panel below displays the ratio to the NLO Born-improved HTL result for the different calculations. The red band indicates the renormalization and factorization scale uncertainties for results including the full NLO QCD corrections.[]{data-label="fig:distrib_14"}](./plots/hh_nlo_mmht_14tev.pdf "fig:"){width="45.00000%"} ![Invariant Higgs-pair-mass distributions for Higgs boson pair production via gluon fusion at the 14 TeV LHC as a function of $Q=m_{HH}$. LO results (in black), HTL results (in blue), HTL results including the full real corrections (in yellow), HTL results including the full virtual corrections (in green, including the numerical errors), and the full NLO QCD results (in red, including the numerical errors). Left: Results with the [MMHT2014]{} PDF set, the panel below displays the K-factors for the different results. Right: Results with the [PDF4LHC15]{} PDF set, the panel below displays the ratio to the NLO Born-improved HTL result for the different calculations. The red band indicates the renormalization and factorization scale uncertainties for results including the full NLO QCD corrections.[]{data-label="fig:distrib_14"}](./plots/hh_nlo_pdf4lhc_14tev.pdf "fig:"){width="45.00000%"}
![Same as Fig. \[fig:distrib\_14\] but for a c.m. energy $\sqrt{s}=27$ TeV.[]{data-label="fig:distrib_27"}](./plots/hh_nlo_mmht_27tev.pdf "fig:"){width="45.00000%"} ![Same as Fig. \[fig:distrib\_14\] but for a c.m. energy $\sqrt{s}=27$ TeV.[]{data-label="fig:distrib_27"}](./plots/hh_nlo_pdf4lhc_27tev.pdf "fig:"){width="45.00000%"}
![Same as Fig. \[fig:distrib\_14\] but for a c.m. energy $\sqrt{s}=100$ TeV.[]{data-label="fig:distrib_100"}](./plots/hh_nlo_mmht_100tev.pdf "fig:"){width="45.00000%"} ![Same as Fig. \[fig:distrib\_14\] but for a c.m. energy $\sqrt{s}=100$ TeV.[]{data-label="fig:distrib_100"}](./plots/hh_nlo_pdf4lhc_100tev.pdf "fig:"){width="45.00000%"}
We have analyzed the structure of the NLO QCD corrections in more detail by comparing the K-factor with the one of the triangle diagrams alone, i.e. with the K-factor of single-Higgs production with mass $M_H=Q$, in all individual approximations. This will determine the amount of universal NLO top-mass effects, common in the triangle and box diagrams. We define the ratio of the NLO triangle-diagram K-factor to the one including all diagrams as K-fac$^\triangle$/K-fac. This is shown, as a function of $Q=m_{HH}$, in Fig. \[fg:ktriafull\] (left). It is visible that the triangle-diagram K-factor provides an acceptable approximation to the full NLO K-factor only for $Q$ values below about 500–600 GeV if maximal deviations of about 15% are allowed (red histogram). The break down into the different mass effects of the virtual (green histogram) and real (yellow histogram) corrections singles out the origin of non-universal mass effects in the virtual corrections, while the non-universal mass effects beyond the single-Higgs case of the real corrections are limited to less than about 5% (apart from the virtual $t\bar t$-threshold region). In comparison to the contribution of the triangle diagrams alone, we also present the ratio of the K-factor obtained by including only the continuum diagrams (box diagrams of the virtual corrections and all box and pentagon diagrams of the real corrections without trilinear Higgs couplings) to the full K-factor in Fig. \[fg:ktriafull\] (right). The different curves show the results for the various approximations, i.e. the blue curves for the Born-improved HTL, the yellow ones with the inclusion of the NLO mass effects of the real corrections, the green curves with only the virtual NLO mass effects and the red curves the full NLO results. The right figure shows that the full NLO K-factor (red curve) is well-described (within 5%) by the one for the continuum diagrams alone which coincides with the observation that the continuum diagrams play a significant role for small values of $Q$ (where the K-factor does not deviate much from the single-Higgs case) and are dominant for large $Q$. This result shows that the K-factor cannot be approximated well by the one of single-Higgs production for large values of $Q$ due to the large mass effects of the virtual corrections.
![Ratios of the K-factor including (left) only triangle diagrams and (right) only continuum diagrams to the full K-factor of Higgs-pair production as a function of the invariant Higgs-pair mass $Q=m_{HH}$ for the LHC with a c.m. energy $\sqrt{s}=14$ TeV and using [MMHT2014]{} parton densities.[]{data-label="fg:ktriafull"}](./plots/hh_nlo_mmht_kfactors_ratio_triangle-to-full.pdf "fig:"){width="45.00000%"} ![Ratios of the K-factor including (left) only triangle diagrams and (right) only continuum diagrams to the full K-factor of Higgs-pair production as a function of the invariant Higgs-pair mass $Q=m_{HH}$ for the LHC with a c.m. energy $\sqrt{s}=14$ TeV and using [MMHT2014]{} parton densities.[]{data-label="fg:ktriafull"}](./plots/hh_nlo_mmht_kfactors_ratio_continuum-to-full.pdf "fig:"){width="45.00000%"}
Total cross section
-------------------
The total cross section has been obtained from the invariant Higgs-pair mass distribution by means of a numerical integration of the bins in $Q$ with the trapezoidal method for $Q>300$ GeV. For a reliable result, we used a Richardson extrapolation [@Richardson] in terms of the bin size in $Q$ also for this step. For $Q<300$ GeV, we have adopted the extension of Boole’s rule to six nodes [@Abramowitz]. We obtain the following values for the total cross section at various c.m. energies, $$\begin{aligned}
\sqrt{s} = 13~{\rm TeV}: \quad
\sigma_{tot} & = & 27.73(7)^{+13.8\%}_{-12.8\%}~{\rm fb}, \nonumber \\
\sqrt{s} = 14~{\rm TeV}: \quad
\sigma_{tot} & = & 32.81(7)^{+13.5\%}_{-12.5\%}~{\rm fb}, \nonumber \\
\sqrt{s} = 27~{\rm TeV}: \quad
\sigma_{tot} & = & 127.0(2)^{+11.7\%}_{-10.7\%}~{\rm fb}, \nonumber \\
\sqrt{s} = 100~{\rm TeV}: \quad
\sigma_{tot} & = & 1140(2)^{+10.7\%}_{-10.0\%}~{\rm fb},
\label{eq:signlo}\end{aligned}$$ where we have used the [PDF4LHC]{} parton densities with $\alpha_s(M_Z)=0.118$ and added for completeness also the value for a c.m. energy of 13 TeV. The numbers in brackets show the numerical errors, while the upper and lower per-centage entries determine the (asymmetric) renormalization and factorization scale dependences. The corresponding results in the Born-improved HTL with [PDF4LHC]{} PDFs, obtained with the program [Hpair]{} [@hpair], read $$\begin{aligned}
\sqrt{s} = 13~{\rm TeV}: \quad
\sigma_{HTL} & = & 32.51^{+18\%}_{-15\%}~{\rm fb}, \nonumber \\
\sqrt{s} = 14~{\rm TeV}: \quad
\sigma_{HTL} & = & 38.65^{+18\%}_{-15\%}~{\rm fb}, \nonumber \\
\sqrt{s} = 27~{\rm TeV}: \quad
\sigma_{HTL} & = & 156.2^{+17\%}_{-13\%}~{\rm fb}, \nonumber \\
\sqrt{s} = 100~{\rm TeV}: \quad
\sigma_{HTL} & = & 1521^{+16\%}_{-13\%}~{\rm fb}.
\label{eq:sightl}\end{aligned}$$ Comparing the results of Eqs. (\[eq:signlo\]) and (\[eq:sightl\]), we observe a reduction of the total cross section by about 15% due to the top-mass effects at NLO and a reduction of the scale uncertainty. These numbers, as well as the differential distributions presented in Section \[sec:diffxs\], agree with the results of Refs. [@Borowka:2016ehy; @Borowka:2016ypz][^13]. It should be noted that a comparison of the full virtual corrections with the analytical large top-mass expansion presented in Ref. [@Grigo:2015dia] was performed in Refs. [@Borowka:2016ehy; @Borowka:2016ypz] and shows a convergence to the full result below the $t\bar t$-threshold, as expected.
Uncertainties originating from the top-mass definition
------------------------------------------------------
An uncertainty that has been neglected or underestimated often previously is the intrinsic uncertainty due to the scheme and scale choice of the virtual top mass. This does not play a large role for single on-shell Higgs-boson production via gluon fusion, $gg\to H$, since the Higgs mass is small and thus the HTL works well, i.e. top-mass effects are suppressed. This uncertainty, however, plays a significant role for the larger values of $Q$ in Higgs-pair production. Top-mass effects are already sizeable at LO, but the NLO corrections add additional relevant top-mass dependences on top of the LO result as we have discussed in the previous subsection. The top mass is a scheme and scale dependent quantity so that the related uncertainties need to be estimated for a reliable determination of the total theoretical uncertainties. For this analysis, we have evaluated the differential cross section for the top mass defined in the on-shell scheme (default) and in the $\overline{\rm MS}$-scheme at the scale $\mu_t$, i.e. adjusting the counterterms and input parameters to the choices $\overline{m}_t(\overline{m}_t)$ and $\overline{m}_t(\mu_t)$ with $\mu_t$ in the range between $Q/4$ and $Q$ according to Section \[sc:renorm\][^14]. Since the scale dependence on $\mu_t$ is a monotonously falling function, we evaluated the differential cross section for four choices of the top mass, $m_t$, $\overline{m}_t(\overline{m}_t)$, $\overline{m}_t(Q/4)$ and $\overline{m}_t(Q)$, for each bin in $Q$.
For the three c.m. energies of 14, 27 and 100 TeV the differential cross sections are presented in Figs. \[fig:distribmt\_1\], \[fig:distribmt\_2\] as a function of $Q=m_{HH}$ for the various definitions of the top mass. The lower panels exhibit the ratios of the differential cross sections to the ones in terms of the top pole mass (OS scheme).
![The differential Higgs-pair production cross section at NLO as a function of the invariant Higgs-pair mass for a c.m. energy of 14 TeV for four different choices of the scheme and scale of the top mass. The lower panel shows the ratio of all results to the default results with the top pole mass (OS scheme). [PDF4LHC]{} PDFs have been used and the renormalization and factorization scales of $\alpha_s$ and the PDFs have been fixed at our central scale choice $\mu_R=\mu_F=Q/2$.[]{data-label="fig:distribmt_1"}](./plots/hh_nlo_pdf4lhc_mtplot_14.pdf){width="60.00000%"}
![Same as Fig. \[fig:distribmt\_1\] but for c.m. energies of 27 (left) and 100 (right) TeV.[]{data-label="fig:distribmt_2"}](./plots/hh_nlo_pdf4lhc_mtplot_27.pdf "fig:"){width="45.00000%"} ![Same as Fig. \[fig:distribmt\_1\] but for c.m. energies of 27 (left) and 100 (right) TeV.[]{data-label="fig:distribmt_2"}](./plots/hh_nlo_pdf4lhc_mtplot_100.pdf "fig:"){width="45.00000%"}
It is clearly visible that the scale and scheme dependence of the top mass induces sizeable variations of the NLO Higgs-pair production cross section and thus contributes to the theoretical uncertainties. For small $Q$ values, the size pattern of the differential cross section due to the different scale and scheme choices is varying. For large values of $Q$, the maximum is always given by the on-shell scheme and the minimum in terms of the $\overline{\rm MS}$-top mass $\overline{m}_t(Q)$ with sizeable differences to the on-shell scheme. Adopting the related uncertainties as the envelope of the cross sections for our four choices, we arrive at the following uncertainties of the differential cross section for a c.m. energy $\sqrt{s}=14$ TeV, $$\begin{aligned}
\frac{d\sigma_{NLO}}{dQ}\Big|_{Q=300~{\rm GeV}} & = &
0.02978(7)^{+6\%}_{-34\%}\
{\rm fb/GeV},\nonumber\\
\frac{d\sigma_{NLO}}{dQ}\Big|_{Q=400~{\rm GeV}} & = &
0.1609(4)^{+0\%}_{-13\%}\
{\rm fb/GeV},\nonumber\\
\frac{d\sigma_{NLO}}{dQ}\Big|_{Q=600~{\rm GeV}} & = &
0.03204(9)^{+0\%}_{-30\%}\
{\rm fb/GeV},\nonumber\\
\frac{d\sigma_{NLO}}{dQ}\Big|_{Q=1200~{\rm GeV}} & = &
0.000435(4)^{+0\%}_{-35\%}\
{\rm fb/GeV} \, .\end{aligned}$$ Since these uncertainties are given relative to the on-shell results, the upper uncertainty vanishes for $Q\geq 400$ GeV, because the on-shell results provide the maximal values. These uncertainties turn out to be significant and at a similar level as the usual renormalization and factorization scale uncertainties. Thus, they constitute an additional contribution to the total theoretical uncertainties that has to be taken into account. The uncertainties due to the top-mass scheme and scale are about a factor of two smaller than at LO, $$\begin{aligned}
\frac{d\sigma_{LO}}{dQ}\Big|_{Q=300~{\rm GeV}} & = &
0.01656^{+62\%}_{-2.4\%}\, \mathrm{fb/GeV},\nonumber\\
\frac{d\sigma_{LO}}{dQ}\Big|_{Q=400~{\rm GeV}} & = &
0.09391^{+0\%}_{-20\%}\, \mathrm{fb/GeV},\nonumber\\
\frac{d\sigma_{LO}}{dQ}\Big|_{Q=600~{\rm GeV}} & = &
0.02132^{+0\%}_{-48\%}\, \mathrm{fb/GeV},\nonumber\\
\frac{d\sigma_{LO}}{dQ}\Big|_{Q=1200~{\rm GeV}} & = &
0.0003223^{+0\%}_{-56\%}\, \mathrm{fb/GeV}\end{aligned}$$ that have been obtained for a c.m. energy of 14 TeV and using PDF4LHC15 NLO parton densities with a NLO strong coupling normalized to $\alpha_s(M_Z)=0.118$[^15]. Their reduction from LO to NLO underlines that the NLO QCD corrections stabilize the theoretical prediction for the Higgs-pair production cross section. The large size of the residual uncertainties is just a consequence of the large NLO QCD corrections as is the case for the renormalization and factorization scale dependences, too. Adopting the envelope for each $Q$-bin individually and integrating over $Q$, we arrive at the impact of these uncertainties on the total cross section for various c.m. energies, $$\begin{aligned}
\sqrt{s} = 13~{\rm TeV}: \quad
\sigma_{tot} & = & 27.73(7)^{+4\%}_{-18\%}~{\rm fb}, \nonumber \\
\sqrt{s} = 14~{\rm TeV}: \quad
\sigma_{tot} & = & 32.81(7)^{+4\%}_{-18\%}~{\rm fb}, \nonumber \\
\sqrt{s} = 27~{\rm TeV}: \quad
\sigma_{tot} & = & 127.0(2)^{+4\%}_{-18\%}~{\rm fb}, \nonumber \\
\sqrt{s} = 100~{\rm TeV}: \quad
\sigma_{tot} & = & 1140(2)^{+3\%}_{-18\%}~{\rm fb}\end{aligned}$$ using [PDF4LHC]{} PDFs. A further reduction of these uncertainties can only be achieved by the determination or reliable estimate of the full mass effects at NNLO.
Since these uncertainties are sizeable, one may wonder why this has not been observed already for single-Higgs boson production $gg\to H$. The measured value of the Higgs mass $M_H=125$ GeV is small compared to the top mass so that for single on-shell Higgs production we are close to the HTL, i.e. finite top-mass effects are small and thus the related uncertainties, too. However, going to larger virtualities $Q$ for off-shell Higgs production $gg\to H^*$ (or larger Higgs masses for on-shell Higgs production), we arrive at similar uncertainties for $\sqrt{s}=14$ TeV, $$\begin{aligned}
\sigma_{NLO}\Big|_{Q=125~{\rm GeV}} & = 42.17^{+0.4\%}_{-0.5\%}\,
\mathrm{pb}, \qquad
\sigma_{NLO}\Big|_{Q=300~{\rm GeV}} & = 9.85^{+7.5\%}_{-0.3\%}\,
\mathrm{pb},\nonumber \\[0.5cm]
\sigma_{NLO}\Big|_{Q=400~{\rm GeV}} & = 9.43^{+0.1\%}_{-0.9\%}\,
\mathrm{pb}, \qquad
\sigma_{NLO}\Big|_{Q=600~{\rm GeV}} & = 1.97^{+0.0\%}_{-15.9\%}\,
\mathrm{pb},\nonumber \\[0.5cm]
\sigma_{NLO}\Big|_{Q=900~{\rm GeV}} & = 0.230^{+0.0\%}_{-22.3\%}\,
\mathrm{pb}, \quad
\sigma_{NLO}\Big|_{Q=1200~{\rm GeV}} & = 0.0402^{+0.0\%}_{-26.0\%}\,
\mathrm{pb}\end{aligned}$$ using PDF4LHC PDFs. This has been known for a long time since there are sizeable effects on the virtual corrections due to the scale choice of the top mass for larger values of $Q$ or the Higgs mass (see Fig. 7a of Ref. [@Spira:1995rr]). For the single off-shell Higgs case, a reduction of the top-mass scale dependence by roughly a factor of two by going from LO to NLO has been observed, too, as can be inferred from the comparison with the explicit LO numbers for $\sqrt{s}=14$ TeV, $$\begin{aligned}
\sigma_{LO}\Big|_{Q=125~{\rm GeV}} & = 18.43^{+0.8\%}_{-1.1\%}\,
\mathrm{pb}, \qquad
\sigma_{LO}\Big|_{Q=300~{\rm GeV}} & = 4.88^{+23.1\%}_{-1.1\%}\,
\mathrm{pb}, \nonumber \\[0.5cm]
\sigma_{LO}\Big|_{Q=400~{\rm GeV}} & = 4.94^{+1.2\%}_{-1.8\%}\,
\mathrm{pb}, \qquad
\sigma_{LO}\Big|_{Q=600~{\rm GeV}} & = 1.13^{+0.0\%}_{-26.2\%}\,
\mathrm{pb}, \nonumber \\[0.5cm]
\sigma_{LO}\Big|_{Q=900~{\rm GeV}} & = 0.139^{+0.0\%}_{-36.0\%}\,
\mathrm{pb}, \quad
\sigma_{LO}\Big|_{Q=1200~{\rm GeV}} & = 0.0249^{+0.0\%}_{-41.1\%}\,
\mathrm{pb}\end{aligned}$$ that have been obtained with PDF4LHC PDFs as in the Higgs-pair case. On the other hand, the uncertainties for $Q=125$ GeV confirm that they are small for on-shell Higgs production via gluon fusion (already at LO) in agreement with the analysis of the LHC Higgs Cross Section Working Group [@deFlorian:2016spz; @Anastasiou:2016cez].
A relevant issue is the theoretical background of the different scale choices for the top mass. For small values of $Q$, the matrix element will be closer to the HTL such that the NLO corrections get closer to the HTL calculation. The HTL on the other hand can be treated by starting from the effective Lagrangian of Eq. (\[eq:leff\]) which is the residual effective coupling of Higgs bosons to gluons after integrating out the top quark. Thus, the corresponding Wilson coefficients $C_1$ and $C_2$ are determined by matching the full SM with the top quark to the effective theory without the top quark. The matching scale is naturally given by the top mass. Performing the proper matching at the scale of the top mass, i.e. using either the top pole mass or the top $\overline{\rm MS}$ mass at the scale of the top mass itself leads to non-logarithmic (in the top mass) matching contributions \[see Eq.(\[eq:leffcoeff\]) for $\mu_R=m_t$\] also for higher powers in $1/m_t^2$, i.e. higher-dimensional operators contributing to the gluonic Higgs couplings at the subleading level. This implies that the top mass is the preferred scale choice for small values of $Q$. This is confirmed by the heavy top expansion of the form factors of Refs. [@Grigo:2013rya; @Grigo:2015dia; @Davies:2018qvx].
At large $Q$ values, on the other hand, we can use the results for the high-energy expansion of Ref. [@Davies:2018qvx]. In the regime of large $Q$, the triangle-diagram contributions are suppressed by the $s$-channel Higgs propagator so that the box diagrams provide the dominant contributions. In our normalization, the explicit results of the virtual box-form factors in the high-energy limit ($Q\gg m_t, M_H$) in terms of the top pole mass $m_t$ are given by[^16] $$\begin{aligned}
F_i & = & F_{i,LO} + \Delta F_i \, , \nonumber \\
\Delta F_i & = & \Delta F_{i,HTL} + \Delta F_{i,mass} \, , \nonumber \\
F_{1,LO} & \to & 4\frac{m_t^2}{\hat s} \, , \nonumber \\
F_{2,LO} & \to & -\frac{m_t^2}{\hat s \hat t (\hat s+\hat t)} \Big\{
(\hat s+\hat t)^2 L_{1ts}^2 + \hat t^2 L_{ts}^2 + \pi^2 [(\hat s+\hat
t)^2+\hat t^2] \Big\} \, , \nonumber \\
\Delta F_{1,mass} & \to & \frac{\alpha_s}{\pi}\left\{ 2 F_{1,LO}
\log\frac{m_t^2}{\hat s} + \frac{m_t^2}{\hat s} G_1(\hat s, \hat t)
\right\} \, , \nonumber \\
\Delta F_{2,mass} & \to & \frac{\alpha_s}{\pi}\left\{ 2 F_{2,LO}
\log\frac{m_t^2}{\hat s} + \frac{m_t^2}{\hat s} G_2(\hat s, \hat t)
\right\} \, ,
\label{eq:ffhe}\end{aligned}$$ where $G_{1,2}(\hat s, \hat t)$ denote explicit and lengthy functions of the kinematical variables $\hat s$ and $\hat t$ that do [*not*]{} depend on the top mass [@Davies:2018qvx]. The logarithms $L_{ts},
L_{1ts}$ are defined as $$L_{ts} = \log \left(-\frac{\hat t}{\hat s}\right) + i\pi \, , \qquad
L_{1ts} = \log \left(1+\frac{\hat t}{\hat s}\right) + i\pi \, .$$ Transforming the top pole mass $m_t$ into the $\overline{\rm MS}$ mass $\overline{m}_t(\mu_t)$, we arrive at the LO expressions for $F_{1/2,LO}$ with $m_t$ replaced by $\overline{m}_t(\mu_t)$ and the appropriately transformed NLO coefficients $$\begin{aligned}
F_{1,LO} & \to & 4\frac{\overline{m}_t^2(\mu_t)}{\hat s} \, , \nonumber \\
F_{2,LO} & \to & -\frac{\overline{m}_t^2(\mu_t)}{\hat s \hat t (\hat
s+\hat t)} \Big\{ (\hat s+\hat t)^2 L_{1ts}^2 + \hat t^2 L_{ts}^2 +
\pi^2 [(\hat s+\hat t)^2+\hat t^2] \Big\} \, , \nonumber \\
\Delta F_{1,mass} & \to & \frac{\alpha_s}{\pi}\left\{ 2 F_{1,LO} \left[
\log\frac{\mu_t^2}{\hat s} + \frac{4}{3} \right] +
\frac{\overline{m}_t^2(\mu_t)}{\hat s} G_1(\hat s, \hat t) \right\} \, ,
\nonumber \\
\Delta F_{2,mass} & \to & \frac{\alpha_s}{\pi}\left\{ 2 F_{2,LO} \left[
\log\frac{\mu_t^2}{\hat s} + \frac{4}{3} \right] +
\frac{\overline{m}_t^2(\mu_t)}{\hat s} G_2(\hat s, \hat t) \right\} \, .\end{aligned}$$ To minimize the logarithms of $\mu_t$, a dynamical scale of the order of $\sqrt{\hat s}=Q$ has to be chosen, but [*not*]{} the top mass. A coefficient $\kappa$ in front of the dynamical scale choice $\mu_t =
\kappa Q$ is still arbitrary (but should not be large) since additional finite parts of the functions $G_{1,2}(\hat s, \hat t)$ may be absorbed in the scale choice. Thus, the dynamical scale $Q$ can be identified as the preferred central scale choice of the Yukawa couplings for large $Q$ values.
The uncertainties originating from the scheme and scale dependence of the top mass can be reduced by calculating the NNLO mass effects. Such a three-loop calculation is beyond everything that has been performed so far with current methods, but for $Q$ values close to threshold a large-mass expansion at NNLO could be used to reach an approximate estimate of the finite top-mass effects at NNLO. As a first step, partial results of the NNLO top-mass effects are known in the soft+virtual approximation [@Grigo:2015dia]. For $Q$ values around the virtual $t\bar t$ threshold $Q\sim 2m_t$, non-relativistic Green’s functions could be used that allow the introduction of higher-order corrections to the QCD potential [@Fadin:1987wz; @Fadin:1988fn; @Fadin:1990wx; @Strassler:1990nw; @Melnikov:1994jb]. This may lead to an improved description of the threshold region. However, for the triangle diagrams, the threshold behaviour is determined by $P$-wave contributions, since the $t\bar t$-ground state appears as a ${\cal
CP}$-odd configuration that does not mix with the virtual ${\cal
CP}$-even threshold state of the triangle diagrams. For the box diagrams, the $P$–wave contributions have to be considered, too. Moreover, it is unclear how large the impact of top-mass effects of the remainder beyond the non-relativistic Green’s functions will be. Finally, for the high-energy tail, the approximate calculation of Ref. [@Davies:2018qvx] could be extended to NNLO.
Variation of the cross section with $\lambda_{H^3}$
---------------------------------------------------
Higgs-pair production at the LHC is directly sensitive to the trilinear Higgs coupling. The dependence of the total and differential cross sections on the trilinear coupling $\lambda_{H^3}$ is modified by the NLO QCD corrections and in particular by the finite mass effects at LO and NLO. Finite top-mass effects result in a non-vanishing matrix element at threshold, while in the HTL the matrix element of Eq. (\[eq:lomat\]) vanishes exactly [@Glover:1987nx; @Plehn:1996wb; @Li:2013rra], $$\begin{aligned}
{\cal A}^{\mu\nu} & \to & F_1 T_1^{\mu\nu} \, , \nonumber \\
F_1 & \to & \frac{2}{3}(C_\triangle - 1)
\to \frac{2}{3}~\left(\frac{3M_H^2}{4 M_H^2-M_H^2} - 1\right) = 0 \qquad
\mbox{for $Q^2\to (2M_H)^2$} \, ,\end{aligned}$$ where we have used that the second form factor $G_\Box$ vanishes in the HTL \[see Eq. (\[eq:ffhtl\])\]. The cancellation is induced by the destructive interference between the triangle and box diagrams at LO. This property is modified by finite subleading ${\cal O}(1/m_t^2)$ terms but explains why the matrix element itself is suppressed at the production threshold. As a function of $\lambda_{H^3}$, the cross section develops a minimum at $\lambda_{H^3}$-values around 2.4 times the SM-value in the Born-improved HTL [@Dawson:1998py; @Baglio:2012np] since the phase-space integration adds contributions from above the production threshold. The NLO QCD corrections will shift the minimum of the cross section as a function of $\lambda_{H^3}$ and finite top-mass effects play a prominent role in the amount of these cancellations. For the determination of the trilinear coupling, the variation of the cross section with $\lambda_{H^3}$ is of interest. As mentioned in the introduction, the total cross section behaves approximately as $\Delta\sigma/\sigma \sim -\Delta\lambda_{H^3}/\lambda_{H^3}$ for $\lambda_{H^3}$ close to the SM value.
In the following, we will analyze the NLO results, where only the trilinear coupling has been varied. In general, however, several coupling modifications contribute to the Higgs-pair production cross section. This could be treated consistently by extending the SM Lagrangian by all contributing dimension-6 operators as has been studied in Ref. [@Grober:2015cwa] in the HTL at NLO and in Ref. [@deFlorian:2017qfk] at NNLO. Recently the HTL analysis has been extended to the inclusion of finite top-mass effects at NLO [@Buchalla:2018yce]. However, we will neglect all dimension-6 operators but the one modifying the Higgs self-interactions. A proper and consistent effective model of this type has been discussed in Ref. [@Degrassi:2017ucl] that adds higher-dimension operators to the scalar Higgs sector only. Thus, a sole variation of the Higgs self-interactions could be realized within Higgs portal models with additional heavy scalar states that couple only to the SM-like Higgs field and are integrated out.
![The total Higgs-pair production cross section at NLO as a function of the trilinear self-coupling $\lambda_{H^3}$ in units of the SM value for a c.m. energy of 14 TeV. The blue curve shows the Born-improved HTL, the yellow includes the NLO mass effects of the real corrections in addition and the green curve those of the virtual corrections in addition. The full NLO result is presented by the red curve. The lower panel shows the ratio of all results to the Born-improved HTL. [PDF4LHC]{} PDFs have been used and the renormalization and factorization scales of $\alpha_s$ and the PDFs have been fixed at our central scale choice $\mu_R=\mu_F=Q/2=m_{HH}/2$.[]{data-label="fig:lambdavar_1"}](./plots/hh_nlo_lambda_pdf4lhc_14tev.pdf){width="60.00000%"}
In Figs. \[fig:lambdavar\_1\] and \[fig:lambdavar\_2\], the dependence of the total Higgs-pair production cross section is shown as a function of the trilinear Higgs coupling $\lambda_{H^3}$ in units of the SM coupling for three c.m. energies, 14, 27 and 100 TeV. The blue curves display the results in the Born-improved HTL, the yellow curves include the mass effects of the real corrections and the green curves the mass effects of virtual corrections in addition. The red curves exhibit the complete NLO results. The comparison of the blue and red curves indicates that the minimum of the $\lambda_{H^3}$-variation is shifted from about 2.4 times the SM value to about 2.3 times the SM value due to the NLO mass effects. The yellow and green curves imply that the main origin of this shift emerges from the mass effects of the real corrections. The lower panels of Figs. \[fig:lambdavar\_1\] and \[fig:lambdavar\_2\] present the ratios of the individual contributions to the Born-improved HTL. While the NLO mass effects are of moderate size for negative values of $\lambda_{H^3}$, where the triangle and box diagrams interfere constructively, they turn out to be more relevant in the region of destructive interference, in particular around the minima of the cross sections. The significantly varying NLO mass effects have to be taken into account when determining the value of $\lambda_{H^3}$ from the experimental data at the HL-LHC. This agrees with the findings of Ref. [@Buchalla:2018yce]. The NLO mass effects on the variation of the total cross section with $\lambda_{H^3}$ become larger with rising c.m. energy of the hadron collider.
![Same as Fig. \[fig:lambdavar\_1\] but for c.m. energies of 27 (left) and 100 (right) TeV.[]{data-label="fig:lambdavar_2"}](./plots/hh_nlo_lambda_pdf4lhc_27tev.pdf "fig:"){width="45.00000%"} ![Same as Fig. \[fig:lambdavar\_1\] but for c.m. energies of 27 (left) and 100 (right) TeV.[]{data-label="fig:lambdavar_2"}](./plots/hh_nlo_lambda_pdf4lhc_100tev.pdf "fig:"){width="45.00000%"}
In Fig. \[fig:lambdakfac\], we display the consistently defined K-factors $K=\sigma_{NLO}/\sigma_{LO}$ as a function of $\lambda_{H^3}$ in units of the SM coupling. The full curves show the NLO K-factors including the NLO top-mass effects for various c.m. energies. The dotted curves exhibit the corresponding K-factors in the Born-improved HTL as computed in Refs. [@Dawson:1998py; @Grober:2015cwa]. The impact of the NLO mass effects on the K-factors ranges at the level of 10–15% for negative $\lambda_{H^3}$ values, where the triangle and box diagrams interfere constructively. For positive values of $\lambda_{H^3}$ (destructive interference), the size and sign of the NLO mass effects is changing considerably as can be inferred from the comparison to the dotted curves. The full K-factors develop a larger dependence on $\lambda_{H^3}$ than the Born-improved HTL due to the NLO top-mass effects. This confirms the findings of Ref. [@Buchalla:2018yce]. The NLO top-mass effects of the total cross section increase with rising collider energy in general except for the regions of destructive interference between the triangle and box diagrams (positive $\lambda_{H^3}$).
![K-factors of Higgs-pair production at NLO as functions of the trilinear self-coupling $\lambda_{H^3}$ in units of the SM value $\lambda_{H^3}^{SM}$ for various c.m. energies of 13 TeV (red curves), 14 TeV (blue curves), 27 TeV (green curves) and 100 TeV (grey curves). The full NLO result is presented by the full curves with the error bars indicating our numerical errors. The dotted curves show the corresponding K-factors of the Born-improved HTL. [MMHT2014]{} PDFs have been used and the renormalization and factorization scales of $\alpha_s$ and the PDFs have been fixed at our central scale choice $\mu_R=\mu_F=Q/2=m_{HH}/2$.[]{data-label="fig:lambdakfac"}](./plots/hh_nlo_lambda_kfactor_mmht.pdf){width="60.00000%"}
The full NLO cross section as a function of $\lambda_{H^3}$ can be parametrized as $$\sigma_{NLO} = \sigma_1 + \sigma_2
\frac{\lambda_{H^3}}{\lambda^{SM}_{H^3}} + \sigma_3 \left(
\frac{\lambda_{H^3}}{\lambda^{SM}_{H^3}} \right)^2 \, .$$ The coefficients $\sigma_{1\ldots 3}$ depend on the c.m. energy of the hadron collider and on the PDFs used in their evaluation. For the various c.m. energies, we obtain the following NLO values for [PDF4LHC]{} PDFs and our central scale choices $\mu_R=\mu_F=Q/2$, $$\begin{aligned}
\sqrt{s} = 13~{\rm TeV}: \quad \sigma_1 & = & 61.35(6)~{\rm fb}\, ,
\quad \sigma_2 = -43.26(5)~{\rm fb}\, , \quad \sigma_3 = 9.62(8)~{\rm
fb} \, , \nonumber \\
\sqrt{s} = 14~{\rm TeV}: \quad \sigma_1 & = & 72.27(7)~{\rm fb}\, ,
\quad \sigma_2 = -50.70(6)~{\rm fb}\, , \quad \sigma_3 = 11.23(9)~{\rm
fb} \, , \nonumber \\
\sqrt{s} = 27~{\rm TeV}: \quad \sigma_1 & = & 270.9(3)~{\rm fb}\, ,
\quad \sigma_2 = -183.1(2)~{\rm fb}\, , \quad \sigma_3 = 39.5(4)~{\rm
fb} \, , \nonumber \\
\sqrt{s} = 100~{\rm TeV}: \quad \sigma_1 & = & 2323(2)~{\rm fb}\, ,
\quad\; \sigma_2 = -1496(2)~{\rm fb}\, , \quad\; \sigma_3 = 313(3)~{\rm
fb} \, ,
\label{eq:siglam1}\end{aligned}$$ where the numbers in brackets denote our numerical errors. The corresponding coefficients with [MMHT2014]{} PDFs read $$\begin{aligned}
\sqrt{s} = 13~{\rm TeV}: \quad \sigma_1 & = & 62.45(7)~{\rm fb}\, ,
\quad \sigma_2 = -44.13(5)~{\rm fb}\, , \quad \sigma_3 = 9.83(9)~{\rm
fb} \, , \nonumber \\
\sqrt{s} = 14~{\rm TeV}: \quad \sigma_1 & = & 73.60(8)~{\rm fb}\, ,
\quad \sigma_2 = -51.75(6)~{\rm fb}\, , \quad \sigma_3 = 11.5(1)~{\rm
fb} \, , \nonumber \\
\sqrt{s} = 27~{\rm TeV}: \quad \sigma_1 & = & 277.4(3)~{\rm fb}\, ,
\quad \sigma_2 = -187.9(2)~{\rm fb}\, , \quad \sigma_3 = 40.6(4)~{\rm
fb} \, , \nonumber \\
\sqrt{s} = 100~{\rm TeV}: \quad \sigma_1 & = & 2401(2)~{\rm fb}\, ,
\quad\; \sigma_2 = -1550(2)~{\rm fb}\, , \quad\; \sigma_3 = 325(3)~{\rm
fb} \, .
\label{eq:siglam2}\end{aligned}$$ It should be noted that the final numerical errors of the cross sections as shown in Figs. \[fig:lambdavar\_1\]–\[fig:lambdakfac\] are smaller than the ones emerging from using the coefficients of Eqs. (\[eq:siglam1\], \[eq:siglam2\]) since the combinations of each bin in $Q$ [*before*]{} integration reduces them.
Conclusions \[sc:conclusions\]
==============================
In this work, we have discussed the full QCD corrections to Higgs-pair production at NLO. We have explained the details of our numerical approach to solve the multi-scale two-loop integrals involving ultraviolet and infrared singularities. The ultraviolet singularities could be extracted from the finite parts by suitable end-point subtractions, while the infrared singularities have been isolated by means of dedicated subtraction terms. The ultraviolet singularities have been absorbed by the proper renormalization of the strong coupling and the top mass, while the infrared ones cancel against the one-loop real corrections involving an additional gluon or quark in the final state of the Higgs-boson pair. We have performed the evaluation of the virtual corrections diagram by diagram without tensor reduction.
The emerging integrals develop thresholds if the virtual $t\bar
t$-threshold is crossed, but also at small virtualities due to the presence of purely gluonic intermediate states. The numerical stabilization of the virtual two-loop integrals has been achieved through integrations by parts of the integrands such that the power of the threshold-singular denominators is reduced. The narrow-width limit of the virtual top quarks has been obtained by a Richardson extrapolation of the results for different sizes of an auxiliarly introduced width parameter. This has allowed a numerical integration of the virtual two-loop corrections with an accuracy of less than one per cent.
The matrix elements for the real corrections have been generated with [FeynArts]{} and [FormCalc]{} and integrated using the library [Collier]{}. The collinear region of the phase-space integration has been regularized numerically by a technical cut.
We have subtracted the Born-improved HTL from the virtual and real corrections individually so that we have been left with the pure NLO top-mass effects beyond the Born-improved HTL that is implemented in the public tool [Hpair]{}. Thus, the final NLO results have been obtained by adding back the numbers from [Hpair]{}.
The final results have been analyzed in detail for the differential cross section in the invariant Higgs-pair mass and the total cross section. Finite top-mass effects beyond the Born-improved HTL decrease the total cross section by about 15% at the LHC. However, the negative mass effects are larger for the differential cross section reaching a level of $-30\%$ or $-40\%$ for large invariant Higgs-pair masses. This implies that the inclusion of the NLO top-mass effects is crucial for a reliable analysis at the LHC and future proton colliders. We have discussed the usual renormalization and factorization scale uncertainties that are in agreement with previous calculations. However, we have identified an additional scale and scheme uncertainty due to the virtual top mass. This uncertainty reaches a level of 15% for the total cross section but can be larger (up to 35%) for the differential cross section. Based on the heavy-top and high-energy expansions, we have discussed the preferred scale choices of the running top mass and identified a large dynamical scale as the proper choice for large invariant Higgs-pair masses. This additional uncertainty has to be combined with the usual renormalization and factorization scale uncertainties. Since the (relative) scheme and scale uncertainties originating from the top mass only mildly depend on the renormalization and factorization scale choice, the addition of this uncertainty may lead to about a [*linear*]{} addition to the other uncertainties, if the total uncertainty is defined as the envelope. This, however, has to be analyzed in more detail which is left for future work.
We have investigated the total cross section as a function of the trilinear coupling varied from its SM value. We have found significant NLO mass effects beyond the Born-improved HTL that result in a shift of the minimum of the cross section at various present and future c.m. energies of the hadron colliders. While the main effect of shifting the minimum originates from the NLO top-mass effects of the real corrections, the more symmetric virtual mass effects mainly affect the size of the total cross section as a function of $\lambda_{H^3}$. The full K-factors develop a larger dependence on $\lambda_{H^3}$ than those of the Born-improved HTL due to the NLO top-mass effects.\
[**Acknowledgements**]{}\
We are indebted to S. Dittmaier for providing us with a copy of his [mathematica]{} program for the QCD corrections in the HTL as constructed for the work of Ref. [@Dawson:1998py] and to R. Gröber for useful discussions. The work of S. G. is supported by the Swiss National Science Foundation (SNF). The work of S. G. and M. M. is supported by the DFG Collaborative Research Center TRR 257 “Particle Physics Phenomenology after the Higgs Discovery”. F. C. and J. R. acknowledge financial support by the Generalitat Valenciana, Spanish Government and ERDF funds from the European Commission (Grants No. RYC-2014-16061, SEJI-2017/2017/019, FPA2017-84543- P,FPA2017-84445-P, and SEV-2014-0398). We acknowledge support by the state of Baden-Württemberg through bwHPC and the German Research Foundation (DFG) through grant no INST 39/963-1 FUGG (bwForCluster NEMO).
Two-loop box diagrams of the virtual corrections
================================================
Here we present the two-loop box diagrams (omitting the ones with reversed fermion flow):
![\[fg:boxdia1\] *Two-loop box diagrams: topologies 1 and 2.*](./plots/topology1.pdf "fig:"){width="75.00000%"} \
![\[fg:boxdia1\] *Two-loop box diagrams: topologies 1 and 2.*](./plots/topology2.pdf "fig:"){width="75.00000%"}
![\[fg:boxdia4\] *Two-loop box diagrams: topologies 5 and 6.*](./plots/topology34.pdf){width="80.00000%"}
\
![\[fg:boxdia4\] *Two-loop box diagrams: topologies 5 and 6.*](./plots/topology56.pdf "fig:"){width="80.00000%"}
[^1]: Note that Higgs pair production will provide indirect constraints on the quartic Higgs coupling [@Liu:2018peg; @Bizon:2018syu; @Borowka:2018pxx].
[^2]: Throughout this work, we will neglect the total Higgs width $\Gamma_H$ in the coefficient $C_\triangle$.
[^3]: Note that we distinguish triangle and box diagrams also at the two-loop level in terms of the number of particles attached to the generic loop, i.e. three particles (two gluons and an off-shell Higgs for the triangle and two gluons and two on-shell Higgs bosons for the box diagrams). The one-particle-reducible diagrams are a special class.
[^4]: The finite part of the complex virtual coefficient ${\cal C}^H_{virt}$ has been shown in Fig. 7a of Ref. [@Spira:1995rr] after renormalization. We define the top mass on-shell, i.e. use the coefficient for $\mu_Q=m_Q$ of this figure for the triangle-diagram contribution to our central prediction.
[^5]: Since $V_{eff}$ is symmetric with respect to $\hat t \leftrightarrow \hat u$ the additional factor 2 emerges from the second term in the numerator of $C_{\triangle\triangle}$ in Eq. (\[eq:coeffvirt\]).
[^6]: Note that $s$ denotes a Feynman parameter here and not the squared hadronic c.m. energy. The same holds for $z$.
[^7]: For the bottom loops, additional stabilization of the numerical integration is required. This is left for future work.
[^8]: Again $z,s$ denote Feynman parameters here.
[^9]: The symmetrization of the integrand $f(\hat t_1, \hat u_1)$ for the $y$ integration is a straightforward result of this substitution.
[^10]: Note that also the individual LO box diagrams are not finite with respect to the $\hat t$ integration, but the sum of all three LO boxes is.
[^11]: Finite top-width effects have been estimated to amount to $\sim -2\%$ [@Maltoni:2014eza]. The effects are slightly larger in the vicinity of the virtual $t\bar t$ threshold, $Q^2\sim 4m_t^2$.
[^12]: Note that a Richardson extrapolation of the integrand [*before*]{} integration provides an alternative to stabilize the numerical integration.
[^13]: The small differences of the total cross sections at the few-per-mille level between the results originate from the slightly different values of the top mass ($m_t=172.5$ GeV in our analysis, $m_t=173$ GeV in Refs. [@Borowka:2016ehy; @Borowka:2016ypz]).
[^14]: We do not separate the treatment of the top-Yukawa couplings and the propagator-top mass, since both are linked by the sum rule emerging from the electroweak $SU(2)\times
U(1)$ symmetry, $y_t - \sqrt{2} m_t/v = 0$, which is needed for the cancellation of divergences in electroweak corrections.
[^15]: Note that these choices are incompatible with a consistent LO prediction, but the relative uncertainties related to the scheme and scale choice of the top mass will be hardly affected by this inconsistency. These uncertainties are just parametric at LO.
[^16]: The NLO form factors of Eq. (\[eq:ffhe\]) correspond to the infrared-subtracted ones according to Ref. [@Davies:2018qvx] plus the additional subtraction of the HTL. The piece related to the latter is absorbed in the functions $G_{1,2}$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A mechanism is proposed for the growth of vesicles dispersed in a liquid solvent and a size distribution function is obtained for the vesicles, both from the first principles calculations. This distribution function is shown to be positively skewed and evolving in time obeying a Fokker-Planck type equation. The critical size of the spherical vesicles is shown to grow in time with an exponent of 0.25. A constant is suggested that characterizes how easily a vesicle can absorb amphiphiles into its periphery.'
---
[**Kinetics of Vesicle Growth**]{}\
Bijit Singha [^1]\
*Department of Physics, Carnegie Mellon University, Pittsburgh, PA 15213*\
Date:
Introduction {#sct1}
============
Amphiphilic molecules can aggregate in a solvent to form different interesting structures. The structure of our interest is called vesicles, in which the amphiphilic bilayer can form hollow sphere. The medium inside such a vesicle can have the same or different concentration than the solvent outside but, for our purpose, we will consider same concentration of the solvent both inside and outside the vesicle. The natural assemblance of the amphiphilic bi-molecules to form such symmetric structures, due to its simplicity and homogeneity in composition, can work as a potential tool to understand the dynamics at the mesoscopic scale. Consequently, these systems have been a subject of wide theoretical studies.\
Vesicles were observed for the first time more than fifty years ago[@VesicleDiscovery1965] and since then, they have been used to improve our understanding of a number of cellular processes[@PantazatosDP1999; @CansAS2003; @LeiG2008; @Lapinski2007], and also in various industrial and commercial applications[@Fernandez2005]. Along with experimental progress, there have been many important theoretical development in understanding the formation and growth of this biological system. Helfrich, in his famous work on the lipid bilyer elasticity[@Helfrich1973], derived the curvature contribution to the free energy and worked on ellipsoidal deformation of the spherical vesicular structure by magnetic fields. Morse and Milner, in their work [@Morse1994], added the finite size contributions to the free energy derived by Helfrich. They showed that the free energy depends on the vesicle size logarithmically. There were several important works before this as well, arguing the existence of the $c\ln N$ contribution to the free energy[@Helfrich1986; @Huse1988], with $c$ being a constant and is less than zero and $N$ being the number of bi-molecules in the vesicle. Morse and Milner found that $c$ is positive with a value of $7/6$ (we will show in a later section how the signature of $c$ will be constrained from the drift of the vesicle size distribution). Several models have been proposed in literature on the positively-skewed size distribution of the vesicles in a solvent[@Morse1994; @Guida2010] as well.\
This work is motivated by a recent observation that the mode of the vesicle size distribution grows in time with a scaling exponent of $1/4$ for some vesicles[@Paulaitis2018]. This was found by assuming that the characteristic scale associated with this process has a power-law dependence on time, $l(t) \sim t^\xi$, and then $\xi$ was fitted to experimental data. In this work, we find out, from first principles calculations that, for an ideal spherical vesicle, the scaling exponent of the exosome size growth is indeed $1/4$ for the critical radius. This helps us to develop a formalism to understand the equilibrium size distribution of vesicles at a given time and its time-evolution of such a system that contains vesicles in aggregate of different sizes. We obtain a Fokker-Planck (FP) equation that will dictate the dynamics of the vesicle-growth process. FP equation has been previously used in [@Zapata2009] to understand the formation of spherical vesicles but we will focus primarily on vesicle-growth in this work. To characterize how easily an amphiphilic bi-molecule can assimilate into the periphery of a given vesicle, we also propose a constant in this work. This constant will be determined by the detailed chemical and structural properties of any vesicle.\
We assume in this work that the dissolution or the growth of the vesicles proceeds through the emission or assimilation of a single amphiphile bi-molecule from the periphery of the vesicles. We have also used the principles of detailed balancing which will hold only when the system is in thermal equilibrium states at all time. Both the assumptions are justified because the vesicle-growth, as found from the experiments, is a sufficietly slow process. Moreover, several interesting properties of the system have direct correspondences to the specific features of the thermodynamic functions, such as free energy. This helps us in two ways: (i) we can come up with the constraints that these functions should satisfy by simply analyzing their consequences to a real system and comparing with the observations, (ii) finding the effect of such functions on the size and time-evolution of the system gives us an intuitive understanding of the corresponding system. This also allows to specify the constants characterizing such processes ($e.g.$ coefficiient of assimilation) from the constraints imposed by observations about the vesicle sizes and their time-evolution.\
The structure of this paper is as follows. In Section \[sct2\], we will find out that the chemical potential corresponding to the vesicle growth processes is effectively a constant for large size of vesicles. In Section \[sct3\], we will assume a size distribution function for the vesicles in the growth phase and will derive a Fokker-Planck equation using principles of detailed balancing. This will give us an FP equation for the time evolution of the vesicle size distribution in general. In Section \[sct4\], we will use this equation for the special case of spherical vesicles to derive the scaling exponent $0.25$ of time for the growth process. We will discuss our findings in Section \[sct5\].
Thermodynamics of vesicle formation and growth {#sct2}
==============================================
\(O) at (0,0,0); (A) at (2,0,0); (B) at (0,1,0); (C) at (2,1,0);
at (8.2,1);
(7,0) rectangle (4,1);
(4.5,0.5) – (5.5,0.5); (4.4,0.5) circle (0.2cm); (5.6,0.5) – (6.6, 0.5); (6.7, 0.5) circle (0.2cm); (-0.8,0.5) – (0.2, 0.5); (0.3, 0.5) circle (0.2cm); (-0.9,0.5) – (-1.9, 0.5); (-2.0, 0.5) circle (0.2cm);
(2.4,2.5) – (2.4, 3.5); (2.4,2.5) circle (0.2cm); (2.4,3.6) – (2.4, 4.6); (2.4,4.8) circle (0.2cm); (2.4, -1.5) – (2.4, -2.5); (2.4,-1.5) circle (0.2cm); (2.4, -2.6) – (2.4, -3.6); (2.4,-3.6) circle (0.2cm);
(4.2,1.25) – (5.2, 1.6); (4.2,1.25) circle (0.2cm); (5.22,1.62) – (6.22,1.9); (6.45,1.9) circle (0.2cm); (0.6,1.25) – (-0.4, 1.6); (0.5,1.25) circle (0.2cm); (-0.38,1.62) – (-1.7,2); (-1.7,2) circle (0.2cm);
(-0.4,-1) – (-1.4, -1.4); (-1.5,-1.4) circle (0.2cm); (0.5,-0.53) – (-0.38,-0.98); (0.56,-0.53) circle (0.2cm); (4.2,-0.25) – (5.2,-0.6); (4.3,-0.25) circle (0.2cm); (5.22, -0.62) – (6.22, -0.92); (6.48,-0.92) circle (0.2cm); (3.7, -0.95) – (4.7, -1.7); (3.8, -0.95) circle (0.2cm); (4.6,-1.71) – (5.6,-2.49); (5.6,-2.47) circle (0.2cm); (3.2,-1.45) – (3.7, -2.2); (3.2,-1.35) circle (0.2cm); (3.8, -2.3) – (4.6,-3.3); (4.6,-3.3) circle (0.2cm); (1.6,-1.45) – (1.3,-2.2); (1.6,-1.4) circle (0.2cm); (1.31,-2.3) – (1.0,-3.4); (1,-3.6) circle (0.2cm); (1.1,-1.2) – (0.3,-1.8); (1.1,-1.1) circle (0.2cm); (0.29,-1.81) – (-0.60, -2.4); (-0.60,-2.55) circle (0.2cm); (0.9,2) – (0.0,2.6); (0.9,1.9) circle (0.2cm); (-0.1, 2.8) – (-0.8,3.3); (-0.9,3.4) circle (0.2cm); (1.5,2.3) – (1,3.2); (1.5,2.3) circle (0.2cm); (0.9,3.2) – (0.5, 4.0); (0.3,4.15) circle (0.2cm);
(3.3,2.3) – (3.8,3.2); (3.3,2.3) circle (0.2cm); (3.9,3.2) – (4.5,4.0); (4.6,4.2) circle (0.2cm); (3.9,1.9) – (4.6,2.5); (3.9,1.9) circle (0.2cm); (4.7,2.6) – (5.4,3.2); (5.55,3.4) circle (0.2cm);
at (11.6,3.5); at (11.6,-0.1);
(11.6,2.5) – (11.6,0); (11.6,2.2) circle(0.3cm);
When a large number of amphiphile bi-molecules are dissolved, vesicles can be formed in the solvent. In this process, entropy of the solute amphiphiles increases along with an increase in their interaction free energy. The former is due to the increase in the count of all possible microstates of the system, while the second effect is attributed to the fact that the amphiphiles like to be surrounded by similar bi-molecules than the solvent molecules. Structure formation by these amphiphiles becomes possible when the interaction energy dominates the energy associated with the increase in entropy. In what follows in this section, we will show that when such vesicles are formed, the chemical potential of the vesicles is the same for the sufficiently large size.\
Let’s imagine that vesicles in aggregate of different sizes $\{i\}$ are present in the solvent, with $i$ being the number of amphiphile bi-molecules in a given vesicle, and let’s say there are $N_i$ number of vesicles in aggregate-$i$. From the first law of thermodynamics, we can write $$\begin{aligned}
\label{1stLaw}
{d\hspace*{-0.08em}\bar{}\hspace*{0.1em}}Q = dU + PdV,\end{aligned}$$ where ${d\hspace*{-0.08em}\bar{}\hspace*{0.1em}}Q$ is the heat gained by the system, $U$ is the internal energy, $P$ is the pressure and $V$ is the volume. Because the system is not thermally isolated in general, second law gives us $$\begin{aligned}
dU + P dV \leq TdS,\end{aligned}$$ where $S$ is the entropy of the system. Because we have different sizes of vesicles, the energy density of the system will depend on the relative amounts of the aggregates and thus, we have to add one more term to the above expression to write $$\begin{aligned}
\label{InEq}
d U + PdV - \sum_i \mu_i dN_i \leq T dS,\end{aligned}$$ Here $\mu_i$ denotes the chemical potential of a vesicle that belongs to the aggregate-$i$. The change in Gibbs Free Energy $dG = dU + pdV - TdS$ should be less than zero for any spontaneous process (and is zero for processes at equilibrium). Combining this with Eq. (\[InEq\]), we get $$\begin{aligned}
\sum_i \mu_i dN_i \leq 0.\end{aligned}$$ Because vesicle growth is a very slow process, we assume that this is in equilibrium: $$\begin{aligned}
\sum_i \mu_i dN_i = 0.\end{aligned}$$
Now, if some vesicles of aggregate-$i$ absorb the same number of vesicles of aggregate-$j$ to form vesicles of aggregate-$(i+j)$, then $-dN_i = -dN_j = +dN_{i+j}$. Using this condition in the above equation, we get $$\begin{aligned}
\label{AdditionOfPotential}
\mu_i + \mu_j = \mu_{i+j}.\end{aligned}$$ From the above expression, we can write $$\begin{aligned}
\mu_i + \mu_{j + \delta n} = \mu_{i + j + \delta n}.\end{aligned}$$ Taylor-expanding the second term on the left hand side and the term on the right hand side and ignoring the higher order terms $\mathcal{O}(\delta n^2)$ for $\delta n \ll i, j$, we get $$\begin{aligned}
\mu_i + \mu_j + \delta n \frac{d \mu_j}{d n} = \mu_{i+j} + \delta n\frac{d \mu_{i+j}}{d n}\end{aligned}$$ Using Eq. (\[AdditionOfPotential\]) again gives us $$\begin{aligned}
\label{MuNConstancy}
\mu_n = \rm{constant}\end{aligned}$$ for sufficiently large $n$. This is a very important relation and will determine the equilibrium size distribution of the vesicles. This tells us that the large-sized vesicles will be in equibrium and will not like to exchange amphiphiles, but they can grow in size by absorbing bi-molecules from small vesicles. This will lead to a specific scaling exponent in time-dependence of the distribution function, as will be shown in the next section.
Size distribution function of the vesicles {#sct3}
==========================================
We assume a distribution function for a vesicle of volume $V$, containing $n$ bi-molecules at time $t$. This is similar to a probability density function at a given time if normalized to unity: $$\begin{aligned}
\label{SizeDist}
f(n,V,t) = \psi(n,t) \delta(V - V(n)),\end{aligned}$$ where the delta function is basically implying that the volume of the vesicle is determined by the number of bi-molecules on its surface. The size-dependence and the time evolution of this function will be dictated by thermodynamics. The size of the vesicle changes by aggregation or dissociation of amphiphile molecules and can be expressed by a kinetic equation $$\begin{aligned}
\frac{\partial f(n,t)}{\partial t}
=
\alpha^{(+)}_{n-1,n} f(n-1,t) - \alpha^{(-)}_{n,n-1} f(n,t) + \alpha^{(-)}_{n+1,n} f(n+1,t) - \alpha^{(+)}_{n,n+1} f(n,t),\end{aligned}$$ where $\alpha^{(+)}_{i,i+1}$ is the average number of events per unit time that one bi-molecule is absorbed by the vesicle to increase the total number from $i$ to $i+1$, $\alpha^{(-)}_{i,i-1}$ is the average number of events per unit time that one bi-molecule is released from the vesicle to decrease the total number from $i$ to $i -1$. Defining the flux $J_n$ as $$\begin{aligned}
J_n = \alpha^{(+)}_{n,n+1}f(n,t) - \alpha^{(-)}_{n+1,n}f(n,t),\end{aligned}$$ we can write $$\begin{aligned}
\frac{\partial f}{\partial t} &=&
- \left(
J_n - J_{n-1}
\right).\end{aligned}$$ The absorption and emission coefficients $\alpha^{(+)}$ and $\alpha^{(-)}$ can be determined from the kinetics of the corresponding process. For a vesicle of critical radius, aggregation and dissociation cancels out each other exactly and assuming the following equilibrium distribution of the vesicle $$\begin{aligned}
f^{(eq)}(n)
&\sim&
\exp \left[
- \frac{R_{rev}(n)}{k_B T}
\right],\end{aligned}$$ where $R_{rev}(n)$ is the work of formation of a vesicle consisting of bi-molecules done in a reversible process. For constant temperature and pressure, this is basically the Gibbs free energy: $$\begin{aligned}
R_{rev}(n) = \Delta G(n).\end{aligned}$$ From this point onwards, we will call $\Delta G$ as free energy. Now, we use the principle of detailed balancing to write $$\begin{aligned}
\frac{\alpha^{(+)}_{n,n+1}}{\alpha^{(-)}_{n+1,n}}
&=&
\frac{f^{(eq)}(n+1)}{f^{(eq)}(n)} \nonumber\\
&=&
\exp \left[
- \left(
\frac{\Delta G(n+1) - \Delta G(n)}{k_B T}
\right)
\right].\end{aligned}$$ Using these will eventually lead us to a Fokker-Planck equation that can describe time-evolution of a vesicle during the nucleation and growth processes: $$\begin{aligned}
\label{FokkerPlanck}
\frac{\partial f(n,t)}{\partial t}
= - \frac{\partial J}{\partial n},\end{aligned}$$ where $$\begin{aligned}
\label{DefJ}
J(n,t) = - \alpha ^{(+)}_{n,n+1} \left[
\frac{\partial f(n,t)}{\partial n}
+ \frac{f(n,t)}{k_B T} \frac{\partial \Delta G(n)}{\partial n}
\right].\end{aligned}$$ Similar expressions are used for a lot of other thermodynamic systems, as shown in [@SlezovBook]. The first term in Eq. (\[DefJ\]) is the diffusion flow rate and the second term corresponds to thermodynamic flow rate or the drift flow rate. The first term in Eq. (\[DefJ\]) describes the size distribution cloud spreading diffusively while the second term denotes the peak of the same cloud drifting towards a different $n$-value. In the absence of the second term, this becomes a zero-drift equation with constant diffusion and the peak of the size distribution simply spread out but doesn’t move. Thus, we can see from Eq. (\[DefJ\]) that it is the size-dependence of the free energy which is pushing the peak of the distribution function to the right and thus dictates the critical growth rate of the vesicles.\
Let’s recall briefly a few ideas about stochastic processes such as the one we are dealing with in this work. We can consider a stochastic differential equation of the form $$\begin{aligned}
\label{DefEto}
d\textbf{n} = \boldsymbol{\mu}(\textbf{n},t) dt + \boldsymbol{\sigma}(\textbf{n},t) d\textbf{W}_t\end{aligned}$$ where $\boldsymbol{\mu}$ is a $K$-dimensional vector, $\boldsymbol{\sigma}$ is an $K \times K$ matrix and $d\textbf{W}_t$ is a $M$-dimensional vector that denotes the infinitesimal increment in the Wiener process with the property that $$\begin{aligned}
\langle \textbf{W}_t \rangle &=& 0,\\
\langle \textbf{W}_t \textbf{W}_{t'} \rangle &=& {\rm{min}}(t,t').\end{aligned}$$ Any process corresponding to Eq. (\[DefEto\]) is called Itô process. Corresponding to this process, there exists a Fokker-Planck equation satisfied by the conditional probability density $P(\textbf{n},t\vert \textbf{n}_0, t_0)$ of the random variable $N_t$: $$\begin{aligned}
\label{FPfromIto}
\frac{\partial P(\textbf{n},t)}{\partial t} &=&
- \sum_{i = 1}^K \frac{\partial}{\partial n_i} \left[
\mu_i(\textbf{n},t)P(\textbf{n},t)
\right]
+ \sum_{i = 1}^K \frac{\partial^2}{\partial n_i \partial n_j} \left[
D_{ij}(\textbf{n},t) P(\textbf{n},t)
\right]\end{aligned}$$ In Eq. (\[FPfromIto\]), $\boldsymbol{\mu}$ is called the drift vector and $\textbf{D} = \frac{1}{2}\boldsymbol{\sigma} \boldsymbol{\sigma}^T$ is called diffusion tensor. We have used tensor indices in Eq. (\[FPfromIto\]) to show the dimensionality of each term explicitly. Defining the diffusion current using Fick’s law $$\begin{aligned}
J^i_{\rm{diff}}(\textbf{n},t) = - D^{ij} \frac{\partial P(\textbf{n},t)}{\partial n^i}\end{aligned}$$ and the drift current $$\begin{aligned}
\label{DefDrift}
J_{\rm{drift}}^i(\textbf{n},t) = \mu^i(\textbf{n}) P(\textbf{n},t),\end{aligned}$$ we can write the Fokker-Planck equation, Eq. ([\[FPfromIto\]]{}), as $$\begin{aligned}
\label{FPinDriftAndDiffusion}
\frac{\partial P(\textbf{n},t)}{\partial t} &=&
- \frac{\partial}{\partial n^i} \left[
J^i_{\rm{drift}}(\textbf{n},t) + J^i_{\rm{diff}}(\textbf{n},t)
\right]\end{aligned}$$ Comparing Eq. (\[DefDrift\]) and (\[FPinDriftAndDiffusion\]) with Eq. (\[FokkerPlanck\]) and Eq. (\[DefJ\]), we get that the drift current $$\begin{aligned}
J_{\rm{drift}}(n,t) = \frac{\alpha_{n,n+1}}{k_B T} \frac{\partial \Delta G}{\partial n} f(n,t)\end{aligned}$$ and the drift vector $$\begin{aligned}
\label{DriftinFP}
\mu(n,t) = \frac{\alpha_{n,n+1}}{k_B T} \frac{\partial \Delta G}{\partial n}.\end{aligned}$$ Notice that this framework can be applied generally for any vesicle systems. In the next section, we are going to assume spherical vesicles and derive a more specific equation governing the time-evolution of such a system.
Time-evolution of the size distribution {#sct4}
=======================================
We can see from Eq. (\[DefJ\]) that we need the free energy of the vesicle in order to find out the size distribution and the dynamic scaling of the vesicle. Free energy is different for different structures but we consider here, for simplicity, an aggregate of spherical vesicles. We use the same form of the free energy for a spherical vesicle as obtained in [@Morse1994]: $$\begin{aligned}
\label{FreeEnergy}
\Delta G(n) = 8 \pi (\kappa + \frac{1}{2}\bar{\kappa}) + k_B T ({\rm{const.}}) - c k_BT \ln n~.\end{aligned}$$ In the above expression, $\kappa$ and $\bar{\kappa}$ are the vesicle’s bending rigidity and gaussian rigidity, $T$ is the temperature of the system, $n$ is the number of bi-molecules in a given vesicle, $k_B$ is the Boltzmann constant and $c$ is another constant. Because the first term in the right hand side is scale-independent, the phase behavior is dictated only by size-dependent contribution coming from the last term in that expression. Notice that, the logarithmic dependence of that term tells us that the chemical potential is effectively a constant near the critical radius of the vesicle, as expected from Eq. (\[MuNConstancy\]). Using Eq. (\[FokkerPlanck\]) and (\[DefJ\]), and assuming that the diffusivity remains constant during the growth process ($\alpha_{n,n+1} = D_0$ for sufficiently large $n$), we get $$\label{FP}
\frac{\partial \psi}{\partial t}
=
D_0 \frac{\partial}{\partial n} \left[
\frac{\partial \psi}{\partial n} + \frac{\psi}{k_B T} \frac{\partial \Delta G}{\partial n}
\right]~.$$ Comparing this expression with Eq. (\[FPfromIto\]), (\[DefDrift\]), (\[DriftinFP\]) with the above equation, we get the drift speed of the size distribution function assumed in Eq. (\[SizeDist\]): $$\begin{aligned}
u(n,t) =
\frac{\partial n_M}{\partial t}
&=&
\frac{c D_0}{ n_M}~.\end{aligned}$$ Here $n_M$ denotes the mode size at a given time. Thus, $$\begin{aligned}
n_M(t) = \sqrt{2cD_0}~ t^{1/2} \sim t^{1/2}~.\end{aligned}$$
Comparing the above expression with the experimental result, we can get the diffusivity constant corresponding to this process. We can express the same differential equation in terms of the radius of the vesicle using $$\label{NRRelation}
n = \frac{4 \pi R^2}{\pi a^2} = \frac{4R^2}{a^2},$$ where $R$ is the radius of the vesicle and can be measured experimentally, and $a$ is the radius of the cross-section of the hydrophilic head of the vesicle. This gives the the dependence of the critical radius on time: $$\begin{aligned}
R_c = \frac{a}{2} \left( \frac{7 D_0}{3} t \right)^{1/4} \sim~ t^{1/4}~.\end{aligned}$$ The scaling exponent of time in the above equation is $0.25$. It is worth mentioning here that, in [@Paulaitis2018], the authors found a similar result by assuming a power-law scaling of the mean diameter and fitting the exponent to their experimental data. Also, our observable is the size of the vesicles and it will be more helpful to express Eq. (\[FP\]) in terms of the radius of the vesicles. $$d n = \frac{8R dR}{a^2},$$ and thus, from Eq. (\[FP\]), $$\begin{aligned}
\frac{\partial \psi}{\partial t} &=&
\frac{D_0a^4}{64 R^2}\left[
\frac{\partial^2 \psi}{\partial R^2} - \left(\frac{1+2c}{R}\right)\frac{\partial \psi}{\partial R} + \frac{4c\psi}{R^2}
\right]
.\end{aligned}$$
Conclusion {#sct5}
==========
We have given a framework under which we can calculate the time evolution of size distribution of the vesicles, and showed that the critical size of sphericle vesicles will grow with time with a scaling exponent of $0.25$. This is derived in a straightforward way, first by assuming that the free energy is only logarithmically (*i.e.* slowly) varying with the number of bi-molecules, thus the chemical potential being a constant effectively for sufficiently large size of vesicles, and then finding out how the size distribution will evolve by writing down the Fokker-Planck equation corresponding to the process of bi-molecule exchange between different aggregates. Furthermore, we have proposed a constant to characterize the ease of diffusivity of an amphiphile bi-molecule into the periphery of the vesicle. The chemistry and the detailed structural properties will be encoded in this constant while the growth process will be dictated by thermodynamics. This means any deviation from the ideal value of this constant (*i.e.* value obtained from this work for a spherical vesicle with same solvent with similar concentration inside and outside) can give us information about the physical or chemical interactions or any other form of perturbations playing a role in the system. All these behooves further investigation with more complicated vesicle systems. The scaling exponent will vary similarly for non-ideal structures but we can use the same underlying theory mentioned in this article to comment on the compositions and geometry of the corresponding vesicles, scope for future work.
Acknowledgements {#acknowledgements .unnumbered}
================
The author is thankful to Saigopalakrishna Yerneni and Sushil Lathwal for all the valuable discussions about vesicles and some other interesting biological systems, and to Prof. Changjin Huang and Prof. Markus Deserno for their valuable comments on this work. This work was supported by Carnegie Mellon University Department of Physics.
[10]{} A. D. Bangham, M. M. Standish, J. C. Watkins. *Diffusion of univalent ions across the lamellae of swollen phospholipids*. the Journal of Molecular Biology **13** (1965) 238 D. P. Pantazatos and R. C. Macdonald. *Directly Observed Membrane Fusion Between Oppositely Charged Phospholipid Bilayers.* The Journal of Membrane Biology **170** (1999) 27 A-S. Cans *et. al.* *Artificial cells: Unique insights into exocytosis using liposomes and lipid nanotubes.* Proceedings of the National Academy of Sciences **100 2** (2003) 400 G. Lei and R. C. Macdonald. *Effects on interactions of oppositely charged phospholipid vesicles of covalent attachment of polyethylene glycol oligomers to their surfaces: adhesion, hemifusion, full fusion and “endocytosis".* The Journal of Membrane Biology **97** (2008) 221 M. M. Lapinski, A. Castro-Forero, A. J. Greiner, R. Y. Ofoli and G. J. Blanchard. *Comparison of liposomes formed by sonication and extrusion: rotational and translational diffusion of an embedded chromophore.* Langmuir **23** (2007) 11677 P. Fernandez, N. Willenbacher, T. Frechen, A. Kuhnle. *Vesicles as rheology modifier*. Colloids and Surfaces A: Physicochemical and Engineering Aspects **262** (2005) 204 W. Helfrich. *Elastic Properties of Lipid Bilayers: Theory and Possible Experiments*. Z. Naturjiorsch. C **28** (1973) 693 D. C. Morse and S. T. Milner. *Fluctuations and Phase Behavior of Fluid Membrane Vesicles*. Europhysics Letters **26** (1994) 565 W. Helfrich. *Size distributions of vesicles : the role of the effective rigidity of membranes.* Journal of Physics (Paris) **47** (1986) 321 D. Huse and S. Leibler. *Phase behaviour of an ensemble of nonintersecting random fluid films.* Journal of Physics (Paris) **49** (1988) 605 V. Guida. *Thermodynamics and kinetics of vesicles formation processes*. Advances in Colloid and Interface Science **161** (2010) 77 M. Paulaitis, K. Agarwal, P. Nana-Sinkam. *Dynamic Scaling of Exosome Sizes*. Langmuir **34** (2018) 9387 E. H. Zapata, L. M. Balbuena, I. S. Holek. *Thermodynamics and dynamics of the formation of spherical lipidic vesicles.* Journal of Biological Physics **35** (2009) 297 V. V. Slezov. *Kinetics of First-order Phase Transitions*. Wiley-VCH (2009)
[^1]: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Stein’s method for concentration inequalities was introduced to prove concentration of measure in problems involving complex dependencies such as random permutations and Gibbs measures. In this paper, we provide some extensions of the theory and three applications: (1) We obtain a concentration inequality for the magnetization in the Curie–Weiss model at critical temperature (where it obeys a nonstandard normalization and super-Gaussian concentration). (2) We derive exact large deviation asymptotics for the number of triangles in the Erdős–R'' enyi random graph $G(n,p)$ when $p \ge0.31$. Similar results are derived also for general subgraph counts. (3) We obtain some interesting concentration inequalities for the Ising model on lattices that hold at all temperatures.'
address: |
Department of Statistics\
University of California, Berkeley\
367 Evans Hall \#3860\
Berkeley, California 94720-3860\
USA\
\
author:
-
-
title: 'Applications of Stein’s method for concentration inequalities'
---
and
.
.
Introduction {#sec:intro}
============
In his seminal 1972 paper [@stein72], Charles Stein introduced a method for proving central limit theorems with convergence rates for sums of dependent random variables. This has now come to be known as *Stein’s method*. The technique is primarily used for proving distributional limit theorems (both Gaussian and non-Gaussian). Stein’s attempts [@stein86] at devising a version of the method for large deviations did not prove fruitful. Some progress for sums of dependent random variables was made by Raič [@raic04]. The problem was finally solved in full generality in [@chatterjee05]. A selection of results and examples from [@chatterjee05] appeared in the later papers [@chatterjee07; @chatterjee07a]. In this paper, we extend the theory and work out three further examples. The paper is fully self-contained.
The sections are organized as follows. In Section \[sec:results\], we state the main results, the examples, and some proof sketches. The complete proofs are in Section \[sec:proof\].
Results and examples {#sec:results}
====================
The following abstract theorem is quotedfrom [@chatterjee07]. It summarizes a collection of results from [@chatterjee05]. This is a generalization of Stein’s method of exchangeable pairs to the realm of concentration inequalities and large deviations.
\[thm:conc\] Let ${\mathcal{X}}$ be a separable metric space and suppose $(X,X')$ is an exchangeable pair of ${\mathcal{X}}$-valued random variables. Suppose $f\dvtx{\mathcal{X}}\to
{\mathbb{R}}$ and $F\dvtx{\mathcal{X}}\times{\mathcal{X}}\to{\mathbb{R}}$ are square-integrable functions such that $F$ is antisymmetric \[i.e., $F(X,X')=-F(X',X)$ a.s.\], and ${\mathbb{E}}(F(X,X'){|}X) = f(X)$ a.s. Let $$\Delta(X) := \tfrac{1}{2}{\mathbb{E}}\bigl(\bigl|\bigl(f(X)-f(X')\bigr)F(X,X')\bigr| |
X\bigr).$$ Then ${\mathbb{E}}(f(X)) = 0$, and the following concentration results hold for $f(X)$:
If ${\mathbb{E}}(\Delta(X)) <\infty$, then ${\operatorname{Var}}(f(X)) = \frac
{1}{2}{\mathbb{E}}((f(X)-f(X'))F(X,X'))$.
Assume that ${\mathbb{E}}(e^{\theta f(X)} |F(X,X')|) < \infty$ for all $\theta$. If there exists nonnegative constants $B$ and $C$ such that $\Delta(X) \le B f(X) + C$ almost surely, then for any $t\ge0$, $${\mathbb{P}}\{f(X) \ge t\}\le\exp\biggl(-\frac{t^2}{2C + 2Bt} \biggr)
\quad\mbox
{and}\quad {\mathbb{P}}\{f(X) \le-t\}\le\exp\biggl(-\frac{t^2}{2C} \biggr).$$
For any positive integer $k$, we have the following exchangeable pairs version of the Burkholder–Davis–Gundy inequality: $${\mathbb{E}}(f(X)^{2k}) \le(2k-1)^k {\mathbb{E}}(\Delta(X)^k).$$
Note that the finiteness of the exponential moment for all $\theta$ ensures that the tail bounds hold for all $t$. If it is finite only in a neighborhood of zero, the tail bounds will hold for $t$ less than a threshold.
One of the contributions of the present paper is the following generalization of the above result for non-Gaussian tail behavior. We apply it to obtain a concentration inequality with the correct tail behavior in the Curie–Weiss model at criticality.
\[thm:nong1\] Suppose $(X,X')$ is an exchangeable pair of random variables. Let $F(X,X'), f(X)$ and $\Delta(X)$ be as in Theorem \[thm:conc\]. Suppose that we have $$\Delta(X) \le\psi(f(X)) \qquad\mbox{almost surely}$$ for some nonnegative symmetric function $\psi$ on ${\mathbb{R}}$. Assume that $\psi$ is nondecreasing and twice continuously differentiable in $(0,\infty)$ with $$\label{eq:al}
\alpha := \sup_{x>0}x\psi'(x)/\psi(x)<2$$ and $$\label{eq:gd}
\delta := \sup_{x>0}x\psi''(x)/\psi(x)<\infty.$$ Assume that ${\mathbb{E}}(|f(X)|^{k}) < \infty$ for all positive integer $k\ge
1$. Then for any $t\ge0$, we have $${\mathbb{P}}\bigl(\vert f(X) \vert> t\bigr) \le c\exp\biggl(-\frac{t^2}{2\psi(t)} \biggr)$$ for some constant $c$ depending only on $\alpha,\delta$. Moreover, if $\psi$ is only once differentiable with $\alpha<2$ as in (\[eq:al\]), then the tail inequality holds with exponent $t^2/4\psi(t)$.
An immediate corollary of Theorem \[thm:nong1\] is the following.
\[cor:nong\] Suppose $(X,X')$ is an exchangeable pair of random variables. Let $F(X,X'), f(X)$ and $\Delta(X)$ be as in Theorem \[thm:conc\]. Suppose that for some real number $\alpha\in(0,2)$ we have $$\Delta(X) \le B \vert f(X) \vert^\alpha+ C \qquad\mbox{almost surely},$$ where $B>0,C\ge0$ are constants. Assume that ${\mathbb{E}}(|f(X)|^{k}) < \infty$ for all positive integer $k\ge1$. Then for any $t\ge0$ we have $${\mathbb{P}}\bigl(\vert f(X) \vert> t\bigr) \le c_{\alpha}\exp\biggl(-\frac{1}{2}\cdot
\frac
{t^2}{Bt^\alpha+C} \biggr)$$ for some constant $c_{\alpha}$ depending only on $\alpha$.
The result in Theorem \[thm:nong1\] states that the tail behavior of $f(X)$ is essentially given by the behavior of $f(X)^{2}/\Delta(X)$. Condition (\[eq:al\]) implies that $\psi(x)< \psi(1)(1+x^2)$ for all $x\in{\mathbb{R}}$. Moreover, the constant $c_{\alpha}$ appearing in Theorem \[thm:nong1\] can be written down explicitly but we did not attempt to optimize the constant. The proof of Theorem \[thm:nong1\] is along the same lines as Theorem \[thm:conc\], but somewhat more involved. Deferring the proof to Section \[sec:proof\], let us move on to examples.
Example: Curie–Weiss model at criticality {#subsec:curie}
-----------------------------------------
The “Curie–Weiss model of ferromagnetic interaction” at inverse temperature $\beta$ and zero external field is given by the following Gibbs measure on $\{+1,-1\}^{n}$. For a typical configuration ${\bolds{\sigma}}=(\sigma_{1},\sigma_{2},\ldots,\sigma_{n})\in\{+1,-1\}^{n}$, the probability of ${\bolds{\sigma}}$ is given by $$\mu_{\beta}(\{{\bolds{\sigma}}\}) := Z_{\beta}^{-1} \exp\biggl( \frac{\beta
}{n}\sum_{i<j}
\sigma_{i}\sigma_{j} \biggr),$$ where $Z_{\beta}=Z_{\beta}(n)$ is the normalizing constant. It is well known that the Curie–Weiss model shows a phase transition at $\beta_{c}=1$. For $\beta<\beta_{c}$, the magnetization $m({\bolds{\sigma}}):=\frac
{1}{n}\sum_{i=1}^{n}\sigma_{i}$ is concentrated at $0$ but for $\beta
>\beta
_{c}$ the magnetization is concentrated on the set $\{-x^{*},x^{*}\}$ where $x^{*}>0$ is the largest solution of the equation $x=\tanh(\beta
x)$. In fact, using concentration inequalities for exchangeable pairs it was proved in [@chatterjee05] (Proposition 1.3) that for all $\beta
\ge0,h\in{\mathbb{R}}, n\ge1, t\ge0$ we have $${\mathbb{P}}\biggl(|m-\tanh(\beta m+h)|\ge\frac{\beta}{n} + \frac{t}{\sqrt
{n}}
\biggr) \le2\exp\biggl(-\frac{t^{2}}{4(1+\beta)} \biggr),$$ where $h$ is the external field, which is zero in our case. Although a lot is known about this model (see Ellis [@ellis85], Section IV.4, for a survey), the above result—to the best of our knowledge—is the first rigorously proven concentration inequality that holds at all temperatures. (See also [@chazottes06] for some related results.)
Incidentally, the above result shows that when $\beta< 1$, the magnetization is at most of order $n^{-1/2}$. It is known that at the critical temperature the magnetization $m({\bolds{\sigma}})$ shows a non-Gaussian behavior and is of order $n^{-1/4}$. In fact, at $\beta=1$ as $n\to
\infty
$, $n^{1/4}m({\bolds{\sigma}})$ converges to the probability distribution on ${\mathbb{R}}$ having density proportional to $\exp(-t^{4}/12)$. This limit theorem was first proved by Simon and Griffiths [@simon73] and error bounds were obtained recently [@chatterjeeshao09; @eichelsbacherlowe09]. The following concentration inequality, derived using Theorem \[thm:nong1\], fills the gap in the tail bound at the critical point.
\[pr:crit\] Suppose ${\bolds{\sigma}}$ is drawn from the Curie–Weiss model at the critical temperature $\beta=1$. Then, for any $n\ge1$ and $ t\ge0$, the magnetization satisfies $${\mathbb{P}}\bigl(n^{1/4}|m({\bolds{\sigma}})| \ge t\bigr) \le2e^{-ct^4},$$ where $c>0$ is an absolute constant.
Here, we may remark that such a concentration inequality probably cannot be obtained by application of standard off-the-shelf results (e.g., those surveyed in Ledoux [@ledoux01], the famous results of Talagrand [@talagrand95] or the recent breakthroughs of Boucheron, Lugosi and Massart [@blm03]), because they generally give Gaussian or exponential tail bounds. There are several recent remarkable results giving tail bounds different from exponential and Gaussian. The papers [@bcr06; @lo00; @ggm05; @chazottes06] deal with tails between exponential and Gaussian and [@bcr05; @bob07] deal with subexponential tails. Also in [@bl00; @gozlan07; @gozlan09], the authors deal with tails (possibly) larger than Gaussian. However, it seems that none of the techniques given in these references would lead to the result of Proposition \[pr:crit\].
It is possible to derive a similar tail bound using the asymptotic results of Martin-Löf [@martinlof82] about the partition function $Z_{\beta}(n)$ (see also Bolthausen [@bolthausen87]). An application of their results gives that $$\sum_{{\bolds{\sigma}}\in\{-1,+1\}^{n}}e^{{n}/{2}m({\bolds{\sigma}})^{2}+n\theta
m({\bolds{\sigma}})^{4}}\simeq\frac{2^{n+1}\Gamma(5/4)}{\sqrt{2\pi}} \biggl(\frac
{12n}{1-12\theta} \biggr)^{1/4}$$ for $\theta<1/12$ in the sense that the ratio of the two sides converges to one as $n$ goes to infinity and from here the tail bound follows easily (without an explicit constant). However, this approach depends on a precise estimate of the partition function \[e.g., large deviation estimates or finding the limiting free energy $\lim n^{-1}\log Z_{\beta}(n)$ are not enough\] and this precise estimate is hard to prove. Our method, on the other hand, depends only on simple properties of the Gibbs measure and is not tied specifically to the Curie–Weiss model.
The idea used in the proof of Proposition \[pr:crit\] can be used to prove a tail inequality that holds for all $0\le\beta\le1$. We state the result below without proof. Note that the inequality gives the correct tail bound for all $0\le\beta\le1$.
Suppose ${\bolds{\sigma}}$ is drawn from the Curie–Weiss model at inverse temperature $\beta$ where $0\le\beta\le1$. Then, for any $n\ge1$ and $
t\ge0$ the magnetization satisfies $${\mathbb{P}}\bigl(3(1-\beta)m({\bolds{\sigma}})^{2}+\beta^{3}m({\bolds{\sigma}})^{4} \ge t\bigr) \le2e^{-nt/160}.$$
It is possible to derive similar non-Gaussian tail inequalities for general Curie–Weiss models at the critical temperature. We briefly discuss the general case below. Let $\rho$ be a symmetric probability measure on ${\mathbb{R}}$ with $\int x^{2} \,d\rho(x)=1$ and $\int\exp(\beta
x^{2}/2) \,d\rho(x)<\infty$ for all $\beta\ge0$. The general Curie–Weiss model CW$(\rho)$ at inverse temperature $\beta$ is defined as the array of spin random variables ${\mathbf{X}}=(X_{1},X_{2},\ldots,X_{n})$ with joint distribution $$d\nu_{n}(\mathbf{x})=Z_{n}^{-1}\exp\biggl(\frac{\beta}{2n}
(x_{1}+x_{2}+\cdots
+x_{n} )^{2} \biggr)\prod_{i=1}^{n}d\rho(x_{i})$$ for $\mathbf{x}=(x_{1},x_{2},\ldots,x_{n})\in{\mathbb{R}}^{n}$ where $$Z_{n}=\int\exp\biggl(\frac{\beta}{2n}(x_{1}+x_{2}+\cdots
+x_{n})^{2}
\biggr)\prod_{i=1}^{n}d\rho(x_{i})$$ is the normalizing constant. The magnetization $m(\mathbf{x})$ is defined as usual by $m(\mathbf{x})=n^{-1}\sum_{i=1}^{n}x_{i}$. Here, we will consider the case when $\rho$ satisfies the following two conditions:
(A) $\rho$ has compact support, that is, $\rho([-L,L])=1$ for some $L<\infty$.
(B) The equation $h'(s)=0$ has a unique root at $s=0$ where $$h(s):=\frac{s^{2}}{2}-\log\int\exp(sx) \,d\rho(x)\qquad \mbox{for }
s\in{\mathbb{R}}.$$
The second condition says that $h(\cdot)$ has a unique global minima at $s=0$ and $|h'(s)|>0$ for $|s|>0$. The behavior of this model is quite similar to the classical Curie–Weiss model and there is a phase transition at $\beta=1$. For $\beta
<1$, $m({\mathbf{X}})$ is concentrated around zero while for $\beta>1, m({\mathbf{X}})$ is bounded away from zero a.s. (see Ellis and Newman [@ellis78; @ellis78a]). We will prove the following concentration result.
\[pr:critgen\] Suppose ${\mathbf{X}}\sim\nu_{n}$ at the critical temperature $\beta=1$ where $\rho
$ satisfies conditions and . Let $k$ be such that $h^{(i)}(0)=0$ for $0\le i<2k$ and $h^{(2k)}(0)\neq0$, where $$h(s):=\frac{s^{2}}{2}-\log\int\exp(sx) \,d\rho(x) \qquad\mbox{for }
s\in{\mathbb{R}},$$ and $h^{(i)}$ is the $i$th derivative of $h$. Then, $k>1$ and for any $n\ge1$ and $ t\ge0$ the magnetization satisfies $${\mathbb{P}}\bigl(n^{1/2k}|m({\mathbf{X}})| \ge t\bigr) \le2e^{-ct^{2k}},$$ where $c>0$ is an absolute constant depending only on $\rho$.
Here, we mention that in Ellis and Newman [@ellis78], convergence results were proved for the magnetization in CW$(\rho)$ model under optimal condition on $\rho$. Under our assumption, their result says that $n^{1/2k}m({\mathbf{X}})$ converges weakly to a distribution having density proportional to $\exp(-\lambda x^{2k}/(2k)!)$ where $\lambda:=h^{(2k)}(0)$. Hence, the tail bound gives the correct convergence rate.
Let us now give a brief sketch of the proof of Proposition \[pr:crit\]. Suppose ${\bolds{\sigma}}$ is drawn from the Curie–Weiss model at the critical temperature. We construct ${\bolds{\sigma}}'$ by taking one step in the heat-bath Glauber dynamics: a coordinate $I$ is chosen uniformly at random, and $\sigma_{I}$ is replace by $\sigma'_{I}$ drawn from the conditional distribution of the $I$th coordinate given $\{\sigma_{j}\dvtx j\neq I\}$. Let $$F({\bolds{\sigma}},{\bolds{\sigma}}') := \sum_{i=1}^{n}(\sigma_{i}-\sigma'_{i})=\sigma
_{I}-\sigma'_{I}.$$ For each $i=1,2,\ldots,n$, define $
m_{i}=m_{i}({\bolds{\sigma}}) = n^{-1}\sum_{j\neq i} \sigma_{j}.
$ An easy computaion gives that ${\mathbb{E}}(\sigma_{i}| \{\sigma_{j}, j\neq i\}
) =
\tanh( m_{i})$ for all $i$ and so we have $$f({\bolds{\sigma}}):= {\mathbb{E}}(F({\bolds{\sigma}},{\bolds{\sigma}}')|{\bolds{\sigma}}) = m - \frac{1}{n} \sum_{i=1}^{n}
\tanh(m_{i})=\frac{m}{n} + \frac{1}{n} \sum_{i=1}^{n} g(m_{i}),$$ where $g(x):=x-\tanh(x)$. Note that $|m_i - m|\le1/n$, and hence $f({\bolds{\sigma}}) = m - \tanh m + O(1/n)$. A simple analytical argument using the fact that, for $x\approx0$, $x- \tanh x = x^3/3 + O(x^5)$ then gives $$\Delta({\bolds{\sigma}}) \le\frac{6}{n}|f({\bolds{\sigma}})|^{2/3}+\frac{12}{n^{5/3}}$$ and using Corollary \[cor:nong\] with $\alpha=2/3, B= 6/n$ and $C=12/n^{5/3}$ we have $${\mathbb{P}}(|m-\tanh m| \ge t+n^{-1}) \le{\mathbb{P}}\bigl(|f({\bolds{\sigma}})| \ge t\bigr) \le2e^{- cnt^{4/3}}$$ for all $t\ge0$ for some constant $c>0$. It is easy to see that this implies the result. The critical observation, of course, is that $x -
\tanh(\beta x) = O(x^3)$ for $\beta= 1$, which is not true for $\beta
\ne1$.
Example: Triangles in Erdős–Rényi graphs
----------------------------------------
Consider the Erdős–Rényi random graph model $G(n,p)$ which is defined as follows. The vertex set is $[n]:=\{1,2,\ldots,n\}$ and each edge $(i,j), 1\le i< j\le n$, is present with probability $p$ and not present with probability $1-p$ independently of each other. For any three distinct vertex $i<j<k$ in $[n]$ we say that the triple $(i,j,k)$ forms a triangle in the graph $G(n,p)$ if all the three edges $(i,j),(j,k),(i,k)$ are present in $G(n,p)$ (see Figure \[fig:Tn\]). Let $T_{n}$ be the number of triangles in $G(n,p)$, that is, $$T_{n}:=\sum_{1\le i<j<k\le n}{\mathbf{1}}\{(i,j,k)\mbox{ forms a triangle
in }
G(n,p)\}.$$
![A graph on $6$ vertices with $8$ edges and $3$ triangles. The triangles being $(1,2,3), (1,3,4)$ and $(1,3,6)$.[]{data-label="fig:Tn"}](542f01.eps)
Let us define the function $I(\cdot,\cdot)$ on $(0,1)\times(0,1)$ as $$\label{kl}
I(r,s) := r\log\frac{r}{s} + (1-r)\log\frac{1-r}{1-s}.$$ Note that $I(r,s)$ is the Kullback–Leibler divergence of the measure $\nu_{s}$ from $\nu_{r}$ and also the relative entropy of $\nu_{r}$ w.r.t. $\nu_{s}$ where $\nu_{p}$ is the Bernoulli$(p)$ measure. We have the following result about the large deviation rate function for the number of triangles in $G(n,p)$.
\[largedev\] Let $T_n$ be the number of triangles in $G(n,p)$, where $p > p_{0}$ where $p_{0}=2/(2+e^{3/2})\approx0.31$. Then for any $r\in(p,1]$, $$\label{eq:trtail}
{\mathbb{P}}\left(T_n \ge\pmatrix{n \cr3} r^3 \right) = \exp\biggl(-\frac{n^2
I(r,p)}{2}\bigl(1+O(n^{-1/2})\bigr) \biggr).$$ Moreover, even if $p \le p_{0}$, there exist $p',p''$ such that $p<p'\le p''<1$ and the same result holds for all $r\in(p,p')\cup(p'',1]$. For all $p$ and $r$ in the above domains, we also have the more precise estimate $$\label{eq:trbound}\quad
{\mathbb{P}}\left( \left|T_n- \pmatrix{n \cr3} r^3 \right| \le
C(p,r)n^{5/2} \right)
= \exp\biggl(-\frac{n^2 I(r,p)}{2}\bigl(1+O(n^{-1/2})\bigr) \biggr),$$ where $C(p,r)$ is a constant depending on $p$ and $r$.
The behavior of the upper tail of subgraph counts in $G(n,p)$ is a problem of great interest in the theory of random graphs (see [@bollobas85; @jansonetal00; @jansonrucinski02; @vu01; @kimvu04], and references contained therein). The best upper bounds to date were obtained by Kim and Vu [@kimvu04] (triangles) and Janson, Oleszkiewicz, and Ruci' nski [@jansonolesz04] (general subgraph counts). For triangles, the results of these papers essentially state that for a fixed $\epsilon>0$, $$\exp\bigl(-\Theta\bigl(n^2 p^2\log(1/p)\bigr)\bigr) \le{\mathbb{P}}\bigl(T_n \ge{\mathbb{E}}(T_n) + \epsilon n^3
p^3\bigr) \le\exp(-\Theta(n^2 p^2)).$$ Clearly, our result gives a lot more in the situations where it works (see Figure \[fig:pr\]). The method of proof can be easily extended to prove similar results for general subgraph counts and are discussed in Section \[subsec:small\]. However, there is an obvious incompleteness in Theorem \[largedev\] (and also for general subgraphs counts), namely, that it does not work for all $(p,r)$.
![The set (colored in gray) of $(p,r), r\ge p$, for which we are able to show that the large deviation result holds.[]{data-label="fig:pr"}](542f02.eps)
In this context, we should mention that another paper on large deviations for subgraph counts by Bolthausen, Comets and Dembo [@bolthausenetal09] is in preparation. As of now, to the best of our knowledge, the authors of [@bolthausenetal09] have only looked at subgraphs that do not complete loops, like $2$-stars. Another related article is the one by Döring and Eichelsbacher [@doringeichelsbacher09], who obtain moderate deviations for a class of graph-related objects, including triangles.
Unlike the previous two examples, Theorem \[largedev\] is far from being a direct consequence of any of our abstract results. Therefore, let us give a sketch of the proof, which involves a new idea.
The first step is standard: consider tilted measures. However, the appropriate tilted measure in this case leads to what is known as an “exponential random graph,” a little studied object in the rigorous literature. Exponential random graphs have become popular in the statistical physics and network communities in recent years (see the survey of Park and Newman [@parknewman04]). The only rigorous work we are aware of is the recent paper of Bhamidi et al. [@bhamidi08], who look at convergence rates of Markov chains that generate such graphs.
We will not go into the general definition or properties of exponential random graphs. Let us only define the model we need for our purpose.
Fix two numbers $\beta\ge0$ and $h\in{\mathbb{R}}$. Let $\Omega=\{0,1\}
^{n\choose2}$ be the space of all tuples like $\mathbf
{x}=(x_{ij})_{1\le
i<j\le n}$, where $x_{ij}\in\{0,1\}$ for each $i,j$. Let ${\mathbf{X}}=
(X_{ij})_{1\le i<j\le n}$ be a random element of $\Omega$ following the probability measure proportional to $e^{H(\mathbf{x})}$, where $H$ is the Hamiltonian: $$H(\mathbf{x}) = \frac{\beta}{n}\sum_{1\le i<j <k\le n}
x_{ij}x_{jk} x_{ik}
+ h\sum_{1\le i<j\le n} x_{ij}.$$ Note that any element of $\Omega$ naturally defines an undirected graph on a set of $n$ vertices. For each $\mathbf{x}\in\Omega$, let $T(\mathbf{x})=\sum
_{i<j<k}x_{ij}x_{jk}x_{ik}$ denote the number of triangles in the graph defined by $\mathbf{x}$, and let $E(\mathbf{x})=\sum_{i<j}x_{ij}$ denote the number of edges. Then the above Hamiltonian is nothing but $$\frac{\beta T(\mathbf{x})}{n}+ hE(\mathbf{x}).$$ For notational convenience, we will assume that $x_{ij}=x_{ji}$. Let $Z_n(\beta,h)$ be the corresponding partition function, that is, $$Z_n(\beta,h) = \sum_{\mathbf{x}\in\Omega} e^{H(\mathbf{x})}.$$ Note that $\beta=0$ corresponds to the Erdős–Rényi random graph with $p=e^{h}/(1+e^{h})$. The following theorem “solves” this model in a “high temperature region.” Once this solution is known, the computation of the large deviation rate function is just one step away.
\[pressure\] Suppose we have $\beta\ge0$, $h\in{\mathbb{R}}$, and $Z_n(\beta,h)$ defined as above. Define a function $\varphi\dvtx[0,1]\to{\mathbb{R}}$ as $$\varphi(x) = \frac{e^{\beta x + h}}{1+e^{\beta x + h}}.$$ Suppose $\beta$ and $h$ are such that the equation $u = \varphi(u)^2$ has a unique solution $u^*$ in $[0,1]$ and $2\varphi(u^{*})\varphi
'(u^{*})<1$. Then $$\lim_{n\to\infty}\frac{\log Z_n(\beta,h)}{n^2} = -\frac
{1}{2}I(\varphi
(u^*),\varphi(0)) - \frac{1}{2} \log\bigl(1-\varphi(0)\bigr) +\frac{\beta
\varphi
(u^*)^3}{6},$$ where $I(\cdot,\cdot)$ is the function defined in (\[kl\]). Moreover, there exists a constant $K(\beta,h)$ that depends only on $\beta$ and $h$ (and not on $n$) such that difference between $n^{-2}\log Z_n(\beta
,h)$ and the limit is bounded by $K(\beta,h)n^{-1/2}$ for all $n$.
Incidentally, the above solution was obtained using physical heuristics by Park and Newman [@parknewman05] in 2005. Here, we mention that, in fact, the following result is always true.
\[lem:lbd\] For any $\beta\ge0, h\in{\mathbb{R}}$, we have $$\begin{aligned}
\label{eq:lbd}
&&\liminf_{n\to\infty}\frac{\log Z_n(\beta,h)}{n^2} \nonumber\\
&&\qquad\ge \sup_{r\in
(0,1)} \biggl\{-\frac{1}{2}I(r,\varphi(0)) - \frac{1}{2} \log
\bigl(1-\varphi
(0)\bigr) +\frac{\beta r^3}{6} \biggr\}\\
&&\qquad= \sup_{u\dvtx\varphi(u)^{2}=u} \biggl\{-\frac{1}{2}I(\varphi
(u),\varphi
(0)) - \frac{1}{2} \log\bigl(1-\varphi(0)\bigr) +\frac{\beta\varphi
(u)^3}{6} \biggr\}.\nonumber\end{aligned}$$
We will characterize the set of $\beta,h$ for which the conditions in Theorem \[pressure\] hold in Lemma \[lem:pcond\]. First of all, note that the appearance of the function $\varphi(u)^{2}-u$ is not magical. For each $i < j$, define $$L_{ij} = \frac{1}{n}\sum_{k\notin\{i,j\}} X_{ik} X_{jk}.$$ This is the number of “wedges” or $2$-stars in the graph that have the edge $ij$ as base. The key idea is to use Theorem \[thm:conc\] to show that these quantities approximately satisfy the following set of “mean field equations”: $$\label{meanfieldeqs}
L_{ij} \simeq\frac{1}{n}\sum_{k\notin\{i,j\}} \varphi
(L_{ik})\varphi
(L_{jk}) \qquad\mbox{for all } i<j.$$ (The idea of using Theorem \[thm:conc\] to prove mean field equations was initially developed in Section 3.4 of [@chatterjee05].) The following lemma makes this notion precise. Later, we will show that under the conditions of Theorem \[pressure\], this system has a unique solution.
\[lem:mf\] Let $\varphi$ be defined as in Theorem \[pressure\]. Then for any $1\le i<j\le n$, we have $${\mathbb{P}}\biggl(
\sqrt{n} \biggl|L_{ij} - \frac{1}{n}\sum_{k\notin\{i,j\}}
\varphi
(L_{ik})\varphi(L_{jk}) \biggr|
\ge t \biggr) \le2\exp\biggl(- \frac{t^{2}}{8(1+\beta)} \biggr)$$ for all $t\ge8\beta/n$. In particular, we have $$\label{eq:lcon}
{\mathbb{E}}\biggl|L_{ij} - \frac{1}{n}\sum_{k\notin\{i,j\}} \varphi
(L_{ik})\varphi(L_{jk}) \biggr| \le\frac{C(1+\beta)^{1/2}}{n^{1/2}},$$ where $C$ is a universal constant.
In fact, one would expect that $L_{ij}\simeq u^{*}$ for all $i<j$, if the equation $$\label{eq:sol}
\psi(u):=\varphi(u)^{2}-u=0$$ has a unique solution $u^{*}$ in $[0,1]$. The intuition behind is as follows. Define $L_{\max} = \max_{i,j} L_{ij}$ and $L_{\min}=\min
_{i,j} L_{ij}$. It is easy to see that $\varphi$ is an increasing function. Hence, from the mean-field equations (\[meanfieldeqs\]), we have $L_{\max}\le\varphi(L_{\max})^{2} + o(1)$ or $\psi(L_{\max})
\ge o(1)$. But $\psi(u)\ge0$ iff $u\le u^{*}$. Hence, $L_{\max}\le
u^{*}+o(1)$. Similarly, we have $L_{\min}\ge u^{*}-o(1)$ and thus all $L_{ij}\simeq u^{*}$. Lemma \[mean2\] formalizes this idea. Here, we mention that one can easily check that equation (\[eq:sol\]) has at most three solutions. Moreover, $\psi(0)>0>\psi(1)$ implies that $\psi
'(u^{*})\le0$ or $2\varphi(u^{*})\varphi'(u^{*})\le1$ if $u^{*}$ is the unique solution to (\[eq:sol\]).
\[mean2\] Let $u^*$ be the unique solution of the equation $u=\varphi(u)^{2}$. Assume that $2\varphi(u^{*})\varphi'(u^{*})<1$. Then for each $1\le
i<j\le n$, we have $${\mathbb{E}}|L_{ij} - u^*| \le\frac{K(\beta,h)}{n^{1/2}},$$ where $K(\beta,h)$ is a constant depending only on $\beta,h$. Moreover, if $2\varphi(u^{*})\varphi'(u^{*}) =1$ then we have $${\mathbb{E}}|L_{ij} - u^*| \le\frac{K(\beta,h)}{n^{1/6}} \qquad\mbox{for all
}1\le
i<j\le n.$$
Now observe that the Hamiltonian $H({\mathbf{X}})$ can be written as $$H({\mathbf{X}}) = \frac{\beta}{6} \sum_{1\le i<j\le n} X_{ij} L_{ij} + h\sum
_{1\le i<j\le n} X_{ij}.$$ The idea then is the following: once we know that the conclusion of Lemma \[mean2\] holds, each $L_{ij}$ in the above Hamiltonian can be replaced by $u^*$, which results in a model where the coordinates are independent. The resulting probability measure is presumably quite different from the original measure, but somehow the partition functions remain comparable.
The following lemma (Lemma \[lem:pcond\]) characterizes the region $S\in{\mathbb{R}}\times[0,\infty)$ such that the equation $u = \varphi(u)^2$ has a unique solution $u^{*}$ in $[0,1]$ and $2\varphi(u^{*})\varphi
'(u^{*})< 1$ for $(h,\beta)\in S$ (see Figure \[fig:unisol\]).
![The set $S$ (colored in gray) of $(h,\beta)$ for which the conditions of Theorem \[pressure\] hold.[]{data-label="fig:unisol"}](542f03.eps)
Let $h_{0}=\log2-\frac{3}{2}<0$. For $h<h_{0}$ there exist exactly two solutions $0<a_{*}=a_{*}(h)<1/2<a^{*}=a^{*}(h)<\infty$ to the equation $$\log x + \frac{1+x}{2x}+h=0.$$ Define $a_{*}(h)=a^{*}(h)=1/2$ for $h=h_{0}$ and $$\label{eq:bstar}
\beta_{*}(h)= \frac{(1+a_{*})^{3}}{2a_{*}} \quad\mbox{and}\quad \beta^{*}(h)=
\frac
{(1+a^{*})^{3}}{2a^{*}}$$ for $h\le h_{0}$.
\[lem:pcond\] Let $S$ be the set of pairs $(h,\beta)$ for which the function $\psi
(u):=\varphi(u)^{2}-u$ has a unique root $u^{*}$ in $[0,1]$ and $2\varphi(u^{*})\varphi'(u^{*})< 1$ where $
\varphi(u):= e^{\beta u +h}/(1+e^{\beta u +h}).
$ Then we have $$S^{c}=\{(h,\beta)\dvtx h\le h_{0} \mbox{ and } \beta_{*}(h) \le\beta
\le\beta
^{*}(h)\},$$ where $\beta^{*},\beta_{*}$ are as given in equation (\[eq:bstar\]). In particular, $(h,\beta)\in S$ if $\beta\le(3/2)^{3}$ or $h>h_{0}$.
The point $h=h_{0}, \beta=\beta_{0}:=(3/2)^{3}$ is the critical point and the curve $$\label{def:gc}
\gamma(t)= \biggl( -\log t - \frac{1+t}{2t}, \frac
{(1+t)^{3}}{2t} \biggr)$$ for $t>0$ is the phase transition curve. It corresponds to $\psi
(u^{*})=0$ and $ 2\psi(u^{*})\psi'(u^{*})= 1$. In fact, at the critical point $(h_{0},\beta_{0})$ the function $\psi(u)=\varphi(u)^{2}-u$ has a unique root of order three at $u^{*}=4/9$, that is, $\psi(u^{*})=\psi
'(u^{*})=\psi''(u^{*})=0$ and $\psi'''(u^{*})< 0$. The second part of Lemma \[mean2\] shows that all the above conclusions (including the limiting free energy result) are true for the critical point but with an error rate of $n^{-1/6}$. Define the “energy” function $$e(r)=\frac{1}{2}I(r,\varphi(0)) + \frac{1}{2} \log\bigl(1-\varphi(0)\bigr)
-\frac
{\beta r^3}{6}$$ appearing in of the r.h.s. of equation (\[eq:lbd\]). The “high temperature” regime corresponds to the case when $e(\cdot)$ has a unique minima and no local maxima or saddle point. The critical point corresponds to the case when $e(\cdot)$ has a nonquadratic global minima. The boundary corresponds to the case when $e(\cdot)$ has a unique minima and a saddle point. In the “low temperature” regime, $e(\cdot)$ has two local minima. In fact, one can easily check that there is a one-dimensional curve inside the set $S^{c}$, starting from the critical point, on which $e(\cdot)$ has two global minima and outside one global minima. Below, we provide the solution on the boundary curve. Unfortunately, as of now, we don not have a rigorous solution in the “low temperature” regime.
For $(h,\beta)$ on the phase transition boundary curve (excluding the critical point), the function $\psi(\cdot)$ has two roots and one of them, say $v^{*}$, is an inflection point. Let $u^{*}$ be the other root. Here, we mention that $u^{*}$ is a minima of $e(\cdot)$ while $v^{*}$ is a saddle point of $e(\cdot)$. On the lower part of the boundary, which corresponds to $\{\gamma(t)\dvtx t<1/2\}$, the inflection point $v^{*}=(1+t)^{-2}$ is larger than $u^{*}$, while on the upper part of the boundary corresponding to $\{\gamma(t)\dvtx t>1/2\}$, the inflection point $v^{*}=(1+t)^{-2}$ is smaller than $u^{*}$. The following lemma “solves” the model at the boundary point $\gamma(t)$ \[see (\[def:gc\])\].
\[lem:boundary\] Let $\gamma(\cdot),u^{*},v^{*}$ be as above and $(h,\beta)=\gamma(t)$ for some $t\neq1/2$. Then, for each $1\le i<j\le n$, we have $${\mathbb{E}}(|L_{ij}-u^{*}|)\le\frac{K(\beta,h)}{n^{1/2}}$$ for some constant $K(\beta,h)$ depending on $\beta,h$. Moreover, we have $$\frac{\log Z_n(\beta,h)}{n^2} = -\frac{1}{2}I(\varphi(u^*),\varphi(0))
- \frac{1}{2} \log\bigl(1-\varphi(0)\bigr) +\frac{\beta\varphi
(u^*)^3}{6}+O(n^{-1/2})$$ and $$\begin{aligned}
&&{\mathbb{P}}\left( \left|T_n({\mathbf{Y}})- \pmatrix{n \cr3} \varphi(u^*)^3 \right|
\le
C(\beta,h)n^{5/2} \right)\nonumber\\[-8pt]\\[-8pt]
&&\qquad = \exp\biggl(-\frac{n^2 I(\varphi(u^*),\varphi
(0))}{2}\bigl(1+O(n^{-1/2})\bigr) \biggr),\nonumber\end{aligned}$$ where ${\mathbf{Y}}=((Y_{ij}))_{i<j}$ follows G$(n,\varphi(0))$ and the constant appearing in $O(\cdot)$ and $C(\beta,h)$ depend only on $\beta,h$.
In the next subsection, we will briefly discuss about the results for general subgraph counts that can be proved using similar ideas.
Example: General subgraph counts {#subsec:small}
--------------------------------
Let $F=(V(F), E(F))$ be a fixed finite graph on ${\mathbf{v}}_{F} :=|V(F)|$ many vertices with ${\mathbf{e}}_{F} :=|E(F)|$ many edges. Without loss of generality, we will assume that $V(F)=[{\mathbf{v}}_{F}]:=\{1,2,\ldots,{\mathbf{v}}_{F}\}$. Let $\alpha_{F}=|{\operatorname{Aut}}(F)|$ be the number of graph automorphism of the graph $F$. Let $N_{n}$ be the number of copies of $F$, not necessarily induced, in the Erdős–Rényi random graph $G(n,p)$ (so the number of $2$-stars in a triangle will be three). We have the following result about the large deviation rate function for the random variable $N_{n}$.
\[largedevgen\] Let $N_n$ be the number of copies of $F$ in $G(n,p)$, where $$p > p_{0}:=\frac{{\mathbf{e}}_{ F}-1}{{\mathbf{e}}_{F}-1+\exp({{\mathbf{e}}_{F}}/({{\mathbf{e}}_{F}-1}) )}.$$ Then for any $r\in(p,1]$, $${\mathbb{P}}\left(N_n \ge\frac{{\mathbf{v}}_{F}!}{\alpha_{F}}\pmatrix{n \cr{\mathbf{v}}_{F}}
r^{{\mathbf{e}}_{F}} \right) = \exp\biggl(-\frac{n^2
I(r,p)}{2}\bigl(1+O(n^{-1/2})\bigr) \biggr).$$ Moreover, even if $p \le p_{0}$, there exist $p',p''$ such that $p<p'\le p''<1$ and the same result holds for all $r\in(p,p')\cup(p'',1]$. For all $p$ and $r$ in the above domains, we also have the more precise estimate $$\begin{aligned}
&&{\mathbb{P}}\biggl( \biggl| N_n - \frac{{\mathbf{v}}_{F}!}{\alpha_{F}}\pmatrix{n \cr{\mathbf{v}}_{F}}
r^{{\mathbf{e}}_{F}} \biggr| \le C(p,r)n^{{\mathbf{v}}_{F}-1/2} \biggr)\\
&&\qquad = \exp\biggl(-\frac{n^2 I(r,p)}{2}\bigl(1+O(n^{-1/2})\bigr) \biggr),\end{aligned}$$ where $C(p,r)$ is a constant depending on $p$ and $r$.
Note that $p_{0}$ as a function of ${\mathbf{e}}_{F}$ is increasing and converges to $1$ as number of edges goes to infinity (see Figure \[fig:p0\]). So there is an obvious gap in the large deviation result, namely the proof does not work when $r\ge p, p\le p_{0}$ and the gap becomes larger as the number of edges in $F$ increases. Note that $p_{0}\to1$ as ${\mathbf{e}}_{F}\to\infty$.
![The curve $p_{0}$ vs. ${\mathbf{e}}_{F}$ where for a graph $F$ with ${\mathbf{e}}_{F}$ many edges our large deviation result holds when $p>p_{0}$.[]{data-label="fig:p0"}](542f04.eps)
The proof of Theorem \[largedevgen\] uses the same arguments that were used in the triangle case. Here, the tilted measure leads to an exponential random graph model where the Hamiltonian depends on number of copies of $F$ in the random graph. Let $\beta\ge0, h\in{\mathbb{R}}$ be two fixed numbers. As before, we will identify elements of $\Omega:=\{0,1\}
^{n\choose2}$ with undirected graphs on a set of $n$ vertices. For each $\mathbf{x}\in\Omega$, let $N(\mathbf{x})$ denote the number of copies of $F$ in the graph defined by $\mathbf{x}$, and let $E(\mathbf{x})=\sum
_{i<j}x_{ij}$ denote the number of edges. Let ${\mathbf{X}}= (X_{ij})_{1\le i<j\le n}$ be a random element of $\Omega$ following the probability measure proportional to $e^{H(\mathbf{x})}$, where $H$ is the Hamiltonian $$H(\mathbf{x})=\frac{\beta}{(n-2)_{{\mathbf{v}}_{F}-2}}N(\mathbf{x})+
hE(\mathbf{x}),$$ where $(n)_{m}=\frac{n!}{(n-m)!}$. Recall that ${\mathbf{v}}_{F}$ is the number of vertices in the graph $F$. The scaling was done to make the two summands comparable. Also we used $(n-2)_{{\mathbf{v}}_{F}-2}$ instead of $n^{{\mathbf{v}}_{F}}$ to make calculations simpler. Let $Z_{n}(\beta,h)$ be the partition function. Note that $N(\mathbf{x})$ can be written as $$N(\mathbf{x}) = \frac{1}{\alpha_{F}}\mathop{\sum_{1\le
t_{1},t_{2},\ldots,t_{{\mathbf{v}}_{F}}\le n,}}_{t_{i}\neq t_{j}\ \mathrm{for}\
i\neq j} \prod_{(i,j)\in E(F)}x_{t_{i}t_{j}}.$$ For $\mathbf{x}\in\Omega, 1\le i<j\le n$, define $\mathbf
{x}_{(i,j)}^{1}$ as the element of $\Omega$ which is same as $\mathbf{x}$ in every coordinate except for the $(i,j)$th coordinate where the value is $1$. Similarly, define $\mathbf{x}
_{(i,j)}^{0}$. For $i<j$, define the random variable $$L_{ij}:= \frac{N({\mathbf{X}}_{(i,j)}^{1})-N({\mathbf{X}}_{(i,j)}^{0})}{(n-2)_{{\mathbf{v}}_{F}-2}}.$$ The main idea is as in the triangle case. We show that $L_{ij}$’s satisfy a system of “mean-field equations” similar to (\[meanfieldeqs\]) which has a unique solution under the condition of Theorem \[pressuregen\]. In fact, we will show that $L_{ij}$$u^{*}$ for all $i<j$ and $E({\mathbf{X}})$${n\choose2}\varphi(u^{*})$ under the condition of Theorem \[pressuregen\]. Now note that we can write the Hamiltonian as $$H({\mathbf{X}})= \frac{\beta}{{\mathbf{e}}_{F}}\sum_{i<j}X_{ij}L_{ij}+h\sum_{i<j}X_{ij},$$ which is approximately equal to $ h^{*} E({\mathbf{X}})$ where $h^{*}=h+\beta
u^{*}/{\mathbf{e}}_{F}$. Now the remaining is a calculus exercise.
So the first step in proving the large deviation bound is the following theorem, which gives the limiting free energy in the “high temperature” regime. Note the similarity with the triangle case.
\[pressuregen\] Suppose we have $\beta\ge0$, $h\in{\mathbb{R}}$, and $Z_n(\beta,h)$ defined as above. Define a function $\varphi\dvtx[0,1]\to{\mathbb{R}}$ as $$\varphi(x) = \frac{e^{\beta x + h}}{1+e^{\beta x + h}}.$$ Suppose $\beta$ and $h$ are such that the equation $\alpha_{F} u =
2{\mathbf{e}}_{F}\varphi(u)^{{\mathbf{e}}_{F}-1}$ has a unique solution $u^*$ in $[0,1]$ and $2{\mathbf{e}}_{F}({\mathbf{e}}_{F}-1)\varphi(u^{*})^{{\mathbf{e}}_{F}-2}\varphi
'(u^{*})<\alpha
_{F}$. Then $$\lim_{n\to\infty}\frac{\log Z_n(\beta,h)}{n^2} = -\frac
{1}{2}I(\varphi
(u^*),\varphi(0)) - \frac{1}{2} \log\bigl(1-\varphi(0)\bigr) +\frac{\beta
\varphi
(u^*)^{{\mathbf{e}}_{F}}}{\alpha_{F}},$$ where $I(\cdot,\cdot)$ is the function defined in (\[kl\]). Moreover, there exists a constant $K(\beta,h)$ that depends only on $\beta$ and $h$ (and not on $n$) such that difference between $n^{-2}\log Z_n(\beta
,h)$ and the limit is bounded by $K(\beta,h)n^{-1/2}$ for all $n$.
Here also we can identify the region where the conditions in Theorem \[pressuregen\] hold. Let $$\label{eq:h0}
h_{0}=\log({\mathbf{e}}_{F}-1)-\frac{{\mathbf{e}}_{F}}{{\mathbf{e}}_{F}-1}.$$ For $h<h_{0}$, there exist exactly two solutions $0<a_{*}=a_{*}(h)<1/2<a^{*}=a^{*}(h)<\infty$ of the equation $$\log x + \frac{1+x}{({\mathbf{e}}_{F}-1)x}+h=0.$$ Define $a_{*}(h)=a^{*}(h)=1/({\mathbf{e}}_{F}-1)$ for $h=h_{0}$ and $$\label{eq:beta}
\beta_{*}(h)= \frac{\alpha_{F}(1+a_{*})^{{\mathbf{e}}_{F}}}{2{\mathbf{e}}_{F}({\mathbf{e}}_{F}-1)a_{*}} \quad\mbox{and}\quad \beta^{*}(h)= \frac{\alpha
_{F}(1+a^{*})^{{\mathbf{e}}_{F}}}{2{\mathbf{e}}_{F}({\mathbf{e}}_{F}-1)a^{*}}$$ for $h\le h_{0}$.
\[lem:pcondgen\] Let $S$ be the set of pairs $(h,\beta)$ for which the function $$\psi(u):=2{\mathbf{e}}_{F}\varphi(u)^{{\mathbf{e}}_{F}-1}-\alpha_{F} u$$ has a unique root $u^{*}$ in $[0,1]$ and $2{\mathbf{e}}_{F}({\mathbf{e}}_{F}-1)\varphi
(u^{*})^{{\mathbf{e}}_{F}-2}\varphi'(u^{*})<\alpha_{F}$ where $
\varphi(u):= e^{\beta u +h}/(1+e^{\beta u +h}).
$ Then we have $$S^{c}=\{(h,\beta)\dvtx h\le h_{0} \mbox{ and } \beta_{*}(h) \le\beta
\le\beta^{*}(h)\},$$ where $h_{0}, \beta^{*},\beta_{*}$ are as given in (\[eq:h0\]), (\[eq:beta\]). In particular, $(h,\beta)\in S$ if $$\beta\le\frac{\alpha_{F}{\mathbf{e}}_{F}^{{\mathbf{e}}_{F}-1}}{2({\mathbf{e}}_{F}-1)^{{\mathbf{e}}_{F}}}
\quad\mbox{or}\quad h>h_{0}.$$
In fact, Lemma \[lem:pcondgen\] identifies the critical point and the phase transition curve where the model goes from ordered phase to a disordered phase. But the results above does not say what happens at the boundary or in the low temperature regime. However, note that the mean-field equations hold for all values of $\beta$ and $h$.
Example: Ising model on ${\mathbb{Z}}^{d}$
------------------------------------------
Fix any $\beta\ge0, h\in{\mathbb{R}}$ and an integer $d\ge1$. Also fix $n\ge2$. Let ${\mathbb{B}}= \{1,2,\ldots,n+1\}^{d}$ be a hypercube with $(n+1)^{d}$ many points in the $d$-dimensional hypercube lattice ${\mathbb{Z}}^{d}$. Let $\Omega
$ be the graph obtained from ${\mathbb{B}}$ by identifying the opposite boundary points, that is, for $x=(x_{1},x_{2},\ldots,x_{d}),
y=(y_{1},y_{2},\ldots,y_{d})\in{\mathbb{B}}$ we have $x$ is identified with $y$ if $x_{i}-y_{i}\in\{-n,0,n\}$ for all $i$. This identification is known in the literature as periodic boundary condition. Note that $\Omega$ is the $d$-dimensional lattice torus with linear size $n$. We will write $x\sim y$ for $x,y\in\Omega$ if $x,y$ are nearest neighbors in $\Omega$. Also, let us denote by $N_{x}$ the set of nearest neighbors of $x$ in $\Omega$, that is, $N_{x}=\{y\in\Omega\dvtx y\sim x\}$.
Now, consider the Gibbs measure on $\{+1,-1\}^{\Omega}$ given by the following Hamiltonian $$H({\bolds{\sigma}}):= \beta\sum_{x\sim y, x,y\in\Omega} \sigma_{x}\sigma
_{y} + h\sum
_{x\in\Omega
} \sigma_{x},$$ where ${\bolds{\sigma}}=(\sigma_x)_{x\in\Omega}$ is a typical element of $\{
+1,-1\}
^{\Omega
}$. So the probability of a configuration ${\bolds{\sigma}}\in\{+1,-1\}^{\Omega
}$ is $$\label{eq:gibbs}
\mu_{\beta,h}(\{{\bolds{\sigma}}\})
:= Z_{\beta,h}^{-1}\exp( H({\bolds{\sigma}}) )
=Z_{\beta,h}^{-1}\exp\biggl(\beta\sum_{x\sim y, x,y\in\Omega} \sigma
_{x}\sigma_{y} +
h\sum_{x\in\Omega} \sigma_{x}\biggr),\hspace*{-28pt}$$ where $Z_{\beta,h}=\sum_{{\bolds{\sigma}}\in\{+1,-1\}^{\Omega}}e^{H({\bolds{\sigma}})}$ is the normalizing constant. Here $\sigma_{x}$ is the spin of the magnetic particle at position $x$ in the discrete torus $\Omega$. This is the famous Ising model of ferromagnetism on the box ${\mathbb{B}}$ with periodic boundary condition at inverse temperature $\beta$ and external field $h$.
The one-dimensional Ising model is probably the first statistical model of ferromagnetism to be proposed or analyzed [@ising25]. The model exhibits no phase transition in one dimension. But for dimensions two and above the Ising ferromagnet undergoes a transition from an ordered to a disordered phase as $\beta$ crosses a critical value. The two-dimensional Ising model with no external field was first solved by Lars Onsager in a ground breaking paper [@onsager44], who also calculated the critical $\beta$ as $\beta_{c}=\sinh^{-1}(1)$. For dimensions three and above the model is yet to be solved, and indeed, very few rigorous results are known.
In this subsection, we present some concentration inequalities for the Ising model that hold for all values of $\beta$. These “temperature-free” relations are analogous to the mean field equations that we obtained for subgraph counts earlier.
The magnetization of the system, as a function of the configuration ${\bolds{\sigma}}$, is defined as $m({\bolds{\sigma}}):= \frac{1}{|\Omega|}\sum_{x\in
\Omega}
\sigma_{x}$. For each integer $k\in\{1,2,\ldots,2d\}$, define a degree $k$ polynomial function $r_{k}({\bolds{\sigma}})$ of a spin configuration ${\bolds{\sigma}}$ as follows: $$\label{def:rk}
r_{k}({\bolds{\sigma}}):= \biggl(\pmatrix{2d\cr k}|\Omega| \biggr)^{-1}\sum_{x\in\Omega
}\sum
_{S\subseteq N_{x},|S|=k}\sigma_{S},$$ where $\sigma_{S}=\prod_{x\in S}\sigma_{x}$ for any $S\subseteq
\Omega$. In particular $r_{k}({\bolds{\sigma}})$ is the average of the product of spins of all possible $k$ out of $2d$ neighbors. Note that $r_{1}({\bolds{\sigma}})\equiv m({\bolds{\sigma}})$. We will show that when $h = 0$ and $n$ is large, $m({\bolds{\sigma}})$ and $r_{k}({\bolds{\sigma}})$’s satisfy the following “mean-field relation” with high probability under the Gibbs measure: $$\label{eq:mfrel}
\bigl(1-\theta_{0}(\beta)\bigr)m({\bolds{\sigma}}) \approx\sum_{k=1}^{d-1}\theta
_{k}(\beta)
r_{2k+1}({\bolds{\sigma}}).$$ These relations hold for all values of $\beta\ge0$. Here, $\theta_{k}$’s are explicit rational functions of $\tanh(2\beta)$ for $k=0,1,\ldots
,d-1$, defined in (\[def:thetak\]) below. \[Later, we will prove in Proposition \[pr:isingh\] that an external magnetic field $h$ will add an extra linear term in the above relation (\[eq:mfrel\]).\] The following proposition makes this notion precise in terms of finite sample tail bound. It is a simple consequence of Theorem \[thm:conc\].
\[pr:ising\] Suppose ${\bolds{\sigma}}$ is drawn from the Gibbs measure $\mu_{\beta,0}$. Then, for any $\beta\ge0, n\ge1$ and $ t \ge0$ we have $${\mathbb{P}}\Biggl( \sqrt{|\Omega|} \Biggl|\bigl(1-\theta_{0}(\beta)\bigr)m({\bolds{\sigma}}) - \sum
_{k=1}^{d-1}\theta_{k}(\beta) r_{2k+1}({\bolds{\sigma}}) \Biggr| \ge t \Biggr)
\le2\exp\biggl( - \frac{t^{2} }{4 b(\beta)} \biggr),$$ where $m({\bolds{\sigma}}):= \frac{1}{|\Omega|}\sum_{x\in\Omega} \sigma
_{x}$ is the magnetization, $r_{k}({\bolds{\sigma}})$ is as given in (\[def:rk\]) and for $k=0,1,\ldots,d-1$ $$\begin{aligned}
\label{def:thetak}
\theta_{k}(\beta) &=& \frac{1}{4^{d}}\pmatrix{2d\cr2k+1}\sum
_{{\bolds{\sigma}}\in
\{
-1,+1\}^{2d}} \tanh\Biggl(\beta\sum_{i=1}^{2d}\sigma_{i} \Biggr)\prod
_{j=1}^{2k+1}\sigma_{j}\quad\mbox{and} \nonumber\\[-8pt]\\[-8pt]
b(\beta) &=& |1-\theta_{0}(\beta)| +
\sum_{k=1}^{d-1}(2k+1)|\theta_{k}(\beta)|.\nonumber\end{aligned}$$ Moreover, we can explicitly write down $\theta_{0}(\beta)$ as $$\theta_{0}(\beta) = \frac{1}{4^{d-1}}\sum_{k=1}^{d}k\pmatrix{2d\cr d+k}
\tanh
(2k\beta)$$ and for $d\ge2$ there exists $\beta_{1}\in(0,\infty)$, depending on $d$, such that $1-\theta_{0}(\beta)>0$ for $\beta<\beta_{1}$ and $1-\theta
_{0}(\beta)<0$ for $\beta>\beta_{1}$.
Here, we may remark that for any fixed $k$, $\theta_{k}(\beta/2d)$ converges to the coefficient of $x^{2k+1}$ in the power series expansion of $\tanh(\beta x)$ and $2d\beta_{1}(d)\downarrow1$ as $d\to
\infty$. For small values of $d$, we can explicitly calculate the $\theta_{k}$’s. For instance, in $d=2$, $$\theta_{0}(\beta)=\tfrac{1}{2} \bigl( \tanh(4\beta) + 2\tanh(2\beta
) \bigr),\qquad
\theta_{1}(\beta)=\tfrac{1}{2} \bigl( \tanh(4\beta) - 2\tanh(2\beta
) \bigr).$$ For $d=3$, $$\begin{aligned}
\theta_{0}(\beta)&=&\tfrac{3}{16} \bigl( \tanh(6\beta) + 4\tanh(4\beta
)+5\tanh
(2\beta) \bigr),\\
\theta_{1}(\beta)&=&\tfrac{10}{16} \bigl( \tanh(6\beta) - 3\tanh(2\beta
)
\bigr),\\
\theta_{2}(\beta)&=&\tfrac{3}{16} \bigl( \tanh(6\beta) - 4\tanh(4\beta
)+5\tanh
(2\beta) \bigr).\end{aligned}$$ For $d=4$, $$\begin{aligned}
\theta_{0}(\beta)&=&\tfrac{1}{16} \bigl(\tanh(8\beta) + 6\tanh(6\beta) +
14\tanh
(4\beta)+14\tanh(2\beta) \bigr),\\
\theta_{1}(\beta)&=&\tfrac{7}{16} \bigl(\tanh(8\beta) + 2\tanh(6\beta) -
2\tanh
(4\beta) - 6\tanh(2\beta) \bigr),\\
\theta_{2}(\beta)&=&\tfrac{7}{16} \bigl(\tanh(8\beta) - 2\tanh(6\beta) -
2\tanh
(4\beta)+6\tanh(2\beta) \bigr),\\
\theta_{3}(\beta)&=&\tfrac{1}{16} \bigl(\tanh(8\beta) - 6\tanh(6\beta) +
14\tanh
(4\beta)-14\tanh(2\beta) \bigr).\end{aligned}$$
For the Ising model on $\Omega$ at inverse temperature $\beta$ with no external magnetic field for all $ t\ge0$ we have:
if $d=1$, $${\mathbb{P}}\bigl( |m({\bolds{\sigma}})|\ge t\bigr) \le2\exp\bigl( - \tfrac{1}{4}|\Omega|\bigl(1-\tanh
(2\beta
)\bigr)t^{2} \bigr);$$
if $d=2$, $${\mathbb{P}}\bigl( |[(1-u)^{2}-u^{3}]m({\bolds{\sigma}}) + u^{3}r_{3}({\bolds{\sigma}})|\ge t\bigr)\le2\exp
\biggl( - \frac{|\Omega|t^{2}}{32} \biggr),$$ where $u=\tanh(2\beta)$ and $r_{3}({\bolds{\sigma}})= \frac{1}{4|\Omega|}\sum^{*}
\sigma
_{x}\sigma_{y}\sigma_{z}$ where the sum $\sum^{*}$ is over all $x,y,z\in
\Omega$ such that $|x-y|=2,|z-y|=2,|x-z|=2$;
if $d=3$, $${\mathbb{P}}\bigl( |g(u)m({\bolds{\sigma}}) + 5u^{3}(1+u^{2})r_{3}({\bolds{\sigma}}) -3u^{5}r_{5}({\bolds{\sigma}})|\ge t\bigr)
\le2\exp( - c|\Omega|t^{2} ),$$ where $c$ is an absolute constant, $g(u) = 1-3u
+4u^{2}-9u^{3}+3u^{4}-3u^{5}$, $u=\tanh(2\beta)$ and $r_{3}, r_{5}$ are as defined in (\[def:rk\]).
Although we do not yet know the significance of the above relations, it seems somewhat striking that they are not affected by phase transitions. The exponential tail bounds show that many such relations can hold simultaneously. For completeness, we state below the corresponding result for nonzero external field.
\[pr:isingh\] Suppose ${\bolds{\sigma}}$ is drawn from the Gibbs measure $\mu_{\beta,h}$. Let $r_{k}({\bolds{\sigma}})$, $\theta_{k}(\beta)$, $b(\beta)$ be as in proposition (\[pr:ising\]). Then, for any $\beta\ge0, h\in{\mathbb{R}}, n\ge1$ and $ t
\ge0$ we have $${\mathbb{P}}\bigl( \bigl|\bigl(1-\theta_{0}(\beta)\bigr)m({\bolds{\sigma}}) - g({\bolds{\sigma}}) \bigr|
\ge
t \bigr)\le2\exp\biggl( - \frac{|\Omega|t^{2} }{4 b(\beta)(1+\tanh|h|)}
\biggr),$$ where $$g({\bolds{\sigma}}):= \sum_{k=1}^{d-1}\theta_{k}(\beta) r_{2k+1}({\bolds{\sigma}})
+\tanh(h)
\Biggl(1-\sum_{k=0}^{d-1}\theta_{k}(\beta) s_{2k+1}({\bolds{\sigma}}) \Biggr)$$ and $$s_{k}({\bolds{\sigma}}):= \biggl(\pmatrix{2d\cr k}|\Omega| \biggr)^{-1}\sum_{x\in\Omega
}\sum
_{S\subseteq N_{x}, |S|=k}\sigma_{S\cup\{x\}}$$ is the average of products of spins over all $k$-stars for $k=1,2,\ldots
,2d$ and $\Omega$ is the discrete torus in ${\mathbb{Z}}^{d}$ with $n^{d}$ many points.
Proofs {#sec:proof}
======
Proof of Proposition 4
----------------------
Instead of proving Theorem \[thm:nong1\] first, let us see how it is applied to prove the result for the Curie–Weiss model at critical temperature. The proof is simply an elaboration of the sketch given at the end of Section \[subsec:curie\].
Suppose ${\bolds{\sigma}}$ is drawn from the Curie–Weiss model at critical temperature. We construct ${\bolds{\sigma}}'$ by taking one step in the heat-bath Glauber dynamics: a coordinate $I$ is chosen uniformly at random, and $\sigma_{I}$ is replace by $\sigma'_{I}$ drawn from the conditional distribution of the $I$th coordinate given $\{\sigma_{j}\dvtx j\neq I\}$. Let $$F({\bolds{\sigma}},{\bolds{\sigma}}') := \sum_{i=1}^{n}(\sigma_{i}-\sigma'_{i})=\sigma
_{I}-\sigma'_{I}.$$ For each $i=1,2,\ldots,n$, define $
m_{i}=m_{i}({\bolds{\sigma}}) = n^{-1}\sum_{j\neq i} \sigma_{j}.
$ An easy computaion gives that ${\mathbb{E}}(\sigma_{i}| \{\sigma_{j}, j\neq i\}
) =
\tanh( m_{i})$ for all $i$ and so we have $$f({\bolds{\sigma}}):= {\mathbb{E}}(F({\bolds{\sigma}},{\bolds{\sigma}}')|{\bolds{\sigma}}) = m - \frac{1}{n} \sum_{i=1}^{n}
\tanh(m_{i})=\frac{m}{n} + \frac{1}{n} \sum_{i=1}^{n} g(m_{i}),$$ where $g(x):=x-\tanh(x)$. By definition, $m_{i}({\bolds{\sigma}})-m({\bolds{\sigma}}) = \sigma_{i}/n$ and $m_{i}({\bolds{\sigma}}')-m({\bolds{\sigma}}) = (\sigma_{i}+\sigma_{I}-\sigma'_{I})/n$ for all $i$. Hence, using Taylor’s expansion up to first degree and noting that $|g'(x)|=\tanh
^{2}(x)\le x^{2}$ we have $$\begin{aligned}
|f({\bolds{\sigma}})-f({\bolds{\sigma}}')| &\le& \frac{2}{n}|g'(m({\bolds{\sigma}}))|+ \frac{2+5\max
_{|x|\le1}|g''(x)|}{n^{2}}\\
&\le& \frac{2}{n}m({\bolds{\sigma}})^{2} +\frac{6}{n^{2}}.\end{aligned}$$ Clearly, $|F({\bolds{\sigma}},{\bolds{\sigma}}')|\le2$. Thus, we have $$\Delta({\bolds{\sigma}}) := \frac{1}{2}{\mathbb{E}}[ |f({\bolds{\sigma}})-f({\bolds{\sigma}}')|\cdot
|F({\bolds{\sigma}},{\bolds{\sigma}}')| {|}{\bolds{\sigma}}]
\le\frac{2}{n}m({\bolds{\sigma}})^{2}+\frac{6}{n^{2}}.$$ Now it is easy to verify that $|x|^{3} \le5|x-\tanh x|$ for all $|x|\le1$. Note that this is the place where we need $\beta=1$. For $\beta
\neq1$, the linear term dominates in $m-\tanh(\beta m)$. Hence, it follows that $$m({\bolds{\sigma}})^{2} \le5^{2/3}|m ({\bolds{\sigma}})-\tanh m({\bolds{\sigma}}) |^{2/3}\le
3|f({\bolds{\sigma}})|^{2/3} +{3}{n^{-2/3}},$$ where in the last line we used the fact that $|f({\bolds{\sigma}}) - (m - \tanh
m)|\le1/n$ and $5^{2/3}< 3$. Thus, $$\Delta({\bolds{\sigma}}) \le\frac{6}{n}|f({\bolds{\sigma}})|^{2/3}+\frac{12}{n^{5/3}}$$ and using Corollary \[cor:nong\] with $\alpha=2/3, B= 6/n$ and $C=12/n^{5/3}$ we have $${\mathbb{P}}(|m-\tanh m| \ge t+n^{-1}) \le{\mathbb{P}}\bigl(|f({\bolds{\sigma}})| \ge t\bigr) \le2e^{- cnt^{4/3}}$$ for all $t\ge0$ for some constant $c>0$. This clearly implies that $${\mathbb{P}}(|m|\ge t) \le{\mathbb{P}}(|m-\tanh m|\ge t^{3}/5) \le2e^{-cnt^{4}}$$ for all $t\ge0$ and for some absolute constant $c>0$. Thus, we are done.
Proof of Proposition 6
----------------------
The proof is along the lines of proof of Proposition \[pr:crit\]. Suppose ${\mathbf{X}}$ is drawn from the distribution $\nu_{n}$. We construct ${\mathbf{X}}'$ as follows: a coordinate $I$ is chosen uniformly at random, and $X_{I}$ is replace by $X'_{I}$ drawn from the conditional distribution of the $I$th coordinate given $\{X_{j}\dvtx j\neq I\}$. Let $$F({\mathbf{X}},{\mathbf{X}}') := \sum_{i=1}^{n}(X_{i}-X'_{i})=X_{I}-X'_{I}.$$ For each $i=1,2,\ldots,n$, define $
m_{i}({\mathbf{X}}) = n^{-1}\sum_{j\neq i} X_{j}.
$ An easy computaion gives that ${\mathbb{E}}(X_{i}| \{X_{j}, j\neq i\}) =
g(m_{i})$ for all $i=1,2,\ldots,n$ where $g(s)=\frac{d}{ds}(\log\int
\exp(x^{2}/2n+sx) \,d\rho(x))$ for $s\in{\mathbb{R}}$. So we have $$f({\mathbf{X}}):= {\mathbb{E}}(F({\mathbf{X}},{\mathbf{X}}')|{\mathbf{X}}) = m({\mathbf{X}}) - \frac{1}{n} \sum_{i=1}^{n}
g(m_{i}({\mathbf{X}})).$$ Define the function $$\label{def:h}
h(s)=\frac{s^{2}}{2}-\log\int\exp(sx) \,d\rho(x) \qquad\mbox{for }s\in
{\mathbb{R}}.$$ Clearly, $h$ is an even function. Recall that $k$ is an integer such that $h^{(i)}(0)=0$ for $0\le i<2k$ and $h^{(2k)}(0)\neq0$. We have $k\ge2$ since $h''(0)=1-\int x^{2} \,d\rho(x)=0$.
Now using the fact that $\rho([-L,L])=1$ it is easy to see that $|f({\mathbf{X}})-h'(m({\mathbf{X}}))|\le c/n$ for some constant $c$ depending on $L$ only. In the subsequent calculations, $c$ will always denote a constant depending only on $L$ that may vary from line to line. Similarly, we have $$\begin{aligned}
|f({\mathbf{X}})-f({\mathbf{X}}')|&\le&\frac{|X_{I}-X'_{I}|}{n} \biggl( |1-g'(m({\mathbf{X}}))|+\frac
{c(1+\sup_{|x|\le L}|g''(x)|)}{n} \biggr)\\
&\le&\frac{2L}{n}|h''(m({\mathbf{X}}))| + \frac{c}{n^{2}}.\end{aligned}$$ Note that $|h''(s)|\le cs^{2k-2}$ for some constant $c$ for all $s\ge
0$. This follows since $\lim_{s\to0}h''(s)/s^{2k-2}$ exists and $h''(\cdot)$ is a bounded function. Also $\lim_{s\to
0}|h'(s)|/\break|s|^{2k-1}=|h^{(2k)}(0)|\neq0$ and $|h'(s)|> 0$ for $s> 0$. So we have $|h'(s)|\ge c|s|^{2k-1}$ for some constant $c>0$ and all $|s|\le L$. From the above results, we deduce that $$\begin{aligned}
|f({\mathbf{X}})-f({\mathbf{X}}')|&\le&\frac{c}{n}|(m({\mathbf{X}}))|^{2k-2} + \frac{c}{n^{2}}
\le\frac{c}{n}|h'(m({\mathbf{X}}))|^{({2k-2})/({2k-1})} + \frac{c}{n^{2}}\\
&\le&\frac{c}{n}|f({\mathbf{X}})|^{({2k-2})/({2k-1})} + \frac{c}{n^{2-{1}/{(2k-1)}}}.\end{aligned}$$ Now the rest of the proof follows exactly as for the classical Curie–Weiss model.
Proof of Theorem 7
------------------
First, let us state and prove a simple technical lemma.
\[techlmm\] Let $x_1,\ldots,x_k,y_1,\ldots,y_k$ be real numbers. Then $$\max_{1\le i\le n} \biggl|\frac{e^{x_i}}{\sum_{j=1}^k e^{x_j}} -
\frac
{e^{y_i}}{\sum_{j=1}^k e^{y_j}} \biggr|\le2\max_{1\le i\le n} |x_i - y_i|$$ and $$\Biggl|\log\sum_{i=1}^k e^{x_i} - \log\sum_{i=1}^k e^{y_i}
\Biggr|\le
\max_{1\le i\le k} |x_i - y_i|.$$
Fix $1\le i\le k$. For $t\in[0,1]$, let $$h(t) = \frac{e^{tx_i + (1-t)y_i}}{\sum_{j=1}^k e^{tx_j + (1-t)y_j}}.$$ Then $$h'(t) = \biggl[(x_i -y_i) - \frac{\sum_{j=1}^k (x_j-y_j) e^{tx_j +
(1-t)y_j}}{\sum_{j=1}^ke^{tx_j + (1-t)y_j}} \biggr] h(t).$$ This shows that $|h'(t)| \le2\max_i |x_i-y_i|$ for all $t\in[0,1]$ and completes the proof of the first assertion. The second inequality is proved similarly.
[Proof of Lemma \[lem:mf\]]{} Fix two numbers $1\le i < j\le n$. Given a configuration ${\mathbf{X}}$, construct another configuration ${\mathbf{X}}'$ as follows. Choose a point $k
\in\{1,\ldots,n\}\setminus\{i,j\}$ uniformly at random, and replace the pair $(X_{ik}, X_{jk})$ with $(X'_{ik}, X'_{jk})$ drawn from the conditional distribution given the rest of the edges. Let $L_{ij}'$ be the revised value of $L_{ij}$. From the form of the Hamiltonian, it is now easy to read off that for $x,y\in\{0,1\}$, $$\begin{aligned}
\hspace*{-3pt}&&{\mathbb{P}}(X'_{ik} = x, X'_{jk} = y {|}{\mathbf{X}}) \\
\hspace*{-3pt}&&\qquad\propto\exp\biggl( \beta x L_{ik} + \beta y L_{jk} + hx + hy -
\frac
{\beta}{n} x X_{ij} X_{jk} - \frac{\beta}{n} y X_{ij} X_{ik} + \frac
{\beta}{n} xyX_{ij} \biggr).\end{aligned}$$ An application of Lemma \[techlmm\] shows that the terms having $\beta
/n$ as coefficient can be “ignored” in the sense that for each $x,y\in
\{0,1\}$, $$\biggl|{\mathbb{P}}(X'_{ik} = x, X'_{jk} = y {|}{\mathbf{X}}) - \frac{e^{\beta x L_{ik}
+ \beta y L_{jk} + hx + hy}}{(1+e^{\beta L_{ik} + h}) (1+e^{\beta
L_{jk} + h})} \biggr|\le\frac{2\beta}{n}.$$ In particular, $$\label{diff}
|{\mathbb{E}}(X'_{ik} X'_{jk} {|}{\mathbf{X}}) - \varphi(L_{ik})\varphi(L_{jk})| \le
\frac{2\beta}{n}.$$ Now, $$\begin{aligned}
\label{fx}
{\mathbb{E}}(L_{ij} - L_{ij}'{|}{\mathbf{X}}) &=& \frac{1}{n(n-2)} \sum_{k\notin\{
i,j\}
} \bigl(X_{ik}X_{jk} - {\mathbb{E}}(X'_{ik}X'_{jk}{|}{\mathbf{X}})\bigr)\nonumber\\[-8pt]\\[-8pt]
&=& \frac{1}{n-2} L_{ij} - \frac{1}{n(n-2)}\sum_{k\notin\{i,j\}} {\mathbb{E}}(X'_{ik} X'_{jk}{|}{\mathbf{X}}).\nonumber\end{aligned}$$ Let $F({\mathbf{X}},{\mathbf{X}}') = (n-2) (L_{ij} - L'_{ij})$ and $f({\mathbf{X}}) = {\mathbb{E}}(F({\mathbf{X}},{\mathbf{X}}'){|}{\mathbf{X}})$. Let $$g({\mathbf{X}}) = L_{ij} - \frac{1}{n}\sum_{k\notin\{i,j\}}\varphi
(L_{ik})\varphi(L_{jk}).$$ From (\[diff\]) and (\[fx\]), it follows that $$\label{eq:fg}
|f({\mathbf{X}}) - g({\mathbf{X}})| \le\frac{2\beta}{n}.$$ Since $X'$ has the same distribution as $X$, the same bound holds for $|f(X')-g(X')|$ as well. Now clearly, $|F(X,X')| \le1$. Again, $|g(X)
- g(X')| \le2/n$, and therefore $$|f(X) - f(X')|\le\frac{4(1+\beta)}{n}.$$ Combining everything, and applying Theorem \[thm:conc\] with $B=0$ and $C=2(1+\beta)/n$, we get $${\mathbb{P}}\bigl(|f({\mathbf{X}})|\ge t\bigr)\le2\exp\biggl( - \frac{nt^{2}}{4(1+\beta)} \biggr)$$ for all $t\ge0$. From (\[eq:fg\]), it follows that $${\mathbb{P}}\bigl(|g({\mathbf{X}})|\ge t\bigr)\le{\mathbb{P}}\bigl(|f({\mathbf{X}})|\ge t-2\beta/n\bigr)\le
2\exp\biggl( - \frac{nt^{2}}{8(1+\beta)} \biggr)$$ for all $t\ge8\beta/n$. This completes the proof of the tail bound. The bound on the mean absolute value is an easy consequence of the tail bound.
[Proof of Lemma \[mean2\]]{} The proof is in two steps. In the first step, we will get an error bound of order $n^{-1/2}\sqrt{\log n}$. In the second step, we will improve it to $n^{-1/2}$. Define $$\Delta= \max_{1\le i<j\le n} \biggl|L_{ij} - \frac{1}{n}\sum_{k\notin\{
i,j\}} \varphi(L_{ik})\varphi(L_{jk}) \biggr|.$$ By Lemma \[lem:mf\] and union bound, we have $${\mathbb{P}}(
\Delta\ge t ) \le n^{2}\exp\biggl(- \frac{nt^{2}}{8(1+\beta)}
\biggr)$$ for all $t\ge8\beta/n$. Intuitively, the above equation says that $\Delta$ is of the order of $\sqrt{\log n/n}$, in fact we have ${\mathbb{E}}( \Delta
^{2} )=
O(\log n/n)$. Clearly, $\varphi$ is an increasing function. Hence, we have $$\varphi(L_{\min})^{2} - \Delta\le L_{\min}\le L_{\max}\le\varphi
(L_{\max
})^{2} + \Delta,$$ where $L_{\max}=\max_{1\le i<j\le n}L_{ij}$ and $L_{\min}=\min
_{1\le
i<j\le n}L_{ij}$.
Now assume that there exists a unique solution $u^{*}$ of the equation $\varphi(u)^{2}=u$ with $2\varphi(u^{*})\varphi'(u^{*})<1$. For ease of notation, define the function $\psi(u)=\varphi(u)^2-u$. We have $\psi
(0)>0>\psi(1)$, $u^{*}$ is the unique solution to $\psi(u)=0$ and $\psi
'(u^{*})<0$. It is easy to see that $\psi'(u)=0$ has at most three solution \[$\psi'(u)=2\beta\varphi(u)^{2}(1-\varphi(u))-1$ is a third degree polynomial in $\varphi(u)$ and $\varphi$ is a strictly increasing function\].
Hence, there exist positive real numbers ${\varepsilon}, \delta$ such that $|\psi
(u)|>{\varepsilon}$ if $|u-u^{*}|>\delta$. Note that $\psi(u)>0$ if $u<u^{*}$ and $\psi(u)<0$ is $u>u^{*}$. Decreasing ${\varepsilon},\delta$ without loss of generality, we can assume that $$\label{eq:deriv}
\inf_{0<|u-u^{*}|\le\delta} \biggl[\frac{u-u^{*}}{-\psi(u)}
\biggr]= c >0.$$ This is possible because $\psi'(u^{*})<0$. Note that $\psi(L_{\max
})\ge
-\Delta$ and $\psi(L_{\min})\le\Delta$. Thus, we have $$u^{*}-\delta\le L_{\min}\le L_{\max}\le u^{*}+\delta,$$ when $\Delta<{\varepsilon}$. Using (\[eq:deriv\]), $u^{*}\le L_{\max}\le
u^{*}+\delta
$ implies that $|L_{\max}-u^{*}|\le c\Delta$ and $u^{*}-\delta\le
L_{\min}\le
u^{*}$ implies that $|L_{\min}-u^{*}|\le c\Delta$. Thus, when $\Delta< {\varepsilon}$, we have $|L_{\max}-u^{*}|\le c{\Delta}$ and $|L_{\min}-u^{*}|\le c{\Delta}$ and in particular, $|L_{ij}-u^{*}|\le
c{\Delta
}$ for all $i<j$. So we can bound the $L^{2}$ distance of $L_{ij}$ from $u^{*}$ by $${\mathbb{E}}(L_{ij}-u^{*})^{2}\le c^{2}{\mathbb{E}}(\Delta^{2}) + {\mathbb{P}}(\Delta\ge{\varepsilon})\le
K(\beta
,h)\frac{\log n}{n}$$ for all $i<j$.
Now let us move to the second step. Recall from (\[eq:lcon\]) that $$\label{eq:lconcopy}
{\mathbb{E}}\biggl|L_{ij} - \frac{1}{n}\sum_{k\notin\{i,j\}} \varphi
(L_{ik})\varphi(L_{jk}) \biggr| \le\frac{C(1+\beta)^{1/2}}{n^{1/2}}$$ for all $i<j$. Let $D_{ij}=L_{ij}-u^{*}$. Using Taylor’s expansion around $u^{*}$ up to degree one, we have $$\begin{aligned}
\varphi(L_{ik})\varphi(L_{jk}) - \varphi(u^{*})^{2}
&=& \varphi(u^{*})\bigl(\varphi(L_{ik}) - \varphi(u^{*})\bigr) +\varphi
(u^{*})\bigl(\varphi(L_{jk}) - \varphi(u^{*})\bigr)\\
&&{} + \bigl(\varphi(L_{ik}) - \varphi(u^{*})\bigr)\bigl(\varphi(L_{jk}) -
\varphi
(u^{*})\bigr) \\
&=& \varphi(u^{*}) \varphi'(u^{*})(D_{ik}+D_{jk}) + R_{ijk},\end{aligned}$$ where ${\mathbb{E}}(|R_{ijk}|)\le C{\mathbb{E}}(D_{ij}^{2})\le C n^{-1}\log n$ for some constant $C$ depending only on $\beta,h$. Thus, $$\begin{aligned}
\label{eq:combl2}
&&{\mathbb{E}}\biggl|L_{ij} - \frac{1}{n}\sum_{k\notin\{i,j\}} \varphi
(L_{ik})\varphi(L_{jk}) - D_{ij} + \frac{\varphi(u^{*}) \varphi
'(u^{*})}{n}\sum_{k\notin\{i,j\}}(D_{ik}+D_{jk}) \biggr|\hspace*{-28pt}\nonumber\\[-8pt]\\[-8pt]
&&\qquad\le\frac{2u^{*}}{n} +\frac{1}{n}\sum_{k\notin\{i,j\}}{\mathbb{E}}|R_{ijk}|\le\frac{C\log
n}{n}.\nonumber\end{aligned}$$ Here, we used the fact that $u^{*}=\varphi(u^{*})^{2}$. Combining (\[eq:lconcopy\]) and (\[eq:combl2\]), we have $${\mathbb{E}}\biggl|D_{ij} - \frac{\varphi(u^{*}) \varphi'(u^{*})}{n}\sum
_{k\notin\{i,j\}}(D_{ik}+D_{jk}) \biggr|\le\frac{C}{\sqrt{n}}$$ for all $i<j$. By symmetry, ${\mathbb{E}}|D_{ij}|$ is the same for all $i,j$. Thus, finally we have $${\mathbb{E}}|L_{ij}-u^{*}|={\mathbb{E}}|D_{ij}|\le\frac{1}{1-2\varphi(u^{*}) \varphi
'(u^{*})}\cdot\frac{C}{\sqrt{n}}=\frac{K(\beta,h)}{\sqrt{n}},$$ where $K(\beta,h)$ is a constant depending on $\beta,h$.
When $\psi(u)=0$ has a unique solution at $u=u^{*}$ with $2\psi
(u^{*})\psi'(u^{*})=1$, which happens at the critical point $\beta
=(3/2)^{3}, h=\log2 -3/2$, instead of (\[eq:deriv\]) we have $$\inf_{0<|u-u^{*}|\le\delta} \biggl[\frac{(u-u^{*})^{3}}{-\psi
(u)} \biggr]=
c >0$$ since $\psi(u^{*})=\psi'(u^{*})=\psi''(u^{*})=0$ and $\psi
'''(u^{*})<0$. Then using a similar idea as above one can easily show that $${\mathbb{E}}|L_{ij}-u^{*}|\le K(\beta,h){n^{-1/6}}$$ for some constant $K$ depending on $\beta,h$. This completes the proof of the lemma.
The proof becomes lot easier if we have $$\label{ccond}
c:=\varphi(1)\cdot\sup_{0\le x\le1}\frac{|\varphi(x) - \varphi
(u^*)|}{|x-u^*|}<\frac{1}{2}.$$ This is because, by the triangle inequality, we have $$\begin{aligned}
\label{maineq}
\sum_{i<j} |L_{ij} - u^*| &\le& \sum_{i<j} \biggl|L_{ij} - \frac
{1}{n}\sum_{k\notin\{i,j\}}\varphi(L_{ik})\varphi(L_{jk})
\biggr|\nonumber\\[-8pt]\\[-8pt]
&&{} + \sum_{i<j} \biggl(\frac{1}{n}\sum_{k\notin\{i,j\}}
|\varphi(L_{ik})\varphi(L_{jk}) - u^*| +
\frac{2u^*}{n} \biggr).\nonumber\end{aligned}$$ Now recall that condition (\[ccond\]) says that $\varphi(1)|\varphi
(x)- \varphi(u^*)|\le c|x-u^*|$ for all $x\in[0,1]$. Moreover, $L_{ij}\in[0,1]$ for all $i,j$, and $u^* = \varphi(u^*)^2$. Thus, $$|\varphi(L_{ik})\varphi(L_{jk}) - u^*| \le c|L_{ik} -
u^*| +
c|L_{jk} - u^*|.$$ Combining everything, we get $$\sum_{i<j} |L_{ij} - u^*| \le\frac{\sum_{i<j}|L_{ij} -
{1}/{n}\sum_{k\notin\{i,j\}}\varphi(L_{ik})\varphi(L_{jk})|
+ n
u^*}{1-2c}.$$ Taking expectation on both sides, and applying Lemma \[lem:mf\], we get $$\sum_{i<j} {\mathbb{E}}|L_{ij} - u^*| \le\frac{C(1+\beta) n^{3/2}}{1-2c}.$$ And this gives the required result. In fact, using basic calculus results one can easily check that condition (\[ccond\]) is satisfied when $h\ge0$ or $\beta\le2$.
Now, we will prove that in the exponential random graph model, the number of edges and number of triangles also satisfy certain “mean-field” relations.
\[lem:ex\] Recall that $E(\mathbf{x})$ and $T(\mathbf{x})$ denote the number of edges and number of triangles in the graph defined by the edge configuration $\mathbf
{x}\in\Omega
$. If ${\mathbf{X}}$ is drawn from the Gibbs’ measure in Theorem \[pressure\], we have the bound $$\begin{aligned}
{\mathbb{E}}\biggl| E({\mathbf{X}}) - \sum_{i<j}\varphi(L_{ij}) \biggr| &\le& C(1+\beta
)^{1/2}n,\\
{\mathbb{E}}\biggl| \frac{T({\mathbf{X}})}{n} - \frac{1}{3}\sum_{i<j}L_{ij}\varphi
(L_{ij}) \biggr| &\le& C(1+\beta)^{1/2}n,\end{aligned}$$ where and $C$ is a universal constant.
It is not difficult to see that $${\mathbb{E}}\bigl(X_{ij}{|}(X_{kl})_{ (k,l)\ne(i,j)}\bigr) = \varphi(L_{ij}).$$ Let us create ${\mathbf{X}}'$ by choosing $1\le i<j\le n$ uniformly at random and replacing $X_{ij}$ with $X_{ij}'$ drawn from the conditional distribution of $X_{ij}$ given $(X_{kl})_{(k,l)\ne(i,j)}$. Let $F({\mathbf{X}},{\mathbf{X}}') = {n\choose2}(X_{ij}-X'_{ij})$. Then $$f({\mathbf{X}}) = {\mathbb{E}}(F({\mathbf{X}},{\mathbf{X}}')|{\mathbf{X}}) = \sum_{k<l} \bigl(X_{kl} - \varphi
(L_{kl})\bigr) =
E({\mathbf{X}}) - \sum_{k<l}\varphi(L_{kl}).$$ Now $|F({\mathbf{X}},{\mathbf{X}}')| \le{n \choose2}$ and $|f({\mathbf{X}}) - f({\mathbf{X}}')| \le
1+\beta$. Here we used the fact that $|\varphi'(x)|\le\beta/4$. Combining the above result and Theorem \[thm:conc\] with $B=0,
C=\frac
{1}{2}(1+\beta){n\choose2}$, we get the required bound.
Similarly, if we define $F({\mathbf{X}},{\mathbf{X}}') = {n\choose2}(X_{ij}L_{ij} -
X_{ij}'L_{ij})$. Then $$\begin{aligned}
f({\mathbf{X}}) &=& {\mathbb{E}}(F({\mathbf{X}},{\mathbf{X}}')|{\mathbf{X}}) = \sum_{k<l} \bigl(X_{kl}L_{kl} -
\varphi
(L_{kl})L_{kl}\bigr)\\
&=& \frac{3}{n}T({\mathbf{X}}) - \sum_{k<l}\varphi(L_{kl})L_{kl}.\end{aligned}$$ Again, $|F({\mathbf{X}},{\mathbf{X}}')| \le{n\choose2}$ and $|f({\mathbf{X}})-f({\mathbf{X}}')| \le
C(1+\beta)$. The bound follows easily as before.
The following result is an easy corollary of Lemmas \[mean2\] and \[lem:ex\].
\[cor:ex\] Suppose the conditions of Theorem \[pressure\] are satisfied. Then we have $${\mathbb{E}}\biggl\vert E({\mathbf{X}}) - \frac{n^{2}\varphi(u^{*})}{2} \biggr\vert\le Cn^{3/2}
\quad\mbox{and}\quad
{\mathbb{E}}\biggl\vert\frac{T({\mathbf{X}})}{n} - \frac{n^{2}\varphi(u^{*})^{3}}{6}
\biggr\vert\le Cn^{3/2},$$ where $C$ is a constant depending only on $\beta,h$.
\[large1\] Suppose the conditions of Theorem \[pressure\] are satisfied. Let $T_n$ be the number of triangles in the Erdős–Rényi graph $G(n,
\varphi(0))$. Then there is a constant $K(\beta,h)$ depending only on $\beta$ and $h$ such that for all $n$ $$\biggl|\frac{\log{\mathbb{P}}(|T_n - {n \choose 3}\varphi(u^*)^3| \le
K(\beta
,h) n^{5/2})}{n^2} - \frac{-I(\varphi(u^*),\varphi(0))}{2}
\biggr|\le
\frac{K(\beta,h)}{\sqrt{n}}.$$
Let $X$ be drawn from the Gibbs’ measure in Theorem \[pressure\] with parameters $\beta,h$. From Corollary \[cor:ex\], we see that there exists a constant $K(\beta,h)$ such that (for all $n$) $${\mathbb{P}}\biggl( \biggl|E(X) - \frac{n^2\varphi(u^*)}{2} \biggr|\le
K(\beta,h)
n^{3/2} \biggr) \ge\frac{3}{4}$$ and $${\mathbb{P}}\biggl( \biggl|\frac{T(X)}{n} - \frac{n^2\varphi(u^*)^3}{6}
\biggr|\le K(\beta,h) n^{3/2} \biggr) \ge\frac{3}{4}.$$ Now let $$A = \biggl\{x\in\{0,1\}^n \dvtx \biggl|\frac{T(x)}{n} - \frac
{n^2\varphi
(u^*)^3}{6} \biggr|\le K(\beta,h) n^{3/2} \biggr\}$$ and $$B = A \cap\biggl\{x\in\{0,1\}^n\dvtx \biggl|E(x) - \frac{n^2\varphi
(u^*)}{2} \biggr|\le K(\beta,h) n^{3/2} \biggr\}.$$ Now suppose $Y=(Y_{ij})_{1\le i<j\le n}$ is a collection of i.i.d. random variables satisfying ${\mathbb{P}}(Y_{ij} = 1) = 1 - {\mathbb{P}}(Y_{ij} = 0) =
\varphi(0)$ and $Z=(Z_{ij})_{1\le i<j\le n}$ is another collection of i.i.d. random variables with ${\mathbb{P}}(Z_{ij}= 1)=1-{\mathbb{P}}(Z_{ij}=0) =
\varphi
(u^*)$. Without loss of generality, we can assume that $K(\beta,h)$ was chosen large enough to ensure that (again, for all $n$) ${\mathbb{P}}(Z\in A)
\ge1/2$ and ${\mathbb{P}}(Z\in B) \ge1/2$. Now, it follows directly from the definition of $A$ and Lemma \[techlmm\] that $$\begin{aligned}
\label{eq1}
&& \biggl|\log\sum_{x\in A} e^{hE(x)} - \log\sum_{x\in A} e^{
{\beta
T(x)}/{n} + hE(x)} + \frac{\beta n^2\varphi(u^*)^3}{6}
\biggr|\nonumber\\
&&\qquad= \biggl|\log\sum_{x\in A} e^{hE(x) + {\beta n^2\varphi
(u^*)^3}/{6}} - \log\sum_{x\in A} e^{{\beta T(x)}/{n} + hE(x)}\biggr|
\\
&&\qquad\le\beta\max_{x\in A} \biggl|\frac{T(x)}{n} - \frac{n^2\varphi
(u^*)^3}{6} \biggr|
\le\beta K(\beta,h) n^{3/2}.\nonumber\end{aligned}$$ Next, observe that $$\begin{aligned}
\label{eq11}
&& \biggl|\log\sum_{x\in A} e^{{\beta T(x)}/{n} + hE(x)} - \log
\sum
_{x\in\Omega} e^{{\beta T(x)}/{n} + hE(x)}
\biggr|\nonumber\\[-8pt]\\[-8pt]
&&\qquad = |{\log}{\mathbb{P}}(X\in A) |\le|{\log}(3/4)|.\nonumber\end{aligned}$$ Similarly, we have $$\begin{aligned}
\label{eq12}
&& \biggl|\log\sum_{x\in B} e^{{\beta T(x)}/{n} + hE(x)} - \log
\sum
_{x\in\Omega} e^{{\beta T(x)}/{n} + hE(x)} \biggr| \nonumber\\[-8pt]\\[-8pt]
&&\qquad= |{\log}{\mathbb{P}}(X\in B) |\le|{\log}(1/2)|,\nonumber\end{aligned}$$ where we used the fact that ${\mathbb{P}}(X\in A\cap C)\ge{\mathbb{P}}(X\in A)+{\mathbb{P}}(X\in C)-1$. Combining the last two inequalities, we get $$\label{eq2}
\biggl|\log\sum_{x\in A} e^{{\beta T(x)}/{n} + hE(x)} - \log
\sum
_{x\in B} e^{{\beta T(x)}/{n} + hE(x)} \biggr| \le\log(8/3).$$ Next, note that by the definition of $B$ and Lemma \[techlmm\], we have that for any $h'$, $$\begin{aligned}
\label{eq3}
&& \hspace*{20pt}\biggl|\log\sum_{x\in B} e^{{\beta T(x)}/{n} + hE(x)} -\frac
{n^2(h-h')\varphi(u^*)}{2} - \frac{\beta n^2 \varphi(u^*)^3}{6} -
\log
\sum_{x\in B} e^{h'E(x)} \biggr| \nonumber\hspace*{-20pt}\\
&&\hspace*{20pt}\qquad\le\sup_{x\in B} \biggl| \frac{\beta T(x)}{n} + hE(x) -\frac
{n^2(h-h')\varphi(u^*)}{2} - \frac{\beta n^2 \varphi(u^*)^3}{6} -
h'E(x) \biggr|
\\
&&\hspace*{20pt}\qquad\le(\beta+|h-h'|)K(\beta,h)n^{3/2}.\nonumber\end{aligned}$$ Now, choose $h' = \log\frac{\varphi(u^*)}{1-\varphi(u^*)}$. Then $$\label{eq4}
\biggl|\log\sum_{x\in B} e^{h'E(x)} - \log\sum_{x\in\Omega}
e^{h'E(x)} \biggr| = |{\log{\mathbb{P}}}(Z\in B)|\le\log2.$$ Adding up (\[eq1\]), (\[eq2\]), (\[eq3\]) and (\[eq4\]), and using the triangle inequality, we get $$\label{eq5}\quad
\biggl|\log\sum_{x\in A} e^{hE(x)} - \frac{n^2(h-h')\varphi(u^*)}{2}
-\log\sum_{x\in\Omega}e^{h'E(x)} \biggr|\le K'(\beta, h) n^{3/2},$$ where $K'(\beta,h)$ is a constant depending only on $\beta,h$. For any $s\in{\mathbb{R}}$, a trivial verification shows that $$\log\sum_{x\in\Omega} e^{sE(x)} = \pmatrix{n\cr2}\log(1+e^s).$$ Again, note that $\log{\mathbb{P}}(Y\in A) = \log\sum_{x\in A} e^{hE(x)} -
\log
\sum_{x\in\Omega} e^{hE(x)}$. Therefore, it follows from inequality (\[eq5\]) that $$\biggl|\frac{\log{\mathbb{P}}(Y\in A)}{n^2} - \frac{(h-h')\varphi(u^*) +
\log
(1+e^{h'}) - \log(1+e^h)}{2} \biggr|\le\frac{K'(\beta, h)}{\sqrt{n}}.$$ Now $h = \log\frac{\varphi(0)}{1-\varphi(0)}$ and $h' = \log\frac
{\varphi(u^*)}{1-\varphi(u^*)}$. Also, $\log(1+e^h) = -\log
(1-\varphi
(0))$ and $\log(1+e^{h'}) = - \log(1-\varphi(u^*))$. Substituting these in the above expression, we get $$\biggl|\frac{\log{\mathbb{P}}(Y\in A)}{n^2} - \frac{-I(\varphi
(u^*),\varphi
(0))}{2} \biggr|\le\frac{K'(\beta,h)}{\sqrt{n}}.$$ This completes the proof of the lemma.
We are now ready to finish the proof of Theorem \[pressure\].
[Proof of Theorem \[pressure\]]{} Note that by adding the terms in (\[eq12\]), (\[eq3\]) and (\[eq4\]) from the proof of Lemma \[large1\], and applying the triangle inequality, we get $$\biggl|\frac{\log Z_n(\beta,h)}{n^2} - \frac{(h-h')\varphi(u)}{2}
-\frac
{\beta\varphi(u)^3}{6} - \frac{1}{2}\log(1+e^{h'}) \biggr|\le\frac
{K(\beta,h)}{\sqrt{n}}.$$ This can be rewritten as $$\biggl|\frac{\log Z_n(\beta,h)}{n^2} + \frac{I(\varphi(u),\varphi
(0)) +
\log(1-\varphi(0))}{2} -\frac{\beta\varphi(u)^3}{6} \biggr|\le
\frac
{K(\beta,h)}{\sqrt{n}}.$$ This completes the proof of Theorem \[pressure\].
Note that the proof of Theorem \[pressure\] contains a proof for the lower bound in the general case. We provide the proof below for completeness.
[Proof of Lemma \[lem:lbd\]]{} Fix any $r\in(0,1)$. Define the set $B_{r}$ as $$B_{r}= \biggl\{x\in\{0,1\}^n\dvtx \biggl|\frac{T(x)}{n} - \frac{n^2r^3}{6}
\biggr|\le K(r) n^{3/2}, \biggl|E(x) - \frac{n^2r}{2} \biggr|\le K(r)
n^{3/2} \biggr\},$$ where $K(r)$ is chosen in such a way that ${\mathbb{P}}(Z\in B_{r})\ge1/2$ where $Z=((Z_{ij}))_{i<j}$ and $Z_{ij}$’s are i.i.d. Bernoulli$(r)$. From the proof of Lemma \[large1\], it is easy to see that $$\biggl|\log\sum_{x\in B_r}e^{{\beta T(x)}/{n} + hE(x)} - \frac
{n^{2}}{2} \biggl( (h-h')r +\frac{\beta r^3}{3} +\log(1+e^{h'})
\biggr) \biggr| \le K'n^{3/2},$$ where $h'=\log\frac{r}{1-r}$ and $K'$ is a constant depending on $\beta
,h,r$. Simplifying, we have $$\begin{aligned}
\label{eq:lowerbd}
\frac{2}{n^{2}}\log Z_{n}(\beta,h)&\ge&\frac{2}{n^{2}}\log\sum
_{x\in
B_r}e^{{\beta T(x)}/{n} + hE(x)} \nonumber\\[-8pt]\\[-8pt]
&\ge&\frac{\beta r^3}{3} +\log(1-p) -I(r,p) -
\frac{K'}{\sqrt{n}}\nonumber\end{aligned}$$ for all $r$ where $p=e^{h}/(1+e^{h})$. Now taking limit as $n\to\infty$ and maximizing over $r$ we have the first inequality (\[eq:lbd\]). Given $\beta,h$, define the function $$f(r)=\frac{\beta r^3}{3} +\log(1-p) -I(r,p),$$ where $p=e^{h}/(1+e^{h})$. One can easily check that $f'(r)\gtreqqless
0$ iff $\varphi(u)^{2}-u\gtreqqless0$ for $u=r^{2}$. From this fact, the second equality follows.
\[large2\] Let $T_n$ be the number of triangles in the Erdős–Rényi graph $G(n, \varphi(0))$. Then there is a constant $K(\beta,h)$ depending only on $\beta$ and $h$ such that for all $n$ $$\frac{\log{\mathbb{P}}(T_n \ge{n \choose3}\varphi(u^*)^3)}{n^2} \le
\frac
{-I(\varphi(u^*),\varphi(0))}{2} + \frac{K(\beta,h)}{\sqrt{n}}.$$
By Markov’s inequality, we have $$\frac{\log{\mathbb{P}}(T_n \ge{n\choose3}\varphi(u^*)^3)}{n^2} \le
-\frac
{\beta}{n^3}\pmatrix{n \cr3}\varphi(u^*)^3 + \frac{{\mathbb{E}}(e^{\beta
T_n/n})}{n^2}.$$ From the last part of Theorem \[pressure\], it is easy to obtain an optimal upper bound of the second term on the right-hand side, which finishes the proof of the lemma.
[Proof of Theorem \[largedev\]]{} Given $p$ and $r$, if for all $r'$ belonging to a small neighborhood of $r$ there exist $\beta$ and $h$ satisfying the conditions of Theorem \[pressure\] such that $\varphi(0) = p$ and $\varphi(u^*) = r'$, then a combination of Lemma \[large1\] and Lemma \[large2\] implies the conclusion of Theorem \[largedev\]. If $p\ge p_{0}=2/(2+e^{3/2})$, we can just choose $h\ge h_{0}=-\log2 - 3/2$ such that $p = e^h/(1+e^h)$ and conclude, from Theorem \[pressure\], Lemma \[large1\] and Lemma \[lem:pcond\], that the large deviations limit holds for any $\beta
\ge0$. Varying $\beta$ between $0$ and $\infty$, it is possible to get for any $r \ge p$ a $\beta$ such that $\varphi(u^*) = r$.
For $p\le p_{0}$, we again choose $h$ such that $\varphi(0) = p$. Note that $h\le h_{0}$. The large deviations limit should hold for any $r\ge
p$ for which there exists $\beta>0$ such that $r = \varphi(u^*) =
\sqrt
{u^*}$ and $(h,\beta)\in S$. It is not difficult to verify that given $h$, $u^*$ is a continuously increasing function of $\beta$ in the regime for which $(h,\beta)\in S$. Recall the settings of Lemma \[lem:pcond\]. Thus, the values of $r$ that is allowed is in the set $(p,p_{*})\cup(p^{*},1]$, where $p^{*},p_{*}$ are the unique nontouching solutions to the equations $$\sqrt{p^{*}}=\frac{e^{\beta_{*}(h)p^{*} +h}}{1+e^{\beta_{*}(h)p^{*}
+h}},\qquad
\sqrt{p_{*}}=\frac{e^{\beta^{*}(h)p_{*} +h}}{1+e^{\beta^{*}(h)p_{*}+h}}.$$ This completes the proof of Theorem \[largedev\].
Finally, let us round up by proving Lemma \[lem:pcond\].
[Proof of Lemma \[lem:pcond\]]{} Fix $h\in{\mathbb{R}}$. Define the function $$\psi(x;h,\beta):= \varphi(x;h,\beta)^{2}-x,$$ where $$\varphi(x;h,\beta) = \frac{e^{\beta x +h}}{1+e^{\beta x +h}} \qquad\mbox{for }
x\in[0,1].$$ For simplicity, we will omit $\beta,h$ in $\varphi(x;\beta,h)$ and $\psi
(x;\beta,h)$ when there is no chance of confusion. Note that $\psi
(0)>0>\psi(1)$. Hence, the equation $\varphi(x;\beta,h)=0$ has at least one solution. Also we have $\psi'(x) = 2\beta\varphi
(x)^{2}(1-\varphi
(x))-1$ and $\varphi$ is strictly increasing. Hence, the equation $\psi
'(x) =0$ has at most three solutions. So either the function $\psi$ is strictly decreasing or there exist two numbers $0<a<b<1$ such that $\psi
$ is strictly decreasing in $[0,a]\cup[b,1]$ and strictly increasing in $[a,b]$. From the above observations, it is easy to see that the equation $\psi(x)=0$ has at most three solutions for any $\beta,h$. If $\psi(x)=0$ has exactly two solutions, then $\psi'=0$ at one of the solution.
Let $u_{*}=u_{*}(h,\beta)$ and $u^{*}=u^{*}(h,\beta)$ be the smallest and largest solutions of $\psi(x;h,\beta)=0$, respectively. If $u_{*}=u^{*}$, we have a unique solution of $\psi(x)=0$. From the fact that $\frac
{\partial}{\partial\beta}\psi(x;h,\beta)>0$ for all $x\in
[0,1],\beta
\ge0,h\in
{\mathbb{R}}$, we can deduce that given $h$, $u_{*}(h,\beta)$ and $u^{*}(h,\beta)$ are increasing functions of $\beta$. Note that $u_{*}$ is left continuous and $u^{*}$ is right continuous in $\beta$ given $h$. Also note that given $h\in{\mathbb{R}}$, $u^{*}=u_{*}$ if $\beta>0$ is very small or very large. So, we can define $\beta_{*}(h)$ and $\beta^{*}(h)$ such that for $\beta<
\beta
_{*}(h)$ and for $\beta> \beta^{*}(h)$ we have $u_{*}(h,\beta
)=u^{*}(h,\beta)$. $\beta_{*}$ is the largest and $\beta^{*}$ is the smallest such number.
Therefore, we can deduce that at $\beta=\beta_{*}(h),\beta^{*}(h)$ the equation $\psi(x;h,\beta)=0$ has exactly two solutions. Thus, we have two real numbers $x_{*},x^{*}\in[0,1]$ such that $$\varphi(x)^{2}=x \quad\mbox{and}\quad 2\beta\varphi(x)^{2}\bigl(1-\varphi(x)\bigr)=1$$ for $(x,\beta)=(x_{*},\beta_{*})$ or $(x^{*},\beta^{*})$. Thus, we have $2\beta
x(1-\sqrt{x})=1$ and $$h= \log\frac{\sqrt{x}}{1-\sqrt{x}} - \frac{1}{2(1-\sqrt{x})}$$ for $x=x_{*},x^{*}$. Define $a_{*}= x_{*}^{-1/2}-1$ and $a^{*}=
(x^{*})^{-1/2}-1$. Note that $x=(1+a)^{-2}, \beta=(1+a)^{3}/2a^{2}$ for $(x,a,\beta)=(x_{*},a_{*},\beta_{*})$ or $(x^{*},a^{*},\beta^{*})$ and we have $$\label{eq:h}
h= -\log a -\frac{1+a}{2a}$$ for $a=a_{*},a^{*}$. Now the function $g(x) = -\log x -({1+x})/{2x}$ is strictly increasing for $x\in(0,1/2]$ and strictly decreasing for $x\ge
1/2$. So (\[eq:h\]) has no solution for $h\ge g(1/2)= \log2
-3/2=:h_{0}$. For $h<h_{0}$, equation (\[eq:h\]) has exactly two solutions and for $h=h_{0}$ equation (\[eq:h\]) has one solution. One can easily check that $\beta_{*}\le\beta^{*}$ implies that $a_{*}\le
a^{*}$. Also from the fact that (\[eq:h\]) has at most two solutions, we have that for $\beta\in(\beta_{*},\beta^{*})$ the equation $\psi(u)=0$ has exactly three solutions.
Proof of Lemma 13
-----------------
For simplicity, we will prove the result only for the lower boundary part, that is, for $(h,\beta)=\gamma(t)$ with $t<1/2$. The proof for the upper boundary is similar. Fix $t<1/2$. Let us briefly recall the setup. The function $\psi(u)=\varphi(u)^{2}-u$ has two roots at $0<u^{*}<v^{*}<1$ and $\psi'(u_{*})<0$ while $\psi'(v^{*})=0, \psi
''(v^{*})<0$. See Figure \[fig:boundary\] for the graph of the function $\psi$ when $t=1/4$.
![The function $\psi(\cdot)$ for $(h,\beta)=\gamma(1/4)$.[]{data-label="fig:boundary"}](542f05.eps)
Define the function $$f(r)=\frac{\beta r^3}{3} +\log(1-p) -I(r,p) \qquad\mbox{for } r\in(0,1).$$ From the proof of Lemma \[lem:lbd\] and the fact that $\psi'(u)< 0$ for $u\in(u^{*},v^{*})$, it is easy to see that $f(\varphi
(u^{*}))>f(\varphi(v^{*}))$ and $$\label{bdlower}
\frac{2}{n^{2}}\log Z_{n}(\beta,h)
\ge f(\varphi(u^{*})) - \frac{K}{\sqrt{n}},$$ where $K$ depends on $\beta,h$. Now, using the same idea used in the proof of Lemma \[mean2\], we have $${\mathbb{P}}(
\Delta\ge t ) \le n^{2}\exp\biggl(- \frac{nt^{2}}{8(1+\beta)}
\biggr)$$ for all $t\ge8\beta/n$ and $
\psi(L_{\max})\ge-\Delta, \psi(L_{\min})\le\Delta
$ where $$\Delta= \max_{1\le i<j\le n} \biggl|L_{ij} - \frac{1}{n}\sum_{k\notin\{
i,j\}} \varphi(L_{ik})\varphi(L_{jk}) \biggr|.$$ Hence, there exists ${\varepsilon}_{0}>0,c>0$ such that whenever $\Delta<{\varepsilon}_{0}$ we have $L_{\min}\ge u^{*}-c\Delta$ and either $L_{\max}\le
u^{*}+c\Delta
$ or $|L_{\max}-v^{*}|\le c\sqrt{\Delta}$. Define $$U=\{L_{\max}< (u^{*}+v^{*})/2\}.$$ Then again using the idea used in Lemma \[mean2\] one can easily show that $${\mathbb{E}}({\mathbh{1}}_{U}\cdot|L_{ij}-u^{*}|)\le\frac{K(\beta,h)}{n^{1/2}}
\qquad\mbox{for
all }i<j.$$ We will show that ${\mathbb{P}}(U^{c})\le(\log n)^{2}/n$ and it will imply that $${\mathbb{E}}(|L_{ij}-u^{*}|)\le{\mathbb{E}}({\mathbh{1}}_{U}\cdot|L_{ij}-u^{*}|) + {\mathbb{P}}(U^{c})\le
\frac{K(\beta,h)}{n^{1/2}} \qquad\mbox{for all }i<j.$$ Then the rest of the assertions follow using the steps in the proof of Theorem \[large1\].
Hence, let us concentrate on the event $U^{c}$. It is enough to restrict to the event $U^{c}\cap\{|L_{\max}-v^{*}|\le c\sqrt{\Delta
}\}
\cap
\{L_{\min}\ge u^{*}-c\Delta\}$. Here, the rough idea is that, a large fraction of $L_{ij}$’s has to be near $v^{*}$ in order to make $L_{\max
}\simeq v^{*}$. Suppose $L_{\max}=L_{i_{0}j_{0}}$. Define the set $$A=\{k\dvtx L_{i_0k}<L_{\max}-\delta_{1}\},$$ where $\delta_{1}$ will be chosen later such that $\delta_{1}+c\sqrt
{\Delta
}<v^{*}-u^{*}$. Note that $\varphi(u)^{2}\le\max\{u,u^{*}\}$ for all $u$ and by assumption $|L_{\max}-v^{*}|\le c\sqrt{\Delta}$. Thus, $\varphi
(L_{ij})\le\sqrt{L_{\max}}$ for all $i,j$ and $\varphi
(L_{i_{0}k})\le
\sqrt{L_{\max}-\delta_{1}}\le\sqrt{L_{\max}}(1-\delta_{1}/2)$ for $k\in A$. Thus, we have $$L_{\max}=L_{i_{0}j_{0}}\le\Delta+ \frac{1}{n}\sum_{k\neq
i_{0},j_{0}}\varphi(L_{i_{0}k})\varphi(L_{j_{0}k})
\le\Delta+ L_{\max} - \frac{|A|\delta_{1}}{2n},$$ which clearly implies that $
\frac{|A|}{n}\le\frac{2\Delta}{\delta_{1}}.
$ Similarly, define the set $
A_{j}=\{k\dvtx L_{jk}<L_{\max}-\delta_{2}\}
$ where $\delta_{2}$ will be chosen later such that $\delta_{2}+c\sqrt
{\Delta
}<v^{*}-u^{*}$. Using the same idea as before, for $j\notin A$ we have $$L_{\max}-\delta_{1}\le L_{i_{0}j}\le\Delta+ L_{\max} - \frac
{|A_{j}|\delta_{2}}{2n}
\quad\mbox{or}\quad \frac{|A_{j}|}{n}\le\frac{2(\Delta+\delta
_{1})}{\delta_{2}}:=M\qquad
(\mbox{say}).$$ Choose $\delta_{2}=\Delta^{1/5}, \delta_{1}=\Delta^{3/5}$. Then we have $$\begin{aligned}
\sum_{i<j}|L_{ij}-L_{\max}|^{2}
&\le&\frac{n|A|+ nM + n^{2}\delta_{2}^{2}}{2} \\
&\le&\frac{n^{2}\Delta}{\delta_{1}} + \frac{n^{2}(\Delta+\delta
_{1})}{\delta_{2}} +
\frac{n^{2}\delta_{2}^{2}}{2} \\
&\le& 4n^{2}\Delta^{2/5}.\end{aligned}$$ Thus, by symmetry and Hölder’s inequality, we have $$\begin{aligned}
\label{vstar}
{\mathbb{E}}({\mathbh{1}}_{U^{c}}\cdot|L_{ij}-v^{*}|^{2}) &\le& K{\mathbb{E}}({\mathbh{1}}_{U^{c}}\cdot
\Delta
^{2/5}) \le K{\mathbb{P}}(U^{c})^{9/10}\cdot{\mathbb{E}}(\Delta^{4})^{1/10}
\nonumber\\[-8pt]\\[-8pt]
&\le& \frac{K(\log n)^{1/5}}{n^{1/5}}{\mathbb{P}}(U^{c})^{9/10}\nonumber\end{aligned}$$ for some constant $K$. Now using Lemma \[lem:ex\] and (\[vstar\]) we have $$\begin{aligned}
\label{ucomp}
{\mathbb{E}}\biggl[\biggl| E({\mathbf{X}}) - \frac{n^{2}\varphi(v^{*})}{2}\biggr|
\Big|U^{c} \biggr]&\le&\frac{Cn^{9/5}(\log
n)^{1/5}}{{\mathbb{P}}(U^{c})^{1/10}}\quad\mbox{and}\nonumber\\[-8pt]\\[-8pt]
{\mathbb{E}}\biggl[\biggl|\frac{T({\mathbf{X}})}{n} - \frac{n^{2}\varphi
(v^{*})^{3}}{6}\biggr| \Big| U^{c} \biggr] &\le&\frac
{Cn^{9/5}(\log
n)^{1/5}}{{\mathbb{P}}(U^{c})^{1/10}}.\nonumber\end{aligned}$$ If ${\mathbb{P}}(U^{c})>(\log n)^{2}/n$, from inequality (\[ucomp\]) we have $${\mathbb{P}}\biggl( \biggl| E({\mathbf{X}}) - \frac{n^{2}\varphi(v^{*})}{2}\biggr|\ge
Kn^{19/10} \big| U^{c} \biggr)\le\frac{1}{4}$$ and $${\mathbb{P}}\biggl( \biggl| \frac{T({\mathbf{X}})}{n} - \frac
{n^{2}\varphi
(v^{*})^{3}}{6}\biggr|\ge Kn^{19/10} \big| U^{c} \biggr)\le
\frac{1}{4}$$ for some large constant $K$ depending on $\beta,h$. Now define the set $$B= \biggl\{x\in\{0,1\}^n\dvtx \biggl|\frac{T(x)}{n} - \frac{n^2\varphi
(v^{*})^3}{6} \biggr|\le Kn^{19/10}, \biggl|E(x) - \frac{n^2\varphi
(v^{*})}{2} \biggr|\le Kn^{19/10} \biggr\}.$$ Using the same idea used in the proof of Lemma \[large1\], one can again show that $$\biggl|\frac{2}{n^{2}}\log(Z_{n}{\mathbb{P}}(U^{c})) - f(\varphi(v^{*}))
\biggr|\le\frac{K}{n^{1/10}}$$ for some constant $K$ depending on $\beta,h$. The crucial fact is that ${\mathbb{P}}(\{L_{\max}({\mathbf{Z}})>(u^{*}+v^{*})/2\}\cap\{{\mathbf{Z}}\in B\})$ is bounded away from zero when ${\mathbf{Z}}=((Z_{ij}))_{i<j}\sim$ G$(n,\varphi(v^{*}))$. Thus, we have $$\biggl|\frac{2}{n^{2}}\log Z_{n} - f(\varphi(v^{*})) \biggr|\le
\frac
{K}{n^{1/10}}.$$ But this leads to a contradiction, since by (\[bdlower\]) we have $$\frac{2}{n^{2}}\log Z_{n}(\beta,h)
\ge f(\varphi(u^{*})) - \frac{K}{\sqrt{n}}$$ and $f(\varphi(u^{*})) > f(\varphi(v^{*})) $. Thus, we have ${\mathbb{P}}(U^{c})\le(\log n)^{2}/n$ and we are done.
Proof of Theorem 15
-------------------
The proof is almost an exact copy of the proof of Theorem \[pressure\]. Recall the definition of $L_{ij}$, $$L_{ij}:= \frac{N({\mathbf{X}}_{(i,j)}^{1})-N({\mathbf{X}}_{(i,j)}^{0})}{(n-2)_{{\mathbf{v}}_{F}-2}} \qquad\mbox{for } i<j.$$ In fact, we can write $L_{ij}$ explicitly as a horrible sum $$L_{ij}= \frac{1}{\alpha_{F}(n-2)_{{\mathbf{v}}_{F}-2}}\mathop{\sum
_{t_{1}<t_{2}<\cdots<t_{{\mathbf{v}}_{F}-2}}}_{t_{l}\in[n]\setminus\{i,j\}
\ \mathrm{for}\ \mathrm{all}\ l} \sum_{(a,b)\in E(F)}{\sum_{\pi}}^{\prime}
\mathop{\prod_{(k,l)\in E(F)}}_{(k,l)\neq(a,b)} X_{\pi_{k}\pi_{l}},$$ where the sum $\sum'$ is over all one-one onto map $\pi$ from $V(F)=[{\mathbf{v}}_F]$ to $\{a,b,t_{1},\ldots,t_{{\mathbf{v}}_{F}-2}\}$ where $\{
\pi
(a),\pi(b)\}=\{i,j\}$. Now, we briefly state the main steps. First, we have ${\mathbb{E}}(X_{ij}{|}\mbox{rest})=\varphi(L_{ij})$. Moreover, using Lemma \[techlmm\] it is easy to see that $|{\mathbb{E}}( \prod_{j=1}^{k}
X_{i_{2j-1}i_{2j}}{|}\mbox{rest})-\prod_{j=1}^{k}\varphi
(L_{i_{2j-1}i_{2j}})|\le C\beta/n$ for every distinct pairs $(i_{1},i_{2}),\ldots,(i_{2k-1},i_{2k})$ where $C$ is an universal constant.
Now, fix $1\le i<j\le n$. Given a configuration ${\mathbf{X}}$, construct another one ${\mathbf{X}}'$ in the following way. Choose ${\mathbf{v}}_{F}-2$ distinct points uniformly at random without replacement from $[n]\setminus\{
i,j\}
$. Replace the coordinates in ${\mathbf{X}}$ corresponding to the edges in the complete subgraph formed by the chosen points including $i,j$ (except that we do not change $X_{ij}$) by values drawn from the conditional distribution given the rest of the edges. Call the new configuration ${\mathbf{X}}'$. Define the antisymmetric function $
F({\mathbf{X}},{\mathbf{X}}'):= (n-2)_{{\mathbf{v}}_{F}-2}(L_{ij}-L'_{ij}).
$ and $f({\mathbf{X}}):={\mathbb{E}}(F({\mathbf{X}},{\mathbf{X}}'){|}{\mathbf{X}})$. Using the same idea as before and Theorem \[thm:conc\], we have $${\mathbb{P}}( \vert L_{ij} - g_{ij} \vert\ge t ) \le\exp\bigl(
-cnt^{2}/(1+\beta)\bigr),$$ where $c$ is an absolute constant and $g_{ij}$ is obtained from $L_{ij}$ by replacing $X_{kl}$ by $\varphi(L_{kl})$ for all $k<l$. Note that there is a slight difference with the calculation in the triangle case, since we have to consider collections of edges where some are modified and some are not. But their contribution will be of the order of $n^{-1}$. Also the conditions on $\varphi$ arises in the following way, if all the $L_{ij}$’s are constant, say equal to $u$, then from the “mean-field equations” for $L_{ij}$’s we must have $$\begin{aligned}
u&\approx&\frac{1}{\alpha_{F}(n-2)_{{\mathbf{v}}_{F}-2}}\mathop{\sum
_{t_{1}<t_{2}<\cdots<t_{{\mathbf{v}}_{F}-2}}}_{t_{l}\in[n]\setminus\{i,j\}
\ \mathrm{for}\ \mathrm{all}\ l} \sum_{(a,b)\in E(F)}{\sum_{\pi}}' \varphi
(u)^{{\mathbf{e}}_{F}-1}\\
&=& \frac{2{\mathbf{e}}_{F}}{\alpha_{F}} \varphi(u)^{{\mathbf{e}}_{F}-1}.\end{aligned}$$
The next step is to show that under the conditions on $\varphi$, we have ${\mathbb{E}}|L_{ij}-u^{*}|\le Kn^{-1/2}$ for all $i<j$ where $K=K(\beta,h)$ is a constant depending only on $\beta,h$. The crucial fact is that the behavior of the function $\varphi(u)^k-au$ where $a>0$ is a positive constant and $k\ge2$ is a fixed integer, is same as the behavior of the function $\varphi(u)^2-u$.
Now it will follow (using the same proof used for Lemma \[lem:ex\]) that $${\mathbb{E}}\biggl\vert E({\mathbf{X}}) - \frac{n^{2}\varphi(u^{*})}{2} \biggr\vert\le Cn^{3/2}$$ and $${\mathbb{E}}\biggl\vert N({\mathbf{X}}) - \frac{(n)_{{\mathbf{v}}_{F}}\varphi(u^{*})^{{\mathbf{e}}_{F}}}{\alpha_{F}} \biggr\vert\le Cn^{{\mathbf{v}}_{F}-1/2},$$ where $C$ is a constant depending only on $\beta,h$. The rest of the proof follows using the arguments used in the proof of Theorem \[pressure\].
[Proof of Theorem \[largedevgen\]]{} Using the method of proof for the triangle case and the result from Theorem \[pressuregen\], the proof follows easily.
[Proof of Lemma \[lem:pcondgen\]]{} The proof is same as the proof of Lemma \[lem:pcond\] except for the constants.
Proof of Theorem 17
-------------------
Suppose ${\bolds{\sigma}}$ is drawn from the Gibbs distribution $\mu_{\beta
,h}$. We construct ${\bolds{\sigma}}'$ by taking one step in the heat-bath Glauber dynamics as follows: choose a position $I$ uniformly at random from $\Omega$, and replace the $I$th coordinate of ${\bolds{\sigma}}$ by an element drawn from the conditional distribution of the $\sigma_I$ given the rest. It is easy to see that $({\bolds{\sigma}},{\bolds{\sigma}}')$ is an exchangeable pair. Let $$F({\bolds{\sigma}},{\bolds{\sigma}}') := |\Omega|\bigl(m({\bolds{\sigma}})-m({\bolds{\sigma}}')\bigr)=\sigma_{I}-\sigma'_{I}$$ be an antisymmetric function in ${\bolds{\sigma}},{\bolds{\sigma}}'$. Since the Hamiltonian is a simple explicit function, one can easily calculate the conditional distribution of the spin of the particle at position $x$ given the spins of the rest. In fact, we have ${\mathbb{E}}(\sigma_{x}| \{\sigma_{y}, y\neq x\}])= \tanh
(2\beta d
m_{x}({\bolds{\sigma}}))$ where $m_{x}({\bolds{\sigma}}) := \frac{1}{2d}\sum_{y\in
N_{x}}\sigma
_{y}$ is the average spin of the neighbors of $x$ for $x\in\Omega$. Now, using Fourier–Walsh expansion we can write the function $\tanh(2\beta d
m_{x}({\bolds{\sigma}}))$ as sums of products of spins in the following way. We have $$\label{eq:fwalsh}
\tanh(2d\beta m_{x}({\bolds{\sigma}})) = \sum_{k=0}^{2d}a_k(\beta)\sum
_{|S|=k,S\subseteq N_x}\sigma_{S},$$ where $$a_{k}(\beta):= \frac{1}{2^{2d}}\sum_{{\bolds{\sigma}}\in\{-1,+1\}^{2d}} \tanh
\Biggl(\beta\sum_{i=1}^{2d}\sigma_{i} \Biggr)\prod_{j=1}^{k}\sigma_{j}$$ for $k=0,1,\ldots,2d$. It is easy to see that $a_{k}(\beta)=0$ if $k$ is even and $a_{k}(\beta)$ is a rational function of $\tanh(2\beta)$ if $k$ is odd. Note that the dependence of $a_{k}$ on $d$ is not stated explicitly. Thus, using (\[eq:fwalsh\]) and the definitions in (\[def:rk\]) we have $$\begin{aligned}
f({\bolds{\sigma}}) &=& {\mathbb{E}}[F({\bolds{\sigma}},{\bolds{\sigma}}')|{\bolds{\sigma}}] = \frac{1}{|\Omega|}\sum
_{x\in\Omega}
E[ \sigma_{x} -\sigma'_{x} |{\bolds{\sigma}}]\\
&=& m({\bolds{\sigma}}) - \frac{1}{|\Omega|}\sum_{x\in\Omega}\tanh(2\beta d
m_{x}({\bolds{\sigma}}))\\
&=& \bigl(1-2da_{1}(\beta)\bigr)m({\bolds{\sigma}}) - \sum_{k=1}^{d-1}\pmatrix{2d\cr
2k+1}a_{2k+1}(\beta) r_{2k+1}({\bolds{\sigma}}).\end{aligned}$$ Define $\theta_{k}(\beta):={2d\choose2k+1}a_{2k+1}(\beta)$ for $k=0,1,\ldots,d-1$. Note that we can explicitly calculate the value of $\theta_{0}(\beta)$ as follows: $$\theta_{0}(\beta)= \frac{1}{4^{d}} \sum_{{\bolds{\sigma}}\in\{-1,+1\}^{2d}} \tanh\Biggl(\beta
\sum
_{i=1}^{2d}\sigma_{i} \Biggr)\sum_{i=1}^{2d}\sigma_{i}
= \frac{2}{4^{d}}\sum_{k=1}^{d}2k\pmatrix{2d\cr d+k} \tanh(2k\beta).$$ Now, we have $|F({\bolds{\sigma}},{\bolds{\sigma}}')|\le2$ and $$|f({\bolds{\sigma}}) - f({\bolds{\sigma}}')| \le\frac{2}{|\Omega|} \Biggl( |1-\theta
_{0}(\beta)| +
\sum_{k=1}^{d-1}(2k+1)\theta_{k}(\beta) \Biggr) = \frac{2}{|\Omega|}
b(\beta)$$ for all values of ${\bolds{\sigma}},{\bolds{\sigma}}'$. Hence, the condition of Theorem \[thm:conc\] is satisfied with $B=0$, $C=2|\Omega|^{-1}b(\beta)$. So by part (ii) of Theorem \[thm:conc\], we have $${\mathbb{P}}\Biggl( \sqrt{|\Omega|} |\bigl(1-\theta_{0}(\beta)\bigr)m({\bolds{\sigma}}) - \sum
_{k=1}^{d-1}\theta_{k}(\beta) r_{2k+1}({\bolds{\sigma}}) | \ge t
\Biggr)\le
2\exp\biggl( - \frac{t^{2} }{4b(\beta) } \biggr)$$ for all $t>0$. Obviously, $\theta_{0}(\cdot)$ is a strictly increasing function of $\beta
$. Also, we have $\theta_{0}(0)=0$ and $$\theta_{0}(\infty):=\lim_{\beta\to\infty}\theta_{0}(\beta) =
\frac
{1}{4^{d-1}}\sum_{k=1}^{d}k\pmatrix{2d\cr d+k}.$$ For $d=1$, we have $\theta_{0}(\infty)=1$ and for $d\ge2$ we have $$\begin{aligned}
\theta_{0}(\infty)&\ge&\frac{1}{4^{d-1}} \Biggl[ 2\sum
_{k=1}^{d}\pmatrix{2d\cr d+k} - \pmatrix{2d\cr d+1} \Biggr]\\
&=& \frac{1}{4^{d-1}} \biggl[ 2^{2d} - \pmatrix{2d\cr d}-\pmatrix{2d\cr
d+1} \biggr]\\
&=& 4 - \frac{8}{2^{2d+1}}\pmatrix{2d+1\cr d+1}\end{aligned}$$ and from the fact that $\sum_{k=d-1}^{d+2}{2d+1\choose k}\le2^{2d+1}$ we have $$\frac{1}{2^{2d+1}}\pmatrix{2d+1\cr d+1} \le\frac{d+2}{4(d+1)}\le
\frac
{1}{3} \qquad\mbox{for }d\ge2.$$ Hence, for $d\ge2$ we have $\theta_{0}(\infty)>1$ and there exists $\beta
_{1}\in(0,\infty)$, depending on $d$, such that $1-\theta_{0}(\beta)>0$ for $\beta<\beta_{1}$ and $1-\theta_{0}(\beta)<0$ for $\beta>\beta
_{1}$. This completes the proof.
Proof of Proposition 19
-----------------------
The proof is almost same as the proof of Proposition \[pr:ising\]. Define ${\bolds{\sigma}},{\bolds{\sigma}}'$ as before. Define the antisymmetric function $F({\bolds{\sigma}},{\bolds{\sigma}}')$ as follows: $$\begin{aligned}
F({\bolds{\sigma}},{\bolds{\sigma}}') :\!&=& |\Omega|\bigl(1+\tanh(h)\tanh(2\beta d m_{I}({\bolds{\sigma}}))\bigr)\bigl(m({\bolds{\sigma}})-m({\bolds{\sigma}}')\bigr)\\
&=&\bigl(1+\tanh(h)\tanh(2\beta d m_{I}({\bolds{\sigma}}))\bigr)(\sigma_{I}-\sigma'_{I}).\end{aligned}$$ Recall that $m_{x}({\bolds{\sigma}}) := \frac{1}{2d}\sum_{y\in N_{x}}\sigma
_{y}$ is the average spin of the neighbors of $x$ for $x\in\Omega$. Now under $\mu
_{\beta,h}$, we have $$\begin{aligned}
{\mathbb{E}}(\sigma_{x}| \{\sigma_{y}, y\neq x\}) &=& \tanh\bigl(2\beta d
m_{x}({\bolds{\sigma}})+h\bigr)\\
&=&\frac{\tanh(h)+ \tanh(2\beta d m_{x}({\bolds{\sigma}}))}{1+\tanh(h)\tanh
(2\beta d
m_{x}({\bolds{\sigma}}))}.\end{aligned}$$ Thus, we have $$\begin{aligned}
f({\bolds{\sigma}}) &=& {\mathbb{E}}(F({\bolds{\sigma}},{\bolds{\sigma}}')|{\bolds{\sigma}}) \\
&=& \frac{1}{|\Omega|}\sum_{x\in\Omega}\bigl(1+\tanh(h)\tanh(2\beta d
m_{x}({\bolds{\sigma}}))\bigr)
{\mathbb{E}}( \sigma_{x} -\sigma'_{x} |{\bolds{\sigma}})\\
&=& m({\bolds{\sigma}}) -\tanh(h)+ \frac{1}{|\Omega|}\sum_{x\in\Omega}\bigl(\tanh
(h)\sigma
_{x}-1\bigr)\tanh(2\beta d m_{x}({\bolds{\sigma}})).\end{aligned}$$ After some simplifications and using the definitions of the functions, $r,s$ we have $$\begin{aligned}
f({\bolds{\sigma}})&=& \bigl(1-\theta_{0}(\beta)\bigr)m({\bolds{\sigma}}) - \sum_{k=1}^{d-1}\theta
_{k}(\beta
) r_{2k+1}({\bolds{\sigma}})\\
&&{} -\tanh(h) \Biggl(1-\sum_{k=0}^{d-1}\theta_{k}(\beta)
s_{2k+1}({\bolds{\sigma}}) \Biggr).\end{aligned}$$ Now for all values of ${\bolds{\sigma}},{\bolds{\sigma}}'$ we have $$|f({\bolds{\sigma}}) - f({\bolds{\sigma}}')|\le\frac{2}{|\Omega|} b(\beta)(1+\tanh|h|)$$ and the proof henceforth is exactly as in the proof of Proposition \[pr:ising\].
Proof of Theorem 2
------------------
Assume that $\psi(0)>0$. We will handle the case $\psi(0)=0$ later. Note that condition (\[eq:al\]) implies that $x^\alpha/\psi(x)$ is a nondecreasing function for $x>0$. Define the function $$\varphi(x):=\frac{x^2}{\psi(x)} \quad\mbox{and}\quad \gamma(x):= 2-\frac
{x\psi
'(x)}{\psi(x)} \qquad\mbox{for } x\neq0$$ and $\varphi(0)=0,\gamma(0)=2$. Clearly, we have $2-\alpha\le\gamma(x)\le2$ for all $x\in{\mathbb{R}}$. Now, $\limsup_{x\to0}\varphi(x)\le\lim_{x\to0+}x^{2-\alpha}/\psi
(1)=0=\varphi(0)$ as $\alpha<2$. Also $\varphi(x)$ is differentiable in ${\mathbb{R}}\setminus\{0\}$ with $$\label{eq:lphi}
\varphi'(x)= \frac{x \gamma(x)}{\psi(x)} > 0 \qquad\mbox{for } x\neq0.$$ Hence, $\varphi$ is absolutely continuous in ${\mathbb{R}}$ and is increasing for $x\ge0$.
Define $Y=f(X)$. First, we will prove that all moments of $\varphi(Y)$ are finite. Next, we will estimate the moments which will in turn show that $\varphi(Y)^{1/2}$ has finite exponential moment in ${\mathbb{R}}$. Finally, using Chebyshev’s inequality we will prove the tail probability.
By monotonicity of $\psi$ in $[0,\infty)$ and definition of $\alpha$, we have $$0\le\frac{x\psi'(x)}{\psi(x)} \le\alpha\qquad\mbox{for all } x\ge0.$$ It also follows from (\[eq:lphi\]) that $0\le(\log\varphi(x))'\le
2/x$ for $x>0$ and integrating we have $\varphi(x)\le\varphi(1)x^2$ for all $x\ge1$. Hence, $\varphi(x)=\varphi(|x|)\le\varphi(1)(1+x^2)$ for all $x\in{\mathbb{R}}$ and this, combined with our assumption that ${\mathbb{E}}(|f(X)|^{k}) < \infty$ for all $k\ge1$, implies that $
{\mathbb{E}}(\varphi(Y)^k)<\infty\mbox{ for all } k\ge1.
$
Define $$\beta:= \biggl\lceil\frac{5(2-\alpha)+\delta+1/4}{(2-\alpha)^2}
\biggr\rceil\ge3.$$ Fix an integer $k\ge\beta$ and define $$g(x)=\frac{x^{2k-1}}{\psi^k(x)} \quad\mbox{and}\quad h(x)=\frac
{x^{2k-2}}{\psi
^k(x)} \qquad\mbox{for } x\in{\mathbb{R}}.$$ Clearly, ${\mathbb{E}}(|Yg(Y)|)<\infty$. Note that $g,h$ are continuously differentiable in ${\mathbb{R}}$ as $k\ge3$. Moreover, for $x\in{\mathbb{R}}$ we have, $|g'(x)|= h(x) \vert k\gamma(x)-1 \vert\le(2k-1)h(x)$, $ h'(x) =
(k\gamma(x)
-2 ) {x^{2k-3}}/{\psi^k(x)}$ and $$h''(x)= \bigl[ \bigl(k\gamma(x) -2 \bigr) \bigl(k\gamma(x)
-3 \bigr) + kx\gamma
'(x)\bigr]\frac{x^{2k-4}}{\psi^k(x)}.$$ We also have $$x\gamma'(x)=-\frac{x\psi'(x)}{\psi(x)} \biggl(1-\frac{x\psi
'(x)}{\psi
(x)} \biggr)-\frac{x\psi''(x)}{\psi(x)}\ge-1/4-\delta$$ for $x\in{\mathbb{R}}$. Now $k\ge\beta$ implies that $$\begin{aligned}
&& \bigl(k\gamma(x) -2 \bigr) \bigl(k\gamma(x) -3 \bigr) + kx\gamma
'(x)\\
&&\qquad
\ge\bigl(k(2-\alpha) -2 \bigr) \bigl(k(2-\alpha) -3 \bigr) - k(\delta
+1/4)\ge0\end{aligned}$$ for all $x$. Thus, $h''(x)\ge0$ for all $x$ and $h$ is convex in ${\mathbb{R}}$.
Let $X',F(X,X')$ be as given in the hypothesis. Define $Y'=f(X')$. Recall that $(X,X')$ is an exchangeable pair and so is $(Y,Y')$. Using the fact that $f(X)={\mathbb{E}}(F(X,X')|X)$ almost surely, exchangeability of $(X,X')$ and antisymmetry of $F$, we have $$\begin{aligned}
\label{eq:ln1}
{\mathbb{E}}(Yg(Y))&=&{\mathbb{E}}(f(X)g(Y))={\mathbb{E}}(F(X,X')g(Y))\nonumber\\[-8pt]\\[-8pt]
&=&\tfrac{1}{2} {\mathbb{E}}\bigl(F(X,X')\bigl(g(Y)-g(Y')\bigr) \bigr).\nonumber\end{aligned}$$ Now, for any $x< y$ we have $$\biggl\vert\frac{g(x)-g(y)}{x-y} \biggr\vert=\biggl\vert\int_0^1 g'\bigl(tx+(1-t)y\bigr) \,dt
\biggr\vert\le(2k-1)
\int_0^1 h\bigl(tx+(1-t)y\bigr) \,dt$$ and convexity of $h$ implies that $$\int_0^1 h\bigl(tx+(1-t)y\bigr) \,dt \le\int_0^1 \bigl(t h(x)+(1-t)h(y)\bigr) \,dt =\bigl(h(x)+h(y)\bigr)/2.$$ Hence, from (\[eq:ln1\]), we have $$\begin{aligned}
\label{eq:gdh}
{\mathbb{E}}(Yg(Y))&\le&\frac{2k-1}{4} {\mathbb{E}}\bigl(|(Y-Y')F(X,X')|\bigl(h(Y)+h(Y')\bigr)\bigr)
\nonumber\\[-8pt]\\[-8pt]
&=&(2k-1){\mathbb{E}}(\Delta(X)h(Y)) \le(2k-1){\mathbb{E}}(\psi(Y)h(Y)),\nonumber\end{aligned}$$ where the equality follows by definition of $\Delta(X)$ and exchangeability of $(Y,Y')$. Thus, for any $k\ge\beta$ we have, from (\[eq:gdh\]), $$\label{eq:rec}
{\mathbb{E}}(\varphi(Y)^k) \le
(2k-1){\mathbb{E}}(\varphi(Y)^{k-1}).$$ Using induction for $k\ge\beta$, we have $${\mathbb{E}}(\varphi(Y)^k) \le\frac{(2k)! 2^\beta\beta!}{2^k k!(2\beta
)!}{\mathbb{E}}(\varphi
(Y)^{\beta}) \qquad\mbox{for } k\ge\beta.$$ Also Hölder’s inequality applied to (\[eq:rec\]) for $k=\beta$ implies that ${\mathbb{E}}(\varphi(Y)^{\beta}) \le(2\beta-1)^\beta$. Thus, we have $$\label{eq:rec2}
{\mathbb{E}}(\varphi(Y)^k)\le\cases{
\dfrac{(2k)! 2^\beta\beta!}{k! 2^k(2\beta)!}{\mathbb{E}}(\varphi(Y)^{\beta
}), &\quad if $k>
\beta$,\vspace*{2pt}\cr
(2\beta-1)^k, &\quad if $0\le k\le\beta$.}$$ Note that we have $e^{x}\le e^{x}+e^{-x}=2\sum_{k\ge0}x^{2k}/(2k)!$ for all $x\in{\mathbb{R}}$. Combining everything, we finally have $$\begin{aligned}
{\mathbb{E}}(\exp(\theta\varphi(Y)^{1/2}))
&\le&2\sum_{k=0}^\infty\frac{\theta^{2k} }{(2k)!}{\mathbb{E}}(\varphi
(Y)^{k})\\
&\le&\frac{2^{\beta+1} \beta!}{(2\beta)!}{\mathbb{E}}(\varphi(Y)^{\beta})
\sum
_{k=\beta
}^\infty\frac{\theta^{2k} }{2^k k!} +
\sum_{k=0}^{\beta-1} \frac{2(2\beta-1)^k\theta^{2k}}{(2k)!} \\
&\le& C_{\beta}\exp(\theta^{2}/2)\end{aligned}$$ for all $\theta\ge0$ where the constant $C_{\beta}$ is given by $$C_{\beta}:= \max\biggl\{ \frac{ 2(2\beta-1)^{k} 2^{k} k! }{(2k)!} \Big|
0\le k \le\beta\biggr\}.$$ Here, we used the fact that $(2k)!\ge2^{2k-1}k!^2/k$. Now recall that $\varphi$ is an increasing function in $[0,\infty)$. Thus, using Chebyshev’s inequality for $\exp(\theta\varphi(x)^{1/2})$ with $\theta
=\varphi(t)^{1/2}$ we have $${\mathbb{P}}\bigl(|f(X)|\ge t\bigr)\le C_{\beta}e^{-\theta\varphi(t)^{1/2}+\theta^{2}/2}=
C_{\beta}e^{-\varphi(t)/2}.$$
Now suppose that $\psi(0)=0$. For ${\varepsilon}>0$ fixed, define $\psi_{\varepsilon}(x)=\psi(x)+{\varepsilon}$. Clearly, we have $\Delta(X)\le\psi_{{\varepsilon}}(f(X))$ a.s. and $\psi_{{\varepsilon}}$ satisfies all the other properties of $\psi$ including $$x\psi'_{{\varepsilon}}(x)/\psi_{{\varepsilon}}(x) = x\psi'(x)/\psi(x) \cdot{\psi
(x)}/{\bigl(\psi(x)+{\varepsilon}\bigr)}\le\alpha$$ and $$x\psi''_{{\varepsilon}}(x)/\psi_{{\varepsilon}}(x) = x\psi''(x)/\psi(x)
\cdot{\psi(x)}/{\bigl(\psi(x)+{\varepsilon}\bigr)}\le\delta$$ for all $x>0$. Hence, all the above results hold for $\psi_{\varepsilon}$ and $\varphi_{\varepsilon}(x)=x^{2}/\psi_{{\varepsilon}}(x)$. Now $\varphi_{\varepsilon}\uparrow\varphi$ as ${\varepsilon}\downarrow0$. Letting ${\varepsilon}\downarrow0$, we have the result.
When $\psi$ is once differentiable with $\alpha<2$, it is easy to see that the function $h$ is nondecreasing (need not be convex) in $[0,\infty)$ for $k\ge\beta:= \lceil2/(2-\alpha)\rceil$. In that case, we have $$\int_{0}^{1}h\bigl(tx+(1-t)y\bigr)\,dy\le\max_{z\in[x,y]}h(z)\le h(x)+h(y)$$ for $x\le y$. Hence, we have the recursion $${\mathbb{E}}(\varphi(Y)^k) \le
2(2k-1){\mathbb{E}}(\varphi(Y)^{k-1})$$ for $k\ge\beta$. Using the same proof as before, it then follows that $${\mathbb{P}}\bigl(|f(X)|\ge t\bigr)\le Ce^{-\varphi(t)/4},$$ where $C$ depends only on $\alpha$.
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors thank Amir Dembo, Erwin Bolthausen, Ofer Zeitouni and Persi Diaconis for various helpful discussions and comments. They would also like to thank an anonymous referee for a careful reading of this article and for constructive criticism that resulted in an improved exposition.
[39]{}
, (). . .
, (). . .
, (). . In . , .
(). . .
(). . .
(). , ed. . , .
(). . .
, (). .
, (). . .
(). . [arXiv:math/0507526](http://arxiv.org/abs/arXiv:math/0507526).
(). . .
(). . .
(). [arXiv:0907.4450](http://arxiv.org/abs/arXiv:0907.4450).
, , (). . .
(). . .
(). [arXiv:0908.1909](http://arxiv.org/abs/arXiv:0908.1909).
(). . .
(). . .
(). . . , .
, (). . .
(). . .
(). .
(). . .
, (). . , .
, (). . .
(). . .
(). . .
(). . In . . , .
(). . . , .
(). . .
(). . .
(). . .
(). . .
(). . .
(). . .
(). . In . , .
(). . , .
(). . .
(). . .
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'For a connected locally $(G,s)$-arc-transitive graph $\Ga$ with $s\geqslant 2$ and an edge $\{v,w\}$, determining the amalgam $(G_v,G_w,G_{vw})$ is a fundamental problem in the area of symmetrical graph theory, but it is very difficult. In this paper, we give a classification of $(G_v,G_w,G_{vw})$ in the case where the vertex stabilisers $G_v$ and $G_w$ are faithful on their neighbourhoods, which shows that except for the case $G_v\cong G_w$, there are exactly 16 such triples.'
address: |
School of Mathematics and Statistics\
The University of Western Australia\
Crawley, WA 6009, Australia
author:
- Shu Jiao Song
title: 'On the Stabilisers of Locally $2$-Transitive Graphs'
---
[^1] [^2]
Introduction {#Intro}
============
Let $\Ga=(V,E)$ be a connected undirected simple graph. An [*$s$-arc*]{} of $\Ga$ is an $(s+1)$-tuple $(v_0,v_1,\dots,v_s)$ of vertices such that $\{v_{i-1},v_i\}\in E$ for $1\le i\le s$ and $v_{i-1}\not=v_{i+1}$ for $1\le i\le s-1$. For a group $G\le\Aut\Ga$, $\Ga$ is called [*locally $(G,s)$-arc-transitive*]{} if, for any vertex $v\in V$, the vertex stabiliser $G_v$ acts transitively on the set of $t$-arcs starting at $v$, for any $t$ with $1{\leqslant}t{\leqslant}s$. A locally $(G,s)$-arc transitive graph is called [*$(G,s)$-arc transitive*]{} if $G$ is also transitive on the vertex set. Let $\Ga(v)$ be the set of vertices adjacent to $v$, and $G_v^{\Ga(v)}$ the permutation group on $\Ga(v)$ induced by $G_v$.
For a connected locally $(G,s)$-arc-transitive graph $\Ga$ and an edge $\{v,w\}$, the triple $(G_v,G_w,G_{vw})$ is called the [*amalgam*]{} of $G$, and also of $\Ga$. A fundamental problem for studying locally $s$-arc-transitive graphs is to determine their amalgams. van Bon [@vanBon-2] proved that if $G_{vw}^{[1]}=1$ then $\Ga$ is not locally $(G,s)$-arc-transitive. Potocnik [@Potocnik] determined the amalgams for valency $\{3,4\}$ in the case where $G_{vw}^{[1]}=1$. This paper is one of a series of papers which aim to classify the amalgams $(G_v,G_w,G_{vw})$ with trivial edge kernel.
Assume that $\Ga$ is locally $(G,s)$-arc-transitive with $s\ge2$. For a vertex $v$ of $\Ga$, denote by $G_v^{[1]}$ the kernel of $G_v$ acting on $\Ga(v)$. Then $G_v^{\Ga(v)}\cong G_v/G_v^{[1]}$, and $G_v^{\Ga(v)}$ is a 2-transitive permutation group. As usual, denote by $X^{(\infty)}$ the smallest normal subgroup of $X$ such that $X/X^{(\infty)}$ is soluble. Let $\Ra(H)$ be the soluble radical of a group $H$, that is, the largest soluble normal subgroup of $H$.
Here we generalise a classical result for primitive permutation groups, see [@Wielandt Section 18], which shows that some information of $(G_v,G_w,G_{vw})$ can be obtained from the permutation groups $G_v^{\Ga(v)}$ and $G_w^{\Ga(w)}$.
\[key-lem\] Let $\Ga$ be a connected $G$-edge-transitive graph, where $G\leqslant\Aut\Ga$. Then, for an edge $\{v,w\}$ of $\Ga$, the following statements hold.
- Each composition factor of $G_v$ is a composition factor of $G_v^{\Ga(v)}$, $G_{vw}^{\Ga(w)}$ or $G_{vw}^{\Ga(v)}$.
- Each composition factor of $G_{vw}$ is a composition factor of $G_{vw}^{\Ga(w)}$ or $G_{vw}^{\Ga(v)}$.
\[faith-stab\] Let $\Ga$ be a connected locally $(G,2)$-arc-transitive graph, and let $\{v,w\}$ be an edge of $\Ga$. Assume that $G_v^{[1]}=G_w^{[1]}=1$. Then either $G_v\cong G_w$ and $\Ga$ is regular, or $(G_v,G_w,G_{vw})$ is one of the sixteen triples listed in Table $\ref{Intro}$.
$$\begin{array}{|c ccccccc |}\hline
&G_v&G_w&G_{vw}&|\Ga(v)|&|\Ga(w)|&G_{w_{-1}vw}&G_{vwv_1} \\ \hline
1 & 3^2{:}\SL_2(3) & 5^2{:}\SL_2(3) & \SL_2(3) & 3^2 & 5^2 & \ZZ_3 & 1\\
2 & 3^2{:}\GL_2(3) & 5^2{:}\GL_2(3) & \GL_2(3) & 3^2 & 5^2 & \S_3& \ZZ_2\\
3 & 3^2{:}\GL_2(3) & 7^2{:}\GL_2(3) & \GL_2(3) & 3^2 & 7^2 & \S_3 & 1\\
4 & 5^2{:}\GL_2(3) & 7^2{:}\GL_2(3) & \GL_2(3) & 5^2 & 7^2 & \ZZ_2 & 1 \\
\hline
5&\A_6&\PSL_2(11)&\A_5&6&11&\A_4&\D_6 \\
6&\PSL_2(11) &2^4{:}\A_5&\A_5&11&16&\D_6&\ZZ_2^2 \\
7&\A_6&2^4.\A_5&\A_5&6&16&\A_4&\ZZ_2^2 \\
8&\S_6&2^4.\S_5&\S_5&6&16&\S_4&\D_8 \\
9&\A_7&2^4.\A_6&\A_6&7&16&\A_5&\S_4 \\
10&\S_7&2^4.\S_6&\S_6&7&16&\S_5&2\times\S_4 \\
11&\A_8&2^4.\A_7&\A_7&8&16&\A_6&\PSL_2(7) \\
12&\A_9&2^4.\A_8&\A_8&9&16&\A_7&2^3{:}\GL_3(2) \\
13& \S_9 & \Sp_6(2) &\S_8 & 9 &36 & \S_7 & (\S_4\times\S_4).2\\ \hline
14&\A_7 & 2^3{:}\GL_3(2) & \PSL_2(7) & 15 & 8 & \A_4 & \S_4\\
15&5^2{:}\SL_2(5)&{11}^2{:}\SL_2(5)&\SL_2(5)&5^2&11^2&\ZZ_5&1\\
16&{13}^2{:}\SL_2(13)&3^6{:}\SL_2(13)&\SL_2(13)&13^2&3^6&\ZZ_{13}&\ZZ_3\\
\hline
\end{array}$$
**Table \[Intro\]**
0.1in
\[pair-2-trans\] Let $P$ and $Q$ be two $2$-transitive permutation groups which are not isomorphic but their stabilisers are isomorphic to the same group $H$. Then $(P,Q,H)$ is one of the triples $(G_v,G_w,G_{uw})$ in Table $\ref{Intro}$.
For locally 3-arc-transitive graphs, there are only six of such amalgams.
\[faith-stab-3-trans\] Let $\Ga$ be a connected locally $(G,3)$-arc-transitive, and let $\{v,w\}$ be an edge of $\Ga$. Assume that $G_v^{[1]}=G_w^{[1]}=1$. Then $(G_v,G_w,G_{vw})$ is one of the following triples: $(\A_7,\A_7,\A_6),\ (\S_7,\S_7,\S_6),\ (\A_7,2^4{:}\A_6,\A_6),\ (\S_7,2^4{:}\S_6,\S_6), (\A_8,2^4{:}\A_7,\A_7),$ $ (\A_9,2^4{:}\A_8,\A_8),$ or $\ (\S_9,\Sp_6(2),\S_8).$
Examples
========
Let $H$ be a 2-transitive permutation group of degree $n$, and let $p$ be a prime which is coprime to the order $|H|$. Let $V=\FF_p^n$ be the permutation module of $H$ over $\FF_p$. Then $H\leqslant\GL_n(p)$. Let $\s$ be the unique involution in $\Z(\GL_n(p))$. Then $\s$ reverses every vector in $V$, and $\l \s,H\r=\l \s\r\times H$ since $\s$ centralizes $H$ and $H$ is a 2-transitive permutation groups. Define a group . $$X=\ZZ_p^n{:}(\l \s\r\times H).$$ Let $a_1,a_2,\dots,a_n$ be a basis for $V=\ZZ_p^n$. Then $a_i^\s=a_i$, and $H$ is 2-transitive on the basis. Let $$g=a_1\s.$$ Then $g$ is an involution. Let $$S=\{a_1\s,a_2\s,\dots,a_n\s\},$$ and let $R=\l S\r$. Then $(a_i\s)(a_j\s)=a_ia_j\in R$, and $H$ acts 2-transitively on $S$. Now $H$ induces a subgroup of $\Aut(R)$. Let $\Ga=\Cay(R,S)$, and $$G=R{:}H.$$ Thus $\Ga$ is a $(G,2)$-arc-transitive Cayley graph of $R$ with vertex stabiliser $H$.
Any $2$-transitive permutation group is the vertex stabiliser of $2$-arc-transitive graph.
[Let $(G_v,G_w,G_{vw})$ be one of the triples listed in Table \[Intro\] except for rows 5, 6, 10, 12 and 13. Then there exists a group $G$ such that $G=G_vG_w$, and thus there is a complete bipartite graph $\Ga=\K_{m,n}$ which is locally $(G,2)$-arc-transitive. These examples given in the next table. ]{}
$$\begin{array}{|c ccc c |}\hline
G & G_v&G_w&G_{vw}& \Ga \\ \hline
(3^2\times 5^2){:}\SL_2(3)& 3^2{:}\SL_2(3) & 5^2{:}\SL_2(3) & \SL_2(3) & \K_{3^2,5^2}\\
(3^2\times 5^2){:}\GL_2(3)& 3^2{:}\GL_2(3) & 5^2{:}\GL_2(3) & \GL_2(3) & \K_{3^2,5^2}\\
(3^2\times 7^2){:}\GL_2(3) & 3^2{:}\GL_2(3) & 7^2{:}\GL_2(3) & \GL_2(3) & \K_{3^2,7^2}\\
(5^2\times 7^2){:}\GL_2(3) & 5^2{:}\GL_2(3) & 7^2{:}\GL_2(3) & \GL_2(3) & \K_{5^2,7^2} \\
2^4{:}\A_6&\A_6&2^4.\A_5&\A_5&\K_{6,2^4} \\
2^4{:}\S_6&\S_6&2^4.\S_5&\S_5&\K_{6,2^4} \\
2^4{:}\A_7&\A_7&2^4.\A_6&\A_6&\K_{7,2^4} \\
2^4{:}\A_8&\A_8&2^4.\A_7&\A_7&\K_{8,2^4} \\
\A_8 &\A_7 & 2^3{:}\GL_3(2) & \GL_3(2) & \K_{7,15}\\
(5^2\times11^2){:}\SL_2(5)&5^2{:}\SL_2(5)&{11}^2{:}\SL_2(5)&\SL_2(5)&\K_{5^2,11^2}\\
(13^2\times 3^6){:}\SL_2(13)&{13}^2{:}\SL_2(13)&3^6{:}\SL_2(13)&\SL_2(13)&\K_{13^2,3^6}\\
\hline
\end{array}$$
0.1in [**Problem.**]{} Construct explicit examples of graphs with the amalgams in row 5, 6, 10, 12, or 13 in Table \[Intro\]. 0.1in
[It is well-known that the triple $(\A_7,\A_7,\A_6)$ is the amalgam for both the complete graph $\K_8$ and Hoffman-Singleton graph. The former is 2-arc-transitive but not 3-arc-transitive, but the latter is 3-arc-transitive. The same phenomena appears for $(\S_7,\S_7,\S_6)$. ]{}
Composition factors of the stabilisers {#B-pty}
======================================
Let $\Ga=(V,E)$ be a connected $G$-edge-transitive bipartite graph with biparts $U$ and $W$. For vertices $v_0,v_1,\dots,v_\ell$ of $\Ga$, let $$G_{v_0v_1\dots v_\ell}^{[1]}=G_{v_0}^{[1]}\cap G_{v_1}^{[1]}\cap\dots\cap G_{v_\ell}^{[1]}.$$ Then, for an edge $\{v,w\}$, the group $G_{vw}^{[1]}$ is the point-wise stabiliser of the subset of vertices $\Ga(v)\cup\Ga(w)$.
\[chain-1\] Let $v_0,v_1,\dots,v_\ell$ be an $\ell$-arc, where $\ell$ is a positive integer. Then we have a chain of normal subgroups: $$G_{v_0v_1\dots v_\ell}^{[1]}\lhd G_{v_0v_1\dots v_{\ell-1}}^{[1]}
\lhd\dots\lhd G_{v_0v_1}^{[1]}\lhd G_{v_0}^{[1]}\lhd G_{v_0}.$$
[*Proof. *]{}
For each integer $i$ with $0\leqslant i\leqslant\ell$, the group $G_{v_0v_1\dots v_i}^{[1]}$ is the kernel of $G_{v_0v_1\dots v_i}$ acting on $\Ga(v_0)\cup\Ga(v_1)\cup\dots\cup\Ga(v_i)$, and hence $$G_{v_0v_1\dots v_i}^{[1]}\lhd G_{v_0v_1\dots v_i}.$$ For $i\geqslant 1$, we have the inclusion relation $G_{v_0v_1\dots v_i}^{[1]}\leqslant G_{v_0v_1\dots v_{i-1}}^{[1]}\leqslant G_{v_0v_1\dots v_i}$, and we thus conclude that $G_{v_0v_1\dots v_i}^{[1]}\lhd G_{v_0v_1\dots v_{i-1}}^{[1]}$. This holds for all possible values of $i$ with $1\leqslant i\leqslant\ell$, giving the chain of normal subgroups as claimed.
\[chain-2\] Let $\{v,w\}$ be an edge, and $v_0,v_1,\dots,v_\ell$ be an $\ell$-arc, where $\ell$ is a positive integer. Then for $0\leqslant i\leqslant \ell-1$, the factor group $G_{v_0v_1\dots v_i}^{[1]}/G_{v_0v_1\dots v_{i+1}}^{[1]}$ is isomorphic to a subnormal subgroup of $G_{vw}^{\Ga(v)}$ or $G_{vw}^{\Ga(w)}$.
[*Proof. *]{}
Observing that $G_{v_0v_1\dots v_i}^{[1]}\lhd\lhd G_{v_i}^{[1]}\lhd G_{v_iv_{i+1}}$, we obtain $$G_{v_0v_1\dots v_i}^{[1]}/G_{v_0v_1\dots v_{i+1}}^{[1]}
\cong (G_{v_0v_1\dots v_i}^{[1]})^{\Ga(v_{i+1})}
\lhd\lhd G_{v_iv_{i+1}}^{\Ga(v_{i+1})}.$$ Since $G$ is edge-transitive on $\Ga$, we have $G_{v_iv_{i+1}}\cong G_{vw}$, and so $$\mbox{ $G_{v_iv_{i+1}}^{\Ga(v_{i+1})}\cong G_{vw}^{\Ga(w)}$ or $G_{vw}^{\Ga(v)}$.}$$ Therefore, $G_{v_0v_1\dots v_i}^{[1]}/G_{v_0v_1\dots v_{i+1}}^{[1]}$ is isomorphic to a subnormal subgroup of $G_{vw}^{\Ga(v)}$ or $G_{vw}^{\Ga(w)}$.
0.1in [**Proof of Theorem \[key-lem\]:**]{} First of all, we claim that a composition factor of $G_{v}^{[1]}$ is isomorphic to a composition factor of $G_{vw}^{\Ga(v)}$ or $G_{vw}^{\Ga(w)}.$ Let $T$ be a composition factor of $G_v^{[1]}$. Since $G\leqslant\Aut\Ga$ and $\Ga$ is finite and connected, there exists a path $v_0=v,v_1,\dots,v_\ell$ such that $G_{v_0v_1\dots v_\ell}^{[1]}=1$. By Lemma \[chain-1\], we have a chain of normal subgroups: $$1=G_{v_0v_1\dots v_\ell}^{[1]}\lhd G_{v_0v_1\dots v_{\ell-1}}^{[1]}
\lhd\dots\lhd G_{v_0v_1}^{[1]}\lhd G_{v_0}^{[1]}=G_v^{[1]}.$$
Let $i$ be the largest integer with $0\leqslant i\leqslant\ell$ such that $T$ is a composition factor of $G_{v_0v_1\dots v_i}^{[1]}$. Then $T$ is not a composition factor of $G_{v_0v_1\dots v_{i+1}}^{[1]}$. Hence $T$ is a composition factor of the quotient group $G_{v_0v_1\dots v_i}^{[1]}/G_{v_0v_1\dots v_{i+1}}^{[1]}.$ By Lemma \[chain-2\], $G_{v_0v_1\dots v_i}^{[1]}/G_{v_0v_1\dots v_{i+1}}^{[1]}$ is isomorphic to a subnormal subgroup of $G_{vw}^{\Ga(w)}$ or $G_{vw}^{\Ga(v)}$, and so $T$ is a composition factor of $G_{vw}^{\Ga(w)}$ or $G_{vw}^{\Ga(v)}$ as we claimed.
Now let $T_1$ be a composition factor of $G_v$. If $T_1$ is a composition factor of $G_v/G_v^{[1]}\cong G_v^{\Ga(v)}$, then we are done. We thus assume that $T_1$ is not a composition factor of $G_v^{\Ga(v)}\cong G_v/G_v^{[1]}$. Then $T_1$ is a composition factor of $G_v^{[1]}$, and so $T_1$ is a composition factor of $G_{vw}^{\Ga(w)}$ or $G_{vw}^{\Ga(v)}$. This proves part (i).
Similarly, let $T_2$ be a composition factor of $G_{vw}$. If $T_2$ is a composition factor of $G_{vw}/G_v^{[1]}\cong G_{vw}^{\Ga(v)}$, then we are done. Now we assume that $T_2$ is a composition factor of $G_v^{[1]}$, and so $T_1$ is a composition factor of $G_{vw}^{\Ga(w)}$ or $G_{vw}^{\Ga(v)}$. Thus part (ii) of Theorem \[key-lem\] hold.
\[key-lem-2\] Let $\Ga$ be a connected graph, and let $G\le\Aut\Ga$ be transitive on the edge set. Let $\{v,w\}$ be an edge of $\Ga$. Then the following statements hold:
- The vertex stabiliser $G_v$ is soluble if and only if $G_v^{\Ga(v)}$ and $G_{vw}^{\Ga(w)}$ are both soluble.
- $G_{vw}$ is soluble if and only if $G_{vw}^{\Ga(v)}$ and $G_{vw}^{\Ga(w)}$ are both soluble; in particular.
- $G_v$, $G_w$ are both soluble if and only if $G_v^{\Ga(v)}$ and $G_w^{\Ga(w)}$ are both soluble.
- $G_v^{[1]}$ and $G_w^{[1]}$ are both soluble if and only if $(G_v^{[1]})^{\Ga(w)}$ and $(G_w^{[1]})^{\Ga(v)}$ are both soluble.
[*Proof. *]{}
Part (i), and part (ii) follow from Theorem \[key-lem\](1), and (2), respectively.
For part (iii), it is obvious that if $G_v$, $G_w$ are both soluble then$G_v^{\Ga(v)}$ and $G_w^{\Ga(w)}$ are both soluble. So it is sufficient to show if both $G_v^{\Ga(v)}$ and $G_w^{\Ga(w)}$ is soluble, then both $G_v$ and $G_w$ are soluble. Suppose that $G_v$ is insoluble. Let $T$ be an insoluble composition factor of $G_v$. By Theorem \[key-lem\] (1), $T$ is a composition factor of $G_v^{\Ga(v)}$ or $G_{vw}^{\Ga(w)}$. Similarly, if $G_w$ is insoluble, then either $G_w^{\Ga(w)}$ or $G_{vw}^{\Ga(v)}$ is insoluble. It implies that if both $G_v^{\Ga(v)}$ and $G_w^{\Ga(w)}$ is soluble, then both $G_v$ and $G_w$ are soluble. This proves part (iii).
Similarly, to prove part [iv]{}, we need only to prove if both $(G_v^{[1]})^{\Ga(w)}$ and $(G_w^{[1]})^{\Ga(v)}$ are soluble then so are $G_v^{[1]}$ and $G_w^{[1]}$. Suppose that $G_v^{[1]}$ is insoluble. Arguing as in the second paragraph with $G_v^{[1]}$ in the place of $G_v$ shows that either $(G_v^{[1]})^{\Ga(w)}$ or $(G_{vw}^{[1]})^{\Ga(v)}$ is insoluble. Similarly, if $G_w^{[1]}$ is insoluble then either $(G_w^{[1]})^{\Ga(v)}$ or $(G_{vw}^{[1]})^{\Ga(w)}$ is insoluble. It implies that if both $(G_v^{[1]})^{\Ga(w)}$ and $(G_w^{[1]})^{\Ga(v)}$ are soluble, then so are $G_v^{[1]}$ and $G_w^{[1]}$.
The next lemma refines Lemma \[chain-2\].
\[chain-3\] Let $\{u,w\}$ be an edge with $u\in U$ and $w\in W$, and let $$u_0,w_0,\dots,w_{i-1},u_i,w_i,u_{i+1},\dots$$ be a path such that $u_0=u$ and $w_0=w$. Then for $i\geqslant 0$, the following are true:
- $G_{u_0\dots u_i}^{[1]}/G_{u_0\dots u_iw_i}^{[1]}$ is isomorphic to a subnormal subgroup of $G_{uw}^{\Ga(w)}$.
- $G_{u_0\dots w_i}^{[1]}/G_{u_0\dots w_iu_{i+1}}^{[1]}$ is isomorphic to a subnormal subgroup of $G_{uw}^{\Ga(u)}$.
[*Proof. *]{}
We notice that $$\begin{array}{l}
G_{u_0\dots u_i}^{[1]}/G_{u_0\dots u_iw_i}^{[1]}\cong (G_{u_0\dots u_i}^{[1]})^{\Ga(w_i)}\lhd\lhd G_{u_iw_i}^{\Ga(w_i)}\cong G_{uw}^{\Ga(w)},\\
G_{u_0\dots w_i}^{[1]}/G_{u_0\dots w_iu_{i+1}}^{[1]}\cong (G_{u_0\dots w_i}^{[1]})^{\Ga(u_{i+1})}\lhd\lhd G_{w_iu_{i+1}}^{\Ga(u_{i+1})}\cong G_{uw}^{\Ga(u)}.
\end{array}$$ Then the proof follows.
2-Transitive permutation groups {#2-trans}
===============================
The proof of Theorem \[faith-stab\] depends on the structural properties of stabilisers of 2-transitive permutation groups. We state the point stabilisers in the following two theorems, refer to [@Cameron-book Tables 7.3 and 7.4] , and [@LiSS].
The [*socle*]{} of a group $X$ is the subgroup of $X$ generated by all minimal normal subgroups of $X$, denoted by $\soc(X)$.
\[AS-2-trans\] Let $X$ be an almost simple $2$-transitive permutation group on $\Omega$ of degree $k$, and let $T=\soc(X)$ and $\o,\o'\in\Omega$. Then $X/T\cong X_\o/T_\o\cong X_{\o\o}/T_{\o\o'}\leqslant \calO\leqslant \Out(T)$, and further, they satisfy the following table:
$$\begin{array}{|c|c|c|c|c|}\hline
\mbox{row} & T & T_\o & T_{\o\o'} & k\\ \hline
1& \A_n,n{\geqslant}5 & \A_{n-1} & \A_{n-2} & n \\ \hline
2& \PSL_n(q) & q^{n-1}{:}(\SL_{n-1}(q).{q-1\over(n,q-1)}) & [q^{2(n-2)}]{:}(\GL_{n-2}(q).{q-1\over(n,q-1)})& {q^n-1\over q-1}
\\ \hline
3& \Sz(q), q>2 & [q^2]{:}\ZZ_{q-1} & \ZZ_{q-1} & q^2+1 \\
4& \Ree(q),q>3 & [q^3]{:}\ZZ_{q-1} & \ZZ_{q-1} & q^3+1 \\
5& \PSU_3(q),q{\geqslant}3 & [q^3]{:}\ZZ_{(q^2-1)/(3,q+1)}
& \ZZ_{(q^2-1)/(3,q+1)} & q^3+1 \\ \hline
6& \Sp_{2n}(2), & \POmega_{2n}^-(2).2 & 2^{2n-2}{:}\POmega_{2n-2}^-(2).2
& 2^{2n-1}-2^{n-1} \\
7& n\geqslant3 & \POmega_{2n}^+(2).2 & 2^{2n-2}{:}\POmega_{2n-2}^+(2).2
& 2^{2n-1}+2^{n-1} \\ \hline
8& \PSL_2(11) & \A_5 & \S_3 & 11 \\
9& \M_{11} & \M_{10} & 3^2{:}\Q_8 & 11 \\
10& \M_{11} & \PSL_2(11) & \A_5 & 12 \\
11& \M_{12} & \M_{11} & \M_{10} & 12 \\
12& \A_7 & \PSL_2(7) & \A_4 & 15 \\
13& \M_{22} & \PSL_3(4) & 2^4{:}\A_5 & 22 \\
14& \M_{23} & \M_{22} & \PSL_3(4) & 23 \\
15& \M_{24} & \M_{23} & \M_{22} & 24 \\
16& \PSL_2(8) & \D_{18} & \ZZ_2 & 28 \\
17& \HS & \PSU_3(5).2 & \A_6.2^2 & 176 \\
18& \Co_3 & \McL{:}2 & \PSU_4(3){:}2 & 276 \\ \hline
\end{array}$$
0.08in A complete list of finite 2-transitive permutation groups was obtained by Hering, see for example [@Liebeck]. Here we list their point stabilisers.
\[HA-2-trans\] Let $X=\ZZ_p^e{:}X_\o$ be an affine $2$-transitive permutation group of degree $p^e$ on a set $\Omega$, where $\o\in\Omega$. Then for a point $\o'\in\Omega\setminus\{\o\}$, either
1. $X_\omega\leqslant \GammaL_1(p^e)\cong\ZZ_{p^e-1}{:}\ZZ_e$, and $X_{\o\o'}\lesssim\ZZ_e$, or
2. $X_\o$ and $X_{\o\o'}$ are as in Table $\ref{2-trans}.2$,
[$$\begin{array}{|c|c|c|c|c|c| c| }\hline
\mbox{\rm row} & X_\o & X_{\o\o'} & p^e & \\ \hline
1 & \SL_n(q).\calO & q^{n-1}{:}\GL_{n-1}(q).\calO & q^n & \calO\leqslant\ZZ_{q-1}{:}\ZZ_e\\
2 & \Sp_{2n}(q).\calO & q.q^{2n-2}{:}(\GL_1(q)\times \Sp_{2n-2}(q)).\calO &q^{2n} & \calO\leqslant \ZZ_e\\
3 & \G_2(q).\calO & [q^5]{:}\GL_2(q).\calO & q^6 & \calO\leqslant \ZZ_{q-1}{:}\ZZ_{(3,q)e}\\ \hline
4& \Q_8{:}3, \ \Q_8.6,\ (\Q_8{:}3).4 & 1,\ \ZZ_2, [4]& 5^2& \\
5 &\Q_8{:}\S_3,\ 3\times(\Q_8.2),\ 3\times(\Q_8{:}\S_3) & 1,\ 1, \ \ZZ_3 & 7^2&\\
6&5\times(\Q_8{:}3), \ 5\times(\Q_8{:}\S_3) & 1, \ \ZZ_2 & 11^2&\\
7&{11}\times(\Q_8{:}\S_3) & 1 & 23^2&\\ \hline
8 &2^{1+4}{:}5,2^{1+4}{:}\D_{10},2^{1+4}{:}(5{:}4),& 2, [4], [8],& 3^4 &\\
&2^{1+4}{:}\A_5,2^{1+4}{:}\S_5 &2.\A_4,2.\S_4&& \\
\hline
9&\SL_2(13) & \ZZ_3 & 3^6 &\\
10&\A_6 & \S_4 & 2^4 &\\
11&\A_7 & \PSL_2(7) & 2^4 & \\
12&\PSU_3(3) & 4.\S_4,\ 4^2.\S_3 & 2^6 &\\ \hline
13&\SL_2(5),\ 5\times\SL_2(5) & 1,\ \ZZ_5 & 11^2 &\\
14& 9\times\SL_2(5) & \ZZ_3 &19^2&\\
15&7\times\SL_2(5),\ 7\times(4\circ\SL_2(5)) & 1,\ \ZZ_2 &29^2 &\\
16&{29}\times\SL_2(5) & 1 & 59^2 & \\ \hline
\end{array}$$ ]{}
[**Table \[2-trans\].2**]{}: Affine 2-transitive groups and their stabilisers
\[2-pt-stab\] Let $X$ be a $2$-transitive group on $\Ome$ of degree $k$, and take $\o,\o'\in\Ome$. Then the stabiliser $X_{\o\o'}$ has a transitive representation of degree $k-1$ if and only if $X=\A_7$ or $\S_7$, and $X_{\o\o'}=\A_5$ or $\S_5$, respectively.
[*Proof. *]{}
Suppose $X_{\o\o'}$ has has a transitive permutation representation of degree $k-1$. Then $k-1$ divides $|X_{\o\o'}|$. It is clear that $k-1$ does not divide $|X_{\o\o'}|$ for candidates in rows 3 to 5 and 8 to 18 in Table \[AS-2-trans\] and rows 4 to 16 in Table \[HA-2-trans\].
Now we inspect the other candidates given in Tables \[AS-2-trans\] and \[HA-2-trans\]. Suppose $\soc(X)=\A_k$ as in row 1 of Table \[AS-2-trans\]. Then $X_{\o\o'}=\A_{k-2}$ or $\S_{k-2}$. If $X_{\o\o'}=\A_{k-2}$ or $\S_{k-2}$ has a transitive permutation representation of degree $k-1$, then $k=7$.
Suppose that $\soc(X)=\PSL_n(q)$, with $k={q^n-1\over q-1}$ as in row 2 of Table \[AS-2-trans\]. Suppose further $X_{\o\o'}$ has a transitive permutation representation of degree $k-1=\frac{q(q^{n-1}-1)}{q-1}$. If $q^{n-1}-1$has a primitive prime divisor $r$ say, then $r$ does not divides the order $|X_{\o\o'}|=|[q^{2(n-2)}]{:}(\GL_{n-2}(q).\ZZ_{q-1\over(n,q-1)}).\calO|$, which is not possible. Thus $q^{n-1}-1$ has no primitive prime divisor, and $n-1=2$ or $(q,n-1)=(2,6)$ by Zsigmondy Theorem. For the former, $k-1=q(q+1)$, and $|X_{\o\o'}|=q^2(q-1)^2/(3,q-1)$ which is not divisible by $q+1$ So $k-1=126$, and $X_{\o\o'}=[2^{10}]:GL(5,2)$. But $X_{\o\o'}$ has no transitive permutation representation of degree 126 contradicts to our assumption.
Similarly, let $X=q^n{:}\SL_n(q).\calO$ be as in the first row of Table \[HA-2-trans\]. If $k-1=q^n-1$ divides the order of $X_{\o\o'}=q^{n-1}{:}\GL_{n-1}(q).\calO$, then $q^n-1$ has no primitive prime divisor. Thus, by Zsigmondy Theorem, $n=2$ or $(q,n)=(2,6)$. For the former, $k-1=q^2-1$, and $|X_{\o\o'}|=q(q-1)$ which is not divided by $k-1$. For the latter, $k-1=63$, $X_{\o\o'}=[2^5]:\GL(4,2)$ which has no transitive permutation representation of degree 63.
For candidates in rows 6,7 in Table \[AS-2-trans\], and rows 2,3 in Table \[HA-2-trans\], we have the following table:
[$$\begin{array}{|c|c|c| }\hline
X_\o & X_{\o\o'} & k-1 \\ \hline
\POmega_{2n}^-(2).2.\calO & 2^{n^2-n+1}(2^{n-1}+1)\Pi_{i=1}^{n-2}(2^{2i}-1).\calO
& 2^{2n-1}-2^{n-1} \\
\POmega_{2n}^+(2).2.\calO & 2^{n^2-n+1}(2^{n-1}-1)\Pi_{i=1}^{n-2}(2^{2i}-1).\calO
& 2^{2n-1}+2^{n-1} \\ \hline
\Sp_{2n}(q).\calO & q^{n^2}(q-1)\Pi_{i=1}^{n-1}(q^{2i}-1).\calO &q^{2n}-1 \\
\G_2(q).\calO & q^6(q-1)^2(q+1).\calO & q^6 -1\\ \hline
\end{array}$$ ]{}
It is easy to see that $k-1$ does not divide the order $|X_{\o\o'}|$. So $X_{\o\o'}$ has no transitive permutation representation of degree $k-1$.
Proof of Theorem \[faith-stab\] {#main-thm}
===============================
Let $\Ga$ be a connected locally $(G,2)$-arc-transitive graph, and let $\{v,w\}$ be an edge of $\Ga$. Assume that $G_v^{[1]}=G_w^{[1]}=1$. In this section, we will give the proof of Theorem \[faith-stab\]. We first notice that, since $G_v\cong G_v^{\Ga(v)}$ and $G_w\cong G_w^{\Ga(w)}$, the stabilisers $G_v,G_w$ are 2-transitive permutation groups which share common point stabilisers $G_{vw}$.
To prove this theorem, we will split it into several cases, given in the following few lemmas.
\[isomorphic socle\] Suppose $G_v$ and $G_w$ have isomorphic socles. Then $G_v\cong G_w$ and $\Ga$ is regular.
[*Proof. *]{}
Since $G_v$ and $G_w$ are 2-arc-transitive groups, we have either they are both affine groups or they are both almost simple groups. Suppose $\soc(G_v)\cong\soc(G_w)\cong\ZZ_p^d$, then $G_v\cong \ZZ_p^d{:}G_{vw}\cong G_w$, and $\Ga$ is regular of valency $p^d$. Suppose that $G_v$ and $G_w$ are almost simple and have isomorphic socles. Since they share point stabiliser, we have $G_v\cong G_w$ by checking the candidates in Table \[2-trans\].1.
From now on, we assume that $G_v$ and $G_w$ have non-isomorphic socles. We need to search for different 2-transitive groups which have the same point stabiliser, by inspecting the 2-transitive permutation groups, listed in Tables \[2-trans\].1 and \[2-trans\].2.
\[both soluble\] Suppose both $G_v$ and $G_w$ are soluble. Then $$\begin{array}{ l | l l l l l |}
G_v & 3^2{:}\SL_2(3) & 3^2{:}\GL_2(3) & 3^2{:}\GL_2(3)& 5^2{:}\GL_2(3) \\ \hline
G_w & 5^2{:}\SL_2(3)& 5^2{:}\GL_2(3) & 7^2{:}\GL_2(3) & 7^2{:}\GL_2(3)\\
\end{array}$$
Table \[main-thm\].1: Soluble stabilisers
[*Proof. *]{}
Suppose that $G_v$ and $G_w$ are soluble. Then by inspecting the 2-transitive permutation groups, listed in Tables \[2-trans\].1 and \[2-trans\].2, we have
$$\begin{array}{ ll } \hline
G_{vw} & |\Ga(v)| \\ \hline
\leqslant \ZZ_{p^d-1}{:}\ZZ_d & p^d \\
\GL_2(2) & 2^2 \\
\SL_2(3),\ \GL_2(3) & 3^2 \\
\Q_8{:}3, \ \Q_8.6,\ (\Q_8{:}3).4 & 5^2 \\
\Q_8{:}\S_3,\ 3\times(\Q_8.2),\ 3\times(\Q_8{:}\S_3) & 7^2 \\
5\times(\Q_8{:}3), \ 5\times(\Q_8{:}\S_3) & 11^2 \\
{11}\times(\Q_8{:}\S_3) & 23^2 \\
2^{1+4}{:}5,2^{1+4}{:}\D_{10},2^{1+4}{:}(5{:}4) & 3^4 \\ \hline
\end{array}$$
Table \[main-thm\].2
0.08in We note that the candidate in the second line, i.e., $G_{vw}=\GL_2(2)$, is the same as the one in the first line with $p^d=2^2$ and $G_{vw}=\ZZ_{p^d-1}{:}\ZZ_2\cong\S_3$.
Assume first that $G_{vw}\leqslant\ZZ_{p^d-1}{:}\ZZ_d$ with $|\Ga(v)|=p^d$. Then $G_{vw}$ is a split metacyclic group. So $G_{vw}$ can not appear as one of rows 3-8 of Table \[main-thm\].2.
We thus have that $G_{vw}$ only appears in one of rows 3-8 of Table \[main-thm\].2. It follows that $G_v,G_w$ are as in Table \[main-thm\].1.
\[one insoluble\] Suppose $G_v$ is insoluble. Then $G_{vw}$ is insoluble.
[*Proof. *]{}
Suppose $G_{vw}$ is soluble. Then one of the candidates in the following table appears: $$\begin{array}{ lll ll } \hline
G_v & G_{vw} & |\Ga(v)| & \calO & \\ \hline
\A_5, \S_5 & \A_4,\S_4 & 5 & & \\
\PSL_3(3).\calO & 3^2{:}\SL_2(3).\calO&26 &&\\
\PSL_2(q).\calO & [q]{:}\ZZ_{(q-1)/(2,q-1)}.\calO & q+1 & \calO\leqslant (2,q-1).f & q=p^f\geqslant4\\
\Sz(q).\calO & [q^2]{:}\ZZ_{q-1}.\calO & q^2+1 & \calO\leqslant f & q=2^f>2, f \mbox{ odd}\\
\Ree(q).\calO & [q^3]{:}\ZZ_{q-1}.\calO & q^3+1 & \calO\leqslant f & q=3^f>3, f \mbox{ odd} \\
\PSU_3(q).\calO & [q^3]{:}\ZZ_{(q^2-1)/(3,q+1)} & q^3+1 & \calO\leqslant (3,q+1).f & q{\geqslant}3\\
\PGammaL_2(8) & 9{:}6 & 28 & &\\ \hline
\end{array}$$
Table \[main-thm\].3: Soluble edges stabilisers
0.08in
We note that, in Table \[main-thm\].2, the candidate in the first line is the same as the one in the second line with $q=4$ due to $\A_5\cong\PSL_2(4)$. The rest are mutually non-isomorphic. Thus $G_w$ should be soluble, appearing in Table \[main-thm\].1, and we conclude that in this case $G_v$ and $G_w$ do not have isomorphic stabilisers. So $G_{vw}$ is insoluble.
\[$G_{vw}$ almost simple\] Suppose $G_{vw}$ is almost simple. Then $(G_v,G_w,G_{vw})$ are listed in rows 5 to 14 of Table \[Intro\].
[*Proof. *]{}
First, we observe that $G_{vw}$ does not have socle isomorphic to one of the following groups:
> $\POmega_{2n}^-(2)$ with $n\geqslant3$, $\POmega_{2n}^+(2)$ with $n\geqslant4$, $\M_{11}$, $\M_{22}$, $\M_{23}$, $\PSU_3(5)$, $\McL$, $\Sp_{2n}(q)$ with $(n,q)\not=(2,2)$, $\G_2(q)$, $\PSU_3(3)$,
since each of them appears only one time as the socle of the stabiliser of 2-transitive permutation groups. Similarly, $G_{vw}$ is not $\M_{10}$. This shows that $\soc(G_{vw})$ is an alternating group of a linear group.
[**Case 1.**]{} Let first $\soc(G_{vw})=\A_m$, where $m\geqslant5$. Notice that the isomorphisms between $\A_m$ and other simple groups: $$\begin{array}{l}
\A_5\cong\PSL_2(5)\cong\PSL_2(4) \cong\POmega_4^-(2), \\
\A_6\cong\PSL_2(9)\cong\Sp_4(2)', \\
\A_8\cong\PSL_4(2)\cong\POmega_6^+(2).
\end{array}$$ Now $G_w$ is one of such groups such that $\soc(G_w)\not\cong\soc(G_v)$. It implies that $m\leqslant8$.
Notice that $\PSL_2(4)\cong\SL_2(4)$, $\PSL_4(2)\cong\SL_4(2)$, and $\Sp_4(2)\cong \S_6$, we have $G_v$ and $G_w$ is one of the following groups:
$$\begin{array}{c |c|c|c|c|c|c}
G_v\mbox{:almost simple} & \A_6,\S_6 &\A_7,\S_7& \A_8,\S_8 &\S_9,\S_9& \Sp_4(2)&\PSL_2(11)\\ \hline
G_{vw} & \A_5,\S_5&\A_6,\S_6&\A_7,\S_7&\A_8,\S_8&\POmega_6^+(2)&\A_5\\
\end{array}$$ 0.1in $$\begin{array}{c |c|c|c|c|c}
G_v\mbox{: affine}&4^2{:}\SL_2(4),4^2{:}\SL_2(4).2 &2^4{:}\SL_4(2)&2^4{:}\Sp_4(2),&2^4{:}\A_6&2^4{:}\A_7\\ \hline
G_{vw} &\SL_2(4),\SL_2(4).2&\SL_4(2)&\Sp_4(2)&\A_6&\A_7\\
\end{array}$$
Table \[main-thm\].4: altnating edge stabilisers
Suppose $m=5$. Then $G_v,G_w\rhd\A_6$, $\PSL_2(11)$, $2^4{:}\SL_2(4)$. It then follows that $\{G_v,G_w\}$ is one of the following four pairs, noticing that $\SL_2(4)\cong\A_5$ and $\SL_2(4).2\cong\SigmaL_2(4)\cong\S_5$: $$\mbox{$\{\A_6,\PSL_2(11)\}$, $\{\PSL_2(11),2^4{:}\A_5\}$,
$\{\A_6,2^4{:}\A_5\}$, or $\{\S_6,2^4{:}\S_5\}$,}$$ as listed in rows 5-8 of Table \[Intro\].
If $m=6$, then each of $ G_v$ and $G_w$ is conjugate to $\A_7$, $\S_7$ $2^4{:}\Sp_4(2)$, or $2^4{:}\A_6$, and hence $\{G_v,G_w\}=\{\A_7,2^4{:}\A_6\}$ or $\{\S_7,2^4{:}\S_6\}$, as in rows 9-10 of Table \[Intro\].
If $m=7$, then $G_v\rhd\A_8$ or $2^4{:}\A_7$, and so is $G_w$. This leads to $\{G_v,G_w\}=\{\A_8,2^4{:}\A_7\}$, as in row 11 of Table \[Intro\].
If $m=8$, then each of $G_v$ and $G_w$ is $\A_9$, $\Sp_6(2)$, or $2^4{:}\GL_4(2)$, leading to $\{G_v,G_w\}=\{\A_9,2^4{:}\A_8\}$ or $\{\S_9,\Sp_6(2)\}$, giving in rows 12-13 of Table \[Intro\].
0.1in [**Case 2.**]{} Assume that $\soc(G_{vw})$ is a linear group which is not isomorphic to $\A_m$ with $m\geqslant5$. We notice that $\PSL_2(7)\cong\SL_3(2)\cong \GL_3(2)$. Then $\soc(G_{vw})=\PSL_2(7)$, and $G_v,G_w\rhd\A_7$ or $2^3{:}\GL_3(2)$. Therefore, we conclude that $\{G_v,G_w\}=\{\A_7,2^3{:}\GL_3(2)\}$, which is listed in row 14 of Table \[Intro\].
\[$G_{vw}$ non-almost simple\] Assume $G_{vw}$ is insoluble but not almost simple. Then $(G_v,G_w,G_{vw})$ are listed in rows 15 and 16 of Table \[Intro\].
[*Proof. *]{}
Assume $G_{vw}$ is insoluble but not almost simple. All the possibilities of $G_{vw}$ are listed in the following table. $$\begin{array}{cc c}\hline
G_{vw} & \soc(G_v) & \\ \hline
q^n{:}\SL_n(q).\calO & \PSL_{n+1}(q) &\\
\SL_n(q).\calO & q^n &\\
\Sp_{2n}(q).\calO & q^{2n} & \\
2^{1+4}{:}\A_5,2^{1+4}{:}\S_5 & 2^4 \\
\SL_2(13) & 3^6 \\
\SL_2(5),\ 5\times\SL_2(5) & 11^2 \\
9\times\SL_2(5) &19^2 \\
7\times\SL_2(5),\ 7\times(4\circ\SL_2(5)) &29^2 \\
{29}\times\SL_2(5) & 59^2 \\ \hline
\end{array}$$ The candidate in the first line or the third line is not isomorphic to any other one in the table. This leads to the following possibilities: $$\{G_v,G_w\}=\{5^2{:}\SL_2(5),11^2{:}\SL_2(5)\},\ \mbox{or}\ \{13^2{:}\SL_2(13), 3^6{:}\SL_2(13)\}.$$ These are listed in rows 15-16 of Table \[Intro\].
[**Proof of Theorem \[Intro\]:**]{} By assumption, $G_{v}^{[1]}=G_w^{[1]}=1$, so $G_v\cong G_v^{\Ga(v)}$, $G_w\cong G_w^{\Ga(w)}$, and the stabilisers $G_v,G_w$ are 2-transitive permutation groups which share common point stabilisers $G_{vw}$. Suppose first that $G_v$ and $G_w$ have isomorphic soles. Then by Lemma \[isomorphic socle\], $G_v\cong G_w$ and $\Ga$ is regular.
Now suppose $G_v$ and $G_w$ have non-isomorphic socles. Suppose further that both of them are soluble. Then by Lemma \[both soluble\], $(G_v,G_w,G_{vw})$ are listed in the first four rows of Table \[Intro\]. Suppose $G_v$ is insoluble. Then by Lemma \[one insoluble\], $G_{vw}$ is insoluble. Thus by Lemmas \[$G_{vw}$ almost simple\] and \[$G_{vw}$ non-almost simple\], $(G_v,G_w,G_{vw})$ are listed in rows 5 to 16 of Table \[Intro\]. Thus the theorem holds.
[**Proof of Corollary \[faith-stab-3-trans\]:**]{} We first notice that $\Ga$ is locally $(G,3)$-arc-transitive if and only if $G_{w_1vw}$ and $G_{vwv_1}$ are transitive on $[G_{w_{-1}vw}:G_{w_{-1}vwv_1}]$ and $[G_{vwv_1}:G_{w_{-1}vwv_1}]$, respectively. This is equivalent to $|G_{w_{-1}vw}:G_{w_{-1}vwv_1}|=|\Ga(w)|-1$ and $|G_{vwv_1}:G_{w_{-1}vwv_1}|=|\Ga(v)|-1$.
Suppose first that $\Ga$ is a regular graph. Then the 2-arc stabiliser $G_{vwv_1}$ is transitive on $\Ga(v)\setminus\{w\}$. Observe that $G_{vwv_1}=(G_w)_{vv_1}$ is the stabiliser of the two points $v,v_1\in\Ga(w)$ in the 2-transitive permutation group $G_w$. By Lemma \[2-pt-stab\], we conclude that $(G_v,G_w,G_{vw})=(\A_7,\A_7,\A_6)$, or $(\S_7,\S_7,\S_6)$.
Now assume that $\Ga$ is not a regular graph. We need analyse the triple of stabilisers $(G_v,G_w,G_{vw})$ listed in Table \[Intro\]. We notice that, for each of the candidates in rows 1-8 and 14-16 of Table \[Intro\], the order $|G_{vwv_1}|$ is not divisible by $|\Ga(v)|-1$. Therefore, we conclude that $(G_v,G_w,G_{vw})$ can only be one of $(\A_7,2^4{:}\A_6,\A_6)$, $(\S_7,2^4{:}\S_6,\S_6)$, $(\A_8,2^4{:}\A_7,\A_7)$, $(\A_9,2^4{:}\A_8,\A_8)$, or $(\S_9,\Sp_6(2),\S_8)$.
For the first candidate $(G_v,G_w,G_{vw})=(\A_7,2^4{:}\A_6,\A_6)$, the intersection $G_{w_{-1}vwv_1}$ of $G_{w_{-1}vw}=\A_5$ and $G_{vwv_1}=\S_4$ is a subgroup of $\ZZ_2^2$. Since $|G_{w_{-1}vw}:G_{w_{-1}vwv_1}|\leqslant|\Ga(w)|-1=15$ and $|G_{vwv_1}:G_{w_{-1}vwv_1}|=|\Ga(v)|-1=6$. We thus have $$G_{w_{-1}vwv_1}=G_{w_{-1}vw}\cap G_{vwv_1}=\A_5\cap\S_4=\ZZ_2^2.$$ Therefore, $|G_{w_{-1}vw}:G_{w_{-1}vwv_1}|=15=|\Ga(w)|-1$ and $|G_{vwv_1}:G_{w_{-1}vwv_1}|=6=|\Ga(v)|-1$, and $\Ga$ is locally $(G,3)$-arc-transitive. Similarly, $(G_v,G_w,G_{vw})=(\S_7,2^4{:}\S_6,\S_6)$ is an amalgam for locally $(G,3)$-arc-transitive graph.
For the case $(G_v,G_w,G_{vw})=(\A_8,2^4{:}\A_7,\A_7)$, we have $$G_{w_{-1}vw}=\A_6,\ G_{vwv_1}=\PSL_2(7).$$ Since $|\Ga(v)|-1=8-1=7$, the stabiliser $G_{w_{-1}vwv_1}$ is of index at most 7 in $G_{vwv_1}=\PSL_2(7)$. It implies that $|G_{vwv_1}:G_{w_{-1}vwv_1}|=7$, and $G_{w_{-1}vwv_1}=\S_4$. Then the index $|G_{w_{-1}vw}:G_{w_{-1}vwv_1}|=|\A_6:\S_4|=15=|\Ga(w)|-1$. We conclude that $G_{w_{-1}vw}$ is transitive on $\Ga(w)\setminus\{v\}$, and $G_{vwv_1}$ is transitive on $\Ga(v)\setminus\{w\}$. Therefore, $\Ga$ is locally $(G,3)$-arc-transitive.
For the triple $(G_v,G_w,G_{vw})=(\A_9,2^4{:}\A_8,\A_8)$, the index of $G_{w_{-1}vwv_1}$ in $G_{w_{-1}vw}=\A_7$ is at most $|\Ga(w)|-1=15$ and in $G_{vwv_1}=2^3{:}\GL_3(2)$ is at most $|\Ga(v)|-1=8$. It follows that $G_{w_{-1}vwv_1}=\GL_3(2)$, and so $|G_{w_{-1}vw}:G_{w_{-1}vwv_1}|=|\Ga(v)|-1$ and $|G_{vwv_1}:G_{w_{-1}vwv_1}|=|\Ga(w)|-1$. Therefore, $(\A_9,2^4{:}\A_8,\A_8)$ is a locally 3-arc-transitive amalgam.
For the case $(G_v,G_w,G_{vw})=(\S_9,\Sp_6(2),\S_8)$, we have $G_{w_{-1}vw}=\S_7,\ G_{vwv_1}=(\S_4\times\S_4).2.$ The intersection $G_{w_{-1}vwv_1}$ of $G_{w_{-1}vw}$ and $G_{vwv_1}$ is $\S_4\times \S_3$ with index $8$ in $G_{vwv_1}$ and index $35$ in $G_{w_{-1}vw}$. Moreover, $|\Ga(v)|-1=9-1=8$ and $\Ga(w)|-1=35$. We conclude that $G_{w_{-1}vw}$ is transitive on $\Ga(w)\setminus\{v\}$, and $G_{vwv_1}$ is transitive on $\Ga(v)\setminus\{w\}$. Therefore, $\Ga$ is locally $(G,3)$-arc-transitive.
[8]{}
J. Bamberg, C. H. Li and E. Stwarze, Anti-flag transitive generalized quadrangles, submitted.
J. van Bon, Thompson-Wielandt-like theorems revisited, [*Bull. London Math. Soc.*]{} [**35**]{} (2003), 30-36.
J. van Bon, On locally s-arc transitive graphs with trivial edge kernel [*Bull. London Math. Soc.*]{}, [**43(4)**]{} (2011), 799-804.
P. Cameron, [*Permutation Groups*]{}, London Math. Soc., Student Texts [**45**]{}, (Cambridge Univ. Press, 1999).
J. H. Conway, R. T. Curtis, S. P. Norton, R. A. Parker and R. A. Wilson, [*Atlas of Finite Groups*]{}, Clarendon Press, Oxford, 1985.
Wen Wen Fan, D. Leemans, C. H. Li, and J. M. Pan, Locally 2-arc-transitive complete bipartite graphs, [*J. Combin. Theory Ser. A*]{} [**120**]{} (2013), no. 3, 683-699.
M. Giudici, C. H. Li and C. E. Praeger, Analysing locally $s$-arc-transitive graphs, [*Trans. Amer. Math. Soc.*]{} [**356**]{} (2004), 291-317.
D. Leemans, Locally $s$-arc-transitive graphs related to sporadic simple groups [J. Algebra]{} [**322**]{} (2009), 882-892.
C. H. Li, A. Seress,and S. J. Song, s-Arc-transitive graphs and normal subgroups. [*J. Algebra*]{} [**421**]{} (2015), 331-348.
M. Liebeck, The affine permutation groups of rank three, [*Proc. London Math. Soc. (3)*]{} [**54**]{} (9187), 477-516.
Luke Morgan, Pablo Spiga and Gabriel Verret, On the order of Borel subgroups of group amalgams and an application to locally-transitive graphs, [*J. Algebra*]{} [**434**]{} 138-152 (2015).
P. Potočnik, Locally arc-transitive graphs of valence [3,4]{} with trivial edge kernel, [*J. Algebraic Combin.*]{} [**38**]{} (2013), 637-651.
V. I. Trofimov, Graphs with projective suborbits. Exceptional cases of characteristic 2. I, [*Izvestiya Mathematics*]{} [**62**]{}(1998), 1221-1279.
R. Weiss, The nonexistence of 8-transitive graphs, [*Combinatorica*]{} [**1**]{} (1981), 309-311.
H. Wielandt, [*Finite Permutation Groups*]{}, (Academic Press, New York, 1964).
[^1]: [*2010 Mathematics subject classification*]{}: 05E18, 20B25
[^2]: This work forms a part of an ARC grant project.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present an overview of current academic curricula for Scientific Computing, High-Performance Computing and Data Science. After a survey of current academic and non-academic programs across the globe, we focus on Canadian programs and specifically on the education program of the SciNet HPC Consortium, using its detailed enrollment and course statistics for the past four to five years. Not only do these data display a steady and rapid increase in the demand for research-computing instruction, they also show a clear shift from traditional (high performance) computing to data-oriented methods. It is argued that this growing demand warrants specialized research computing degrees. The possible curricula of such degrees are described next, taking existing programs as an example, and adding SciNet’s experiences of student desires as well as trends in advanced research computing.'
author:
- Marcelo Ponce
- Erik Spence
- Daniel Gruner
- Ramses van Zon
bibliography:
- 'references\_v2.bib'
title: 'Scientific Computing, High-Performance Computing and Data Science in Higher Education'
---
Introduction {#sec:intro}
============
The computational resources available to scientists and engineers have never been greater. The ability to conduct simulations and analyses on thousands of low-latency-connected computer processors has opened up a world of computational research which was previously inaccessible. Researchers using these resources rely on scientific-computing and high-performance-computing techniques; a good understanding of computational science is no longer optional for researchers in a variety of fields, ranging from bioinformatics to astrophysics.
Similarly, the advent of the internet has resulted in a paradigm where information can be more easily captured, transmitted, stored, and accessed than ever before. Researchers, both in academia and industry [@Radermacher:2013:GIE:2445196.2445351], have been actively developing technologies and approaches for dealing with data of previously-unimaginable scale. Researchers’ ability to analyze data has never been greater, and many branches of science are actively using these newly-developed techniques.
Unfortunately, the skills needed to harness these computational and data-empowered resources are not systematically taught in university courses [@Richards:2011:TRN:2016741.2016801]. Some researchers, postdocs and students may find non-academic programs to fill this void, but others either do not have access to these courses or cannot commit the time to follow them. These researchers typically end up learning by trial and error, or by self-teaching, which is rarely optimal.
A number of academic programs that aim to address this issue have emerged at universities across the world (a few examples are [@Tel-Zur:2014:PEB:2690854.2690857; @Burkhart:2014:THC:2690854.2690859]). Some of these grew out of the training efforts of High Performance Computing (HPC) centres and organizations ([ *e.g*]{}. [@Stewart:1995:HUC:224170.224209]). Recognizing the need for additional skills in their users, computing centres such as those in the [XSEDE](https://www.xsede.org/) partnership [@XSEDE2014] in the U.S., [PRACE](http://www.prace-ri.eu/) in Europe, and [Compute/Calcul Canada](https://www.computecanada.ca/) have been providing local and online HPC training as part of their user support. Universities have also developed graduate programs in both Scientific and High-Performance Computing, to train scientists and engineers in the use of these computational resources.
A more-recent complement to these graduate programs is the development of the degree in Data Science (DS), that is, degrees which focus on the analysis of data, especially at scale. These degrees come in a variety of forms, from multi-year academic graduate programs to specialized private-sector training. These programs are in strong demand at present, as large companies have discovered the value in thoroughly analysing the vast quantities of customer data which they collect. It is expected that this field will continue to grow, and academic programs will continue to be introduced to meet this demand.
The SciNet HPC Consortium [@1742-6596-256-1-012026; @website:SciNet] is the high-performance-computing consortium of the University of Toronto. SciNet provides both computational resources and specialized user support for Canadian academic researchers, and as members of its support team, we are responsible for training researchers, postdocs and graduate students at the University of Toronto in HPC techniques. In this paper we give a review of the current state of graduate-level Scientific Computing, High-Performance Computing and Data Science academic programs, and endeavour to design an ideal HPC and DS graduate program. The paper is organized as follows. In Sec. \[sec:HPC-research-review\] we discuss how computation has become an essential ingredient in many academic research endeavours; in Sec. \[sec:HPC-review\] we review the current status of education in the areas of High-Performance and Scientific Computing. In Sec. \[sec:DS-review\] we present the Data Science education efforts at the academic and non-academic level. Sec. \[sec:futurePrograms\] describes what HPC and DS Master’s programs could look like. We conclude with final remarks and perspectives for the future in Sec. \[sec:concl\].
The role of HPC in current research {#sec:HPC-research-review}
===================================
The breadth of science being as large as it is, it is essentially impossible to give an overview of the uses of computational methods in current scientific research. We will nonetheless attempt a review of at least some computational scientific research, since the way computers are used in research (and other realms of inquiry) influences what should be taught to students.
Astrophysical computational research inherently involves large scale computing, such as the simulation of gravitational systems with many particles, magnetohydrodynamic systems, and bodies involving general relativity. Atmospheric physics requires large weather and climate models with many components to be simulated in a variety of scenarios. High-energy particle physics projects, such as the ATLAS project at CERN, require the analysis of many recorded events from large experiments, while other high-energy physics projects have a need for large scale simulations ([*e.g.*]{} lattice QCD investigations). Condensed matter physics, quantum chemistry and material science projects must often numerically solve quantum mechanical problems in one approximation or another; the approximations make the calculations feasible but still rely on large computing resources. Soft condensed matter and chemical biophysics research often involve molecular dynamics or Monte Carlo simulations, and frequently require sampling a large parameter space. Engineering projects can involve optimizing or analyzing complex airflow or combustion, leading to large fluid dynamics calculations. Bioinformatics often involves vast quantities of genomic input data to be compared or assembled, requiring many small computations. Research in other data-driven sciences such as social science, humanities, health care and biomedical science [@Ihantola:2015:EDM:2858796.2858798], is also starting to outgrow the capacity of individual workstations and standard tools in their respective fields.
Examining these cases in more detail, one can distinguish different ways in which research relies on computational resources:
1. Research that is inherently computational, [*i.e.*]{} it cannot reasonably be done without a computer, but which requires relatively minor resources ([*e.g.*]{} a single workstation).
2. Research that investigates problems that do not fit on a single computer, and therefore rely on multiple computing nodes attached through a low-latency network.
3. Research that requires many relatively small computations.
4. Research that requires access to a large amount of storage, but not necessarily a lot of other resources.
5. Research that requires access to a lot of storage, on which many relatively small calculations are performed.
The distinction between the various types of research determines the appropriate systems and tools to use. Graduate students that are just starting their research often do not have enough knowledge to make the distinction (as nobody has taught them about this), let alone select and ask for the resources that they will need [@Richards:2011:TRN:2016741.2016801].
Note that all five categories fall under “Advanced Research Computing” (ARC). The categories are not mutually exclusive, but research of the second and third kind are usually associated with HPC, while the fourth and fifth, and sometimes the first, are associated with Data Science (DS). Although there is a lot of overlap between HPC and DS, these fields require somewhat different techniques. For that reason, we will consider separate programs for HPC and DS.
Programs in High-Performance Computing {#sec:HPC-review}
======================================
Much of the research presented in the previous section falls in the category *Scientific Computing* (SC). The growth in the computational approach to research, both academic and industrial, has prompted some institutions to develop graduate-level programs designed to teach the skills needed to design, program, debug and run such calculations. These programs, having been in development for more than two decades, are now fairly wide spread and mature, and are known by the names “Scientific Computing” or “Computational Science and Engineering”. Scientific Computing graduate degrees are offered internationally in several graduate education hubs around the world (U.S., England, Germany, Switzerland, [*etc.*]{}, — lists of which can be found at the [SIAM](https://www.siam.org/students/resources/cse_programs.php) and [HPC University](http://hpcuniversity.org/students/graduatePrograms) [@Lathrop:2013:HUG:2484762.2484771] websites). Canada is no exception here either, with at least eight universities offering graduate-level programs in Computational Science. These programs include one-year and two-year Master’s programs, as well as Ph.D. programs. Most of these programs ([*e.g.*]{} the ones shown in Tables \[table:HPCProgram\] and \[table:BioProgram\]) require a final thesis. The projects and theses are faculty-guided research projects and are usually one-term long, though, as with all research, these projects sometimes take longer.
A typical curriculum for a two-year Master’s program in Scientific Computing (in this case from San Diego State University) is presented in Table \[table:HPCProgram\]. It clearly shows that Scientific Computing has its roots in research in the physical sciences; the programs heavily emphasize numerical analysis and scientific modelling. In some ways this is not surprising: computers are very apt at solving such problems, and the formalism of the physical sciences often lends itself easily to computer programming. Other topics of study which are also often encountered in these programs include finite element analysis, matrix computations, optimization, stochastic methods, differential equations and stability.
Course Name Type
-------------------------------------------------- ---------- --
Introduction to Computational Science required
Computational Methods for Scientists required
Computational Modelling for Scientists required
Computational Imaging required
Scientific Computing required
Applied Mathematics for Computational Scientists required
Seminar Problems in Computational Science required
Computational and Applied Statistics elective
Computational Database Fundamentals elective
Research required
Thesis required
: The curriculum for the two-year [Master’s program](http://www.csrc.sdsu.edu/masters.html) at the Computational Science Research Center at San Diego State University [@website:SDSUmasters]; this forms a good example of a typical Scientific Computing graduate program. []{data-label="table:HPCProgram"}
In contrast to Scientific Computing, HPC requires somewhat wider knowledge; its practitioners need to understand more than just the theoretical and numerical principles. They require skills such as serial and parallel programming (often in several languages, and on different platforms) and scripting, as well familiarity with numerics, data handling, statistics, and supercomputers and their technical bottlenecks. In addition, these practitioners are usually not computer science students, so they must cope without that background. This is somewhat unavoidable as they need to have sufficient domain knowledge as well. Much of the same holds for Data Science.
Academic HPC Programs {#section:HPCprograms}
---------------------
There are not many academic programs that focus on HPC. Part of the reason may be that such programs require access to a high-performance-computing machine so that students can develop their skills on real hardware, in a real supercomputing environment. These machines require multiple computing nodes which are connected by a low-latency network. Fortunately, such systems do not need to be local: as long as the machine is accessible through the internet the machine could be used for teaching. Nonetheless, having the hardware local to the students lends advantages, since most of the administrators and analysts of the system are typically available to assist students with optimizing their codes and developing good computational strategies. Not surprisingly, the majority of the currently offered HPC graduate programs seem to have been developed in conjunction with or by supercomputer centres.
As examples of High Performance Computing programs, the University of Edinburgh (UK) offers an [MSc in High Performance Computing](https://www.epcc.ed.ac.uk/msc), the Universitat Politècnica de Catalunya/BarcelonaTech (Spain) offers a [Master in High Performance Computing](http://masters.fib.upc.edu/masters/master-high-performance-computing) and a [Master program in Data Mining and Business Intelligence](http://masters.fib.upc.edu/masters/master-data-mining-and-business-intelligence), SISSA/ICTP in Italy offers a [Master in High Performance Computing](http://www.mhpc.it/), while a collaboration between the University ITMO (Russia) and the University of Amsterdam (Netherlands) offers a [Double-Degree Master Programs in Applied Mathematics and Informatics (Computational Science)](http://en.hpc-magistr.escience.ifmo.ru/magistr/9/). Note that many of these programs emerged from locations with a very strong tradition and consolidated background in HPC.
SciNet’s HPC Programs {#section:nonacadHPCprograms}
---------------------
Many HPC centers provide training for their users to fill the computational-skills gap for the wider scientific community, such as, [SDSC](http://www.sdsc.edu/), [PSC](http://www.psc.edu/), [TACC](https://www.tacc.utexas.edu), [NCSA](http://www.ncsa.illinois.edu/), [BSC](https://www.bsc.es/), [EPCC](https://www.epcc.ed.ac.uk/), [CSCS](http://www.cscs.ch), [SHARCNET](https://www.sharcnet.ca/), [AceNet](http://www.ace-net.ca/), [Calcul Québec](http://calculquebec.ca/en/), among many others. In its capacity as an HPC centre based at the University of Toronto, SciNet has developed several education and training classes [@website:SciNet-Edu] aimed at helping students and users obtain the skills and knowledge required to get the most out of advanced-research-computing resources. SciNet’s training events and courses are currently taken by researchers, postdocs, and graduate students across many different departments and even from outside of the University of Toronto (UofT). Some of these courses are considered part of the official curricula and count as graduate level courses within the Ph.D. programs at UofT.
Initially SciNet provided training specifically oriented toward Scientific Computing, with the purpose of maximizing user productivity. These early classes focused on parallel programming (MPI and OpenMP), best coding practices, debugging, and other scientific computing needs. Over the years the breadth of classes has grown, with classes offered in Linux shell programming, parallel input/output, advanced C++ and Fortran coding, accelerator programming, and visualization. This is in addition to the annual HPC Summer School which SciNet runs in collaboration with two other HPC centres within Compute Ontario[@website:CO], i.e., CAC[@website:cac] and SHARCNET[@website:sharcnet]. This summer school is a week-long intensive workshop on HPC topics, and more recently, also Data Science topics[^1].
Table \[table:SciNet-courses\] shows the training events and courses that SciNet has already been teaching in the areas of HPC and Data Science.
Course Name Certificate Credits
------------------------------------------------------------- ------------- ---------
*Data Analysis with R*$^\ddag$ DS/SC 12
Intro to Apache Spark DS 3
Machine Learning Workshop DS/SC 6
Hadoop Workshop DS 3
Scalable Data Analysis Workshop DS 12
Relational Database Basics DS/SC 6
Storage and Input/Output in Large Scale Scientific Projects DS/SC 6
Workflow Optimization for Large Scale Bioinformatics DS/HPC/CS 6
Python for High Performance Computing DS/HPC/SC 12
Parallel R DS/HPC/SC 3
Python GUIs with Tkinter DS/SC 2
Scientific Visualization DS/SC 6
Visualizing Data with Paraview DS/SC 6
*Scientific Computing for Physicists*$^\ddag$[^2] HPC/SC 36
*Intro to Research Computing with Python*$^\ddag$ HPC/SC 12
Intro to High Performance Computing HPC/SC 3
Intro to Scientific C++ HPC/SC 6
Intro to Scientific Programming with Modern FORTRAN HPC/SC 7
Intro to Parallel Programming HPC/SC 7
Programming Clusters with Message Passing Interface HPC/SC 12
Programming Shared Memory Systems with OpenMP HPC/SC 6
Practical Parallel Programming Intensive HPC/SC 32
Intro to GPGPU with CUDA HPC/SC 9
Programming GPUs with CUDA HPC/SC 12
SciNet/CITA CUDA GPU Minicourse HPC/SC 12
Coarray Fortran HPC/SC 2
Parallel I/O HPC/SC 6
Debugging, Optimization, Best Practices HPC/SC 6
HPC Best Practices and Optimization HPC/SC 3
HPC Debugging HPC/SC 3
Intro to the Linux Shell HPC/SC 2
Seminars in High Performance Computing HPC/SC 4
Seminars in Scientific Computing HPC/SC 4
The number and types of classes which SciNet teaches have grown significantly [@website:SciNet-Training]. This can be seen in Figure \[fig:classhours\], which presents the total student class-hours taught by SciNet over the last four years and the projection for the current year. This remarkable growth is a testament to the latent need for this material to be taught.
![Attendance hours at SciNet training and education events, per year, for all SciNet classes and Data Science specific classes.[]{data-label="fig:classhours"}](./teaching_stats-2012-2016){width="\columnwidth"}
The need for this training is supported by the enrolment statistics: our students constitute 35% of SciNet’s total users, clearly showing that even in a specialized audience this kind of training is still needed.
For several years the four-week graduate-style classes offered by SciNet have been accepted for *graduate* class credit by the departments of Physics, Chemistry and Astrophysics at UofT. This was possible by accepting the classes as “modular” (or “mini”) courses, one-third semester long, and bundling three such classes into a full-semester course. This arrangement has been so popular with students and faculty that the Physics Department recently listed SciNet’s winter twelve-week HPC class in the course calendar [@website:SciNet-PHY1610H], allowing graduate students from other departments in the university to take the class for university credit.
The skills that SciNet aims to transfer are rare and sought-after, and complement and enhance the skills students learn in regular curricula. That is why SciNet has developed a set of *Certificate Programs* [@website:SciNet-Certificates], that users and students can pursue in *Scientific Computing*, *High Performance Computing*, and/or *Data Science*, once they have completed enough credit-hours. As a document that proves the holder has highly competitive skills, and in lieu of graduate credit for most SciNet courses, the certificates are in high demand. In a resounding endorsement of our teaching, thus far students have completed a total of 78 certificates (52 in Scientific Computing, 19 in High-Performance Computing and 8 in Data Science). According to the current registration and trends, we are projecting to have above a 100 of certificates completed by mid-2016. Moreover, the feedback from some of our students was that their SciNet’s certificates gave them an advantage to get jobs in industry and the financial sector.
Programs in Data Science {#sec:DS-review}
========================
The wide adoption of the internet in the professional and the personal sphere ushered in the age of “Big Data”. The ease of recording of people’s online behaviour, and the ability to rapidly move data, lead to a large, diffuse, complex amount of data waiting to be mined for useful information. Because of the typically large size of the data special hardware and training are often needed. In contrast to Scientific Computing and HPC, there are many applications of Data Science in the private sector, in the medical science, banking, retail, insurance, and internet industries.
Of these industries, Bioinformatics also has a large component in the academic world. Though a more-recent addition to the HPC world, the bioinformatics field is well-populated with graduate programs, a testament to its rapid growth and latent demand. Its emergence as a major user of HPC systems has resulted in the development of “Master’s of Bioinformatics”, and related degrees. A typical Master’s program is outlined in Table \[table:BioProgram\], this one from the Indiana-Purdue University at Indianapolis. While having many features in common with a more-standard SC Master’s program, such as the study of programming and algorithms, it exhibits the particular needs of the bioinformatics community, stressing the importance of genetics and biological processes, and a lesser emphasis on mathematics and programming theory.
Course Name Type
------------------------------------------ ---------- --
Introduction to Bioinformatics required
Seminar in Bioinformatics required
Biological Database Management required
Programming for Life Science required
High Throughput Data in Biology required
Machine Learning in Bioinformatics elective
Computational System Biology elective
Structural Bioinformatics elective
Transitional Bioinformatics Applications elective
Algorithms in Bioinformatics elective
Statistical Methods in Bioinformatics elective
Computational Methods for Bioinformatics elective
Next Generation Genomic Data Analytics elective
Next Generation Sequencing elective
Bioinformatics Project required
: The curriculum for the “Project Track” two-year [Master’s of Science in Bioinformatics](https://soic.iupui.edu/biohealth/graduate/bioinformatics-masters) at the Indiana University-Purdue University in Indianapolis; this forms a good example of a typical Bioinformatics graduate program. []{data-label="table:BioProgram"}
Degrees in Data Science are relatively new, with the first Master’s program only being introduced in the U.S. (by North Carolina State University) in 2007. A sample of some of the classes offered in one such program is given in Table \[table:DataProgram\]. As can be seen, these programs have a strong focus on data, with statistics, machine learning, and databases being their standard focus. Analyzing data that are too big to fit on a standard desktop computer requires specialized equipment; such training is also part of these graduate-level programs, as indicated by the presence of the “Cloud Computing” and “Distributed Systems” classes. Like typical graduate-level programs, these degrees usually require of the student a final project or thesis.
Course Name Type
----------------------------------------- ---------- --
Analysis of Algorithms required
Machine Learning required
Advanced Database Concepts required
Distributed Systems elective
Advanced Database Concepts elective
Cloud Computing elective
Information Retrieval elective
Data Mining elective
Web Mining elective
Applied Machine Learning elective
Complex Networks and Their Applications elective
Relational Probabilistic Models elective
Internship in Data Science elective
: A selection of the courses available for the [Master’s of Data Science](http://www.soic.indiana.edu/graduate/degrees/data-science/ms-data-science/index.html) at the Indiana University. []{data-label="table:DataProgram"}
One could argue that the novelty of methods in Data Science is due to its roots in Business Analytics (BA), where the objective is to make a decision. The field has certainly grown beyond that, and BA is now considered a sub-field of Data Science. Another more-recently developed sub-field is in the realm of health care (“Health Informatics”). Because these sub-fields are directly applicable to the private sector (and the associated revenue streams these present) these have become the most-commonly implemented post-graduate programs. The Business Analytics programs focus on using data to refine business administration, as well as develop marketing strategies. Health Informatics programs concentrate on using clinical data to optimize health care processes.
The practical focus of Data Science is reflected in the presence of an internship in the Data Science curriculum listed in Table \[table:DataProgram\]. Internships in such programs are similar to other co-op-type arrangements: the student works with an employer for a semester, allowing the student to gain hands-on experience applying the skills learnt during such period.
Academic Data Science Programs
------------------------------
Graduate level programs in Data Science are not difficult to find. For instance, programs in bioinformatics (a data-driven field), can be found on the [web site of the International Society for Computational Biology](https://www.iscb.org/iscb-degree-certificate-programs). It speaks to the rapid rise of the field bioinformatics, that there are more bioinformatics programs available than Scientific Computing programs. Examples lists of other Data Science programs can be found [at the NCSU analytics web site](http://analytics.ncsu.edu/?page_id=4184), [the online business analytics programs site of [predictiveanalyticstoday.com]{}](http://www.predictiveanalyticstoday.com/online-business-analytics-programs) and [at [online.coursereport.com]{}](http://www.onlinecoursereport.com/the-50-best-masters-in-data-science). They are not as common as programs in Scientific Computing, due to the fact that Data Science is relatively new field of study. Among those programs about half are offered in the fields of Business Analytics and Health Informatics, with the other half being Data Science programs proper.
Non-academic Data Science training
----------------------------------
The demand for Data Science skills (or “Data Analytics” skills as they are often called in the private sector) is so high [@Radermacher:2013:GIE:2445196.2445351] that the private sector has developed programs to meet the growing demand. A list of such companies can be found [on skilledup.com, which contains a list of data science boot-camps](http://www.skilledup.com/articles/list-data-science-bootcamps). The format of these classes is varied, though they are all oriented toward a “boot-camp” format: some are in person, some online; some are one-week long, others twelve weeks. These programs are very applied, often with one-on-one mentorship with a seasoned Data Analytics expert. They also include direct contact with possible future employers.
Moreover, a great number of these training programs are not focused on developing analytical thinking or problem-solving skills, [@bootcampsnogood] but rather are aimed at graduated Ph.D.s and postdocs, whose problem-solving skills are assumed to have already developed. This allows them to focus on the technical training relevant to the job market. Some of these programs are free, some of them offer fellowships, and many of them charge on the order of 10-30 thousand US-dollars for a training period of, typically, three months. These programs have acquired such a level of popularity among young and recent graduates that the companies offering these programs have started to perform evaluation tests in order to assess which candidates are more suitable to be accepted to their programs. Perhaps the most appealing part for trainees is the networking platform offered by these programs, as in most of the cases they provide the opportunity to interact with actual companies looking for new talent and avoid recruitment layers.
Institutions in the non-profit arena are also starting to offer programs on Data Science. For instance the Fields Institute, a traditional institution for mathematical research, has offered several workshops and courses, and developed a thematic program on Big Data. Other examples include the International Centre for Theoretical Physics (ICTP) and the International School for Advanced Studies (SISSA), prestigious institutions with a well known tradition in theoretical physics, now offering training in “Research Data Science”. HPC centers are also venturing into Data Science training, offering workshops on R, Hadoop, machine learning, [*etc*]{}. SciNet started offering classes with greater data-oriented content ([ *cf*]{}. Table \[table:SciNet-courses\]) in 2013, with a four-week class in scientific analysis using Python. Having now finished its third year, the class remains popular, with about twenty students taking the class each year. The 2015 fall semester also inaugurated SciNet’s first “Data Science with R” class, a class focusing on data analysis techniques using the R language. This class was very popular with over twenty-five students finishing the class, and most students requesting a second installment with more advanced material. Continuing its growth in the Data Science area, in the last year SciNet has held workshops in machine learning, scalable data analysis, and Apache Spark.
Comparing the student- and taught-hours per year shown in Fig. \[fig:DSclasses\], one sees that the Data Science classes have been growing consistently, both absolutely as well as relatively (Data Science related courses roughly constituted less than $2\%$ (2012), $4\%$ (2013), $12\%$ (2014), $15\%$ (2015), and we project around $31\%$ (2016) of the total classes taught respectively in each year.
![Total student hours (left) and taught hours (right) per year, for SciNet’s Data Science related courses.[]{data-label="fig:DSclasses"}](./DS_courses-studentshours "fig:"){width="0.49\columnwidth"} ![Total student hours (left) and taught hours (right) per year, for SciNet’s Data Science related courses.[]{data-label="fig:DSclasses"}](./DS_courses-taughthours "fig:"){width="0.49\columnwidth"}
Designing Master’s programs in HPC and Data Science {#sec:futurePrograms}
===================================================
As mentioned above, scientific computing is used by scientists and engineers as never before, and graduate-level programs in Scientific Computing are numerous in Canada and around the world. In contrast, the development of HPC and Data Science programs is in its early stages, both in academia and the private sector. These programs are being developed to meet the continued shortfall in skill in these areas, with the McKinsey Global Institute estimating that the United States will be short 140,000 to 190,000 data analytics professionals by 2018 [@McKinsey_BigData].
One may wonder whether online learning could not satisfy this need. A few examples of MOOCs (Massively Open Online Courses) in HPC and Data Science do exist. However, seeing the growth in enrolment in SciNet’s in-person courses and the summer school over the years ([ *cf*]{}. Figs. \[fig:classhours\] and \[fig:DSclasses\]) shows that many students still prefer the face-to-face format.
Similarly, one may wonder why certificate programs do not suffice for HPC and DS education. As successful as these programs are, they have a few disadvantages. Firstly, they are mostly collections of fairly specific technical training: this leaves no room for more fundamental material. Secondly, it is also hard to encorporate an internship or thesis into such a certificate. Finally, certificates tend to carry less weight than degrees, and, in line with this, the demand for for-credit courses is larger than that for not-for-credit courses, as our experience with SciNet’s Scientific Computing graduate course has shown.
A degree program in HPC or DS could offer more academic and fundamental education, which would leave the student with the analytical skills and high-level knowledge to stay on top of their field regardless of changes in computational technology.
In the following sections we propose a curriculum for graduate-level HPC and DS programs. One will notice a substantial overlap of topics with the training courses currently taught by SciNet ([*cf*]{}. Table \[table:SciNet-courses\]). This is no coincidence: the training program was developed on the basis of student feedback and requests, and was a primary inspiration for the curricula proposed here. The design was also influenced by the few existing examples of such programs, as described above.
It should also be emphasized that the programs in HPC and DS are both designed to allow students to a follow more industrial/practically-oriented track or an academic/research-oriented track, by selecting the appropriate set of elective courses and the corresponding research project/internship. Additional *advanced* courses could be made available according to the interest and demand of the students.
Design of an HPC Master’s Program {#sec:SciNet-HPCprogram}
---------------------------------
In this section we present a comprehensive and complete curriculum for a two-year *Master’s Program in High-Performance Computing*. Students would complete a total of twelve courses. The five required courses, each one-term long, set the basis of HPC and ARC knowledge (including topics such as modern and professional software development, parallel techniques, performance and optimization, best practices, distributed systems and resources). The seven elective courses allow the student to specialize in a particular area. In addition, a final internship or research project would be carried out.
As in any typical Master’s program, a student entering the program will be expected to possess a Bachelor’s degree (B.Sc. or B.A.). It is desirable that students have some background in sciences and the basics of coding and programming ([*e.g*]{}. Fortran, C, C++); otherwise it is strongly recommended that students take introductory programming classes. Note that courses from the *Data Science* Master’s Program are also eligible to be taken, with consent of the graduate coordinator or adviser.
The course work for the High-Performance Computing program could consist of the following courses.
Software Development (\*):
: The principles of creating modern, maintainable code. Special attention is given to designing modular code, and tackling scientific computational projects. Languages: C++/C/Fortran.
Best Practices (\*):
: Introduction and discussion of techniques and methods to be considered when designing and implementing computational research projects. Topics include: version control, modularity, libraries.
Performance & Optimization (\*):
: Principles and tools for measuring performance, finding bottlenecks, and optimizing existing code.
Basics of Parallel Programming (\*):
: A review of homogeneous and heterogeneous architectures is presented, followed by parallel programming paradigms such as OpenMP, MPI, and hybrid implementations.
HPC Algorithms:
: A review of commonly used algorithms in computational science, such as Monte Carlo, implicit and explicit methods to solve differential equations, timestepping techniques, finite-element and finite-volume methods.
Machine Learning:
: Theory and practice of constructing algorithms that create models from data. Topics: probabilistic foundations, linear and logistic regression, neural networks, Bayesian networks, tree models, support vector machines, density estimation, accuracy estimation, normalization, model selection.
Numerical Methods:
: A review of the commonly used numerical methods in scientific computing, such as linear solvers, fast Fourier transforms. In contrast to the HPC Algorithms, it will cover the basic principles and the theory behind the methods only briefly, and focus mostly on the implementation and utilization via libraries and specific examples.
Scientific Computational Modeling:
: This course will offer an introduction to the basics of Scientific Computing and a review of the most common algorithms and packages used in different fields of research computing and computational sciences (astrophysics, chemistry, genetics, [*etc.*]{} Introduction to computing modeling, such as implementation of complex networks based on relational data sets, and the evaluation of network properties using graph theory elements, among many others.
Programming Accelerators:
: Some computational problems can be computed much more efficiently on accelerators such as graphics cards. This course will present an introduction to the use of hardware accelerators (GPUs, FPGAs, many-core systems) in HPC. Topics will include a review of the hardware, architectures, and programming languages, such as, CUDA, OpenACC, and OpenCL.
Research Data Management:
: Design strategies, storage management and I/O patterns, in order to prevent bottlenecks in massively data-driven projects. Real use cases from various fields (bioinformatics, molecular biophysics, medical physics, biochemistry, quantum-chemistry, geophysics, [*etc*]{}.) have shown that, quite often, approaches that work on a desktop do not perform on a larger scale. A review of the latest policies regarding scientific data availability will also be discussed and presented.
Visualization for Scientific Computing:
: A review of basic visualization concepts and methods with applications to scientific data.
Operating System Environment:
: Scientific and high performance computing is intimately linked to the hardware, OS, and application framework on which they used. This course will help students become comfortable working on \*nix systems. Topics such as the command line, shell scripting and advanced OS topics will be covered.
HPC Hardware and System Administration:
: An integrated view on the technology in HPC, the machines, hardware, network, file systems, and what is involved in getting an HPC system up and running.
Student HPC/ARC Seminar:
: Weekly sessions running throughout the year, with students presenting and discussing current papers and research in the fields of HPC and ARC. Researchers and instructors will providing guidance and supervision during the sessions.
In addition to this course work, the program would include an 4-month internship or independent research project in the final year.
The program presented above is intended to be flexible. In the first year of the HPC program, students might take *HPC Algorithms*, *Parallel Programming*, *Software Development*, *Performance and Optimization*, *Best Practices*, and *Numerical Methods*, as well as attend the HPC/ARC seminar series. In year two, they might take *Research Data Management*, *Scientific Computational Modeling*, *Graph Theory Applications*, *Machine Learning*, *Visualization for Scientific Computing*, and complete the degree with a Research Project.
Eligible courses from the *Data Science* Master’s Program are also possible to take with previous consent of the graduate coordinator or adviser.
Design of a Data Science Master’s Program {#sec:SciNet-DSprogram}
-----------------------------------------
Here we present a comprehensive curriculum for a two-year *Master’s Program in Data Science*. As with the HPC program, students would complete a total of twelve courses. The five required courses set the foundation of data analysis knowledge (including topics such algorithms, databases, statistics, machine learning) and seven elective courses allow the student to specialize in a particular area, such as data mining, machine learning, complex systems. In addition, a final internship or research project would be carried out, in order to obtain real-world experience.
As in any typical Master’s program, the entry level will be a bachelor’s degree (B.Sc. or B.A.). It is desirable that students have some background in sciences and the basics of coding and programming ([*e.g*]{}. Python, R); otherwise it is strongly recommended that students take introductory programming classes. Notice that courses from the *High Performance Computing* Master’s Program may be taken with permission of the graduate coordinator or adviser ([*e.g*]{}. *Operating System Environment*).
The fundamentals of data analysis should be at the core of a program that will produce analysts capable of tackling real-life problems in Data Science. Learning theoretical and practical approaches gives students an advantage in the real world; this program proposes to combine both in a unique fashion (similar to how most SciNet courses are structured). Topics such as statistical analysis, algorithms and large data sets are at the centre of the proposed program and constitute the “core” (required) courses. Additionally, the elective courses allow students to choose a specialization path by gaining expertise in areas such as: social data mining, machine learning, and representation of complex interactions. The course work for the Data Science program could consist of the following courses:
Overview of Data Science:
: An overview of the field of Data Science, covering data-driven problems from several disciplines such as, astronomy, bioinformatics, digital humanities, social sciences, [*etc*]{}.
Basics of Programming:
: An introduction to programming, coding structures, and basic algorithms. This course will focus on languages with data-analytic capabilities, such as *Python* and *R*.
Data Analysis Algorithms (\*):
: An overview of the major classes of algorithms, including comparison-based algorithms: search, sorting, hashing; information extraction algorithms (graphs, databases); graph algorithms: spanning trees, shortest paths, depth and breath-first search.
Machine Learning (\*):
: Theory and practice of constructing algorithms that create models from data. Topics: probabilistic foundations, linear and logistic regression, neural networks, Bayesian networks, tree models, support vector machines, density estimation, accuracy estimation, normalization, model selection. Deep learning algorithms.
Database Theory (\*):
: The mathematical foundations of databases, search and query semantics, relational, complex, object-oriented and semi-structured database models.
Database Applications:
: SQL, relational algebra, index, views, constraints; query complexity; data models, including I/O model, streaming model, query optimization, optimal join algorithm will be presented.
High Dimensional Data Analysis (\*):
: Theory and methods for exploring high-dimensional data will be presented, including linear and non-linear dimension reduction, manifold learning; Euclidean representation of proximity and network data; clustering, statistical pattern recognition.
Applied Statistics (\*):
: A review of statistical techniques, with applications to data analysis problems. Topics will include hypothesis testing, general linear models, generalized linear methods, multivariate statistics.
Best Practices:
: Introduction and discussion of techniques and methods to be considered when designing and implementing computational research projects. Topics include: version control, modularity, libraries.
Data Mining:
: Algorithmic approaches to discovering patterns in large data sets will be covered, including data exploration and cleaning; association rules, clustering, anomaly detection, and classification. Other examples will include applications to text and the web: crawling, indexing, ranking and filtering algorithms; applications to search, classification and recommendation; link analysis, significance tests. Data mining techniques applied to social media. Sentiment analysis, polarity classification; graph properties of social networks, homophily, distance, influence, spectral methods, information diffusion, probabilities models.
Data Security and Integrity:
: This course will cover technique on how to safely protect the integrity and privacy of sensible data, standards on data regulations ([*e.g*]{}. health, patient data, denomic data), privacy, encryption algorithms.
Search and Classification Algorithms:
: Information retrieval theory, with applications, will be explored. Examples of searching algorithms, mathematical representations and matrix applications will be covered.
Distributed Systems:
: Focus on distributed resources and data sets across different architectures and systems. Students will learn the skills and abilities of dealing with retrieving and screening large data sets in *cloud*-type based systems.
Graph Theory Applications:
: An introduction to graph theory and complex systems and its applications will be offered. Material will include the implementation of complex networks based on relational data sets, and the evaluation of network properties.
Visualization Techniques of Unstructured Data sets:
: A review of basic visualization concepts, with applications to unstructured and/or large data sets and complex and dynamical systems, in order to gain insights in the data and visually expose potential correlations.
Data Science Seminar Series:
: Weekly sessions running throughout the year, where students will present and discuss current papers and research in the fields of data science and analytics. Researchers and instructors will provide guidance and supervision during the sessions.
In addition to this course work, the program would include an 4-month internship or research project.
The first year of classes could consist of *Overview of Data Science*, *Data Analysis Algorithms*, *Database Theory*, *Applied Statistics* and *Visualization Techniques of Unstructured Data Sets*, while the second year could contain courses in *High Dimensional Data Analysis*, *Graph Theory Applications*, *Data Mining*, *Search and Classification Algorithms*, and *Distributed Systems*, and would be completed by an Internship or a Research Project.
As with the HPC/ARC track, eligible courses from the *HPC* Master’s Program are also possible to take with previous consent of the graduate coordinator or adviser.
Conclusion {#sec:concl}
==========
We have demonstrated the need for programs in higher education in High-Performance Computing and Data Science. If the qualitative evidence of this seems somewhat limited, it should be understood that existing HPC and DS programs (academic and non-academic) are still relatively new. While some such programs are already in existence, in many cases students must use non-academic options, or teach the material to themselves. Academic programs would offer the benefit of not just teaching specific technical skills, but an education in the fundamentals of HPC and DS and instilling the analytical skills needed to adapt to an ever-changing technological landscape.
We have reviewed existing academic and non-academic education programs, in both HPC and DS. In light of this review, we presented a design for Master’s programs in HPC and DS, based on these examples and drawing from the experience and enrollment statistics in not-for-credit training in HPC and DS by the SciNet HPC Consortium at the University of Toronto.
To get well-founded graduate master’s programs off the ground will not be without challenges. It will likely involve partnerships and discussions with other departments and institutes in order to offer a stronger and multi-disciplinary program. Existing HPC Centers, which already operate between different disciplines, can play a fundamental role in bringing together such programs.
[^1]: Similar initiatives and trends are being carried on by the International HPC Summer School [@website:iHPCss] within the theme of HPC Challenges in Computational Sciences.
[^2]: This includes 3 previously separate module courses: *Scientific Software Development*, *Numerical Tools for Physical Scientist*, and *High Performance Scientific Computing*.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A new version of the phase method of determining the sound velocity is 55proposed and implemented. It utilizes the “Nonius” measurement technique and can give acceptable accuracy ($\leq 1\%)$ in samples of submillimeter size. Measurements of the sound velocity are made in single-crystal samples of the borocarbides RNi$_{2}$B$_{2}$C (${\rm R = Y,Lu,Ho}$). The elastic constants and the Debye temperature are calculated.'
address:
- 'B. Verkin Institute for Low Temperature Physics and Engineering, National Academy of Sciences of Ukraine, pr. Lenina 47, 61103 Kharkov, Ukraine'
- 'Department of Physics, Pohang University of Science and Technology, Pohang 790-784, Korea'
author:
- 'E. A. Masalitin, V. D. Fil,[^1] K. R. Zhekov, A. N. Zholobenko, and T. V. Ignatova'
- 'Sung-Ik Lee'
date: 'Published Low Temp. Phys. [**29**]{}, No. 1 (2003)'
title: 'Elastic constants of borocarbides. New approach to acoustic measurement technique'
---
1. INTRODUCTION {#introduction .unnumbered}
===============
An important problem of physical acoustics is to obtain reliable data on the elastic constants of newly synthesized compounds. These data, while being of independent interest, also serve as tests for theoretical calculations of band structures, force constants, and phonon spectra. As a rule, newly synthesized materials come either in the form of products of solid-phase synthesis (i.e., more or ‘less porous ceramics) or in the form of fine single crystals. Objects of the first group are characterized by appreciable scattering of elastic vibrations, making it practically impossible to use some version of a resonance or quasiresonance (of the long-pulse type) method to determine the absolute values of the sound velocity in them. Single crystals most often are of millimeter or submillimeter size; besides, in layered crystals the characteristic size in the direction perpendicular to the layers is often 100–200 $\mu$m or even less. To determine the elastic constants of such objects the method of ultrasonic resonance spectroscopy[@1] was developed, which consists in measurement of the spectrum of resonance frequencies of a sample and subsequent solution of the inverse problem of recovering all the components of the tensor of elastic constants. The technique is inherently a resonance method, i.e., it applies only to objects with small scattering (damping), a condition which is not always possible to satisfy even in small single crystals, e.g., near points of phase transitions. In addition, it can be implemented only in samples having a definite simple geometric shape (rectangular parallelepiped). The lucidity of this method is compromised by the complexity of the mathematical processing, making it hard to spot possible errors.
We have implemented a new version of the phase method of measuring sound velocities; it is applicable both to ceramic samples with strong scattering and to single crystals of submillimeter size. Utilizing a kind of “Nonius” measurement procedure, the method permits one to achieve acceptable accuracy (as a rule, better than 1%) in both cases. It has been used to measure the sound velocity in MgB$_{2}$ polycrystals[@2] and in VSe$_{2}$ layered single crystals.[@3] Furthermore, being completely independent of the nature of the signals to be analyzed, the instrumental implementation of the method enables one to study the variation of the amplitude and phase of any pulsed high-frequency signals. In particular, it has been used to measure the characteristics of an electric field accompanying a longitudinal sound wave in a metal.[@4]
Section 2 of this paper is devoted to a description of the basic principles of implementation of the “Nonius” method of phase measurements of sound velocity. In Sec. 3 we present the results of measurements of the elastic constants in single crystals of the borocarbides RNi$_{2}$B$_{2}$C (${\rm R = Y,Lu,Ho})$.
2. “NONIUS” METHOD OF SOUND VELOCITY MEASUREMENT. PRINCIPLES AND INSTRUMENTAL IMPLEMENTATION {#nonius-method-of-sound-velocity-measurement.-principles-and-instrumental-implementation .unnumbered}
============================================================================================
A block diagram of a device implementing the technique is presented in Fig. \[f1\].
\[f1\] {width="15cm"}
It is essentially a standard compensation or bridge circuit, depending on the algorithm used for process the pulsed signals, which is set by the pulse code modulation unit. In the bridge mode the signal that passed the sample channel is summarized with the equal in absolute value antiphase comparison one. The amplitude and phase of the latter are regulated by the receiver, which functions as a null device. The unbalance signal is separated into amplitude and phase components by high-frequency synchronous detectors.[@5] In the compensation mode the receiver, with the aid of sampling–storage devices, matches the amplitudes of the signals arriving at its input at different times. In this case the unbalance signals with respect to amplitude and phase are produced through a special code modulation of the pulse trains of the signals in the two channels. In any variant the data input to the computer are the readings of an attenuator (amplitude of the comparison signal) and phase meter (phase difference of the signal to be analyzed and the comparison signal).
Two original developments employed in the implementation of this standard scheme have substantially expanded its operational capabilities: an electronically controlled (linear) phase shifter with a practically unlimited tuning range, and a new data processing algorithm, which maintains a phase shift of 120$^{\circ}$ (or 240$^{\circ})$ between the signals being analyzed. The advantages of the new phase shifter are quite obvious. In particular, in relative measurements this phase shifter provides a practically unlimited dynamic range while maintaining an extremely high accuracy of measurement, which is actually determined by the resolution of the phase meter (at a signal-to-noise ratio $ \geq 5$). Let us discuss the second development in somewhat more detail. In the bridge mode the working algorithm of the circuit consists in maintaining a null signal at the input of the receiver upon changes in the sound velocity and damping in the sample. In inhomogeneous (e.g., polycrystalline) samples, internal re-reflections and mutual conversion of different modes at inhomogeneities lead to nonconstancy of the phase of the signal over the duration of the rf pulse envelope. The same situation is observed in short single crystals due to the superposition of secondary reflections. In this case the length of the time interval during which the sum of the two signals has zero amplitude turns out to be short ($ \leq 10^{-7}$ s). For analysis of such narrow features the receiving system should have a rather wide passband and not allow any overshoots under reproducing steep signal fronts.
For the 120$^{\circ}$ algorithm the sum of two signals of identical amplitude (their equality is maintained by an independent channel) is equal to the amplitude of each of the signals (equilateral triangle). In this case at the time of sampling–storage there are no sharp amplitude drops at the input of the receiver; this substantially improves the working of the system as a whole. A distinct advantage of the 120$^{\circ}$ algorithm is that it is unnecessary to have frequency (phase) modulation of the master oscillator in order to obtain unbalance signals of different polarity upon passage through the compensation point, as one must have for self-balancing of the circuit. Furthermore, the usual amplitude detection used in the 120$^{\circ}$ algorithm allows one to use as the signals of the two channels any two reflections that have traveled different distances in the sample.
The measurement algorithm in part resembles one proposed earlier.[@6] First the phase–frequency (P–F) characteristic of an acoustic circuit consisting of two delay lines is measured at fixed frequency points (step 1). Then the P–F characteristic of a sandwich consisting of the same delay lines but with the sample between them (Fig. \[f2\]) is measured at the same temperature (step 2).[^2]
![Phase–frequency characteristics of the delay lines ([*1*]{}), of a sandwich consisting of the sample (LuNi$_{2}$B$_{2}$C, ${\bf q}\parallel [100]$, ${\bf u}\parallel [100]$, $L = 0.835$ mm) and the delay lines ([*2*]{}), and the difference function, i.e., the P–F characteristic of the sample ([*3*]{}). Note the difference in the scales of the vertical axes. []{data-label="f2"}](ris2a.eps){width="7cm"}
Because the signal circuits contain elements capable of resonating (piezotransducers, imperfectly matched feeders), each of these characteristics is not necessarily a straight line. However, their difference, i.e., the P–F characteristic of the sample, in the absence of interference distortions in it, should form a strictly straight line, the slope of which determines the phase velocity of the sound, $$v=\frac{360 L}{S}$$ where $v$ is the sound velocity (cm/s), $L$ is the thickness of the sample (cm), and $S$ is the slope of the P–F difference characteristic (deg/Hz). It is easily seen by a direct calculation that when the P–F characteristics [*1*]{} and [*2*]{} are approximated by straight lines by the least-squares method (the slopes are $S_{1}$ and $S_{2}$, respectively), then $$S=S_1 - S_2$$ for any deviations of the P–F characteristics [*1*]{} and [*2*]{} from straight lines. This relation is valid only if the frequency points at which the P–F characteristics [*1*]{} and [*2*]{} are measured are coincident. In Ref. essentially the same procedure was used to determine $S$, but since the technique used there did not guarantee the required coincidence, additional errors could have been introduced.
If $S$ is comparable to $S_{1}$ (0.3 or larger), then in homogeneous materials the measurements can be limited to this step with completely acceptable accuracy (0.3% or better).
However, in homogeneous but rather thin samples the superposition of secondary reflections distorts the main part of the measurement signal. Because of this, the parts of the pulse that coincide with the leading edge are customarily used for measurements. An analogous procedure, as a rule, should be used in inhomogeneous materials for the reasons already mentioned, even though the acoustic path length in them may be comparatively large.
As a result of the occurrence of various kinds of transient processes, the rate of which depends on the carrier frequency of the pulses, the slopes of the P–F characteristics [*1*]{} and [*2*]{} become functions of the temporal position of the strobe readout pulse at the leading edge of the measurement signal. The variation of $S_{1,2}$, depending on the type of piezotransducers, is 2–4% (for comparison, in extended samples the variation of $S_{1,2}$ at the steady part of the pulse is at the 0.1% level). This means that in going from step 1 to step 2 the readout pulse should be shifted precisely by the sound delay time $\tau _0$ in the sample. Since the latter is initially unknown and also because of the discreteness of the step for the time shift of the strobe signal ($5\times 10^{-8}$ s in our experiments), it was practically impossible to satisfy this condition. For finding $\tau _0$ (and, hence, the sound velocity) we used the following interpolation procedure.
For each series of measurements with a definite mode (longitudinal or transverse) we calibrated the dependence of $S_{1}$ on the temporal position $t_{x}$ of the readout pulse. Then for a given sample we measured $S_{2}$ at some known position $t_{c}$ of the readout pulse at the leading edge of the signal. It is easy to see from Eqs. (1) and (2) that $\tau _0$ is a solution of the equation $S(x) = 360x$, where $x\equiv t_{c}-t_{x}$ is the time shift of the readout pulse between the set of calibration measurements $S_{1}$ and the measurements with the sample, $S_{2}$. An example of the graphical solution of the interpolation equation for several values of $t_{c}$ is presented in Fig. \[f3\]. The results of the interpolation (the value of $\tau
_0)$ coincide regardless of the choice of $t_{c}$.
![Example of the interpolation procedure for finding the sound delay time $\tau _0$. YNi$_{2}$B$_{2}$C sample (${\bf q}\parallel
[100]$, ${\bf u}\parallel [010]$, $L = 0.885$ mm) for several values of $t_{c}$ (see text). At $x = 0$ the values of $t_{c}$ increase from bottom to top with a step of $5\times 10^{-8}$ s. The linear function $F(x) = 360x$. []{data-label="f3"}](ris3a.eps){width="7cm"}
At this step of the procedure the “rough” determination of the sound velocity is completed. To refine the values we use the “Nonius” method. Let the phase of the signal registered at some definite frequency $f_0$ by the phase meter in step 1 be equal to $\Phi _{1}$. In step 2 at the same frequency the phase of the signal will be $\Phi
_{2}$. The total phase inserted by the sample is $\Phi _0 = 360n+(\Phi
_{2}-\Phi _{1}$), where $n = 0,1,2,\ldots\,\,$. Since $\Phi _0 =
360f_0L/v$, by trying values of $n$ we find the refined value of $v$ that is closest to the “rough” estimate.
In the above discussion it was tacitly assumed that on going from step 1 to step 2 the phase of the signal changes only because of the addition of the sample. Actually, however, besides the sample we also had an additional layer of grease in step 2. During measurements in very thin samples the contribution of the grease layer can become noticeable. In our experiments GKZh-94 silicone oil was used as the bonding agent, forming a layer ${\sim }1$–$2$ $\mu$m thick between the ground surfaces. The passage of an elastic wave through such a thin layer is described by the sum of an infinite geometric progression with the denominator $q = k^{2}{\rm e}^{-2l(\alpha +iq)}$, where $k$ is the reflection coefficient at the boundary (we assume that the wave impedances of the delay line and sample are close in value), $l$ is the thickness of the grease layer, $\alpha $ is the damping coefficient, and $q$ is the wave number.
An estimate of the propagation velocity of sound in the grease gave $v_{l}\sim 2.1\times 10^{5}$ cm/s, $v_{t}\sim 1.2\times 10^{5}$ cm/s, which correspond to reflection coefficients $k\sim 0.85$ for our samples. In Fig. \[f4\] we present the calculated dependence of the phase of the wave passing through the grease layer on the thickness for various damping coefficients. The regions of $ql$ corresponding to the conditions of the experiment are also indicated in Fig. \[f4\].
![Diagram pertaining to the calculation of the additional phase shift inserted by the grease layer. The reflection coefficient at the grease–sample boundary $k = 0.85$, and the numbers on the curves give the sound damping coefficient in the grease layer (neper/cm). The dotted curve corresponds to $\Phi = -ql$. The horizontal lines indicate the regions of the actual values of the parameter $ql$ for the corresponding mode. []{data-label="f4"}](ris4.eps){width="7cm"}
At low damping the correction can be rather large. We were unable to estimate the value of the sound damping in the grease — in thick layers (${\sim }0.5$ mm) it was very large, probably because of cracking — but we assume that its value is found at the 20 dB/cm level or higher, i.e., the phase inserted by the grease layer is close to $ql$. In processing the results of the measurements we introduced a correction for the additional grease layer — 10$^{\circ}$ for longitudinal sound and 20$^{\circ}$ for transverse sound. In thin samples the effect of this correction was not over 1%. We suppose that this correction can be eliminated by making comparative measurements on two samples of different thickness.[@3] In that case the length difference $\delta L$ should be comparable to $L$, since otherwise the contribution of possible nonuniformities of the sound velocity over the whole length of the sample would be attributed to the small difference $\delta L$.
Let us conclude with an estimate of the potential accuracy of a single measurement. Special studies have established that the irreducibility of the phase upon the remounting (regluing) of the acoustic circuit is at the level of 20$^{\circ}$. We estimate the indeterminacy of the correction for the additional grease layer to be 10$^{\circ}$. Assuming that the accuracy of the “rough” estimate of the velocity is sufficient for determining the necessary value of $n$, we obtain for the measurement error (at $f_0\sim 50$ MHz) $$\frac {\delta v} {v} = \frac {30} {\Phi_0} \approx 2 \cdot 10^{-9}
\frac {v} {L}$$
3. ELASTIC CONSTANTS OF BOROCARBIDES ${ \rm RN \lowercase{i}_{2}B_{2}C}$ (${\rm R = Y,Lu, Ho} $) {#elastic-constants-of-borocarbides-rm-rn-lowercasei_2b_2c-rm-r-ylu-ho .unnumbered}
=================================================================================================
In spite of the significant interest in the family of superconducting borocarbides, very little information about their elastic properties can be found in the literature. We know of only one “acoustical” study,[@7] devoted to YNi$_{2}$B$_{2}$C, in which the sound velocity was measured by a time-of-flight method. Single crystals of borocarbides were grown by the method described in Ref. and had the shape of a slab with a maximum dimension along the \[001\] axis of ${\sim }0.8$ mm (${\rm R = Y}$), ${\sim }0.2$ mm (${\rm R =
Ho}$), and ${\sim }0.4$ mm (${\rm R = Lu}$). They were quite brittle, and therefore the mounting of the samples between the delay lines was done with the aid of a special brass ring, which acted as a holder and reinforcer; the ring was ground simultaneously with the preparation of the working faces (Fig. \[f5\]). The diameter of the ring was chosen larger than the diameter of the piezotransducers to prevent spurious signals.
![Diagram of the mounting of the sample: 1 — piezotransducers, 2 — delay lines, 3 — brass support ring, 4 — sample. []{data-label="f5"}](ris5.eps){width="8cm"}
All of the measurements were made at liquid nitrogen temperature. The results are presented in Table \[t1\]. It contains some “superfluous” data, marked by an asterisk $(^{\ast })$. For example, for $C_{44}$ it was sufficient to make a single measurement ${\bf q}\parallel [100]$, ${\bf u}\parallel [001]$ (${\bf u}$ is the polarization vector of the elastic wave). We assume, however, that having the “superfluous” data will make it possible to get an idea of the accuracy of the measurements in this case.
One can also see that certain relations which follow from the general theory of elasticity[@9] are well satisfied. For example, in a tetragonal crystal the sum of the squares of the velocities of the three modes remains constant under rotation of the wave vector ${\bf
q}$ in the (001) plane.
The elastic constants of the single crystals studied are presented in Table \[t2\]. The x-ray densities were used in calculating them. For ${\rm R = Y}$ the agreement with the results of Ref. is poor, although the relationships among the various constants are preserved on the whole. The Debye temperature was calculated according to the formula[@9]
$$\theta_{D} = 1146.8 \biggl( \frac {\rho s} {A I} \biggr) ^{1/3}$$
where $A$ is the molecular weight, $s$ is the number of atoms in the molecule, $\rho $ is the mass density, and $I$ is the sum of the inverse cubes of the phase velocities of the elastic waves, averaged over all directions of the wave normal. For ${\rm R = Ho}$, because of the difficulty of preparing a sample of the required orientation, the elastic constant $C_{13}$ was not measured, and in the calculation of the bulk modulus and Debye temperature it was assumed equal to the value of $C_{13}$ in lutecium borocarbide. For ${\rm R = Y}$ the calculated value of $\theta _{D}$ is close to the thermodynamic estimate.[@10] For ${\rm R = Lu}$ the deviation of the calculated value of $\Theta _{D}$ from the thermodynamic value is, generally speaking, greater than the allowable error. That may be an indication of the existence in lutecium borocarbide of a low-temperature ferroelastic structural transition, accompanied by a significant softening of some elastic constant. Our preliminary measurements in holmium borocarbide have shown that at 5.2 K the velocity of the $C_{66}$ mode falls to ${\sim }3.3\times 10^{5}$ cm/s. When this softening is taken into account, one obtains $\theta _{D} = 383$ K for ${\rm R = Ho}$.
This study was partly supported by the Government Foundation for Basic Research of the Ministry of Education and Science of Ukraine (Grant No.0207/00359).
------------------------------------ ------------------------- ---------------------- ---------------------- ----------------------
$\bf {q \| }$ $\bf {u \|}$ $v \cdot 10^5$cm/sec $v \cdot 10^5$cm/sec $v \cdot 10^5$cm/sec
\[100\] 6.78 (0.885) 5.88 (0.8) 6.04 (0.606)
\[001\] 3.25 (0.885) 2.65 (0.8) 2.73 (0.606)
\[010\] 4.80 (0.885) 4.30 (0.8) 4.33 (0.606)
\[110\]$^{\ast }$ 7.55 (0.59) 6.64 (0.988) 6.86 (0.525)
\[001\]$^{\ast }$ 3.26 (0.59) 2.64 (0.988) -
\[$ 1 \overline{1} 0$\] 3.34 (0.59) 2.77 (0.988) 2.83 (0.525)
\[001\] 6.49 (0.84) 6.01 (0.4) 5.91 (0.23)
\[100\]$^{\ast}$ 3.26 (0.84) 2.70 (0.4) 2.81 (0.23)
\[010\]$^{\ast }$ 3.28 (0.84) 2.70 (0.4) 2.83 (0.23)
45$^{\circ}$ from the \[001\] axis QL$^{\ast}$ 7.28 (0.303) - -
in the (110) plane QT 3.18 (0.465) - -
\[110\]$^{\ast}$ 3.31 (0.303) - -
45$^{\circ}$ from the \[001\] axis
in the (100) plane QT - 2.01 (0.27) -
------------------------------------ ------------------------- ---------------------- ---------------------- ----------------------
: Sound velocity in single crystals of borocarbides ($T = 77$ K). []{data-label="t1"}
Note: The “superfluous” data are denoted by an asterisk ($^{\ast }$). QL and QT are the quasilongitudinal and quasitransverse modes; the thickness of the sample in mm is given in parentheses.
\[t2\]
------------------ ------- ----------- ----------------- -------
C$_{11}$ 27.94 22\[7\] 29.39 29.47
C$_{12}$ 14.39 9.84\[7\] 16.34 16.53
C$_{13}$ 17.81 23.15
C$_{33}$ 25.61 21.1\[7\] 30.68 28.20
C$_{44}$ 6.43 5.42\[7\] 5.97 6.02
C$_{66}$ 14.00 13.1\[7\] 15.71 15.15
B 20.16 20.27 23
$\theta_{D}$, K 501 490\[10\] 409 (360\[10\]) 404
$\rho$, g/cm$^3$ 6.08 6.05\[7\] 8.5 8.08
------------------ ------- ----------- ----------------- -------
: Calculated parameters for borocarbides ($T = 77$ K)
Note: $C_{ik}$ are elastic constants (in units of $10^{11}$ dyn/cm$^{2}$), $\theta _{D}$ is the Debye temperature, and $B$ is the bulk modulus. For Ho the modulus C$_{13}$ was not measured, and in the calculation of $\theta _{D}$ and $B$ it was assumed equal to 23.15 (see text).
A. Migliory, J. L. Sarrao, W. M. Visscher, T. M. Bell, M. Lei, Z. Fisk, and R. G. Leisure, Physica B [**183**]{}, 1 (1993).
T. V. Ignatova, G. A. Zvyagina, I. G. Kolobov, E. A. Masalitin, V. D. Fil, Yu. B. Paderno, A. N. Bykov, V. N. Paderno, and V. I. Lyashenko, Fiz. Nizk. Temp. [**28**]{}, 270 (2002) \[Low Temp. Phys. [**28**]{}, 190 (2002)\].
I. A. Gospodarev, A. V. Eremenko, T. V. Ignatova, G. V. Kamarchuk, I. G. Kolobov, P. A. Minaev, E. S. Syrkin, S. B. Feodos’ev, V. D. Fil, A. Soreau-Leblanc, P. Molinie, and E. C. Faolques, Fiz. Nizk. Temp. \[Low Temp. Phys.\] (in press).
Yu. A. Avramenko, E. V. Bezuglyi, N. G. Burma, V. M. Gokhfeld, I. G. Kolobov, V. D. Fil, and O. A.Shevchenko, Low Temp. Phys. [**28**]{}, 328 (2002).
V. D. Fil’, P. A. Bezuglyĭ, E. A. Masalitin, and V. I. Denisenko, Prib. Tekh. Éksp., No. 3, 210 (1973).
E. V. Bezuglyĭ, N. G. Burma, I. G. Kolobov, V. D. Fil’, I. M. Vitebskiĭ, A. N. Knigavko, N. M. Lavrinenko, S. N. Barilo, D. I. Zhigunov, and L. E. Soshnikov, Fiz. Nizk. Temp. [**21**]{}, 86 (1995) \[Low Temp. Phys. [**21**]{}, 65 (Erratum p. 452) (1995)\].
S. Isida, A. Matsushita, H. Takeya, and M. Suzuki, Physica C [**349**]{}, 150 (2001).
M. O. Mun, S. I. Lee, W. C. Lee, P. C. Canfield, B. K. Cho, and D. C. Johnston, Phys. Rev. Lett. [**76**]{}, 2790 (1996).
F. I. Fedorov, [*Theory of Elastic Waves in Crystals*]{}, Plenum Press, New York (1968), Nauka, Moscow (1965).
H. Michor, T. Holubar, C. Dusek, and G. Hilscher, Phys. Rev. B [**52**]{}, 16165 (1995).
[^1]: E-mail: [email protected]
[^2]: At the frequencies we used the scale of the phase variations of the signal are much greater than 360$^{\circ}$. The phase meter, of course, measures phase differences in the interval 0–360$^{\circ}$, and the absence of discontinuities (360$^{\circ}$ jumps) in Fig. \[f2\] is achieved through programming.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We define and explore the concept of ideal stabilization. The program is ideally stabilizing if its every state is legitimate. Ideal stabilization allows the specification designer to prescribe with arbitrary degree of precision not only the fault-free program behavior but also its recovery operation. Specifications may or may not mention all possible states. We identify approaches to designing ideal stabilization to both kinds of specifications. For the first kind, we state the necessary condition for an ideally stabilizing solution. On the basis of this condition we prove that there is no ideally stabilizing solution to the leader election problem. We illustrate the utility of the concept by providing examples of well-known programs and proving them ideally stabilizing. Specifically, we prove ideal stabilization of the conflict manager, the alternator, the propagation of information with feedback and the alternating bit protocol.'
author:
- 'Mikhail Nesterenko[^1]'
- 'Sébastien Tixeuil[^2]'
bibliography:
- 'ideal.bib'
- '../../biblio/biblio.bib'
title: Ideal Stabilization
---
Introduction {#SecIntro}
============
A program is *self-stabilizing* [@D74j; @D00b; @T09bc] (or just *stabilizing*) if, regardless of the initial state, it eventually satisfies its specification. This elegant property enables the program to recover from transient faults or lack of initialization. During this stabilization period the program behavior is unpredictable. It is tempting to attempt to engineer the specification such that the program behavior during fault-recovery is controlled. For example, the program starts behaving correctly in no more than ten steps, or critical messages are never lost. However, one of the features of classic stabilization is that the program does not have to satisfy the specification for an arbitrary amount of time. That is, the program is free to ignore the recovery constrains built into the specification.
Another disadvantage of stabilizing programs is their poor compositional properties. Stabilization programs are usually composed by layers: an lower level components are not influenced by the higher level components and, after the lower component starts behaving correctly, they higher level, due to stabilization will eventually behave correctly as well. However, if there is non-trivial two-way interaction between components, the stabilization or correct operation of the composed system is not guaranteed. These shortcomings diminish the attractiveness of stabilization as a viable fault-tolerance technique.
In this paper we study the class of programs that always satisfy their specification. We call such programs *ideally stabilizing*. Related concepts are periodically considered by fault-tolerance researchers. However, these approaches are often regarded as theoretical curiosity with a few isolated examples of no practical importance. Our thesis is that the opposite is true. Ideal stabilization retains most of the advantages of classical stabilization while allowing the engineers and program designers to control the program behavior during fault recovery. Moreover, since the ideally stabilizing programs always satisfy their specification, their composition is similar to conventional program composition. Thus, the vast array of established program composition techniques can be applied to ideally stabilizing programs. We are thus hopeful that our approach to stabilization makes the general concept of self-stabilization more attractive to fault-tolerance practitioners.
#### Related work.
Ideal stabilization is close to *snap-stabilization* [@BDPV99c; @DDNT08c] and should be considered complementary. However, snap-stabilization is defined in terms of immediate correct satisfaction of *external* invocations. Such definition may lead to specifications with sequence-based safety and liveness properties. Proving stabilization to such specifications often results in operational proofs that are difficult to verify. Ideal stabilization, on the other hand, does not restrict specifications in any way.
Adding safety properties to self-stabilizing protocols has been a very active recent path of research, however most approaches make some restrictions on the *nature* or the *extent* of the faults in order to guarantee a particular kind of safety. Safe stabilization [@GB03c; @BGR06c] refers to the fact that *few* faults hitting the network should not compromise a particular safety predicate, while preserving global self-stabilization of the system. Also, the line of work about fault-containing stabilization [@GGHP96c] may handle only a *single* transient failure to ensure actual containment and recovery, and super-stabilizing protocols [@DH97j] can withstand one topology change at a time. When faults are simple enough to be detected by checksums [@HP00j], it is also possible to maintain elaborate safety predicates. Other approaches included using *external* entities that enforce safety [@DS03j] or make use of non-corruptible memory [@LS92c]. By contrast, ideal stabilization does not restrict the nature or the extent of the transient faults that can hit the system in *any* way, and does not make use of any external entity.
#### Our contribution.
In this paper we study two approaches to ideal stabilization. The approaches depend on the specification type. Specification itself is ideal if it allows (*i.e.* mentions) all possible states. Specification is not ideal otherwise. Ideal stabilization to specifications that are not ideal hinges on the approach we call *state displacement*. The specification implementer provides such mapping from program states to specification states that none of the possible program states map to disallowed specification states. We identify the necessary condition for such specifications to allow ideal stabilization and explain how two well-known programs: conflict manager and the alternator use state displacement to achieve ideal stabilization.
The second approach relies on stating the specification such that all the possible states are allowed. This allows the engineer to specify precisely what behavior, including failure recovery behavior, is expected of the program. Ideally stabilizing program, by definition, has to follow this specification exactly. We state a proposition that states that such programs are rather common. As an example, we consider the problem specifications for two well-known stabilizing programs: propagation of information with feedback and alternating-bit protocol and provide assertional proofs that the programs ideally stabilize to these specifications.
Model {#SecModel}
=====
This section introduces the notation and terms we use in the rest of the paper. To the person familiar with the literature on self-stabilization, our notation may look fairly conventional. However, we encourage even the specialists to read this section as the understanding of the results in the further sections hinges on the notions defined in this one.
\
**Program.** A *program* consists of a set of $N$ processes. Each process contains a set of variables. Every variable ranges over a fixed domain of values. Variable $v$ of process $p$ is denoted $v.p$. A process *state* is an assignment of a value from its domain to each variable of the process. A program state, in turn, is an assignment of a value to every variable of each process. The Cartesian product of the values of all program variables is program *state universe*. That is, the state universe defines all states that the program can assume.
Each process also contains a set of actions. An action has the form $\langle name\rangle : \langle guard \rangle \longrightarrow \langle
command \rangle$. A *guard* is a predicate over the variables of the process. A *command* is a sequence of assignment and branching statements.
An action whose guard is **true** at some program state is *enabled* in that state. The execution of an enabled action changes the values of program variables and thus transitions the program from one state to another. A program *computation* is a maximal sequence of such transitions. By maximality we mean that the computation is either infinite or it ends in a state where none of the actions are enabled. Note that we do not assume any fairness of action execution for infinite computations.
\
**Communication model, extended state.** Processes that share variables are *neighbors*. The communication model determines the type and access method of the variables shared by neighbor processes. For example, in *shared-memory* communication model, the process may mention the variables of the neighbor processes in its actions. That is, the process may read the state of its neighbor processes. The *extended* state of the process is the state of its local variables and the variables that the process can read.
\
**Problem specifications and program sequence mapping.** *Problem specification* prescribes the program behavior. This is done by defining the program inputs and outputs through *external variables*. External variables are thus either *input* or *output* variables. Input variables are modified by the environment while the program may only read them. The output variables are updated by the program, they are used to display the results of the program computation.
The problem specification is the set of sequences of states of external variables. A program implements the specification. Part of the implementation is the mapping from the program states to the specification states. This mapping does not have to be one-to-one. However, we only consider *unambiguous* programs where each program state maps to only one specification state.
We make another important assumption about program sequence mappings. The mappings have to be *merge-symmetric*. Specifically, let there be a set of program states $pr_1$ through $pr_k$ such that they map to specification states $s_1$ through $s_k$. Then, if there is a specification state $s_m$ such that the extended state of each process $p$ in $s_m$ is the same as in one of the states $s_1$ through $s_k$, then there exists a program state $pr_m$ that maps to $s_m$. In other words, any specification state formed by extended process state-preserving combination of other specification states, has a program state that maps to this new specification state. Most known mappings are merge-symmetric. *Identical* mapping maps every program state to the specification state. That is, the program operates external variables only. Another simple program mapping is *projection*. Each process maintains output variables and internal variables for computations and record keeping. The projection of program states onto specification states removes the internal variables. However, the mapping may not be as straightforward as identical mapping or projection. For example, the specification requires the output variable of a process be boolean while the program maintains an integer variable. The mapping is such that the even values of the integer variable are mapped to **true** while odd values to **false**.
Once the mapping between program and specification states is established, the program computations are mapped to specification sequences as follows. Each program state is mapped to the corresponding specification state. Then, *stuttering*, the consequent identical specification states, is eliminated.
The program does not have to implement all specification sequences. However, the program cannot select implementation sequences so that it ignores environment input. Hence the following notion of input completeness. Given a set of sequences , a subset $\MPROG{B} \subset\MPROG{A}$ is *input-complete* if, for every sequence $\alpha \in \MPROG{A}$, there exists a sequence $\beta \in
\MPROG{B}$ such that: every step $s_1$ of $\alpha$ is also in $\beta$ and for every pair of steps $s_1$ and $s_2$ their order in $\alpha$ and $\beta$ is the same. Informally, the sequences in an input-complete subset preserve the results and the order of the input steps in .
The state universe of the specification is the Cartesian product of the values of all the external variables. The state that is present in one of the specification sequences is *allowed* by the specification. The state is *disallowed* otherwise. The specification is *ideal* if it allows every state in its state universe; otherwise the specification is not ideal.
We only consider specifications that are suffix-closed. That is, every suffix of a specification sequence is also a sequence in this specification. Suffix closure enables us to discuss the correctness of the program on the basis of its current state rather than potentially arbitrary long program history. This facilitates assertional reasoning about program correctness.
\
**Predicates, invariants, stabilization.** A state *predicate* is a boolean expression over program variables. A program state *conforms* to predicate $R$, if $R$ evaluates to **true** in this state; otherwise, the state *violates* $R$. By this definition every state in the program state universe conforms to predicate **true** and none conforms to **false**.
The predicate defines a set of program states that conform to it. In the sequel we use the predicate and the set of states it defines interchangeably. Predicate $R$ is *closed* in a certain program , if every state of every computation of conforms to $R$ provided that the computation starts in a state conforming to $R$.
A closed predicate $I$ is an *invariant* of the program with respect to specification if $I$ has the following property: every computation of that starts in a state conforming to $I$, maps to a sequence that belongs to the specification. A program state is *legitimate* if it conforms to the invariant and *illegitimate* otherwise.
Program *satisfies* (or *solves*) specification , if there exists an invariant $I$ of with respect to such that the mappings of the program computations that start from $I$ form an input-complete subset of . That is, the program does not have to implement all the specification sequences, but it does need to accommodate all possible inputs. Specifically, it needs to implement an input-complete subset of these sequences.
A program is *stabilizing* to specification if every computation that starts in an arbitrary state of the program sate universe contains a state conforming to the invariant with respect to . Therefore, any computation of a stabilizing program contains a suffix that implements a specification sequence.
A program is *ideally* stabilizing if every state in its state universe is legitimate.
That is, **true** is an invariant of an ideally stabilizing program.
State Displacement {#SecNonIdealSpec}
==================
Necessary Condition for Specification
-------------------------------------
A non-ideal specification disallows certain states in its state universe. Yet, every state in the program state universe is legitimate. Thus, for a program to ideally stabilize to such specification, the state mapping should be such that the disallowed states are displaced. That is, none of the states in the program state universe maps to the disallowed states. However, state displacements may not be possible for an arbitrary specification. In this section we provide a theorem that establishes a necessary condition for a specification to be solvable by an ideally-stabilizing program.
\[trmNoIdeal\] An ideal stabilization is possible to non-ideal specification only if the specification contains an input-complete subset of sequences such that in every disallowed specification state there is at least one process whose extended state does not occur in any of the specification states of this subset.
Assume the opposite. There is, a non-ideal specification that disallows state $d$ and for every input-complete subset of and for every process $p_i$, where $i=1,N$, there is a specification state $c_i$ in one of the sequences of such that the extended state of $p_i$ is the same in $c_i$ and in $d$. However, there is a program that ideally stabilizes to .
Since solves , implements an input complete-subset of . Assume, without loss of generality, that implements . That is, for every sequence of there is a computation of that maps to this sequence. This means that for each specification state of , there is a program state of that maps to it. This includes the states $c_i$. For each $i$, let $pr_i$ be the program state of that maps to $c_i$.
Recall that the extended state of each process in $d$ is the same as in one of the states $c_i$. If this is the case then, according to merge-symmetry of program mapping, there exists are program state $pr_d$ that maps to $d$. Let us consider the computation of that starts in $pr_d$. This computation contains a state that maps to a disallowed state, $pr_d$ itself. That is, this computation maps to a sequence outside the specification . This means that $d$ does not conform to an invariant of with respect to . That is, $pr_d$ is an illegitimate state. However, if has an illegitimate state then, contrary to our initial assumption, does not ideally stabilize to . Hence the theorem.
Examples
--------
To illustrate the concept of ideal stabilization to non-ideal specifications and the ramifications of Theorem \[trmNoIdeal\], we provide several examples.
\
**Conflict manager.** The specification we consider is a simplified (unfair) variant of the dining philosophers problem [@CM84; @D68] that we call . The program is adapted from the deterministic conflict manager presented by Gradinariu and Tixeuil [@GT07cb]. The processes are arranged in a chain. Every process has a unique identifier. The specification defines one external output boolean variable $in$ per process. If the value of $in$ is **true**, the process may execute the exclusive critical section of code. The specification defines infinite sequences where $in$ variables alternate between **true** and **false**. The sequences are not necessarily fair as a certain process may never be given a chance to execute the crucial section. That is, an input-complete subset of the specification contains any subset of such sequences.
The specification prohibits concurrent critical section access by neighbor processes. That is, the specification disallows states where $in$ variables of two neighbors are **true** in the same state. Assuming shared memory communication, the extended state of the process contains the state of its neighbors. Hence, none of the allowed specification states contain an extended process state where both the process and one of its neighbors are inside the critical section. The specification thus satisfies the conditions of Theorem \[trmNoIdeal\].
The conflict manager program is as follows. Each process has a single boolean variable *access* and a single action *flip* that is always enabled. The action toggles the value of *access*.
$$\emph{flip}: \textbf{true} \longrightarrow access := \neg access$$
The program mapping is this. For each process $p$ the variable $p.in$ is **true** if $p.access$ is **true** and $p$ has the highest identifier among its neighbors with $access$ set to **true**.
Let us discuss why this mapping is merge-symmetric. Any extended process state-preserving combination of specification states produces a specification state where neighbors are not accessing the critical section. Then, by appropriately setting $access$ variables, we can generate the program state that maps to this specification state.
Let us give an illustration for this reasoning. Assume we have a chain of four processes with identifiers $\langle 2,1,3,4\rangle$. The extended state of each process in this case is its own state plus the state of its left and right neighbors. Consider two specification states $$s_1 \equiv
\langle\textbf{true}, \textbf{false}, \textbf{false}, \textbf{false}\rangle$$ and $$s_2 \equiv
\langle\textbf{false}, \textbf{false}, \textbf{false}, \textbf{true}\rangle$$ Some of the program states that map to $s_1$ and $s_2$ are respectively $pr_1 \equiv
\langle\textbf{true},\textbf{true},\textbf{false},\textbf{false}\rangle$ and $pr_2 \equiv
\langle\textbf{false},\textbf{false},\textbf{false},\textbf{true}\rangle$.
Specification state $s_3
\equiv\langle\textbf{true},\textbf{false},\textbf{false},\textbf{true}\rangle$ is formed by merging states $s_1$ and $s_2$. Note that the extended states of each process in $s_3$ are the same in either $s_1$ or $s_2$. For example, the extended state of process $p_2$ is $\langle
\textbf{true}, \textbf{false}, \textbf{false}\rangle$, which is the same in $s_1$ and $s_3$. Note that there are a number of program states that map to $s_3$. For example, $pr_3 \equiv
\langle\textbf{true},\textbf{false},\textbf{false},\textbf{true}\rangle$.
\[trmCM\] Program ideally stabilizes to the unfair dining philosophers specification .
To prove ideal stabilization we need to show that a computation of from an arbitrary program state satisfies . First, we show that every state of maps to an allowed state of . Indeed, among the neighbors whose $access$ is **true**, $in$ is set to **true** only for the process with the highest identifier. That is, every program state maps to the state of where neighbors do not access the critical section concurrently. Hence, no program state maps to a disallowed state.
Moreover in every computation of the program at least one process, the process with the largest identifier in the system, alternates between setting *access* to **true** and **false**. This means that this process alternates between entering and exiting the CS indefinitely. Such computations satisfy the specification. That is, ideally stabilizes to .
\
**Leader election.** We show a simplified leader election problem as an example of the specification to which ideally stabilizing solutions do not exist. Again, the processes form a chain. In this case, we only consider $N > 3$. In the external state, each process has two boolean variables: an input variable $contend$ and an output variable $leader$. The value of the input variable $contend$ is set to a particular value and does not change throughout the specification sequence. In each specification state $leader$ of at most one process is $\textbf{true}$. To exclude trivial solutions, the specification requires that the leader is elected only out of the processes that contend for leadership. That is, the processes whose $contend$ variable is **true**. Each specification sequence is finite and ends with a state where the leader is elected. Note that the input complete subset of sequences has to contain a sequence for every combination of the contending processes.
There does not exist a program that ideally stabilizes to the simplified leader election specification .
Let us consider state $s_1$ where the first process is the only one contending for leadership. This is the process that has to be elected leader. That is, the following output state has to be in every input-complete subset. $s_1 \equiv
\langle\textbf{true},\textbf{false},\cdots,
\textbf{false},\textbf{false}\rangle$ Similarly, let $s_2$ be the state where the last process is the only one contending: $s_2 \equiv
\langle\textbf{false},\textbf{false},\cdots,\textbf{false},\textbf{true}\rangle$.
We now form the state $s_3$ where both the first and the forth processes are contending for leadership and both of them are elected. That is, $$s_3 \equiv
\langle\textbf{true},\textbf{false},\cdots,\textbf{false},\textbf{true}\rangle$$ This state is disallowed . Yet, the extended state of every process is present in either $s_1$ or $s_2$. Thus, according to Theorem \[trmNoIdeal\], ideal stabilization is not possible to .
\
**Linear alternator.** For another example, we demonstrate how a well-known program called the linear alternator proposed by Gouda and Haddix [@GH99ca] fits into our ideal stabilization model. The alternator provides a solution to the fair variant of the dining philosophers problem . The problem specification is the same as described above except all the sequences are fair with respect to the process critical section access. The modified specification still excludes the states where two neighbors are executing the critical section concurrently and the specification still satisfies the conditions of Theorem \[trmNoIdeal\]. Therefore, the ideal solution is still possible for this specification.
The implementation of is as follows. Similar to manager, each process has a boolean variable $x$. This time though we assume that the processes in the chain are numbered in the increasing order from $1$ to $N$. The numbering is for presentation purposes only as as each process only need to be aware of its right and left neighbor. The process actions are as follows. In the actions, parameter $j$ ranges from $2$ to $N-1$.
$$\begin{array}{lll}
x.p_1=x.p_2 & \longrightarrow & x.p_1 := \neg x.p_1 \\
(x.p_j \neq x.p_{j-1}) \wedge (x.p_j = x.p_{j+1})
&\longrightarrow & x.p_j := \neg x.p_j \\
x.p_{N-1} \neq x.p_{N} &\longrightarrow & x.p_N := \neg x.p_N
\end{array}$$
The program to specification states mapping is as follows. For each process $p_j$, the output variable $in.p_j$ evaluates to **true** if process’ action is enabled.
Program ideally stabilizes to the fair dining philosophers specification .
Gouda and Haddix [@GH99ca] prove that the alternator satisfies the fairness properties of the dining philosophers specification from an arbitrary program state. We only show the displacement of disallowed states. Note that an action of a process $p$ is enabled if $x.p$ is not equal to the left neighbor’s variable and equal to the right neighbor’s variable. This can only hold for one process in the neighborhood. That is, every program state maps to a specification state where none of the neighbors are in the critical sections concurrently. In other words, the program states only map to the allowed states. Hence the theorem.
Stabilizing to Ideal Specifications {#SecIdealSpec}
===================================
Forming Ideal Specifications
----------------------------
Another method of achieving ideal stabilization is by stating the specification such that all the states in its universe are legitimate. That is, stating ideal specification. At first sight this seems difficult to achieve. However, the following proposition demonstrates that such specifications are rather common.
For every program there is an ideal specification to which this program ideally stabilizes.
We provide an informal argument for the validity of this proposition. Consider any program and all the computations produced by this program when it starts from an arbitrary state of its universe. Now define the specification that contains exactly these computations. This specification is ideal as all the states of its state universe are allowed while the program ideally stabilizes to this specification.
Naturally, this kind of specification may not be very useful as it, in essence, defines specification to be whatever the program computes. However, below we describe how a number of stabilizing program published in the literature can be defined as ideally stabilizing to ideal specifications.
Examples
--------
**Propagation of Information with Feedback.** As the first example we describe a program that ideally stabilizes to the propagation of information with feedback [@C82; @S83] specification. The presentation of this program as snap-stabilizing is well-known [@BDPV99c]. The program is presented on rooted trees. However, to simplify the presentation, we describe the operation of this program on a chain.
For ease of exposition we preserve the rooted tree terminology. Similarly to the alternator, we assume that the processes are numbered from $1$ to $N$ from left to right in the chain. We refer to the leftmost processor, with identifier $1$, as *root*; the rightmost, with identifier $N$, as *leaf*; and the processes in between as *intermediate*. For each process $p$, the processes to the left of it are *ancestors* and to the right — *descendants*.
Each process has a state variable $st$. In the intermediate processes, the variable may hold one of the three values: **i**, **rq**, **rp** which stand for *idle*, *requesting* and *replying* respectively. The root can only be idle or requesting while the leaf can be either idle or replying.
The objective of the program is to send a signal from the root to the leaf and in return receive an acknowledgment that matches this signal. Operationally, the program should ensure that after the root makes a request then every process propagates this request along the chain from left to right in causally ordered steps transitioning from idle to requesting. Afterwards, the processes propagate the reply from right to left in causally ordered steps transitioning from requesting to replying.
We define the following state predicates. $RP(k)$ are the specification states where all $k$ processes on the left are requesting $(k=1,N-1)$ and the rest of the processes are replying. $RQ(l,m)$ are the specification states where all $l$ processes on the left are requesting $(l=0,N-1)$ and all $m-l$ processes following them are idle $(m = l+1, N)$ while the remaining processes are replying. Specification includes the sequences where the system satisfies one of the predicates and transitions from one to the other infinitely.
Let us define another pair of predicates. Predicate $RP'(k)$ defines the states where $k$ processes on the left are requesting $(k=1,N-1)$, the process $k+1$ is replying and none of the rest are idle. Notice that $RP(k) \subset RP'(k)$. Predicate $RQ'(l,m)$ are the states where all $l$ processes on the left are requesting $(l=0,N-1)$, all $m-1$ following them are idle $(m = l+1, N)$ and the state of the other processes is arbitrary. Similarly, $RQ(l,m) \subset
RQ'(l,m)$. Specification includes the sequences where the system always satisfies either $RP'(k)$ or $RQ'(l,m)$, each sequence has a sequence in as a suffix and the transition from $RQ'(l,m)$ is only to $RP(k)$.
We now describe program . It has only external variables. The mapping between the program and specification variables is identical. The actions of are shown in Figure \[figPIF\].
$$\begin{array}{llllllll}
\emph{request}:& & &st.p_1 = \textbf{i} &\wedge& st.p_2=\textbf{i}
& \longrightarrow& st.p_1 := \textbf{rq} \\
\emph{clear}: & & &st.p_1 = \textbf{rq} &\wedge& st.p_2 = \textbf{rp}
& \longrightarrow& st.p_1 := \textbf{i} \\
\emph{forward}:& st.p_{j-1} = \textbf{rq} &\wedge& st.p_{j}=\textbf{i} &\wedge& st.p_{j+1} = \textbf{i}
&\longrightarrow& st.p_j :=\textbf{rq}\\
\emph{back}: & st.p_{j-1} = \textbf{rq} &\wedge& st.p_{j}=\textbf{rq} &\wedge& st.p_{j+1}=\textbf{rp}
&\longrightarrow& st.p_{j} :=\textbf{rp}\\
\emph{stop}:& st.p_{j-1} = \textbf{i} &\wedge& st.p_{j} \neq \textbf{i} & &
&\longrightarrow & st.p_j :=\textbf{i} \\
\emph{reflect}:& st.p_{N-1}=\textbf{rq} &\wedge& st.p_N = \textbf{i} &&
&\longrightarrow & st.p_{N} := \textbf{rp} \\
\emph{reset}:& st.p_{N-1} = \textbf{i} &\wedge& st.p_N = \textbf{rp} &&
&\longrightarrow & st.p_{N} := \textbf{i}
\end{array}$$
classically stabilizes to and ideally stabilizes to .
First, we prove that $RQ(l,m) \vee RP(k)$ is an invariant of with respect to . Notice that this disjunction is closed in . That is, none of the actions of violate the predicate. Let us show that a program transitions from $RQ(l,m)$ to $RP(k)$ and back. If the program satisfies $RQ(l,m)$, at least one action is enabled. Indeed, if $l=0, m<N$, *request* is enabled in *stop* is enabled in process $p_{m+1}$. If, $l<m-1,m\leq N$, *request* is enabled in $p_{l+1}$. With the execution of an action either $l$ or $m$ are incremented. If $l=N-1,
m=N$, *reflect* is enabled that transitions the program into $RP(N-1)$. Similarly, if the system satisfies $RP(k)$ for $k=2,N-1$, action *back* is enabled. The execution of this action keeps the system in $RP(k)$ but decrements $k$. If $k=1$, *clear* is enabled. Its execution moves the program back in $RQ(l,m)$. That is, the disjunction of the two predicates is closed and a computation that starts in a state conforming to one of them, satisfies . Hence, $RQ(l,m) \vee RP(k)$ is an invariant of .
Observe that the disjunction of $RQ'(l,m) \vee RP'(k)$ contains the program, as well as specification, state universe. We show that stabilizes to from this predicate as required by . An argument similar to the above demonstrates that an action is enabled if satisfies $RP'(k)$. This action either decrements $k$ or moves the system to $RQ'(l,m)$. Also, similarly, if satisfies $RQ'(l,m)$, then an action is enabled that increments either $l$ or $m$. Moreover, if $l=N-1,m-N$, the execution of *reflect* transitions the system to $RP(k)$. That is, moves from $RP'(k)$ to $RQ'(l,m)$ to $RP(k)$. This means that the program stabilizes to and ideally stabilizes to .
\
**Alternating bit protocol.** Alternating bit protocol is an elementary data-link network protocol. There is a number of classic stabilizing implementations of the protocol. Refer to Howell et al [@HNM99c] for an extensive list of citations. There is also a snap-stabilizing version [@DDNT08c].
The problem is stated as follows. There are two processes: *sender* — $p$, and *receiver* — $q$. The processes maintain boolean sequence numbers $ns.p$ and $nr.q$. The processes exchange messages over communication channels. The channels are reliable and their capacity is one. That is, if the channel is empty, the message is reliably sent. If the channel already contains a message, an attempt to send another message leads to the loss of the new message. The processes exchange two types of messages: $data$ and $ack$. Both carry the sequence numbers.
Specification prescribes infinite sequences of states where there is exactly one message in the channels. The message carries the sequence number of the sender. The state transitions are such that $p$ changes the value of $ns$. This change is followed by the change of the value in $nr$ that matches the value of $ns$.
Specification of is such that for every state in the universe there is a sequence that starts in it and every sequence contains a sequence of as a suffix. Specification of is thus ideal.
The program uses only external variables as described by and . actions are shown in Figure \[figABPcode\]. The sender has two actions: *next* and *timeout*. Action *next* is enabled if there is a message from $q$ in the channel. The timeout action is enabled if there are no messages in either channel. Upon receiving a message from $q$ with matching sequence number, $p$ increments the sequence number and sends the next message. If $p$ times out, it resubmits the same message. The receiver has a single action. When $q$, receives a message, it sends an acknowledgment back to $p$. If the message bears a sequence number different from $rn$, $q$ increments $rn$ signifying the successful receipt of the message.
1234678912341234=*next*: $\textbf{receive}\ ack(nm) \longrightarrow$\
$\textbf{if}\ nm = ns\ \textbf{then} $\
$ns:=\neg ns$\
$\textbf{send}\ data(ns)$\
*timeout*: $\textbf{timeout}() \longrightarrow \textbf{send}\ data(ns)$\
*reply*: $\textbf{receive}\ data(nm) \longrightarrow$\
$\textbf{if}\ nm \neq nr\ \textbf{then}$\
$nr:= nm$\
$ \textbf{send}\ ack(nm) $
classically stabilizes to and ideally stabilizes to .
We prove the correctness of the theorem by enumerating the state transitions of . We classify the state universe of into two groups: (i) $ns$ is equal to $nr$ and (ii) $ns$ is not equal to $nr$. The states are further classified according to the type of messages in the channels. The states and state transitions are shown in Figure \[figABPstates\]. Note that to simplify the diagram we do not show the states that contain more than a single message. However, after a single transition, the program moves from one of those states to a state shown in the figure. The correctness of the theorem claims can be ascertained by examining the states and transitions shown in the figure.
The Impact of Ideal Stabilization Approach {#SecEnd}
==========================================
In this paper we proposed a new way of approaching stabilization. Our approach eliminates two of the most problematic features of classic stabilization: unpredictable behavior during stabilization and poor composability. We hope that this work adds more credence to stabilization as a viable fault-tolerance technique and generates more interest in the subject among both theoretical researches and reliability engineers.
[^1]: This research is supported in part by NSF Career award CNS-0347485.
[^2]: This author is supported in part by the ANR SHAMAN project.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Using a sample of 903 APM clusters we investigate whether their dynamical status, as evidenced by the presence of significant substructures, is related to the large-scale structure of the Universe. We find that the cluster dynamical activity is strongly correlated with the tendency of clusters to be aligned with their nearest neighbour and in general with the nearby clusters that belong to the same supercluster. Furthermore, dynamically active clusters are more clustered than the overall cluster population. These are strong indications that clusters develop in a hierarchical fashion by anisotropic merging along the large-scale filaments within which they are embedded.
[**Keywords:**]{} galaxies: clusters: general - large-scale structure of universe
author:
- |
Manolis Plionis$^1$ & Spyros Basilakos$^2$\
$^1$Institute of Astronomy & Astrophysics, National Observatory of Athens, I.Metaxa & B.Pavlou, Palaia Penteli, Athens 152 36, Greece\
$^2$Astrophysics Group, Imperial College London, Blackett Laboratory, Prince Consort Road, London SW7 2BW, UK
title: 'The Cluster Substructure - Alignment Connection'
---
Introduction
============
An interesting observable, that was thought initially to provide strong constraints on theories of galaxy formation, is the tendency of clusters to be aligned with their nearest neighbour as well as with other clusters that reside in the same supercluster (cf. Binggeli 1981; West 1989; Plionis 1994; Chambers, Melott & Miller 2001). Analytical and numerical work have shown that such alignments, expected naturally to occur in “top-down” scenarios (cf. Zeldovich 1970), are also found in hierarchical clustering models of structure formation like the CDM (Bond 1986; West et al. 1991; Splinter et al. 1997; Onuora & Thomas 2000). This fact could be explained as the result of an interesting property of Gaussian random fields that occurs for a wide range of initial conditions and which is the “cross-talk” between density fluctuations on different scales. Furthermore, there is strong evidence that the brightest galaxy (BCGs) in clusters is aligned with the orientation of its parent cluster and even with the orientation of the large-scale filamentary structure within which they are embedded (cf. Struble 1990; West 1994, Fuller, West & Bridges 1999).
Within the framework of hierarchical clustering, the anisotropic merger scenario of West (1994), in which clusters form by accreting material along the filamentary structure within which they are embedded, provides an interesting explanation of such alignments as well as of the observed strong alignment of BCGs with their parent cluster orientation. In this framework, one should expect that dynamical young clusters, at there early stages of formation in which they are not smooth and spherically symmetric, should show substructures that are aligned with the local large-scale structures, an effect observed also in numerical simulations for a variety of power-spectra (van Haarlem & van de Weygaert 1993; Tormen 1997). Indeed, supporting this view West, Jones & Forman (1995) found, using 43 Einstein clusters that have a neighbour within $\sim 10$ $h^{-1}$ Mpc, that cluster substructures do show a tendency to be aligned with the orientation of the major axis of their parent cluster and with the nearest-neighbouring cluster (see also Novikov et al 1999). In this work we investigate the relation between the strength of cluster-cluster alignments and the large-scale environment in which the clusters are embedded using the APM cluster catalogue (Dalton et al 1997), which is the largest one available.
Methodology
===========
The APM cluster catalogue is based on the APM galaxy survey which covers an area of 4300 square degrees in the southern sky containing about 2.5 million galaxies brighter than a magnitude limit of $b_{J}=20.5$ (for details see Maddox et al. 1990). Dalton et al (1997) applied an object cluster finding algorithm to the APM galaxy data using a search radius of $0.75 \; h^{-1}$ Mpc in order to minimize projection effects, and so produced a list of 957 clusters with $z_{est} {\raise
-2.truept\hbox{\rlap{\hbox{$\sim$}}\raise5.truept\hbox{$<$}\ }}0.13$. Out of these 309 ($\sim 32\%$) are ACO clusters, while 374 ($\sim 39\%$) have measured redshifts (179 of these are ACO clusters). The APM clusters that are not in the ACO list are relatively poorer systems than the Abell clusters, as we have verified comparing their APM richness’s (see Dalton et al 1997 for definition of richness).
For the present analysis we use 903 of the above APM clusters, since 54 clusters are found in the vicinity of plate-holes or crowded regions, a fact which affects severely their shape parameters. The cluster distance is estimated from their redshift using $H_{\circ}=100 h$ km s$^{-1}$ Mpc$^{-1}$ and $q_{\circ}=0.5$.
Cluster Shape Parameters and Alignment Measure
----------------------------------------------
A detailed analysis of the cluster shape determination procedure and of the intrinsic APM cluster shapes can be found in Basilakos, Plionis & Maddox (2000). Here we only sketch the basic procedure which is based on the familiar moments of inertia method with $I_{11}=\sum\ w_{i}(r_{i}^{2}-x_{i}^{2})$, $I_{22}=\sum\ w_{i}(r_{i}^{2}-y_{i}^{2})$, $I_{12}=I_{21}=-\sum\ w_{i}x_{i}y_{i}$, where $x_i$ and $y_i$ are the Cartesian coordinates of the galaxies that their projected separation is such that they are judged as belonging to the cluster (details in Basilakos et al 2000) and $w_i$ is their weight. We, then diagonalize the inertia tensor solving the basic equation: $$\label{eq:diag}
det(I_{ij}-\lambda^{2} \; M_{2})=0 ,$$ where $M_{2}$ is the $2 \times 2$ unit matrix. The cluster ellipticity is given by $\epsilon=1-\frac{\lambda_2}{\lambda_1}$, where $\lambda_i$ are the positive eigenvalues with $(\lambda_1>\lambda_2)$. This method can be applied to the data using either the discrete or smoothed distribution of galaxies. The determination of the cluster orientation is consistent among the two methods but this is not true also for the cluster ellipticity (for details see Basilakos et al. 2000)
\[fig:alin\] =8.5cm
In Figure 1 we present the derived cluster position angle, $\theta_i$, distribution. We see no significant systematic orientation effects, a fact that we quantify estimating the Fourier transform of the galaxy position angles; $C_{n} = \sqrt{\frac{2}{N}} \sum \cos 2n\theta_i$ and $S_{n} = \sqrt{\frac{2}{N}} \sum \sin 2n\theta_i$. If the galaxy position angles are uniformly distributed between $0^{\circ}$ and $180^{\circ}$, then both $C_{n}$ and $S_{n}$ have zero mean and unit standard deviation. Therefore large values ($>2.5-3$) indicate significant deviation from isotropy. We find that $C_{1,2}$ and $S_{1,2}$ have values ${\raise
-2.truept\hbox{\rlap{\hbox{$\sim$}}\raise5.truept\hbox{$<$}\ }}1.2$, which indicates that there is no systematic orientation bias.
In order to investigate the alignment between cluster orientations, we define the relative position angle between the major axis orientation of a cluster and the direction to a neighbouring one by $\phi_{i,j}\equiv |\theta_i - \vartheta_j|$ (where $\vartheta$ is the position angle of the cluster pair separation vector). In an isotropic distribution we will have $\langle \phi_{i,j} \rangle \simeq 45^{\circ}$. A significant deviation from this would be an indication of an anisotropic distribution which can be quantified by (Struble & Peebles 1995): $$\label{eq:alin}
\delta=\sum_{i=1}^{N}\frac{\phi_{i,j}}{N}-45$$ In an isotropic distribution we have $\langle \delta \rangle \simeq 0$, while the standard deviation is given by $\sigma=90/\sqrt{12 N}$. A significantly negative value of $\delta$ would indicate alignment and a positive misalignment. To avoid problems related to ill defined position angles we will use in our analysis only clusters that have ellipticities $>0.05$ (ie., $N=888$)[^1]. It should be noted that systematic biases and projection effects will tend to mask any true alignment signal. For example, the projection of foreground galaxies along the line of sight of a cluster as well as the projection on the plane of the sky of member galaxies, always work in the direction of smearing alignments.
Substructure Measure & Significance
-----------------------------------
Major obstacles in attempting to determine the dynamical state of a cluster is (1) the ambiguity in identifying cluster substructure in 2D or even 3D cluster data and (2) the uncertainty of post-merging relaxation timescales. Evrard et al. (1993) and Mohr et al. (1995) have suggested as an efficient indicator of cluster substructure the shift of the center-of-mass position as a function of density threshold above which it is estimated. The [*centroid-shift*]{} ($sc$) is defined as the distance between the cluster center-of-mass, $(x_{\rm o}, y_{\rm o})$, which may change at different density thresholds and the highest cluster density-peak, $(x_{\rm p}, y_{\rm p})$, ie., $sc = \sqrt{(x_{\rm o}-x_{\rm p})^{2}\,+\,(x_{\rm o}-x_{\rm p})^{2}}$.
Kolokotronis et al. (2001), using in a complementary fashion optical and X–ray data (see also Rizza et al. 1998), since in the X-ray band projection effects are minimal, calibrated various substructure measures using APM data and pointed ROSAT observations of 22 Abell clusters and found that in most cases using X–ray or optical data one can identify substructure unambiguously. Only in $\sim 20\%$ of the clusters that they studied did they find projection effects in the optical that altered the X-ray definition of substructure. An important conclusion of Kolokotronis et al. (2001) was that a large and significant value of $sc$ is a clear indication of substructure in APM optical cluster data.
The significance of such centroid variations to the presence of background contamination and random density fluctuations are quantified using Monte Carlo cluster simulations in which, by construction, there is no substructure. For each APM cluster, 1000 simulated clusters are produced having the observed ellipticity, the observed number of galaxies, following a King’s profile, as well as a random distribution of expected background galaxies, determined by the distance of the cluster and the APM selection function (note that the number of “galaxies” that we allow to follow the King’s profile is the observed number minus the expected random background). The King-like profile is: $$\label{eq:sb}
\Sigma(r) \propto \left[1\,+\,\left(\frac{r}{r_{\rm c}} \right)^{2}
\right]^{-\alpha} \;,$$
where $r_{\rm c}$ is the core radius. We use the weighted, by the sample size, mean of most recent $r_c$ and $\alpha$ determinations (cf. Girardi et al. 1998), i.e., $r_{\rm c} \simeq 0.085 \;h^{-1}$ Mpc and $\alpha \simeq 0.7$. We do test the robustness of our results for a plausible range of these parameters (details can be found in Kolokotronis et al. 2001).
Naturally, we expect the simulated clusters to generate small $sc$’s and in any case insignificant shifts. Therefore, from each set of Monte-Carlo cluster simulations we derive $\langle sc \rangle_{\rm sim}$ as a function of the same density thresholds as in the real cluster case. Then, within a search radius of $0.75 \;h^{-1}$ Mpc from the simulated highest cluster peak, we calculate the quantity: $$\label{eq:sig}
\sigma =\frac{\langle sc \rangle_{\rm o} - \langle sc
\rangle_{\rm sim}}{\sigma_{\rm sim}}\;,$$ which is a measure of the significance of real centroid shifts as compared to the simulated, substructure-free clusters. Note that $\langle sc \rangle_{\rm o}$ is the average, over three density thresholds, centroid shift for the real APM cluster.
A further possible substructure identification procedure is based on a friend-of-friends algorithm, applied on 3 overdensity thresholds of each cluster (for details see Kolokotronis et al. 2001). Three categories are identified, based on the subgroup multiplicity and size: [*(a)*]{} No substructure (unimodal), [*(b)*]{} Weak substructure (multipole groups but with total group mass $\le 25\%$ of main), [*(c)*]{} Strong substructure (multipole groups but with mass $>
25\%$ of main).
Results & Discussion
====================
Cluster Substructure and Alignments
-----------------------------------
Applying the [*centroid shift*]{} substructure identification procedure to the 903 APM clusters we find that about 30% of clusters have significant ($> 3 \sigma$) substructure. Note that defining as having significant substructure those clusters with $\sigma
>2.5$ or 2 increases the fraction to $\sim$ 40% and 50% respectively. Furthermore, changing the structural parameters of the Monte-Carlo clusters changes the actual $\sigma$-values, although their relative significance rank-order remains unaltered. Alternatively if we apply the [*subgroup*]{} categorization procedure we find that $\sim$53% of the APM clusters show strong indications of substructure.
We have tested whether there is any systematic redshift dependent effect of the cluster substructure categorization and found none (using either measured or estimated redshifts). We conclude that irrespectively of the method $\sim 30\%
- 50\%$ of the APM clusters show indications of significant substructure (in accordance with Kolokotronis et al 2001).
We now test whether the well known nearest-neighbour alignment effect, present in the Abell clusters (cf. Bingelli 1982; Plionis 1994), is evident also in the poorer APM clusters. In table 1 we present, for the whole APM cluster sample, our alignment results as a function of maximum intercluster separation, $D_{max}$, for both the nearest-neighbours ($\delta_{nn}$) and for all neighbouring pairs ($\delta_{an}$), within $D_{max}$.
$D_{max}$ $\langle \delta_{nn} \rangle$ $P(>\chi^2)$ $N_{pairs}$ $\langle \delta_{an} \rangle$ $P(>\chi^2)$ $N_{pairs}$
----------- ------------------------------- -------------- ------------- ------------------------------- -------------- -------------
5 -7.7$\pm 2.7$ 0.005 90 -7.5 $\pm 2.5$ 0.002 110
10 -4.6$\pm 1.6$ 0.005 270 -4.1$\pm 1.3$ 0.000 374
15 -3.8$\pm 1.2$ 0.001 444 -2.6$\pm 0.9$ 0.002 848
20 -2.9$\pm 1.0$ 0.004 616 -1.1$\pm 0.7$ 0.023 1602
30 -1.7$\pm 0.9$ 0.069 805 -0.9$\pm 0.4$ 0.112 4012
: Nearest neighbour ([*nn*]{}) and all-neighbour ([*an*]{}) alignment signal as a function of maximum pair separation, $D_{max}$. The $\chi^2$ probabilities are derived by comparing the binned $\phi$-distribution (using 3 bins of 30$^{\circ}$ width) with the Poisson expectation values.
It is evident that there is significant indication of cluster alignments, with the alignment signal dropping in amplitude and significance, as a function of increasing $D_{max}$. In order to test whether this result is dominated by the ACO cluster pairs, and thus whether it is a manifestation of the already known Abell cluster alignment effect, we have excluded such pairs (113/888 for the $nn$-case) to find consistent but more significant alignment results. For example for the $D_{max}=15$ $h^{-1}$ Mpc case, we obtain $\langle \delta_{nn} \rangle \simeq -4.3 \pm 1.3$ with $P(>\chi^2)=0.0004$ and $\langle \delta_{an} \rangle \simeq -2.7 \pm 0.9$ with $P(>\chi^2)=0.001$.
Note that a large number of APM clusters have estimated redshifts and thus the cluster distance uncertainties will tend to hide true alignments. Therefore, the measured alignment signal should be considered rather as a lower limit to the true one.
The Substructure-Alignment Connection
-------------------------------------
We have correlated the alignment signal with the substructure significance indication in order to see whether there is any relation between the large-scale environment, in which the cluster distribution is embedded, and the internal cluster dynamics.
\[fig:alinsig\] =8.4cm
In figure 2 we present the alignment signal, $\langle \delta \rangle$, between all cluster pairs with separations $< 20$ $h^{-1}$ Mpc that have substructure significance above the indicated $\sigma$ value. Evidently, there is a strong correlation between the strength of the alignment signal and the substructure significance level (the two panels are based on slightly different definition of the alignment signal - see caption for details).
In order to assess the statistical significance of obtaining such a $\delta - \sigma$ relation, we perform a large number (10000) of Monte-Carlo simulations in which we reshuffle the measured position angles, assigning them at random to the clusters. Then we derive the corresponding $\delta - \sigma$ relation and count in how many such simulations do we get values of $\delta$ which are as negative, or more, than the corresponding ones of figure 2. We find that the corresponding probability is $<10^{-5}$, while for the less restrictive case, of having negative $\delta$-values (of any amplitude) for all $\sigma$, the probability is again quite small ($\sim 7.5 \times
10^{-3}$).
Note that from the analysis of Kolokotronis et al (2001) it is expected that our procedure will misidentify the dynamical state of $\sim 20\%$ of the APM clusters. However, such misidentification will act as a noise factor and will tend to smear any true alignment-substructure correlation, since there is no physical reason why random projection effects, within 0.75 $h^{-1}$ Mpc of the cluster core, should be correlated with the direction of neighbours within distances up to a few tens of Mpc’s (such a correlation could be expected at some level only for nearest-neighbours in angular space but we have verified that choosing such pairs we obtain an insignificant alignment signal). Therefore our result, based on the largest cluster sample available, supports the hierarchical clustering scenario and in particular the formation of cluster by anisotropic merging along the filamentary structure within which they are embedded (cf. West 1994; West, Jones & Forman 1995).
\[fig:xi\] =9.cm
Local density - Substructure Correlation
----------------------------------------
If the above view is correct then one would expect that clusters with significant substructure should be residing preferentially in high-density environments (superclusters), and this would then have an imprint in their spatial two-point correlation function.
In figure 3 we present the spatial 2-point correlation function of all APM clusters (open symbols) and of those with substructure significance $\sigma \ge 4$ (filled symbols). It is clear that the latter are significantly more clustered. This can be seen also in the insert of figure 3 were we plot the correlation length, $r_{0}$, as a function of $\sigma$, which is clearly an increasing function of cluster substructure significance level. Note that in order to take into account the possible systematic distance dependent effects in the different cluster subsamples we generate random catalogues, used to normalize the number of cluster-cluster pairs in the $\xi(r)$ estimate, using the individual distance distribution of each subsample and not the overall APM cluster selection function. Had we used the latter, the increase of $r_{\circ}$ with $\sigma$ would have been more severe.
Furthermore, we have tested whether this effect could be due to the well-known richness dependence of the correlation strength, and found a weak, if any, such richness trend. In order to further investigate this issue we have correlated the APM richness (see Dalton et al 1997) with $\sigma$ and found a very weak but significant correlation (the Pearson’s coefficient is only 0.26 but with a significance $>$99.9%). Therefore, a component contributing to the increase of the $r_{\circ}(\sigma)$ function could possibly be the cluster richness effect (through the above richness-$\sigma$ weak correlation). However, the increase of $r_{\circ}(\sigma)$ is quite dramatic and cannot be attributed only to this weak richness dependence. This is corroborated also from the fact that there is a larger fraction of clusters with significant substructure (high $\sigma$ values) residing in high-density regions (superclusters), as shown in figure 4 where we plot the fraction of clusters with $\sigma>3$ that belong to superclusters identified by the indicated percolation radius. Evidently the fraction increases inversely with percolation radius. These results are similar to those of Schombert & West (1990) based on Abell clusters, in which they found that flattened clusters, flatness being an indication of dynamical youth, are more frequent in high density environments (see however Herrera & Sanroma 1997 for a different view).
=10.5cm
The conclusion of this correlation function analysis is that indeed the clusters showing evidence of dynamical activity reside in high-density environments, as anticipated from the alignment analysis. It is interesting that such environmental dependence has also been found in a similar study of the BCS and REFLEX clusters (Schüecker et al. 2001) and for the cooling flow clusters with high mass accretion rates (Loken, Melott & Miller 1999).
Conclusions
===========
We have presented evidence, based on the largest available cluster sample, the APM, that there is a strong link between the dynamical state of clusters and their large-scale environment. Clusters showing evidence of dynamical activity are significantly more aligned with their nearest neighbours and they are also much more spatially clustered. This supports the hierarchical clustering models in which clusters develop by accreting matter along the large-scale filamentary structures within which they are embedded.
Acknowledgements {#acknowledgements .unnumbered}
================
M.Plionis acknowledges the hospitality of the Astrophysics Group of Imperial College. This work was partially supported by the EC Network programme ‘POE’ (grant number HPRN-CT-2000-00138). We thank the referee, M.West, for very helpful suggestions that improved our work.
Basilakos S., Plionis M., Maddox S. J., 2000, MNRAS, 315, 779 Bingelli B., 1982, AA, 250, 432 Bond, J.R., 1986, in [*Galaxy Distances and Deviations from the Hubble Flow*]{}, eds. Madore, B.F., Tully, R.B., (Dordrecht: Reidel), p.255 Chambers, S.W., Melott, A.L., Miller, C.J., 2001, ApJ, in press, [*(astro-ph/0102218)*]{} Dalton G. B., Maddox S. J., Sutherland W. J., Efstathiou G., 1997, MNRAS, 289, 263 Evrard A.E., Mohr J.J., Fabricant D.G., Geller M.J.,1993, ApJ, 419, L9 Fuller, T.M., West, M.J. & Bridges, T.J., 1999, ApJ, 519, 22 Girardi M., Giuricin G., Mardirossian F., Mezzetti M., Boschin W., 1998, ApJ, 505, 74 van Haarlem, M., van de Weygaert, R., 1993, ApJ, 418, 544 Herrera, B., Sanroma, M., 1997, AA, 320, 13 Kolokotronis, V., Basilakos, S., Plionis, M., Georgantopoulos, I., 2001, MNRAS, 320, 49 Loken, C., Melott, A.L., Miller, C.J., 1999, ApJ, 520, L5 Maddox S.J. , Sutherland W.J., Efstathiou G., Loveday, J. 1990, MNRAS, 243, 692 Mohr, J.J., Evrard, A.E., Fabricant, D.G., Geller, M.J., 1995, ApJ, 447, 8 Novikov, D. et al., 1999, MNRAS, 304, L5 Onuora, L.I., Thomas, P.A, 2000, MNRAS, 319, 614 Plionis M., 1994, ApJS., 95, 401 Plionis M., 2001, in the proceedings of the [*Clusters and the High-Redshift Universe observed in X-rays*]{}, XXI$^{\rm th}$ Moriond Astrophysics Meeting, eds. Neumann et al., [*in press*]{} Richstone, D., Loeb, A., Turner, E.L., 1992, ApJ, 393, 477 Rizza E., Burns J. O., Ledlow M. J., Owen F. N., Voges, W., Blito M., 1998, MNRAS, 301, 328 Schombert, J. M., West, M. J., 1990, ApJ, 363, 331 Schüecker, P., Böhringer, H., Reiprich, T. H., Feretti, L., AA, 2001, 378, 408 Splinter, R.J., Melott, A.L., Linn, A.M., Buck, C., Tinker, J., 1997, ApJ, 479, 632 Struble, M.F., 1990, AJ, 99, 743 Struble, M.F., Peebles, P.J.E., 1985, AJ, 90, 582 Tormen, G., 1997, MNRAS, 290, 411 West, M. J., 1989, ApJ, 347, 610 West, M. J., Villumsen, J.V., Dekel, A., 1991, ApJ, 369, 287 West, M. J., 1994, MNRAS, 268, 79 West, M. J., Jones C., Forman W., 1995, ApJ, 451, L5 Zeldovich, Y.B., 1970, AA, 5, 84
[^1]: Our results remain unaltered even for higher values of the ellipticity cutoff.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Recent advancements on the fabrication of organic micro- and nanostructures have permitted the strong collective light-matter coupling regime to be reached with molecular materials. Pioneering works in this direction have shown the effects of this regime in the excited state reactivity of molecular systems and at the same time have opened up the question of whether it is possible to introduce any modifications in the electronic ground energy landscape which could affect chemical thermodynamics and/or kinetics. In this work, we use a model system of many molecules coupled to a surface-plasmon field to gain insight on the key parameters which govern the modifications of the ground-state Potential Energy Surface (PES). Our findings confirm that the energetic changes *per molecule* are determined by effects which are essentially on the order of single-molecule light-matter couplings, in contrast with those of the electronically excited states, for which energetic corrections are of a collective nature. Still, we reveal some intriguing quantum-coherent effects associated with pathways of concerted reactions, where two or more molecules undergo reactions simultaneously, and which can be of relevance in low-barrier reactions. Finally, we also explore modifications to nonadiabatic dynamics and conclude that, for our particular model, the presence of a large number of dark states yields negligible effects. Our study reveals new possibilities as well as limitations for the emerging field of polariton chemistry.'
author:
- 'Luis A. Martínez-Martínez'
- 'Raphael F. Ribeiro'
- 'Jorge Campos-González-Angulo'
- 'Joel Yuen-Zhou'
bibliography:
- 'usc.bib'
title: 'Can ultrastrong coupling change ground-state chemical reactions?'
---
Introduction
============
The advent of nano- and microstructures which enable strong confinement of electromagnetic fields in volumes as small as $1\times10^{-7}\lambda^{3}$[@Kim2015], $\lambda$ being a characteristic optical wavelength, allows for the possibility of tuning light-matter interactions that can “dress" molecular degrees of freedom and give rise to novel molecular functionalities. Several recent studies have considered the effects of strong coupling (SC) between confined light and molecular states, and its applications in exciton harvesting and transport[@Gonzalez-Ballestero2015; @Feist2015], charge transfer[@Herrera2016], Bose-Einstein condensation [@Andre2006; @Gerace2012; @Nguyen2015], Raman [@Strashko2016; @delPino2015] and photoluminiscence [@Herrera2016a; @Melnikau2016] spectroscopy, and quantum computing [@DelPino2014; @Hartmann2006; @Raimond2001], among many others [@Bellesa2004; @Laussy2008; @Lidzey1999]. Organic dye molecules are good candidates to explore SC effects due to their unusually large transition dipole moment [@Tischler2005; @Hobson2002; @Bellessa2004; @Salomon2009]. More recently, it has been experimentally and theoretically shown that the rates of photochemical processes for molecules placed inside nanostructures can be substantially modified [@Hutchison2012; @Herrera2016; @Thomas2016; @Galego2016; @Flick2016]. The underlying reason for these effects is that the SC energy scale is comparable to that of vibrational and electronic degrees of freedom, as well as the coupling between them [@Galego2015a]; this energetic interplay nontrivially alters the resulting energetic spectrum and dynamics of the molecule-cavity system. It is important to emphasize that in these examples, SC is the result of a collective coupling between a single photonic mode and $N \gg 1$ molecules; single-molecule SC coupling is an important frontier of current research [@Chikkaraddy2016], but our emphasis in this work will be on the $N$ molecule case. Since the energy scale of this collective coupling is larger than the molecular and photonic linewidths, the resulting eigenstates of the system have a mixed photon-matter character. Understanding these so-called *polariton* states is relevant to develop a physical picture for the emerging energy landscapes which govern the aforementioned chemical reactivities. More specifically, Galego and coworkers [@Galego2015a] have recently provided a comprehensive theoretical framework to explain the role of vibronic coupling and the validity of the Born-Oppenheimer (BO) approximation in the SC regime, as well as a possible mechanism for changes in photochemical kinetics afforded by polaritonic systems [@Galego2016]; another theoretical study that focused on control of electron transfer kinetics was given by Herrera and Spano [@Herrera2016]. Using a model of one or two molecules coupled to a single mode in a cavity, Galego and coworkers noticed that some effects on molecular systems are collective while others are not; similar findings were reported by Cwik and coworkers using a multimode model and $N$ molecules [@Cwik2016]. While prospects of photochemical control seem promising, it is still a relatively unexplored question whether ground-state chemical reactivity can be altered via polaritonic methods, although recently, George and coworkers have shown a proof of concept of such feasibility using vibrational SC [@Thomas2016]. Along this line, ultrastrong coupling regime (USC) seems to also provide the conditions to tune the electronic ground-state energy landscape of molecules and in turn, modify not only photochemistry, but ground-state chemical reactivity. Roughly speaking, this regime is reached when $\Omega/\hbar\omega_{0}\geq0.1$, $\Omega$ being the (collective) SC of the emitter ensemble to the electromagnetic field and $\hbar\omega_{0}$ the energy gap of the molecular transition[@Moroz]. Under USC, the “nonrotating" terms of the light-matter Hamiltonian acquire relevance and give rise to striking phenomena such as the dynamical Casimir effect [@Wilson; @Stassi2013] and Hawking radiation in condensed matter systems [@Stassi2013]. Furthermore, recent experimental advances have rendered the USC regime feasible in circuit QED [@Niemczyk2010], inorganic semiconductors [@Ciuti2005; @Todorov2010], and molecular systems [@Schwartz2011; @george2016multiple], thus prompting us to explore USC effects on ground-state chemical reactivity.\
In this article, we address how this reactivity can be influenced in the USC by studying a reactive model system consisting of an ensemble of thiacyanine molecules strongly coupled to the plasmonic field afforded by a metal, where each of the molecules can undergo cis-trans isomerization by torsional motion. The theoretical model for the photochemistry of the single thiacyanine molecule has been previously studied in the context of coherent control [@Hoki2009a]. As we will show, the prospects of controlling ground-state chemical reactivity or nonadiabatic dynamics involving the ground state are not promising for this particular model, given that the alterations of the corresponding PES are negligible on a per-molecule basis. However, we notice the existence of salient quantum-coherent features associated with concerted reactions that might be worth considering in models featuring lower kinetic barriers.\
This article is organized as follows: in the Theoretical Model section, we describe the polariton system and its quantum mechanical Hamiltonian. In Methods, we describe the methodology to perform the relevant calculations and understand the effects of polariton states on the ground-state PES of the molecular ensemble. In Results and Discussion we describe our main findings, and finally, in the Conclusions section, we provide a summary and an outlook of the problem.
Theoretical model
=================
To begin with, we consider a thiacyanine derivative molecule (Fig. \[Adia\_PESs\]c) and approximate its electronic degrees of freedom as a quantum mechanical two-level system. To keep the model tractable, this electronic system is coupled to only one vibrational degree of freedom $R$, namely, the torsion along the bridge of the molecule (Fig. \[Adia\_PESs\]c) along which cis-trans isomerization occurs. The mathematical description of the PES of the ground and excited states (Fig. \[Adia\_PESs\]a) as well as the transition dipole moment as a function of the reaction coordinate (Fig. \[Adia\_PESs\]b) have been obtained from Ref. [@Hoki2009a]. The adiabatic representation of the electronic states is given by, $$\begin{split}|g(R)\rangle= & \cos\left(\theta(R)/2\right)|\text{trans}\rangle+\sin\left(\theta(R)/2\right)|\text{cis}\rangle\\
|e(R)\rangle= & -\sin\left(\theta(R)/2\right)|\text{trans}\rangle+\cos\left(\theta(R)/2\right)|\text{cis}\rangle
\end{split}
\label{adiabatic_rep}$$ where $|e(R)\rangle$ and $|g(R)\rangle$ are the $R$-dependent adiabatic excited and ground state respectively. $|\text{trans}\rangle$ and $|\text{cis}\rangle$ are the ($R$-independent) crude diabatic electronic states that describe the localized chemical character of each of the isomers. The ground-state PES has a predominant trans (cis) character to the left (right) of the barrier ($\theta(0)=0$, $\theta(\pi)=\pi$) in Fig. \[Adia\_PESs\]a.\
![a) Adiabatic potential energy surfaces (PESs) of the ground and first excited electronic states of the thyacynine-like model molecule. b) Transition dipole moment ($\mathbf{\mu}(R)$) of the model molecule in the adiabatic basis. c) Thiacynine molecule. There exist two geometrical isomers of the molecule, a cis- and a trans-like configuration. The cis-trans isomerization of thiacynine-like molecules occurs via a photo-induced torsion along the bridge which connects the aromatic rings\[Adia\_PESs\].](PESs_merged_arrow.pdf)
Our USC model consists of a setup where an orthorhombic ensemble of thyacyanine molecules is placed on top of a thin spacer which, in turn, is on top of a metallic surface that hosts surface plasmons (SPs) [@Yuen-Zhou2015] (see Fig. \[fig:plexc\_struct\]). The coupling between molecular electronic transitions and plasmons in the metal give rise to polaritons that are often called plexcitons [@Gonzalez-Tudela2013; @Yuen-Zhou2015]. The ensemble is comprised of $N_{z}$ single-molecule layers. The location of each molecule can be defined by the Cartesian coordinates $\mathbf{n}+(0,0,z_{s})$ where $\mathbf{n}=(\Delta_{x}n_{x},\Delta_{y}n_{y},0)$ and $z_{s}=z_{0}+\Delta_{z}s$ for the *s*-th layer. Here, the spacing between molecules along the *i*-th direction is denoted by $\Delta_{i}$, and $z_{0}$ is the width of the spacer (see Fig. \[fig:plexc\_struct\]). We chose a SP electromagnetic environment because its evanescent intensity decreases fast enough with momentum $\mathbf{k}$ (giving rise to vanishing light-matter coupling for large $|\mathbf{k}|$), resulting on a convergent Lamb-shift of the molecular ground-state. As shall be explained below, this circumvents technical complications of introducing renormalization cutoffs, as would be needed for a dielectric microcavity [@Cwik2016]. The Hamiltonian of the plexciton setup is given by $H=H_{el}+T_{nuc}$, where $T_{nuc}=\sum_{i}\frac{\mathbf{P}_{i}^{2}}{2M_{i}}$ is the nuclear kinetic energy operator and
$$\begin{split}H_{el}(\mathbf{R})= & \sum_{\mathbf{k}}\hbar\omega_{\mathbf{k}}a_{\mathbf{k}}^{\dagger}a_{\mathbf{k}}+\sum_{\mathbf{n},s}\left(\hbar\omega_{e}(R_{\mathbf{n},s})-\hbar\omega_{g}(R_{\mathbf{n},s})\right)b_{\mathbf{n},s}^{\dagger}(R_{\mathbf{n},s})b_{\mathbf{n},s}(R_{\mathbf{n},s})\\
& +\sum_{\mathbf{k}}\sum_{\mathbf{n},s}g_{\mathbf{k}}^{\mathbf{n},s}(R_{\mathbf{n},s})\left(a_{\mathbf{k}}^{\dagger}b_{\mathbf{n},s}(R_{\mathbf{n},s})+a_{\mathbf{k}}b_{\mathbf{n},s}^{\dagger}(R_{\mathbf{n},s})+a_{\mathbf{k}}b_{\mathbf{n},s}(R_{\mathbf{n},s})+a_{\mathbf{k}}^{\dagger}b_{\mathbf{n},s}^{\dagger}(R_{\mathbf{n},s})\right)\\
& +\sum_{\mathbf{n},s}\hbar\omega_{g}(R_{\mathbf{n},s}),
\end{split}
\label{Dicke}$$
corresponds to the Dicke Hamiltonian [@garraway2011dicke]. Here $a_{\mathbf{k}}^{\dagger}$ ($a_{\mathbf{k}}$) is the creation (annihilation) operator for the SP mode with in-plane momentum $\mathbf{k}$ which satisfies $[a_{\mathbf{k}},a_{\mathbf{k}'}^{\dagger}]=\delta_{\mathbf{k},\mathbf{k}'}$, and $\mathbf{R}=\{R_{\mathbf{n}s}\}$ is an $N$-dimensional vector that describes the vibrational coordinates of the $N=N_{x}N_{y}N_{z}$ molecules of the ensemble, where $N_{i}$ is the number of molecules along each ensemble axis. $\hbar\omega_{g}(R_{\mathbf{n},s})$ accounts for the ground-state energy of the molecule whose location in the ensemble is defined by $\mathbf{n}$ and $s$. We introduce the (adiabatic $R$-dependent) exciton operator $b_{\mathbf{n},s}^{\dagger}(R_{\mathbf{n},s})\left(b_{\mathbf{n},s}(R_{\mathbf{n},s})\right)$ to label the creation (annihilation) of a Frenkel exciton (electronic excitation) with an energy gap $\hbar\omega_{e}(R_{\mathbf{n},s})-\hbar\omega_{g}(R_{\mathbf{n},s})$ on the molecule located at $\mathbf{n}+z_{s}\mathbf{\hat{z}}$. The coefficients $\hbar\omega_{\mathbf{k}}$ and $g_{\mathbf{k}}^{\mathbf{n},s}(R_{\mathbf{n},s})$ stand for the energy of a SP with in-plane momentum $\mathbf{k}$ and the coupling of the molecule located at $\mathbf{n}+z_{s}\hat{\mathbf{z}}$ with the latter, respectively. The dipolar SP-matter interaction is described by $g_{\mathbf{k}}^{\mathbf{n},s}(R_{\mathbf{n},s})=h_{\mathbf{k}}(R_{\mathbf{n},s})f_{\mathbf{k}}(z_{s})$, where $h_{\mathbf{k}}(R_{\mathbf{n},s})=-\mathbf{\boldsymbol{\mu}}_{\mathbf{n},s}(R_{\mathbf{n},s})\cdot\mathbf{E}_{\mathbf{k}}(\mathbf{n})$ is the projection of the molecular transition dipole $\mathbf{\boldsymbol{\mu}}_{\mathbf{n},s}(R_{\mathbf{n},s})$ onto the in-plane component of the SP electric field $\mathbf{E}_{\mathbf{k}}(\mathbf{n})$ and $f_{\mathbf{k}}(z_{s})=e^{-\alpha_{\mathbf{k}}z_{s}}$ is the evanescent field profile along the $z$ direction, with $\alpha_{\mathbf{k}}$ being the decay constant in the molecular region ($z>0$). The quantized plasmonic field $\hat{\mathbf{E}}_{\mathbf{k}}f_{\mathbf{k}}(\mathbf{z}_{s})$ has been discussed in previous works [@Novotny; @Gonzalez-Tudela2013; @Torma2015; @Yuen-Zhou2015] and reads $\hat{\mathbf{E}}_{\mathbf{k}}(\mathbf{n})f_{\mathbf{k}}(\mathbf{z}_{s})=\sqrt{\frac{\hbar\omega_{\mathbf{k}}}{2\epsilon_{0}SL_{\mathbf{k}}}}a_{\mathbf{k}}\hat{\mathbf{\chi}}_{\mathbf{k}}e^{i\mathbf{k}\cdot\mathbf{n}}e^{-\alpha_{\mathbf{k}}z}+h.c.$, where $\epsilon_{0}$ is the free-space permittivity, $S$ is the coherence area of the plexciton setup, $L_{\mathbf{k}}$ is the quantization length, and $\hat{\mathbf{\chi}}_{\mathbf{k}}=\hat{\mathbf{k}}+i\frac{|\mathbf{k}|}{\alpha_{\mathbf{k}}}\hat{\mathbf{z}}$ is the polarization. Note that the parametric dependence of the exciton operators on $R_{\mathbf{n},s}$ yield residual non-adiabatic processes induced by nuclear kinetic energy that may be relevant to the isomerization in question. We also highlight the fact that Eq. (\[Dicke\]) includes both rotating (“energy conserving”) terms ($a_{\mathbf{k}}^{\dagger}b_{\mathbf{n},s}$ and $a_{\mathbf{k}}b_{\mathbf{n},s}^{\dagger}$) where a photon creation (annihilation) involves the concomitant annhilation (creation) of an exciton; and counterrotating (“non-energy conserving”) terms ($a_{\mathbf{k}}b_{\mathbf{n},s}$ and $a_{\mathbf{k}}^{\dagger}b_{\mathbf{n},s}^{\dagger}$) where there is a simultaneous annhilation (creation) of photon and exciton. These latter terms are ignored in the widely used Rotating Wave Approximation (RWA)[@scully1999quantum], where light-matter coupling is weak compared to the transition energy. Since we are interested in the USC, we shall keep them throughout.
Methods
=======
![*Plexciton setup.* The model consists of a surface-plasmon (SP) metal layer whose width $W_{m}$ can be considered infinite in comparison with the relevant length scales of the structure. The thiacynine molecular ensemble is separated from the metallic surface by a spacer of width $z_{0}$; the balls and sticks represent the molecules, while the arrows denote their transition dipole moments. The molecular layer has a height $W_{z}$ and is extended along the $x$ and $y$ planes. \[fig:plexc\_struct\]](plexc_structure)
For simplicity, we assume that all the transition dipoles are equivalent and aligned along $x$, $\mathbf{\boldsymbol{\mu}}_{\mathbf{n},s}(R_{\mathbf{n},s})=\boldsymbol{\mu}(R_{\mathbf{n},s})=\mu(R_{\mathbf{n},s})\hat{\mathbf{x}}$; a departure of this perfect crystal condition does not affect the conclusions of this article. Furthermore, it is convenient to first restrict ourselves to the cases where all nuclei are fixed at the same configuration ($\mathbf{R}=\tilde{\mathbf{R}}$, which denotes $R_{\mathbf{n},s}=R$ for all $\mathbf{n}$ and $s$), so that we can take advantage of the underlying translational symmetry to introduce a delocalized exciton basis where the in-plane momentum $\mathbf{k}$ is a good quantum number. The creation operator of this delocalized state is defined by $b_{\mathbf{k}}^{\dagger}(R)=\frac{1}{\sqrt{\mathcal{N}_{\mathbf{k}}(R)}}\sum_{\mathbf{n}}\sum_{s}f_{\mathbf{k}}(z_{s})h_{\mathbf{k}}(R)b_{\mathbf{n},s}^{\dagger}(R)$, and the normalization squared is given by $\mathcal{N}_{\mathbf{k}}(R)=\sum_{\mathbf{n}}\sum_{s}|h_{\mathbf{k}}(R)|^{2}|f_{\mathbf{k}}(z_{s})|^{2}$ which, in the continuum limit, can be seen to be proportional to $\rho$, the number density of the molecular ensemble. In this collective basis, the previously introduced $H_{el}(\mathbf{R})$ reads
$$\begin{split}
H_{el}(\tilde{\mathbf{R}}) & =\sum_{\mathbf{k}}\hbar\Delta(R)b_{\mathbf{k}}^{\dagger}(R)b_{\mathbf{k}}(R)+\sum_{\mathbf{k}}\hbar\omega_{\mathbf{k}}a_{\mathbf{k}}^{\dagger}a_{\mathbf{k}}\\
& +\sum_{\mathbf{k}}\sqrt{\mathcal{N}_{\mathbf{k}}(R)}\left(a_{\mathbf{k}}^{\dagger}b_{\mathbf{k}}(R)+a_{\mathbf{k}}b_{\mathbf{k}}^{\dagger}(R)+a_{\mathbf{k}}b_{-\mathbf{k}}(R)+a_{\mathbf{k}}^{\dagger}b_{-\mathbf{k}}^{\dagger}(R)\right)+\sum_{\mathbf{k}}H_{\mathrm{dark},\mathbf{k}}(R)+\sum_{\mathbf{k}}H_{\mathrm{unklapp},\mathbf{k}}(R)+N\hbar\omega_{g}(R)\\
& =\sum_{\mathbf{k}}H_{\mathbf{k}}(R)+\sum_{\mathbf{k}}H_{\mathrm{dark},\mathbf{k}}(R)+\sum_{\mathbf{k}}H_{\mathrm{unklapp},\mathbf{k}}(R)+N\hbar\omega_{g}(R),
\end{split}
\label{collective}$$
where $\Delta(R)=\omega_{e}(R)-\omega_{g}(R)$ is the exciton transition frequency. $$\label{dark_states}
H_{\mathrm{dark},\mathbf{k}}(R) =\hbar\Delta(R)\mathbf{P}_{\mathrm{dark},\mathbf{k}}(R)$$ accounts for the energy of the ($N_{z}-1$)-degenerate exciton states with in-plane momentum $\mathbf{k}$ that do not couple to SPs, and are usually known as *dark* *states*. The latter are orthogonal to the bright exciton $b_{\mathbf{k}}^{\dagger}(R)|G_{m}(\tilde{\mathbf{R}})\rangle$ that couples to the SP field, where $|G_{m}(\tilde{\mathbf{R}})\rangle$ is the bare molecular ground-state ($b_{\mathbf{k}}(R)|G_{m}(\tilde{\mathbf{R}})\rangle=0$). More specifically, $\mathbf{P}_{\mathrm{dark},\mathbf{k}}(R)=\mathbf{I}_{\mathrm{exc},\mathbf{k}}(R)-b_{\mathbf{k}}^{\dagger}(R)b_{\mathbf{k}}(R)$ is a projector operator onto the $\mathbf{k}$-th dark-state subspace, with $\mathbf{I}_{\mathrm{exc}}(R)=\sum_{\mathbf{n},s}b_{\mathbf{n},s}^{\dagger}(R)b_{\mathbf{n},s}(R)=\sum_{\mathbf{k},s}b_{\mathbf{k},s}^{\dagger}(R)b_{\mathbf{k},s}(R)=\sum_{\mathbf{k}}\mathbf{I}_{\mathrm{exc},\mathbf{k}}(R)$ being the identity on the exciton space, and $b_{\mathbf{k},s}^{\dagger}(R)=\frac{1}{\sqrt{N_{x}N_{y}}}\sum_{\mathbf{n}}e^{-i\mathbf{k}\cdot\mathbf{n}}b_{\mathbf{n},s}^{\dagger}(R)$. Finally, $$\begin{split}
H_{\mathrm{unklapp},\mathbf{k}}(R)=&\sum_{\mathbf{q}=(\frac{2\pi q_{x}}{\Delta_{x}},\frac{2\pi q_{y}}{\Delta_{y}})}\sqrt{\mathcal{N}_{\mathbf{k}}(R)}\Big(a_{\mathbf{k+q}}^{\dagger}b_{\mathbf{k}}(R)\\
&+a_{\mathbf{k}+q}^{\dagger}b_{-\mathbf{k}}^{\dagger}(R)+h.c.\Big)
\end{split}$$ stands for the coupling of excitons with momentum $\mathbf{k}$ to SP modes with momentum beyond the first excitonic Brillouin zone. $H_{\mathrm{unklapp},\mathbf{k}}(R)$ is usually ignored given the large off-resonance between the SP energy and the exciton states; however, since this work pertains off-resonant effects, we considered it to acquire converged quantities in the calculations explained below. We also note that the normalization constant $\sqrt{\mathcal{N}_{\mathbf{k}}(R)}$ in Eq. \[collective\] is precisely the collective SP-exciton coupling. As mentioned in the introduction, the condition $\sqrt{\mathcal{N_{\mathbf{k}}}(R)}/\hbar\Delta(R)>0.1$ is often used to define the onset of USC [@Moroz], and it is fulfilled with the maximal density considered in our model (see Fig. \[Collective-coupling\]) taking into account that the largest $\hbar \Delta(R)$ is 3 eV (See Fig. \[Adia\_PESs\]a). We note, as will be evident later, that our main results do not vary significantly by considering ratios $\sqrt{\mathcal{N_{\mathbf{k}}}(R)}/\hbar\Delta(R)$ below the aforementioned threshold.
![$\textit{Upper}$: Polariton dispersion that results from the interaction of a molecular ensemble with the plasmonic field; we chose $\rho=1.0\times10^{9}\mu m^{-3}$. $\textit{Lower}$: Collective SP-exciton coupling at equilibrium geometry $\sqrt{\mathcal{N}_{\mathbf{k}}(0)}$ as a function of $|\mathbf{k}|$, assuming $\mathbf{\mu}(R=0)$ and $\mathbf{k}$ are parallel to the $x$ axis. We consider a slab with $W_{z}=120$ nm and compute couplings as a function of varying molecular densities $\rho$. The range of the resulting couplings is well above the plasmonic linewidth of the order of 10 $meV$ [@Gonzalez-Tudela2013], indicating the polaritonic onset of strong and ultrastrong light-matter coupling. \[Collective-coupling\]](merged_collect_poldisp.pdf)
A Bogoliubov transformation [@Ciuti2005] permits the diagonalization of the Bloch Hamiltonian $H_{\mathbf{k}}$ in Eq. \[collective\] by introducing the polariton quasiparticle operators $$\xi_{\mathbf{k}}^{j}(R)=\alpha_{\mathbf{k}}^{j}a_{\mathbf{k}}+\beta_{\mathbf{k}}^{j}b_{\mathbf{k}}(R)+\gamma_{\mathbf{k}}^{j}a_{-\mathbf{k}}^{\dagger}+\delta_{\mathbf{k}}^{j}b_{-\mathbf{k}}^{\dagger}(R),\label{eq:bogoliubov_quasi}$$ where $j=U,L$ and $U$ ($L$) stands for the upper (lower) Bogoliubov polariton state. Notice that this canonical transformation is valid for a sufficiently large number of molecules $N$, where the collective exciton operators $b_{\mathbf{k}}(R)$, $b_{\mathbf{k}}^{\dagger}(R)$ are well approximated by bosonic operators [@Tassone1999].\
The bare molecular ground-state with no photons in the absence of light-matter coupling $|G_{m}(\tilde{\mathbf{R}});0\rangle$, ($a_{\mathbf{k}}|G_{m}(\tilde{\mathbf{R}});0\rangle=b_{\mathbf{k}}(R)|G_{m}(\tilde{\mathbf{R}});0\rangle=0$ for all $\mathbf{k}$) has a total extensive energy with molecular contributions only $\langle G_{m}(\tilde{\mathbf{R}});0|H_{el}(\tilde{\mathbf{R}})|G_{m}(\tilde{\mathbf{R}});0\rangle=N\hbar\omega_{g}(R)$. Upon inclusion of the counterrotating terms, the ground-state becomes the dressed Bogoliubov vacuum $|G(\tilde{\mathbf{R}})\rangle_{d}$, characterized by $\xi_{\mathbf{k}}^{j}(R)|G(\tilde{\mathbf{R}})\rangle_{d}=0$ for all $\mathbf{k}$ and $j$, with total energy $_{d}\langle G(\tilde{\mathbf{R}})|H_{el}(\tilde{\mathbf{R}})|G(\tilde{\mathbf{R}})\rangle_{d}=E_{0}(\tilde{\mathbf{R}})$, where the zero-point energy is given by $$\begin{split}
E_{0}(\tilde{\mathbf{R}})=&N\hbar\omega_{g}(R)\\
&+\frac{1}{2}\sum_{\mathbf{k}}\left(\sum_{j=U,L}\hbar\omega_{j,\mathbf{k}}(R)-\hbar\omega_{\mathbf{k}}-\hbar\Delta(R)\right),
\end{split}\label{shift_vacuum}$$ $\{\hbar\omega_{j,\mathbf{k}}(R)\}$ being the eigenvalues of the Bogoliubov polariton branches given by $$\omega_{_{_{L}^{U}},\mathbf{k}}(R)=\sqrt{\frac{(\Delta(R))^{2}+\omega_{\mathbf{k}}^{2}\pm\sqrt{B(R)^{2}+16\mathcal{N}_{\mathbf{k}}^{2}(R)\Delta(R)\omega_{\mathbf{k}}}}{2}},\label{eq:bogo_disp}
%\end{align}$$ where we have introduced $B(R)=\omega_{\mathbf{k}}^{2}-\Delta(R)^{2}$. A hallmark of the SC and USC regimes is the anticrossing splitting of the polariton energies at the $\mathbf{k}$ value where the bare excitations are in resonance, $\Delta(R)=\omega_{\mathbf{k}}$ [@Torma2015] (see Fig. \[Collective-coupling\]). The sum in Eq. \[shift\_vacuum\] accounts for the energy shift from the bare molecular energy $N\hbar\omega_{g}(R)$ due to interaction with the infinite number of SP modes in the setup. Using Eq. (\[eq:bogo\_disp\]), it is illustrative to check that this shift vanishes identically when the non-RWA terms are ignored.
It is worth describing some of the physical aspects of the Bogoliubov ground-state $|G(\tilde{\mathbf{R}})\rangle_{d}$. With the numerically computed wavefunctions, we can use the inverse transformation of Eq. \[eq:bogoliubov\_quasi\] to explicitly evaluate its SP and exciton populations [@Ciuti2005],
\[eq:populations\] $$\begin{aligned}
n_{\mathbf{k}}^{SP}={}_{d}\langle G(\tilde{\mathbf{R}})|a_{\mathbf{k}}^{\dagger}a_{\mathbf{k}}|G(\tilde{\mathbf{R}})\rangle_{d} & =\sum_{j}|\gamma_{\mathbf{k}}^{j}|^{2},\label{eq:pop_SP}\\
n_{\mathbf{k}}^{exc}={}_{d}\langle G(\tilde{\mathbf{R}})|b_{\mathbf{k}}^{\dagger}b_{\mathbf{k}}|G(\tilde{\mathbf{R}})\rangle_{d} & =\sum_{j}|\delta_{\mathbf{k}}^{j}|^{2},\label{eq:pop_exc}\end{aligned}$$
which give rise to humble $O(10^{-3})$ values per mode $\mathbf{k}$, considering a molecular ensemble with $\rho=3\times 10^{8} \mu m^{-3}$ and $W_{z}=$120 nm; this calculation is carried out using $N=8\times10^{7}$, although results are largely insensitive to this parameter as long as it is sufficiently large to capture the thermodynamic limit. The consequences of the dressing partially accounted for by Eq. (\[eq:populations\]) (partially since there are also correlations of the form $_{d}\langle G(\tilde{\mathbf{R}})|b_{\mathbf{k}}a_{-\mathbf{k}}|G(\tilde{\mathbf{R}})\rangle_{d}$) are manifested as energetic effects on $|G_{m}(\tilde{\mathbf{R}});0\rangle$: $E_{0}(\tilde{\mathbf{R}})-N\hbar\omega_{g}(R)$ can be interpreted as the energy stored in $|G(\tilde{\mathbf{R}})\rangle_{d}$ as a result of dressing; it is an extensive quantity of the ensemble, but becomes negligible when considering a per-molecule stabilization. For instance, in molecular ensembles with the aforementioned parameters we find $E_{0}(\tilde{\mathbf{0}})-N\hbar\omega_{g}(0)=O(10^{2})$ eV, which implies a $O(10^{-5})$ eV value per molecule; our calculations show that this intensive quantity is largely insensitive to total number of molecules. This observation raises the following questions: to what extent does photonic dressing would impact ground-state chemical reactivity? What are the relevant energy scales that dictate this impact? With these questions in mind, we aim to study the polaritonic effects on ground-state single-molecule isomerization events. To do so, we map out the PES cross section where we set one “free‘" molecule to undergo isomerization while fixing the rest at $R_{\mathbf{n},s}=0$. A similar strategy has been used before in [@Galego2016]. This cross section, described by $E_{0}(R_{\mathbf{n}_{0},0},0,\cdots,0)\equiv E_{0}(R_{\mathbf{n}_{0},0},\mathbf{\tilde{0}}')$ ($R_{\mathbf{n}_{0},0}$ being the coordinate of the unconstrained molecule), should give us an approximate understanding of reactivity starting from thermal equilibrium conditions, since the molecular configuration $\tilde{\mathbf{R}}=\tilde{\mathbf{0}}$ still corresponds to the global minimum of the modified ground-state PES, as will be argued later. By allowing one molecule to move differently than the rest, we weakly break translational symmetry. Rather than numerically implementing another Bogoliubov transformation, we can, to a very good approximation, account for this motion by treating the isomerization of the free molecule as a perturbation on $H_{el}(\tilde{\boldsymbol{0}})$. More precisely, we write $H_{el}(R_{\mathbf{n}_{0},0},\mathbf{\tilde{0}}')|G(R_{\mathbf{n}_{0},0},\mathbf{\tilde{0}}')\rangle_{d}=E_{0}(R_{\mathbf{n}_{0},0},\mathbf{\tilde{0}}')|G(R_{\mathbf{n}_{0},0},\mathbf{\tilde{0}}')\rangle_{d}$, where $H_{el}(R_{\mathbf{n}_{0},0},\mathbf{\tilde{0}}')$ is the sum of a translationally invariant piece $H_{el}(\tilde{\mathbf{0}})$ plus a perturbation due to the free molecule, $$H_{el}(R_{\mathbf{n}_{0},0},\mathbf{\tilde{0}}')=H_{el}(\tilde{\mathbf{0}})+V(R_{\mathbf{n}_{0},0}).\label{def_pert}$$ The perturbation is explicitly given by
$$\begin{split}V(R_{\mathbf{n}_{0},0}) & =H_{el}(R_{\mathbf{n}_{0},0},\mathbf{\tilde{0}}')-H_{el}(\tilde{\mathbf{0}})\\
& =\hbar\Delta(R_{\mathbf{n}_{0},0})b_{\mathbf{n}_{0},0}^{\dagger}(R_{\mathbf{n}_{0},0})b_{\mathbf{n}_{0},0}(R_{\mathbf{n}_{0},0})-\hbar\Delta(0)b_{\mathbf{n}_{0},0}^{\dagger}(0)b_{\mathbf{n}_{0},0}(0)\\
& +\sum_{\mathbf{k}}\Bigg\{ g_{\mathbf{k}}^{\mathbf{n}_{0},0}(R_{\mathbf{n}_{0},0})\Bigg[b_{\mathbf{n}_{0},0}(R_{\mathbf{n}_{0},0})+b_{\mathbf{n}_{0},0}^{\dagger}(R_{\mathbf{n}_{0},0})\Bigg]-g_{\mathbf{k}}^{\mathbf{n}_{0},0}(0)\Bigg[b_{\mathbf{n}_{0},0}(0)+b_{\mathbf{n}_{0},0}^{\dagger}(0)\Bigg]\Bigg\}\Bigg[a_{\mathbf{k}}+a_{\mathbf{k}}^{\dagger}\Bigg]\\
& +\hbar\omega_{g}(R_{\mathbf{n}_{0},0})-\hbar\omega_{g}(0).
\end{split}
\label{perturbation}$$
Notice that we have chosen the free molecule to be located at an arbitrary in-plane location $\mathbf{n}_{0}$ and at the very bottom of the slab at $s=0$, where light-matter coupling is strongest as a result of the evanescent field profile along the $z$ direction. We write an expansion of the PES cross section as $E_{0}(R_{\mathbf{n}_{0},0},\mathbf{\tilde{0}}')=\sum_{q=0}^{\infty}E_{0}^{(q)}(R_{\mathbf{n}_{0},0},\mathbf{\tilde{0}}')$, where $q$ labels the $O(V^{q})$ perturbation correction. The zeroth order term is the Bogoliubov vacuum energy associated to every molecule being at the equilibrium geometry $E_{0}^{(0)}(R_{\mathbf{n}_{0},0},\mathbf{\tilde{0}}')=E_{0}(\tilde{\mathbf{0}})$ as in Eq. (\[shift\_vacuum\]). The $O(V)$ correction corresponds to $\hbar\omega_{g}(R_{\mathbf{n}_{0},0})-\hbar\omega_{g}(0)$, merely describing the PES of the isomerization of the bare molecule in the absence of coupling to the SP field. The contribution of the SP field on the PES cross-section of interest appears at $O(V^{2})$, and it is given by $$\begin{aligned}
E^{(2)}(R_{\mathbf{n}_{0},0},\mathbf{\tilde{0}}')\approx\sum_{_{i,j=UP,LP}^{\mathbf{k}_{1}\leq\mathbf{k}_{2}}}\frac{|\langle\mathbf{k}_{1},i;\mathbf{k}_{2},j|V(R_{\mathbf{n}_{0},0})|G(\tilde{\mathbf{0}})\rangle_{d}|^{2}}{E_{0}(\tilde{\mathbf{0}})-E_{\mathbf{k}_{1},\mathbf{k}_{2},i,j}^{(0)}},\label{second_order}\end{aligned}$$ where $|\mathbf{k}_{1},i;\mathbf{k}_{2},j\rangle\equiv\xi_{\mathbf{k}_{1}}^{i\dagger}(0)\xi_{\mathbf{k}_{2}}^{\dagger j}(0)|G(\tilde{\mathbf{0}})\rangle_{d}$ and $E_{\mathbf{k}_{1},\mathbf{k}_{2},i,j}^{(0)}=\hbar(\omega_{i,\mathbf{k}_{1}}(0)+\omega_{j,\mathbf{k}_{2}}(0))$. As shown in the Appendix, the approximation in Eq. (\[second\_order\]) consists of ignoring couplings between $|G(\tilde{\mathbf{0}})\rangle_{d}$ and states with three and four Bogoliubov polariton excitations, since their associated matrix elements become negligible in the thermodynamic limit compared to their double excitation counterparts. The remaining matrix elements can be calculated by expressing the operators $a_{\mathbf{k}},\,a_{\mathbf{k}}^{\dagger},\,b_{\mathbf{n}_{0},0}(R_{\mathbf{n}_{0},0}),\,b_{\mathbf{n}_{0},0}^{\dagger}(R_{\mathbf{n}_{0},0})$ in Eq. (\[perturbation\]) in terms of the Bogoliubov operators $\xi_{\mathbf{k}}^{j}(0),\,\xi_{\mathbf{k}}^{\dagger j}(0)$ (see Eq. (\[eq:bogoliubov\_quasi\])), leading to
$$\begin{split}\langle\mathbf{k}_{1},i;\mathbf{k}_{2},j|V(R_{\mathbf{n}_{0},0})|G(\tilde{\mathbf{0}})\rangle_{d} & =F^{\mathbf{k}_{2}}(R_{\mathbf{n}_{0},0})D_{\mathbf{k}_{1}}\left(-\delta_{-\mathbf{k}_{1}}^{i}\alpha_{\mathbf{k}_{2}}^{j}+\delta_{-\mathbf{k}_{1}}^{i}\gamma_{-\mathbf{k}_{2}}^{j}-\beta_{\mathbf{k}_{1}}^{i}\gamma_{-\mathbf{k}_{2}}^{j}+\beta_{\mathbf{k}_{1}}^{i}\alpha_{\mathbf{k}_{2}}^{j}\right)\\
& +F^{\mathbf{k}_{1}}(R_{\mathbf{n}_{0},0})D_{\mathbf{k}_{2}}\left(-\delta_{-\mathbf{k}_{2}}^{j}\alpha_{\mathbf{k}_{1}}^{i}+\delta_{-\mathbf{k}_{2}}^{j}\gamma_{-\mathbf{k}_{1}}^{i}-\beta_{\mathbf{k}_{2}}^{j}\gamma_{-\mathbf{k}_{1}}^{i}+\beta_{\mathbf{k}_{2}}^{j}\alpha_{\mathbf{k}_{1}}^{i}\right),
\end{split}
\label{coupling}$$
where $F^{\mathbf{k}}(R)=\cos(\theta(R))g_{\mathbf{k}}^{\mathbf{n}_{0},0}(R)-\cos(\theta(0))g_{\mathbf{k}}^{\mathbf{n}_{0},0}(0)$ depends on the mixing angle that describes the change of character of $b_{\mathbf{n}_{0},0}^{\dagger}(R)$ as a function of $R$ (see Equation (\[adiabatic\_rep\])); it emerges as a consequence of coupling molecular states at different configurations. $D_{\mathbf{k}}=\langle G_{m}(\tilde{\mathbf{R}});0\rangle|b_{\mathbf{n}_{0},0}(0)b_{\mathbf{k}}^{\dagger}(0)|G_{m}(\tilde{\mathbf{R}});0\rangle=\frac{1}{\sqrt{N_{x}N_{y}}}\sqrt{\frac{1-e^{-2\alpha_{\mathbf{k}}\Delta_{z}}}{1-e^{-2\alpha_{\mathbf{k}}\Delta_{z}N_{z}}}}$ accounts for the weight of a localized exciton operator in a delocalized one, such as the participation of $b_{\mathbf{n}_{0},0}^{\dagger}(0)$ in $b_{\mathbf{k}}^{\dagger}(0)$. Eq. (\[coupling\]) reveals that the maximal contribution of each double-polariton Bogoliubov state to the energetic shift of the considered PES cross section $E(R_{\mathbf{n}_{0},0},\mathbf{\tilde{0}}')$ is of the order of $\frac{g_{\mathbf{k}}^{\mathbf{n}_{0},0}(0)}{\sqrt{N_{x}N_{y}}}$. Considering macroscopic molecular ensembles with large $N\approx10^{7}$, we computed Eq. \[second\_order\] by means of an integral approximation over the polariton modes $\mathbf{k}$.
Results and discussion
======================
Energetic effects
-----------------
We carry out our calculations with $\rho$ in the range of $10^{6}$ to $10^{9}$ molecules $\mu m^{-3}$ keeping $W_{z}=120\,\text{nm}$ (see Fig. \[1D\_PES\]); to obtain results in the thermodynamic limit, our calculations take $N=8\times10^{7}$, even though the exact value is unimportant as long as it is sufficiently large to give converged results. The results displayed in Fig. \[1D\_PES\] show that the second order energy corrections to the isomerization PES $E^{(2)}(R_{\mathbf{n}_{0},0},\mathbf{\tilde{0}}')$, and in particular $E^{(2)}(R_{\mathbf{n}_{0},0}=R^{*},\mathbf{\tilde{0}}')\approx-0.25\,\text{meV}$, are negligible in comparison with the bare activation barrier $E_{a}=\hbar\omega_{g}(R^{*})-\hbar\omega_{g}(0)=\hbar\omega_{g}(R^{*})\approx 1.8$ eV, where $R^{*}\approx1.64$ rad corresponds to the transition state. From Fig. \[Adia\_PESs\]b, we notice that there is a substantial difference in SP-exciton coupling between the equilibrium ($R_{\mathbf{n}_{0},0}=0$) and transition state geometries ($R_{\mathbf{n}_{0},0}=R^{*}$). Since the perturbation in Eq. (\[perturbation\]) is defined with respect to the equilibrium geometry, $|E^{(2)}(R_{\mathbf{n}_{0},0},\mathbf{\tilde{0}}')|$ maximizes at the barrier geometry. To get some insight on the order of magnitude of the result, we note that the sum shown in Eq. \[second\_order\] can be very roughly approximated as
$$\begin{aligned}
E^{(2)}(R_{\mathbf{n}_{0},0},\mathbf{\tilde{0}}')&=O\left[-\sum_{\mathbf{k}_{1}\leq\mathbf{k_{2}}}\frac{[g_{\mathbf{k}_{1}}^{\mathbf{n}_{0},0}(R_{\mathbf{n}_{0},0})]^{2}D_{\mathbf{k}_{2}}^{2}+[g_{\mathbf{k}_{2}}^{\mathbf{n}_{0},0}(R_{\mathbf{n}_{0},0})]^{2}D_{\mathbf{k}_{1}}^{2}}{(\hbar\omega_{\mathbf{k}_{1}}+\hbar\omega_{\mathbf{k}_{2}})/2+\hbar\omega_{e}(R_{\mathbf{n}_{0},0})}\right]\nonumber \\
&=O\left[-\frac{1}{N_{x}N_{y}}\sum_{\mathbf{k}_{1}\leq\mathbf{k_{2}}}\frac{[g_{\mathbf{k}_{1}}^{\mathbf{n}_{0},0}(R_{\mathbf{n}_{0},0})]^{2}+[g_{\mathbf{k}_{2}}^{\mathbf{n}_{0},0}(R_{\mathbf{n}_{0},0})]^{2}}{(\hbar\omega_{\mathbf{k}_{1}}+\hbar\omega_{\mathbf{k}_{2}})/2+\hbar\omega_{e}(R_{\mathbf{n}_{0},0})}\right]\nonumber \\
& =O\left[-\sum_{\mathbf{k}}\frac{[g_{\mathbf{k}}^{\mathbf{n}_{0},0}(R_{\mathbf{n}_{0},0})]^{2}}{\hbar\omega_{\mathbf{k}}+\hbar\omega_{e}(R_{\mathbf{n}_{0},0})}\right]\label{eq:lamb_shift_red}\\
& =O\left(E_{LS}(R_{\mathbf{n}_{0},0})\right).\nonumber \end{aligned}$$
In the first line, we used the fact that $\langle\mathbf{k}_{1},i;\mathbf{k}_{2},j|V(R_{\mathbf{n}_{0},0})|G(\tilde{\mathbf{0}})\rangle_{d}\approx [g_{\mathbf{k}_{1}}^{\mathbf{n}_{0},0}(R_{\mathbf{n}_{0},0})]^{2}D_{\mathbf{k}_{2}}^{2}+[g_{\mathbf{k}_{2}}^{\mathbf{n}_{0},0}(R_{\mathbf{n}_{0},0})]^{2}D_{\mathbf{k}_{1}}^{2}$ and averaged the Bogoliubov polariton excitation energies. In the second line, assuming that the $\mathbf{k}\gg0$ values contribute the most, we have $D_{\mathbf{k}}\approx\frac{1}{\sqrt{N_{x}N_{y}}}$. Finally, in the third line, we have used the fact that the sum of terms over $\mathbf{k}_{1},\mathbf{k}_{2}$ is roughly equal to $N_{x}N_{y}$ times a single sum over $\mathbf{k}$ of terms of the same order. The reason why we are interested in the final approximation is because it corresponds to the Lamb shift of a single isolated molecule, which can be calculated to be $E_{LS}(0)=0.16\,\text{meV}$. Typically, Lamb shift calculations require a cutoff to avoid unphysical divergences [@Bethe1950]; we stress that in our plexciton model, this is not necessary due to the decaying $|g_{\mathbf{k}}^{\mathbf{n}_{0},0}(R_{\mathbf{n}_{0},0})|$ as a function of $|\mathbf{k}|$. The fact that the corrections $E^{(2)}(R_{\mathbf{n}_{0},0},\mathbf{\tilde{0}}')$ have a similar order of magnitude to single-molecule Lamb shifts give a pessimistic conclusion of harnessing USC to control ground-state chemical reactions.
![*Upper:* Second order energy correction $E^{(2)}(R_{\mathbf{n}_{0},0},\mathbf{\tilde{0}}')$ of PES for one molecule isomerizing along the torsional coordinate $R_{\mathbf{n}_{0},0}$; the rest of the molecules are fixed at the equilibrium geometry. Calculations are displayed for various densities $\rho$, keeping $W_{z}=120$ nm. Energy corrections are due to SP-exciton (see Eq. \[second\_order\]). Note also that the energy scale of this correction is negligible in comparison with the energy barrier of the reaction (see Fig. \[Adia\_PESs\]a). *Lower:* Same plot as in left, but for the 2D-ground-isomerization PES of two molecules, keeping the configuration of the other molecules at equilibrium ($E^{(2)}(R_{0},R_{1},\mathbf{\tilde{0}}')|$); the density $\rho=3\times10^{8}\,\text{molecules}/\mu m^{3}$. \[1D\_PES\]](drawing.pdf)
Note, however, from calculations in Fig. \[1D\_PES\], that there is variability in $E^{(2)}(R_{\mathbf{n}_{0},0},\mathbf{\tilde{0}}')$ as a function of molecular density (since density alters the character of the Bogoliubov polaritons), although the resulting values are always close to $E_{LS}(0)$. The molecular density cannot increase without bound, since there exists a minimum molecular contact distance determined by a van der Waals radius of the order of $0.3\,\text{nm}$ for organic molecules [@rowland1996intermolecular], giving a maximum density of $\rho\approx10^{10}\,\text{molecules}/\mu m^{3}$.\
The results discussed so far describe the energy profile of the isomerization of a single molecule keeping the rest at equilibrium geometry. It is intriguing to inquire the effects of the SP field in a concerted isomerization of two or more molecules, while keeping the rest fixed at equilibrium geometry. Generalizing Eqs. (\[perturbation\])(\[coupling\]) to a two-molecule perturbation $V(R_{\mathbf{n}_{0},0},R_{\mathbf{n}_{1},0})$, we computed the second order energetic corrections to the 2D-PES that describe the isomerization of two neighbouring molecules at $\mathbf{n}_{0}$ and at $\mathbf{n}_{1}\equiv\mathbf{n}_{0}+\Delta_{x}\mathbf{\hat{x}}$, keeping the other molecules fixed at $R_{\mathbf{n},s}=0$. The results are reported in Fig. \[1D\_PES\] for $\rho=3\times10^{8}\,\text{molecules}/\mu m^{3}$, although outcomes of the same order of magnitude are obtained for the other densities considered in the one-dimensional case. The two-dimensional PES cross-section $E^{(2)}(R_{\mathbf{n}_{0},0},R_{\mathbf{n}_{1},0},0,\cdots,0)\equiv E^{(2)}(R_{\mathbf{n}_{0},0},R_{\mathbf{n}_{1},0},\mathbf{\tilde{0}}')$ shows the existence of an energetic enhancement for the concerted isomerization with respect to two independent isomerizations, *i.e.* $E^{(2)}(R_{\mathbf{n}_{0},0}=R^{*},R_{\mathbf{n}_{1},0}=R^{*},\mathbf{\tilde{0}}')\approx4E^{(2)}(R_{\mathbf{n}_{0},0}=R^{*},\mathbf{\tilde{0}}')$. This enhacement is due to a constructive interference arising at the amplitude level, $\langle\mathbf{k}_{1},i;\mathbf{k}_{2},j|V(R_{\mathbf{n}_{0},0}=R^{*},R_{\mathbf{n}_{1},0}=R^{*})|G(\tilde{\mathbf{0}})\rangle_{d}\approx2\langle\mathbf{k}_{1},i;\mathbf{k}_{2},j|V(R_{\mathbf{n}_{0},0}=R^{*})|G(\tilde{\mathbf{0}})\rangle_{d}$ for values of $\text{\ensuremath{\mathbf{k}}}_{1},\,\text{\ensuremath{\mathbf{k}}}_{2}\ll\frac{1}{\Delta_{x}}$, such that the phase difference between the isomerizing molecules is negligible. Interestingly, choosing the neighbouring molecules along the $x$ direction is important for this argument; if instead we consider neighbours along $z$ (molecular positions $\mathbf{n}_{0}$ and $\mathbf{n}_{0}+\Delta_{z}\mathbf{\hat{z}}$), these interferences vanish and we approximately get the independent molecules result $E^{(2)}(R_{\mathbf{n}_{0},0}=R^{*},R_{\mathbf{n}_{0},1}=R^{*},\mathbf{\tilde{0}}')\approx2E^{(2)}(R_{\mathbf{n}_{0},0}=R^{*},\mathbf{\tilde{0}}')$.\
In light of the nontrivial energetic shift of the two-molecule case, it is pedagogical to consider the SP effects on the cross-section of the concerted isomerization of the whole ensemble, even though it is highly unlikely that this kinetic pathway will be of any relevance, especially considering the large barrier for the isomerization of each molecule. Notice that the conservation of translational symmetry in this scenario allows for the exact (nonperturbative) calculation of the energetic shift $E_{0}(\tilde{\mathbf{R}})-N\hbar\omega_{g}(R)$ by means of Eq. \[shift\_vacuum\]. Our numerical calculations reveal an energetic stabilization profile, which is displayed in Fig. \[concerted\] for a molecular ensemble with $\rho=3\times 10^{8}$ molecules $\mu m^{-3}$. As expected, we observe a stabilization of reactant and product regions of the ground-state PES. This is a consequence of the transition dipole moment being the strongest at those regions, as opposed to the transition state, see Fig. \[Adia\_PESs\]b. However, even though these energetic effects are of the order of hundreds of eV, they are negligible in comparison with the total ground-state PES $N\hbar\omega_{g}(R)$, or more specifically, to the transition barrier $NE_{a}=N\hbar\omega_{g}(R^{*})$ for the concerted reaction.
![*Main:* profile of the energy stabilization of the concerted isomerization ($E_{0}(\tilde{\mathbf{R}})-N\hbar\omega_{g}(R)$, see Eq. \[shift\_vacuum\]) of the whole molecular ensemble discussed in the main text, due to the interaction with the plasmonic field. We consider a molecular macroscopic ensemble ($N=8\times10^{7}$) with density $\rho=3\times10^{8}\,\text{molecules}/\mu m^{3}$. *Inset:* molecular-density dependence of the the energy shift of the energy barrier per molecule $|\Delta E_{a}|$ (see main text) due to the plasmonic field, in this concerted scenario. \[concerted\]](collect_iso_merged.pdf)
Importantly, the change in activation energy per molecule in the concerted isomerization with respect to the bare case $|\Delta E_{a}|=\left|\left(\frac{E_{0}(\tilde{\mathbf{R}}^{*})-E_{0}(\tilde{\mathbf{0}})}{N}-E_{a}\right)\right|\approx0.009\,\text{meV}$ is more than one order of magnitude smaller than the corresponding quantity $\left|E^{(2)}(R_{\mathbf{n}_{0},0}=R^{*},\mathbf{\tilde{0}}')\right|\approx0.25\,\text{meV}$ for the single-molecule isomerization case, see Fig. \[1D\_PES\] and inset of Fig. \[concerted\]. We believe that the reason for this trend is that the isomerization of $n$ molecules, $n\ll N$, translates into a perturbation which breaks the original translational symmetry of the molecular ensemble. This symmetry breaking permits the interaction of the molecular vacuum with the polaritonic $\mathbf{k}$-state reservoir without a momentum-conservation restriction. This is reflected in Eq. \[second\_order\], where the sum is carried out over two not necessarily equal momenta. In contrast, in the case of the concerted isomerization of $N$ molecules, the translational symmetry of the system is preserved, which in turn restricts the coupling of the vacuum $|G(\tilde{\mathbf{0}})\rangle_{d}$ to excited states with $\mathbf{k}_{\mathrm{exc}}=-\mathbf{k}_{\mathrm{phot}}$.\
Another intriguing observation is that, for this concerted isomerization, the SP energetic effect per molecule $\frac{E_{0}(\tilde{\mathbf{R}})}{N}$ diminishes with the width of the slab $W_{z}$. This is the case given that the SP quantization length $L_{\mathbf{k}}$ decays quickly with $|\mathbf{k}|$ so that only the closest layers interact strongly with the field. When we divide the total energetic effects due to the SP modes by $N=N_{x}N_{y}N_{z}$, we obtain that $\frac{E_{0}(\tilde{\mathbf{R}})}{N}=O(\frac{1}{N_{z}})$ for large $W_{z}$.\
The energetic shifts in all the scenarios discussed above are negligible with respect to the corresponding energy barriers and the thermal energy scale at room temperature which, unfortunately, signal the irrelevance of USC to alter ground-state chemical reactivity for this isomerization model. Although there is an overall (extensive) stabilization of the molecular ensemble ground state, this effect is distributed across the ensemble, giving no possibility to alter the chemical reaction kinetics or thermodynamics considerably. However, we highlight the intriguing interferences observed in the concerted isomerization processes. Even though they will likely be irrelevant for this particular reaction, they might be important when dealing with reactions with very low barriers, especially when considering that these concerted pathways are combinatorially more likely to occur than the single-molecule events in the large $N$ limit. This is intriguing in light of the study carried out in [@galego2017many], which discusses a different but related effect of many reactions triggered by a single photon.\
Effects on non-adiabatic dynamics
---------------------------------
Finally, we discuss the importance of the nonadiabatic effects afforded by nuclear kinetic energy. Previous works have considered the nonadiabatic effects between polariton states at the level of SC [@Galego2015a; @Kowalewski2016a]. Alternatively, the consideration of nonadiabatic effects in USC for a single molecule in a cavity was provided in [@Bennett2016a]; here, we address these issues for the many-molecule case and consider both polariton and dark state manifolds. One could expect significantly modified non-adiabatic dynamics about nuclear configurations where the transition dipole moment magnitude $|\mathbf{\mathbf{\mu}}_{\mathbf{n},s}(R_{\mathbf{n},s})|$ is large, given a reduction in the energy gap between the ground and the lower Bogoliubov polariton state. However, as we show below, this energetic effect is not substantial due to the presence of dark states.\
We consider the magnitude of the non-adiabatic couplings (NACs) for the isomerization of a single molecule with reaction coordinate $R_{\mathbf{n}_{0},0}$. For a region about $\tilde{\mathbf{R}}=\tilde{\mathbf{0}}$, we estimate the magnitude of the NAC between $|G(\tilde{\mathbf{0}})\rangle_{d}$ and a state $|\mathbf{k},i\rangle=\xi_{\mathbf{k}}^{i\dagger}(0)|G(\tilde{\mathbf{0}})\rangle_{d}$ as: $$\begin{aligned}
|A_{\mathbf{k},i;g}(0)| & =\left|\langle\mathbf{k},i\big|\frac{\partial}{\partial R_{\mathbf{n}_{0},0}}\big|G(\mathbf{\tilde{0}})\rangle_{d}\right|\label{eq:NAC}\\
& \approx\left|\beta_{\mathbf{k}}^{i}D_{\mathbf{k}}\left\langle e_{\mathbf{n}_{0},0}(0)\left|\frac{\partial}{\partial R_{\mathbf{n}_{0,0}}}\right|g_{\mathbf{n}_{0},0}(0)\right\rangle \right|,\nonumber \end{aligned}$$ where $|g_{\mathbf{n}_{0},0}(0)\rangle$($|e_{\mathbf{n}_{0},0}(0)\rangle$) is the ground (excited) adiabatic state of the single molecule under consideration (see Eq. (\[adiabatic\_rep\])) and we have ignored the derivatives of $\beta_{\mathbf{k}}^{i}$ and $D_{\mathbf{k}}$ with respect to $R_{\mathbf{n}_{0},0}$, assuming they are small at $\tilde{\mathbf{R}}=\tilde{\mathbf{0}}$, where the chemical character of the Bogoliubov polariton states does not change significantly with respect to nuclear coordinate. This is a consequence of the slowly changing transition dipole moment of the model molecule around $R_{\mathbf{n}_{0},0}=0$, see Fig. \[Adia\_PESs\]b. Notice that we have also assumed $\langle\mathbf{k},i|e_{\mathbf{n}_{0},0}(0)\rangle\approx\beta_{\mathbf{k}}^{i}D_{\mathbf{k}}$, where we have used the fact that $\beta_{\mathbf{k}}^{i}\gg\gamma_{\mathbf{k}}^{i}$, thus ignoring counterrotating terms, which as we have seen, give negligible contributions. The time-evolution of a nuclear wavepacket in the ground-state will be influenced by the Bogoliubov polariton states, each of which will contribute with a finite probability of transition out of $\big|G(\mathbf{\tilde{0}})\rangle_{d}$. From semiclassical arguments [@Bohm2003], we can estimate the transition probability $|C_{\mathbf{k}}^{i}(0)|^{2}$ for a nuclear wavepacket on the ground-state PES at $\tilde{\mathbf{R}}=0$ to the state $|\mathbf{k},i\rangle$, $$\begin{aligned}
|C_{\mathbf{k}}^{i}(0)|^{2} & \approx\left|\frac{\hbar v_{nuc}A_{\mathbf{k},i;g}(0)}{\hbar\omega_{i,\mathbf{k}}(0)-\hbar\omega_{g}(0)}\right|^{2}\label{eq:prob_NAC}\\
& =\left|\frac{\hbar v_{nuc}\beta_{\mathbf{k}}^{i}D_{\mathbf{k}}}{\hbar\omega_{i,\mathbf{k}}(0)-\hbar\omega_{g}(0)}\right|^{2}\nonumber \\
& \times\left|\left\langle e_{\mathbf{n}_{0},0}(0)\left|\frac{\partial}{\partial R_{\mathbf{n}_{0,0}}}\right|g_{\mathbf{n}_{0},0}(0)\right\rangle \right|^{2},\nonumber \end{aligned}$$ $v_{nuc}$ being the expectation value of the nuclear velocity. However, the Bogoliubov polariton $\mathbf{k}$-states are only a small subset of the excited states of the problem. As mentioned right after Eq. \[collective\], the plexciton setup contains $N_{z}-1$ dark excitonic states for every $\mathbf{k}$ (eigenstates of $H_{\mathrm{dark},\mathbf{k}}(0)$, see discussion right after Eq. \[collective\]); we ignore the very off-resonant couplings considered in $H_{\mathrm{unklapp},\mathbf{k}}(0)$. The dark states also couple to $\big|G(\mathbf{\tilde{0}})\rangle_{d}$ non-adiabatically, with the corresponding transition probability out of the ground state being, $$\begin{aligned}
|C_{\mathbf{k}}^{\mathrm{dark}}(0)|^{2} & \approx\sum_{Q}\left|\frac{\hbar v_{nuc}A_{\mathbf{k},Q;g}(0)}{\hbar\Delta(0)}\right|^{2}\label{eq:prob_NAC-1}\\
& \approx P_{bare}(0)\left(\frac{1}{N_{x}N_{y}}-|D_{\mathbf{k}}|^{2}\right),\nonumber \end{aligned}$$ Here, we have summed over all dark states $Q$ for a given $\mathbf{k}$ and used $P_{bare}(0)=\left|\frac{v_{nuc}}{\Delta(0)}\right|^{2}\left|\left\langle e_{\mathbf{n}_{0},0}(0)\left|\frac{\partial}{\partial R_{\mathbf{n}_{0,0}}}\right|g_{\mathbf{n}_{0},0}(0)\right\rangle \right|^{2}$ to denote the probability of transition out of the ground state in the absence of coupling to the SP field. In Eq. (\[eq:prob\_NAC-1\]) we used the fact that the projection $|e_{\mathbf{n}_{0},0}(0)\rangle$ onto the dark $\mathbf{k}$ manifold of exciton states is $\left|\mathbf{P}_{\mathrm{dark},\mathbf{k}}(0)|e_{\mathbf{n}_{0},0}(0)\rangle\right|^{2}=\langle e_{\mathbf{n}_{0},0}(0)|\mathbf{I}_{\mathrm{exc},\mathbf{k}}(0)|e_{\mathbf{n}_{0},0}(0)\rangle-|D_{\mathbf{k}}|^{2}=\frac{1}{N_{x}N_{y}}-|D_{\mathbf{k}}|^{2}$, with $\mathbf{P}_{\mathrm{dark},\mathbf{k}}(0)$ being the corresponding projector (see Eq. (\[dark\_states\])). We noticed that when $|\mathbf{k}|\to0$, the quantization volume $L_\mathbf{k}$ of the plasmonic field spans all the molecular-ensemble volume resulting in completely delocalized bright and dark exciton states across the different layers of the slab, $\left|\mathbf{P}_{\mathrm{dark},\mathbf{k}}|e_{\mathbf{n}_{0},0}(0)\rangle\right|^{2}=\frac{N_{z}-1}{N}$, and the dark states give the major contribution to the nonadiabatic dynamics. On the other hand, when $|\mathbf{k}|\to\infty$, the plasmonic field interacts with the molecular layer at the bottom of the slab only and $\left|\mathbf{P}_{\mathrm{dark},\mathbf{k}}|e_{\mathbf{n}_{0},0}(0)\rangle\right|^{2} \to 0$. The dark states do not participate, because the molecule located at $\mathbf{n}_{0}$ only overlaps with the bright state which is concentrated across the first layer of the slab (the dark states, being orthogonal to the bright one, are distributed in the upper layers, and do not overlap with $|e_{\mathbf{n}_{0},0}\rangle$). With these results, we can compute the probability of transition out of the ground-state $P_{out}$ as $$\begin{aligned}
P_{out}(0) & =\sum_{\mathbf{k}}\Bigg[\sum_{i}|C_{\mathbf{k}}^{i}(0)|^{2}+|C_{\mathbf{k}}^{\mathrm{dark}}(0)|^{2}\Bigg].\label{eq:P_out}\end{aligned}$$ In view of the large off-resonant nature of most SP modes with respect to $\hbar\Delta(0)$ (see Fig. \[Collective-coupling\]) and Eq. (\[eq:prob\_NAC-1\]), we have $\sum_{i}|C_{\mathbf{k}}^{i}(0)|^{2}\approx P_{bare}(0)|D_{\mathbf{k}}|^{2}$, such that $P_{out}(0)\approx P_{bare}(0)$. In our model, this is the case, since the plexciton anticrossing occurs at small $|\mathbf{k}|$ and the SP energy quickly increases and reaches an asymptotic value after that point (see Fig. \[Collective-coupling\]).\
Using the parameters in [@Hoki2009a], we obtain $\langle e(R_{\mathbf{n}_{0},0})|\frac{\partial}{\partial R_{\mathbf{n}_{0},0}}|g(R_{\mathbf{n}_{0},0})\rangle\approx0.01\,\text{\AA}^{-1}$, where we have assumed an effective radius of 1 $\text{\AA}$ for the isomerization mode of the model molecule. We get an estimate of $v_{nuc}\approx 1\text{\AA}$ $\omega_{nuc}=1\text{\AA} \sqrt{\frac{k_{B}T}{m}}=9\times10^{10}\,\text{\AA} \text{s}^{-1}$ using $k_{B}=8.62\times10^{-5}\,\text{eV}\,\text{K}^{-1}$, $T=298\,\text{K}$ and $m=2.5\,\text{amu}$ $\text{\AA}^{2}$. Finally, applying $\Delta(0)=3\,$ $\text{eV}$ gives $P_{bare}(0)\approx10^{-7}$, which is a negligible quantity. A more pronounced polariton-effect is expected close to the PES avoided crossing. However, the rapid decay of the transition dipole moment in this region (see Fig. \[Adia\_PESs\]a) precludes the formation of polaritonic states that could have affected the corresponding nonadiabatic dynamics. To summarize this part, even when the USC effects on the nonadiabatic dynamics are negligible for our model, the previous discussion as well as Eq. (\[eq:P\_out\]) distill the design principle that controls these processes in other polariton systems: the plexciton anticrossings should happen at large $\mathbf{k}$ values to preclude the overwhelming effects of the dark states. This principle will be explored in future work in other molecular systems.\
The negligible polariton effect on the NACs, and the magnitude of the energetic effects on the electronic energy landscape are strong evidence to argue that the chemical yields and rates of the isomerization problem in question remain intact with respect to the bare molecular ensemble.
Conclusions
===========
We showed in this work that, for the ground state landscape of a particular isomerization model, there is no relevant collective stabilization effect by USC to SPs which can significantly alter the kinetics or thermodynamics of the reaction, in contrast with previous calculations which show such possibilities in the Bogoliubov polariton landscapes [@Galego2016; @Herrera2016]. The negligible energetic corrections to the ground-state PES per molecule can be approximated and interpreted as Lamb shifts [@Bethe1950] experienced by the molecular states due to the interaction with off-resonant plasmonic modes. The key dimensionless parameter which determines the USC effect on the ground-state PES is the ratio of the individual coupling to the transition frequency $g_{\mathbf{k}}^{\mathbf{n},s}/\hbar\Delta$. This finding is similar to the conclusions of a recent work [@Galego2015a; @Cwik2016]. In particular, it is shown in [@Cwik2016] that the rotational and vibrational degrees of freedom of molecules exhibit a self-adaptation which only depends on light-matter coupling at the single-molecule level. Therefore, more remarkable effects are expected in the regime of USC of a single molecule interacting with an electric field. To date, the largest single molecule interaction energy achievable experimentally is around 90 meV [@Chikkaraddy2016] in an ultralow nanostructure volume. This coupling strength is almost two orders of magnitude larger that those in our model. Also, previous works have shown [@Niemczyk2010; @Jenkins2013] that this regime is achievable for systems with transition frequencies on the microwave range. Additionally, the experimental realization of vibrational USC has been carried out recently [@george2016multiple]. The latter also suggests the theoretical exploration of USC effects on chemical reactivity at the rotational or vibrational energy scales, where the energy spacing between levels is significantly lower than typical electronic energy gaps.\
We highlighted some intriguing quantum-coherent effects where concerted reactions can feature energetic effects that are not incoherent combinations of the bare molecular processes. These interference effects are unlikely to play an important role in reactions exhibiting high barriers compared to $k_{B}T$. However, they might be important for low-barrier processes, where the number of concerted reaction pathways becomes combinatorially more likely than single molecule processes. On the other hand, we also established that, due to the large number of dark states in these many-molecule polariton systems, nonadiabatic effects are not modified in any meaningful way under USC, at least for the model system explored. We provided a rationale behind this conclusion and discussed possibilities of seeing modifications in other systems where the excitonic and the electromagnetic modes anticross at large $\mathbf{k}$ values.
Finally, it is worth noting that even though we considered an ultrastrong coupling regime ( $\sqrt{\mathcal{N}_{\mathbf{k}}(R)}$ reaches more than $10\%$ of the maximum electronic energy gap in our model [@Moroz]), the system does not reach a Quantum Phase Transition (QPT) [@Emary; @Li2006a]. In our model, this regime would require high density ($\sim 10^{10}\,\text{molecules}$ $\mu m^{-3}$) samples, keeping $\mu\approx 2 $ $e\text{\AA}$. The implications of this QPT on chemical reactivity have not been explored in this work, but are currently being studied in our group. To conclude, our present work highlights the limitations but also possibilities of USC in the context of control of chemical reactions using polaritonic systems.\
Acknowledgments
===============
R.F.R., J.C.A., and J.Y.Z. acknowledge support from the NSF CAREER award CHE-1654732. L.A.M.M is grateful to the support of the UC-Mexus CONACyT scholarship for doctoral studies. All authors acknowledge generous startup funds from UCSD. L.A.M.M. and J.Y.Z are thankful to Prof. Felipe Herrera for useful discussions.
Appendix
========
In this appendix we outline the perturbative methodology that leads to the equations shown in the main text. Under the perturbative approach, it is convenient to express the perturbation in Eq. (\[perturbation\]) in terms of the Bogoliubov operators defined by Eq. (\[eq:bogoliubov\_quasi\]). Notice that Eq. \[perturbation\] introduces $R$-dependent exciton operators, while the zeroth order eigenstates (the polariton quasiparticles) are defined for all molecules at the configuration $R_{\mathbf{n},s}=0$. It would be useful to find a relation $b_{\mathbf{n}_{0},0}^{\dagger}(R)=f(b_{\mathbf{n}_{0},0}^{\dagger}(0),b_{\mathbf{n}_{0},0}(0))$ for any $R$, in order to carry out the aforementioned change of basis.
The function $f(b_{\mathbf{n}_{0},0}^{\dagger}(0),b_{\mathbf{n}_{0},0}(0))$ can be found by working on the diabatic basis (see Eq. \[adiabatic\_rep\]). For any operator $b_{\mathbf{n},s}^{\dagger}(R)$, using $b_{\mathbf{n},s}^{\dagger}(R)=|e_{\mathbf{n},s}(R)\rangle\langle g_{\mathbf{n},s}(R)|$, we have $$\begin{aligned}
b_{\mathbf{n},s}^{\dagger}(R) & =\left[-\sin(\theta(R)/2)|\text{trans}_{\mathbf{n},s}\rangle+\cos(\theta(R)/2)|\text{cis}_{\mathbf{n},s}\rangle\right]\nonumber \\
& \times\left[\cos(\theta(R)/2)\langle\text{trans}_{\mathbf{n},s}|+\sin(\theta(R)/2)\langle\text{cis}_{\mathbf{n},s}|\right]\nonumber \\
& =-\sin(\theta(R)/2)\cos(\theta(R)/2)b_{\mathbf{n},s}(0)b_{\mathbf{n},s}^{\dagger}(0)\label{exc_adia}\\
& -\sin(\theta(R)/2)^{2}b_{\mathbf{n},s}(0)+\cos(\theta(R)/2)^{2}b_{\mathbf{n},s}^{\dagger}(0)\nonumber \\
& +\cos(\theta(R)/2)\sin(\theta(R)/2)b_{\mathbf{n},s}^{\dagger}(0)b_{\mathbf{n},s}(0),\nonumber \end{aligned}$$ where we used $b_{\mathbf{n},s}^{\dagger}(R_{eq})=|\text{cis}_{\mathbf{n},s}\rangle\langle\text{trans}_{\mathbf{n},s}|$. In light of Eq. \[exc\_adia\] we notice that the perturbation (\[perturbation\]) in the second row produces chains with up to four exciton operators. In view of the delocalized nature of the zeroth-order eigenstates and the localized character of the exciton operators $b_{\mathbf{n},s}(0)$, we have that the matrix elements that appear in $E^{(2)}(R_{\mathbf{n}_{0},0},\mathbf{\tilde{0}}')$ are of the form $\langle G(\tilde{\mathbf{0}})|_{d}\xi_{\mathbf{k}_{1}}^{i}\xi_{\mathbf{k}_{2}}^{j}\dots\xi_{\mathbf{k}_{m+1}}^{l}F_{m}(b_{\mathbf{n}_{0},0}^{\dagger}b_{\mathbf{n}_{0},0})Z(a_{\mathbf{k}}^{\dagger},a_{\mathbf{k}})|G(\tilde{\mathbf{0}})\rangle_{d}\approx O(1/(N_{x}N_{y})^{m/2})$, where $F_{m}(b_{\mathbf{n}_{0},0}^{\dagger}b_{\mathbf{n}_{0},0})$ stands for a chain with $1\leq m\leq4$ exciton operators and $Z(a_{\mathbf{k}}^{\dagger},a_{\mathbf{k}})$ is a function of a single photonic operator. In the macroscopic limit $1\ll N_{x}N_{y}$, we can neglect chains $F_{m}(b_{\mathbf{n}_{0},0}^{\dagger}b_{\mathbf{n}_{0},0})Z(a_{\mathbf{k}}^{\dagger},a_{\mathbf{k}})$ for $m\geq2$. This leads to the simplification, $$\begin{aligned}
b_{\mathbf{n}_{0},0}^{\dagger}(R)+b_{\mathbf{n}_{0},0}(R) & \approx\left(\cos(\theta(R)/2)^{2}-\sin(\theta(R)/2)^{2}\right)\label{eq:approx_exc}\\
& \times\left(b_{\mathbf{n}_{0},0}^{\dagger}(0)+b_{\mathbf{n}_{0},0}(0)\right),\nonumber \end{aligned}$$ and the perturbation acquires the simple form, $$\begin{aligned}
V(R) & \approx\sum_{\mathbf{k},\mathbf{k}_{1}}D_{\mathbf{k}_{1}}^{\mathbf{n}_{0},0}\left(\cos(\theta(R))g_{\mathbf{k}}^{\mathbf{n}_{0},0}(R)-g_{\mathbf{k}}^{\mathbf{n}_{0},0}(0)\right)\nonumber \\
& \times\left(b_{\mathbf{k}_{1}}(0)a_{\mathbf{k}}^{\dagger}+b_{\mathbf{k}_{1}}^{\dagger}(0)a_{\mathbf{k}}^{\dagger}+a_{\mathbf{k}}b_{\mathbf{k}_{1}}(0)+b_{\mathbf{k}_{1}}^{\dagger}(0)a_{\mathbf{k}}\right).\label{eq:approx_pert}\end{aligned}$$ To write this last expression in terms of the Bogoliubov operators $\{\xi_{\mathbf{k_{1}}}^{i\dagger}(0)\xi_{\mathbf{k}_{2}}^{j\dagger}(0)\}$ we start from the transformation $\vec{\xi}=T\vec{b}$ : $$\begin{bmatrix}\xi_{\mathbf{k}}^{L}(0)\\
\xi_{\mathbf{k}}^{U}(0)\\
\xi_{\mathbf{-k}}^{L\dagger}(0)\\
\xi_{\mathbf{-k}}^{U\dagger}(0)
\end{bmatrix}=\begin{bmatrix}\alpha_{\mathbf{k}}^{L} & \beta_{\mathbf{k}}^{L} & \gamma_{\mathbf{k}}^{L} & \delta_{\mathbf{k}}^{L}\\
\alpha_{\mathbf{k}}^{U} & \beta_{\mathbf{k}}^{U} & \gamma_{\mathbf{k}}^{U} & \delta_{\mathbf{k}}^{U}\\
\gamma_{\mathbf{k}}^{L*} & \delta_{\mathbf{k}}^{L*} & \alpha_{\mathbf{k}}^{L*} & \beta_{\mathbf{k}}^{L*}\\
\gamma_{\mathbf{k}}^{U*} & \delta_{\mathbf{k}}^{U*} & \alpha_{\mathbf{k}}^{U*} & \beta_{\mathbf{k}}^{U*}
\end{bmatrix}\begin{bmatrix}a_{\mathbf{k}}\\
b_{\mathbf{k}}(0)\\
a_{\mathbf{-k}}^{\dagger}\\
b_{-\mathbf{k}}^{\dagger}(0)
\end{bmatrix}.\label{matrix_bogo}$$ From the matrix representation of the normalization $|\alpha_{\mathbf{k}}^{i}|^{2}+|\beta_{\mathbf{k}}^{i}|^{2}-|\gamma_{\mathbf{k}}^{i}|^{2}-|\delta_{\mathbf{k}}^{i}|^{2}=1$ [@Ciuti2005], it follows that, $$TI_{-}T^{\dagger}=I_{-},\label{matrix_norm}$$ where $$I_{-}=\begin{bmatrix}1 & 0 & 0 & 0\\
0 & 1 & 0 & 0\\
0 & 0 & -1 & 0\\
0 & 0 & 0 & -1
\end{bmatrix}.\label{norm}$$ We also have that $T^{-1}=I_{-}T^{\dagger}I_{-}$ and that $$T^{-1}=\begin{bmatrix}\alpha_{\mathbf{k}}^{L*} & \alpha_{\mathbf{k}}^{U*} & -\gamma_{\mathbf{k}}^{L} & -\gamma_{\mathbf{k}}^{U}\\
\beta_{\mathbf{k}}^{L*} & \beta_{\mathbf{k}}^{U*} & -\delta_{\mathbf{k}}^{L} & -\delta_{\mathbf{k}}^{U}\\
-\gamma_{\mathbf{k}}^{L*} & -\gamma_{\mathbf{k}}^{U*} & \alpha_{\mathbf{k}}^{L} & \alpha_{\mathbf{k}}^{U}\\
-\delta_{\mathbf{k}}^{L*} & -\delta_{\mathbf{k}}^{U*} & \beta_{\mathbf{k}}^{L} & \beta_{\mathbf{k}}^{U}
\end{bmatrix}.\label{eq:T-1}$$ Using Eq. \[eq:T-1\], we can readily evaluate $\vec{b}=T^{-1}\vec{\xi}$. From this relationship, the change of the localized operators to the Bogoliubov basis is accomplished. Finally, the matrix elements to compute $E^{(2)}(R_{\mathbf{n}_{0},0},\mathbf{\tilde{0}}')$ can be evaluated by means of Wick’s theorem, $$\begin{aligned}
\langle G(\tilde{\mathbf{0}})|_{d}\xi_{\mathbf{k}_{1}}^{l}\xi_{\mathbf{k}_{2}}^{n}\xi_{\mathbf{k}}^{i\dagger}\xi_{\mathbf{k}'}^{j\dagger}|G(\tilde{\mathbf{0}})\rangle_{d} & =\delta_{l,j}\delta_{\mathbf{k}_{1},\mathbf{k}'}\delta_{i,n}\delta_{\mathbf{k}_{2},\mathbf{k}}\label{wicks}\\
& +\delta_{n,j}\delta_{\mathbf{k}_{2},\mathbf{k'}}\delta_{m,i}\delta_{\mathbf{k}_{1},\mathbf{k}},\nonumber \end{aligned}$$ leading to Eq. \[coupling\].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The synchronized phase of globally coupled nonlinear oscillators subject to noise fluctuations is studied by means of a new analytical approach able to tackle general couplings, nonlinearities, and noise temporal correlations. Our results show that the interplay between coupling and noise modifies the effective frequency of the system in a non trivial way. Whereas for linear couplings the effect of noise is always to increase the effective frequency, for nonlinear couplings the noise influence is shown to be positive or negative depending on the problem parameters. Possible experimental verification of the results is discussed.'
address: |
Departamento de Matemáticas and Grupo Interdisciplinar de Sistemas Complicados,\
Universidad Carlos III de Madrid, E-28911 Leganés, Madrid, Spain
author:
- 'Esteban Moro$^*$ and Angel Sánchez$^{\dag}$'
title: Noise Effects on Synchronized Globally Coupled Oscillators
---
= 10000
Systems of coupled nonlinear oscillators are a generic paradigm of a whole class of problems arising in physics, chemistry, or biology [@Kuramoto]. Examples are Josephson Junction arrays (JJA) [@JJA], charge-density waves [@Fukuyama], thin film fabrication [@CW], chemical reactions and cardiac tissue [@Winfree], neuronal activity [@SGK], and many more (see, e.g., Ref. [@Kuramoto] and references therein). Due to the diversity of individual oscillators or to external (thermal) noise, these systems generally exhibit cooperative dynamical response or incoherent behavior as a function of the relevant parameters. The study of the transition between both regimes and the nature of the so-called [*synchronized*]{} phase requires the calculation of the phase distribution probability density. This involves solving a highly nonlinear, partial differential (Liouville) equation which is almost always a formidable task and only perturbative results can be obtained for practically all models of interest.
In this paper, we approach these problems by introducing a new method, from a completely different viewpoint, that avoids dealing with partial differential equations at all. We focus on the stochastic (Langevin) evolution equation that allows to perturbatively obtain accurate results in a very simple and direct manner. This method can be applied to any system of $N$ coupled homogeneous oscillators evolving in the presence of random forces, i.e., systems of the form $$\label{model}
\dot \phi_i = \omega - f(\phi_i) + \frac{\kappa}{N}
\sum_{j=1}^{N} \Gamma(\phi_i - \phi_j) + \sigma \eta_i(t),$$ where $\phi_i$ is the phase of oscillator $i$, $\omega > 0$ is the oscillator frequency, $\kappa$ is a constant, $f(\phi)$ is any 2$\pi$-periodic function, $\Gamma$ is any separable (see below) function, and the random term $\eta_i$ is local (uncorrelated from site to site), Gaussian and stationary. Aside from these requirements, $\eta_i$ can be any process defined by a stochastic differential equation, e.g., it can be white or colored noise. To illustrate our method, we consider an Ornstein-Uhlenbeck process, given by $\dot \eta_i =
-\gamma \eta_i + \gamma \xi_i(t)$, with $\xi_i(t)$ being uncorrelated Gaussian white noises \[$\langle \xi_i(t) \xi_j(t')\rangle = \delta_{ij}
\delta(t-t')$; $\langle \cdots \rangle$ stands for averages over noise realizations\] and $\tau = 1/\gamma$ being the correlation time.
Far from being academic, model (\[model\]) contains already several important applications. To begin with, the coupling can describe many different systems: For instance, in the case of the phase approximation of globally coupled nonlinear oscillators or of a JJA, $\Gamma(\phi_i - \phi_j) = \sin(\phi_i-\phi_j + \alpha)$, $\alpha$ being a constant, whereas in reaction-diffusion (RD) problems such as those arising in growth models [@CW], the coupling comes from the discretization of some spatial linear operator followed by a mean field approximation, yielding $\Gamma(\phi_i - \phi_j) = \phi_i -\phi_j$. Second, the nonlinear term, $f(\phi)$, for the oscillator can be chosen as required in each problem; as in most coupled oscillator models, we take here $f(\phi)=\sin\phi$, but our method can be applied to other choices as well.
We now begin the study of Eq. (\[model\]) in the case when, as mentioned above, $\Gamma$ is separable, i. e., $\Gamma(\phi_j -
\phi_i) = \sum_n A^n(\phi_i) B^n(\phi_j)$. This technicality is needed to simplify the study but does not not restrict much the applicability of our calculations. Eq. (\[model\]) becomes then $$\label{modelb}
\dot \phi_i = \omega - f(\phi_i) + \frac{\kappa}{N}
\sum_{n} A^n(\phi_i)\left[\sum_{j=1}^N B^n(\phi_j)\right] + \sigma \eta_i(t).$$ The first step is to take the limit $N\to\infty$. In this limit the sum within the brackets above can be computed in terms of the mean value of $\phi$ [@mean]. As way of example, we focus on the RD problems and JJA’s, for which we arrive (respectively) at $$\begin{aligned}
\label{lineal}
\dot \phi&=&\omega - a \sin \phi +
\kappa \left[ \langle \phi \rangle - \phi\right] + \sigma \eta(t), \\
\label{kura}
\dot \phi&=&\omega - a \sin \phi + \nonumber\\
& & + \kappa \left[
\cos (\phi -\alpha) \langle\sin\phi\rangle -
\sin (\phi -\alpha) \langle\cos\phi\rangle \right] + \sigma \eta(t).\end{aligned}$$ We note that we have thus reduced the $N$ equations in (\[model\]) to a single (self-consistent) one.
We deal first with eq. (\[lineal\]) for RD problems, hereafter called [*linear coupling model*]{}. When $\sigma=0$, it can be straightforwardly shown that if $\omega \leq \omega_C = a$ there is a stable time independent pinned solution $\phi =
\sin^{-1} \omega/a$, whereas if $\omega > \omega_C$ the phase increases oscillatorily in time with frequency $\Omega^2\equiv\omega^2 - a^2$. The natural question to ask is whether this scenario, involving a [*depinning transition*]{} at $\omega_C$, changes when noise is switched on, and if so, how. To address this issue we choose $\langle \phi \rangle$ as our order parameter, because in the pinned phase $\langle \phi
\rangle$ is constant whereas in the depinned phase it increases in time. For noise intensities which are small compared to $\omega$ or $\kappa$, we expand $\omega_C$ and $\phi$ in powers of $\sigma$, $\omega_C = a + a^{(1)}
\sigma + a^{(2)} \sigma^2 + {\cal O}(\sigma^3)$ and $\phi = \phi^{(0)} + \sigma
\phi^{(1)} + \sigma^2 \phi^{(2)} + {\cal O}(\sigma^3)$. We first look for possible changes in $\omega_C$, by inserting these two expressions in Eq. (\[lineal\]); collecting powers of $\sigma$ and imposing that $\langle\dot{\phi}^{(i)}\rangle=0$ for $i=0,1,\ldots$ (i.e., that the oscillators remained pinned), we can compute the $a^{(i)}$ corrections to $\omega_C$. The final result is $$\begin{aligned}
\frac{\omega_C}{a}&=&1 - \frac{\sigma^2}{4 \kappa (\kappa \tau + 1)} + \nonumber\\
& & +\frac{\sigma^4}{32 \kappa^4} \left[ \frac{\kappa^2 - a^2}{(\kappa \tau +1)^2}
+{\cal O}(\tau) \right] + {\cal O}(\sigma^6).\end{aligned}$$
From this expression, we immediately see first, that the transition occurs at a value of $\omega$ which is lower than in the deterministic case, which means that the noise is actually helping the oscillators to overcome the barrier and start their motion. Another important conclusion is that for a given, fixed set of the other parameters, $\tau$ controls the state of the system: whereas for low $\tau$ values, the oscillators are depinned, as $\tau$ increases the pinned phase is set on. We thus see the importance of the correlation time in the critical region of the system.
Having found the changes in $\omega_C$, we now turn to the dynamics in the depinned phase, when $\omega >
\omega_C$. We again expand $\phi$ in powers of $\sigma$ and write down equations for each contribution, which read
$$\begin{aligned}
\label{x0}
\langle \dot \phi^{(0)}\rangle&=&\omega - a \sin\langle\phi^{(0)}\rangle;\\
\label{x12}
\langle \dot \phi^{(1)}\rangle&=&0;\\
\langle (\dot\phi^{(1)})^2\rangle&=
&-2 \langle (\phi^{(1)})^2\rangle (\kappa + a \cos\langle\phi^{(0)}\rangle) + 2 \zeta(t), \\
\dot \zeta&=&
\frac{\gamma}{2} - \zeta (\kappa + \gamma + a \cos\langle\phi^{(0)}\rangle);\\
\label{x2}
\langle \dot \phi^{(2)} \rangle&=& -
a \langle \phi^{(2)} \rangle \cos\langle\phi^{(0)}\rangle
+ a \frac{\langle (\phi^{(1)})^2\rangle}{2} \sin\langle\phi^{(0)}\rangle.\end{aligned}$$
where $\zeta \equiv \langle \phi^{(1)} \eta \rangle$. Eq. (\[x2\]) implies that $\langle \phi^{(2)} \rangle$ grows in time faster than $\langle \phi^{(0)} \rangle$, and therefore our expansion is not correct. This problem, very well known in deterministic equations [@kevor], can be cured by realizing that the system dynamics does not depend only on one time scale but on two, $t$ and $t^* = \mu(\sigma)t$. Among the different approaches one can use to deal with this, we choose Linstedt’s method: we assume that $\phi^{(i)}$ depend on time through the combination $t + t^*$; as for $\mu(\sigma)$, dimensional analysis shows that the noise term is of order one when $\mu(\sigma) = \nu \sigma^2$. Introducing this new time scale in equations (\[x0\]-\[x2\]) and imposing the usual solvability condition on equation (\[x2\]) (Fredholm’s alternative, see e. g. [@kevor] for details) we obtain $$\label{nu}
\nu = \frac{a}{2T} \int_0^T
\frac{\langle \phi_1^2 \rangle \sin\langle \phi_0 \rangle}
{\omega - a \sin\langle \phi_0 \rangle} dt ,$$ where $T = 2\pi / \Omega$ is the period of $\langle\dot{\phi}^{(0)}
\rangle$ computed from eq. (\[x0\]). With this value, the effective frequency of the oscillations, defined as $\omega_{\rm eff} = \lim_{t,s\to\infty} \frac{1}{s}[\langle{\phi}(t+s)\rangle -
\langle{\phi}(t)\rangle ]$, is: $$\label{weff}
\omega_{\rm eff} = \Omega (1 + \nu \sigma^2 + {\cal O}(\sigma^4))$$
Figure \[fig1\] compares our analytical, order $\sigma^2$ expression (\[weff\]) for $\omega_{\rm eff}$ with the numerical solution of Eqs. (\[lineal\]) up to $\sigma\sim 0.5$, already a not so small value. The expression for $\omega_{\rm eff}$ above indicates that the noise increases the frequency as compared with the deterministic value, in agreement with our previous conclusion that the noise helps the oscillators jump over the potential barrier: The larger the noise strength, the easier the potential barrier to overcome, effectively suppresing (renormalizing) it in the $\sigma\to\infty$ limit. The fact that $ \nu > 0$, implying that the frequency increases with the noise strength, can be proven rigorously when $\tau=0$, i. e., for white noise, but we conjecture that this result is general for linearly coupled systems.
Another advantage of our approach is that it can be applied to other limits, such as the case when the noise intensity is larger than $\omega$ and $a$. Expanding the solution now in powers of $a$, the same steps as before yield, up to order $a^2$, $$\begin{aligned}
\label{weff1}
& & \omega_{\rm eff}=\omega - \displaystyle{a^2\over2}
\exp\left(-\frac{\tilde\sigma^2}{2\kappa}\right)\Bigg[
\frac{\kappa^2 \omega}{\kappa^2 \omega^2 + \omega^4} +
\nonumber\\
& & %\mbox{ }\!\!\!\!\!\times
\displaystyle{\int}_0^{\infty} \exp \left\{
{\frac{\tilde\sigma^2(\kappa\tau e^{-s/\tau} -
e^{-\kappa s})}{2\kappa(\kappa\tau-1)} - \kappa s}
\right\} \sin \omega s \ d s \Bigg]
%+ {\cal O}(a^4) \end{aligned}$$ where $\tilde\sigma = \sigma/\sqrt{\kappa \tau +1}$. Again, the comparison with the numerical solution (see Fig. \[fig1\]) shows that our approach is very accurate, in this case for values of $\sigma$ down to $\sigma\sim 1.2$, very close to the values of $\omega$ and $a$. We want to stress that, except for a small interval $0.5\leq \sigma\leq 1.2$, our analytical predictions (\[weff\]) and (\[weff1\]) describe to a high degree of accuracy the numerical solution for any noise intensity, providing a global picture of the main features of the linearly-coupled system behavior.
We now move on to the nonlinear problem (\[kura\]) in the case $\alpha=0$; Eqs. (\[model\]) are then known as [*Kuramoto-Sakaguchi model*]{} (KS). In the small noise regime, our technique applied to the KS model leads once again to Eqs. (\[x0\]-\[x2\]). This stems from the fact that, when $\sigma \rightarrow 0$, the phases of individual oscillators are similar, and consequently $\sin(\phi_j-\phi_i) \approx
\phi_j-\phi_i$. However, due to the periodicity of the KS coupling term, the model is invariant under changes $\phi_i-\phi_j \leftrightarrow \phi_i - \phi_j + 2 n \pi$, for any integer $n$. Therefore, $2\pi$ jumps between oscillators are possible without consequence other than the breakdown of the linear approximation. Of course, the larger the noise strength, the more likely such jumps are, and the KS model becomes desynchronized at a finite value of $\sigma$, i. e., it has a true synchronization-desynchronization transition at $\sigma^2 = \kappa$ when $a=0$ [@Kuramoto], and at lower values for $a \not= 0$ [@nota2]. Up to that point, it can be seen from Fig. \[fig2\] that the linear model, the KS model and our analytical prediction Eq. (\[weff\]) are all in excellent agreement with each other. For our parameters, $a = 1$, $\omega = 1.1$, the transition takes place at $\sigma \approx 0.5$.
The order parameter for the synchronization-desynchronization transition is $\rho = \langle \cos \phi \rangle^2 + \langle \sin \phi \rangle^2$; in the synchronized phase, $\rho$ oscillates around a nonzero value, while in the desynchronized phase, $\rho = 0$. The inset in Fig. \[fig2\] compares the numerical value of $r$ with the result of our aproximation, $\rho = 1 - \sigma^2 \langle (\phi^{(1)})^2 \rangle +
{\cal O}(\sigma^3)$, exhibiting the quantitative validity of our approach. The agreement is very good up to the critical value of $\sigma$. As for the limit $\sigma\to\infty$, the computations, although feasible, are quite involved, and therefore we do not present the results here.
In the two cases analyzed so far, in the synchronized phase noise helps the oscillators overcome the nonlinear potential, hence effectively increasing their frequency. However, this intuitively reasonable picture is not the generic situation: Antisymmetric couplings verifying $\Gamma(-\phi)
= -\Gamma(\phi)$ keep the difference between oscillators small, forcing their motion to be approximately the same, and then the above picture holds, but if the coupling lacks this symmetry (like $\Gamma(\phi_i-\phi_j+\alpha)$ for nonzero $\alpha$), the oscillators do not tend to synchronize (they rather have phases separated by a factor $\alpha$). In the latter case, the noise competes with the frustration induced by the coupling and its activation effect disappears. Our calculations for Eqs. (\[kura\]) when $\alpha\neq 0$ and $\sigma \to 0$ show that this is indeed what occurs. Our method leads now to a value of $\nu$, the $\sigma^2$ correction to the frequency, which depends on $\alpha$, given by $$\label{nualfa}
\nu_{\alpha} = \frac{1}{2T} \int_0^T \frac{\langle \phi_1^2 \rangle
(a \sin \langle \phi_0 \rangle
-2 \kappa \sin\alpha)}
{\omega_0 -a\sin\langle \phi_0 \rangle}\ dt$$ where $\omega_0 \equiv \omega + \kappa \sin \alpha$, this parameter playing the role of an effective driving. Eq. (\[nualfa\]) shows that, when $\alpha < \alpha_0 \equiv
\sin^{-1}(1/2\kappa)$, $\nu_{\alpha}$ is positive or negative depending on the value of $\omega$, whereas when $\alpha > \alpha_0$ we have $\nu_{\alpha} < 0$. This divides the $(\omega,\alpha)$-space in two regions, according to the possible signs of $\nu_\alpha$, as depicted in Fig. \[fig3\]. Interestingly, in region I $\alpha$ (or equivalently $\omega$) acts as a switching parameter between noise induced acceleration or deceleration of the oscillator motion, as illustrated in Fig. \[fig4\]. Note that this is a novel, noise intrinsic phenomenon, absent in the deterministic problem, whose origin is the nonlinear coupling: If we add a constant to the coupling in Eq. (\[lineal\]) only the external frequency is changed ($\omega \rightarrow \omega + \kappa \alpha$).
In summary, we have introduced a procedure to study systems of globally coupled oscillators in the synchronized phase, general enough to analyze any stationary random processes generated by a stochastic differential equation, any form of the nonlinearity, and a large family of coupling terms. We have used it to analyze problems relevant in a number of physical contexts, obtaining fairly good results in a much more straightforward way than the approaches proposed so far. For linear coupling, we have computed the noise-induced corrections to the critical value for the depinning transition as well as the effective frequency in the depinned phase for almost all values of the noise. We have shown that the same results hold for the KS model. In the general case of frustrated JJA’s we have found that the effect of noise on the effective frequency is not trivial, leading to an increase or a decrease of the effective frequency depending on the system parameters and the driving frequency $\omega$. This very important result shows the power of our technique as it was not known prior to our work. Our predictions can be directly checked by experiments in very many fields, among which we suggest JJA’s subject to a magnetic field where tuning the voltage for a given temperature one should be able to find the two different behaviors predicted.
We are indebted to J. A. Cuesta and R. Cuerno for discussions. Work at GISC (Leganés) was supported by CICyT (Spain) grant MAT95-0325 and DGES (Spain) grant PB96-0119.
Electronic address: [[email protected]]{} Electronic address: [[email protected]]{}
Y. Kuramoto, [*Chemical Oscillations, Waves, and Turbulence*]{} (Springer, Berlin, 1984); S. H. Strogatz, in [*Frontiers in Mathematical Biology*]{}, edited by S. Levin (Springer, Berlin, 1994).
A. V. Ustinov [*et al.*]{}, Phys. Rev. B [**47**]{}, 8357 (1993); S. Watanabe [*et al.*]{}, Physica D [**97**]{}, 429 (1996).
H. Fukuyama, J. Phys. Soc. Jpn. [**41**]{}, 513 (1976); [**45**]{}, 1474 (1978); H. Fukuyama and P.A. Lee, Phys. Rev. B [**17**]{}, 535 (1977); P. A. Lee and T. M. Rice, Phys. Rev. B [**19**]{}, 3970 (1979).
J. D. Weeks and G. H. Gilmer, Adv. Chem. Phys. [**40**]{}, 157 (1979); A. Sánchez [*et al.*]{}, Phys. Rev. B [**51**]{}, 14664 (1995); E. Moro [*et al.*]{}, Phys. Rev. Lett., 4982 (1997).
A. T. Winfree, [*When Time Breaks Down*]{} (Princeton University, Princeton, 1987).
C. W. Gardiner, [*Handbook of Stochastic Methods*]{} (Springer-Verlag, Berlin, 1996).
H. Sompolinsky [*et al.*]{}, Phys. Rev. A [**43**]{}, 6990 (1991).
When $N\to\infty$, the average of the phase $\phi_i$ over the ensemble of oscillators equals its average over realizations of the noise $\eta_i(t)$.
J. Kevorkian and J. D. Cole, [*Multiple Scale and Singular Perturbations Methods*]{} (Springer, New-York, 1996).
P. E. Kloeden and E. Platen, [*Numerical Solution of Stochastic Differential Equations*]{} (Springer, Berlin, 1992).
S. Shinomoto and Y. Kuramoto, Prog. Theor. Phys. [**75**]{} (1986), 1105.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Analyzing polytomous response from a complex survey scheme, like stratified or cluster sampling is very crucial in several socio-economics applications. We present a class of minimum quasi weighted density power divergence estimators for the polytomous logistic regression model with such a complex survey. This family of semiparametric estimators is a robust generalization of the maximum quasi weighted likelihood estimator exploiting the advantages of the popular density power divergence measure. Accordingly robust estimators for the design effects are also derived. Robust testing of general linear hypotheses on the regression coefficients are proposed using the new estimators. Their asymptotic distributions and robustness properties are theoretically studied and also empirically validated through a numerical example and an extensive Monte Carlo study.'
author:
- |
Castilla, E.$^{1}$; Ghosh, A$^{2}$; Martin, N.$^{1}$ and Pardo, L.$^{1}$\
$^{1}$[Complutense University of Madrid, Madrid, Spain]{}\
$^{2}$[Indian Statistical Institute, Kolkata, India]{}
title: '**Robust semiparametric inference for polytomous logistic regression with complex survey design**'
---
**Keywords**[: Cluster sampling; Design effect; Minimum quasi weighted DPD estimator; Polytomous logistic regression model; Pseudo minimum phi-divergence estimator; Quasi-likelihood; Robustness]{}
Introduction\[sec1\]
====================
In many real-life applications, we come across data that have been collected through a complex survey scheme, like stratified sampling or cluster sampling, etc., rather than the simple random sampling. Such situations commonly arise in large scale data collection, for example, within several states of a country or even among different countries. Suitable statistical methods are required to analyze these data by taking care of the stratified structure of the data; this is because there often exist several inter and intra-class correlations within such stratification and ignoring them often lead to erroneous inference. Further, in many such complex surveys, stratified observations are collected on some categorical responses having two or more mutually exclusive unordered categories along with some related covariates and inference about their relationship is of up-most interest for insight generation and policy making. Polytomous logistic regression (PLR) model is a useful and popular tool in such situations to model categorical responses with associated covariates. However, most of classical literature deals with the cases of simple random sampling scheme (e.g. McCullagh, 1980; Lesaffre and Albert, 1989; Agresti, 2002; Gupta et al., 2006, 2008). The application of PLR model under complex survey setting can be found, for example, in Binder (1983), Roberts et al. (1987), Morel (1989), Morel and Neerchal (2012) and Castilla et al. (2018); most of them, except the last one, are based on the quasi maximum likelihood approach.
Even though the maximum quasi weighted likelihood estimator is the base of most of the existing literature on logistic models under complex survey designs, it is known to be non-robust in the presence of possible outliers in the data. In practice, with such a complex survey design, it is quite natural to have some outlying observations that make the likelihood based inference highly unstable. So, we often may need to make additional efforts to find and discard the outliers from the data before their analyses. A robust method providing stable solution even in presence of the outliers will be really helpful and more efficient in practice. The cited work by Castilla et al (2018) has developed an alternative minimum divergence estimator based on $\phi $-divergences (Pardo, 2006), but the important issue of robustness is still ignored there.
In this paper, we develop a robust estimator under the PLR model with a complex survey based on a minimum quasi weighted divergence approach. In particular, we exploit the nice properties of the density power divergence (DPD) of Basu et al. (1998). This measure become very popular in recent literature for yielding highly robust and efficient estimators under various statistical models; see, for example, Ghosh and Basu (2013, 2016, 2018) and, in particular, the recent paper by Castilla et al (2019) which discussed a PLR model but under the simple random sampling scheme.
We first start with the mathematical description of the PLR model with complex survey set-up and a brief discussion about the maximum quasi weighted likelihood estimator of the underlying parameters in Section \[sec:MLE\]. Then we introduce a class of new robust parameter estimates for the PLR model with complex survey by minimizing a suitably defined DPD measure in Section \[sec3\]. The asymptotic distribution of the resulting estimator and the design effect for the PLR model with complex survey are also described there. In Section \[secWald\], a new family of Wald-type tests is introduced based on our new estimators for testing linear hypotheses about the parameters of the PLR model. In Section \[sec4\] we theoretically study the robustness of the proposed estimators and Wald-type tests through the influence function analysis. After presenting an illustrative example in Section \[secEx\], an extensive simulation study is presented in Section \[sec5\]. The paper ends with a brief concluding remark in Section \[secCon\]. For brevity in presentation, proofs of all the results are given in Appendix \[App\].
Maximum Quasi Weighted Likelihood Estimator \[sec:MLE\]
=======================================================
Let us assume that the whole population is partitioned into $H$ distinct strata and the data consist of $n_{h}$ clusters in stratum $h$ for each $h=1,\ldots ,H$. Further, for each cluster $i=1,\ldots ,n_{h}$ in the stratum $h$, we have observed the values of a categorical response variable ($Y$) for $m_{hi}$ units. Assuming $Y$ has $(d+1)$ categories, we denote these observed responses by a $(d+1)$-dimensional classification vector $$\boldsymbol{y}_{hij}=\left( y_{hij1},....,y_{hij,d+1}\right) ^{T},\text{ }h=1,...,H,\text{ }i=1,...,n_{h},\text{ }j=1,...,m_{hi}, \label{2.1}$$with $y_{hijr}$ $=1$ and $y_{hijl}$ $=0$ for $l\in \{1,...,d+1\}-\{r\}$ if the $j$-th unit selected from the $i$-th cluster of the $h$-th stratum falls in the $r$-th category. We also have data on $(k+1)$ explanatory variables which are common for all the individuals in the $i$-th cluster of the $h$-th stratum (very common with dummy or qualitative explanatory variables) to be denoted as $\boldsymbol{x}_{hi}=\left( x_{hi0},x_{hi1},....,x_{hik}\right) ^{T}$; the first one $x_{hi0}=1$ is associated with the intercept. Let us denote the sampling weight from the $i$-th cluster of the $h$-th stratum by $w_{hi}$. For each $i$, $h$ and $j$, the expectation of the $r$-th element of the random variable $\boldsymbol{Y}_{hij}=(Y_{hij1},...,Y_{hij,d+1})^{T}$, corresponding to the realization $\boldsymbol{y}_{hij}$, is given by the PLR model $$\pi _{hir}\left( \boldsymbol{\beta }\right) =\mathrm{E}\left[ Y_{hijr}|\boldsymbol{x}_{hi}\right] =\Pr \left( Y_{hijr}=1|\boldsymbol{x}_{hi}\right)
=\left\{
\begin{array}{ll}
\dfrac{\exp \{\boldsymbol{x}_{hi}^{T}\boldsymbol{\beta }_{r}\}}{1+{\sum_{l=1}^{d}}\exp \{\boldsymbol{x}_{hi}^{T}\boldsymbol{\beta }_{l}\}}, &
r=1,...,d, \\
\dfrac{1}{1+{\sum_{l=1}^{d}}\exp \{\boldsymbol{x}_{hi}^{T}\boldsymbol{\beta }_{l}\}}, & r=d+1,\end{array}\right. , \label{2.1.0}$$with $\boldsymbol{\beta }_{r}=\left( \beta _{r0},\beta _{r1},...,\beta
_{rk}\right) ^{T}\in \mathbb{R}^{k+1}$, $r=1,...,d$. Note that, the expectation of $\boldsymbol{Y}_{hij}$ does not depends on the unit number $j$ (homogeneity), which is not a strong assumption as we generally have random sampling with the clusters in each stratum. Let $\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) $ denote the $(d+1)$-dimensional probability vector with the elements given in (\[2.1.0\]), i.e., $$\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) =\left( \pi
_{hi1}\left( \boldsymbol{\beta }\right) ,...,\pi _{hi,d+1}\left( \boldsymbol{\beta }\right) \right) ^{T}. \label{2.2}$$Then, the associated parameter space for (\[2.1.0\]) is given by $$\Theta =\{\boldsymbol{\beta }=(\boldsymbol{\beta }_{1}^{T},...,\boldsymbol{\beta }_{d}^{T})^{T},\text{ }\boldsymbol{\beta }_{r}=\left( \beta
_{r0},...,\beta _{rk}\right) ^{T}\in \mathbb{R}^{k+1},\text{ }r=1,...,d\}=\mathbb{R}^{d(k+1)}.$$
For modeling data exhibiting overdispersion, the quasi-likelihood method is an widely used method, originally defined by Wedderburn (1974). The quasi-loglikelihood function is constructed without complete distributional knowledge but through the mean and the variance of independent sampling units, the total of clusters in all the strata $n=\sum_{h=1}^{H}n_{h}$ in the PLR model with complex sampling as given in (\[2.1.0\]). It is semiparametric method, since it only specifies the first two multivariate moments of $\boldsymbol{Y}_{hij}$. Let $f_{\boldsymbol{\beta }}(\boldsymbol{y}_{hij}|\boldsymbol{x}_{hi})$ be the probability mass function of $\boldsymbol{Y}_{hij}$ such that $\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) $ is modeled by the PLR model with complex sampling, i.e. since the support of $\boldsymbol{Y}_{hij}$, say $\mathbb{Y}_{hi}$, is the set of column vectors of identity matrix $\boldsymbol{I}_{d+1}$ it holds $$\begin{aligned}
f_{\boldsymbol{\beta }}(\boldsymbol{y}_{hij}|\boldsymbol{x}_{hi})& =\boldsymbol{\pi }_{hi}^{T}\left( \boldsymbol{\beta }\right) \boldsymbol{y}_{hij}={\sum_{l=1}^{d}}\pi _{hil}\left( \boldsymbol{\beta }\right) y_{hijl},
\\
\log f_{\boldsymbol{\beta }}(\boldsymbol{y}_{hij}|\boldsymbol{x}_{hi})&
=\log \left( \boldsymbol{\pi }_{hi}^{T}\left( \boldsymbol{\beta }\right)
\boldsymbol{y}_{hij}\right) ={\sum_{l=1}^{d}\log }\pi _{hil}\left(
\boldsymbol{\beta }\right) y_{hijl}=\log \boldsymbol{\pi }_{hi}^{T}\left(
\boldsymbol{\beta }\right) \boldsymbol{y}_{hij},\end{aligned}$$where we use the notation $\log \boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) =\left( \log \pi _{hi1}\left( \boldsymbol{\beta }\right)
,...,\log \pi _{hi,d+1}\left( \boldsymbol{\beta }\right) \right) ^{T}$. It is important to be aware that the correct loglikelihood should be $
\mathcal{\ell (}\boldsymbol{\beta })={\sum\limits_{h=1}^{H}}{\sum\limits_{i=1}^{n_{h}}}\log f_{\boldsymbol{\beta }}(\underline{\boldsymbol{y}}_{hi}|\boldsymbol{x}_{hi}),
$ where $\underline{\boldsymbol{y}}_{hi}=(\boldsymbol{y}_{hi1},...,\boldsymbol{y}_{him_{hi}})^{T}$. Under homogeneity assumption within the clusters, since $f_{\boldsymbol{\beta }}(\underline{\boldsymbol{y}}_{hi}|\boldsymbol{x}_{hi}) $ is unknown, the quasi loglikelihood, $\mathcal{\ell (}\boldsymbol{\beta })$, considers the likelihood within each cluster the same as independent case as an approximation, $$\log f_{\boldsymbol{\beta }}(\underline{\boldsymbol{y}}_{hi}|\boldsymbol{x}_{hi})\overset{def}{=}{\sum\limits_{j=1}^{m_{hi}}}\log f_{\boldsymbol{\beta }}(\boldsymbol{y}_{hij}|\boldsymbol{x}_{hi}), \label{def0}$$and thus$$\mathcal{\ell (}\boldsymbol{\beta })={\sum\limits_{h=1}^{H}}{\sum\limits_{i=1}^{n_{h}}}{\sum\limits_{j=1}^{m_{hi}}}\log \boldsymbol{\pi }_{hi}^{T}\left( \boldsymbol{\beta }\right) \boldsymbol{y}_{hij}={\sum\limits_{h=1}^{H}}{\sum\limits_{i=1}^{n_{h}}}\log \boldsymbol{\pi }_{hi}^{T}\left( \boldsymbol{\beta }\right) \widehat{\boldsymbol{y}}_{hi},$$where$$\begin{aligned}
\widehat{\boldsymbol{y}}_{hi}& =\boldsymbol{1}_{m_{hi}}^{T}\underline{\boldsymbol{y}}_{hi}=\sum\limits_{j=1}^{m_{hi}}\boldsymbol{y}_{hij},
\label{2.201} \quad
\widehat{\boldsymbol{y}}_{hi} =(\widehat{y}_{hi1},...,\widehat{y}_{hi,d+1})^{T}=\left(
\sum\limits_{j=1}^{m_{hi}}y_{hij1},...,\sum\limits_{j=1}^{m_{hi}}y_{hij,d+1}\right) ^{T}, \end{aligned}$$denotes the realization value of counts in the $i$-th cluster of the $h$-th stratum.
An important feature of the quasi loglikelihood is that the marginal distributions of $\boldsymbol{Y}_{hij}$ are completely known but the components of $\underline{\boldsymbol{Y}}_{hi}$, jointly, might be correlated. This means that distribution of their total, $\widehat{\boldsymbol{Y}}_{hi}$, might be also unknown, but the expectation is obtained as the total of the expectation of $\boldsymbol{Y}_{hij}$, $j=1,...,m_{hi}$. The most common assumption is to consider that $\widehat{\boldsymbol{Y}}_{hi}$ has a multinomial sampling scheme, which means that $\boldsymbol{Y}_{hij}$, $j=1,...,m_{hi}$ are independent random variables and$$\boldsymbol{\Sigma }_{hi}=\boldsymbol{\Sigma }_{hi}\left( \boldsymbol{\beta }\right) =m_{hi}\boldsymbol{\Delta }(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) ), \label{var1}$$where $\boldsymbol{\Delta }(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right)) =\mathrm{diag}(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) )-\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \boldsymbol{\pi }_{hi}^{T}\left( \boldsymbol{\beta }\right)$. Since (\[def0\]) is not an approximation, the term quasi should be dropped. A weaker assumption is to consider that $\widehat{\boldsymbol{Y}}_{hi}$ has a multinomial sampling scheme with a overdispersion parameter $\nu _{hi}=1+\rho _{hi}^{2}(m_{hi}-1)$, which means that $\boldsymbol{Y}_{hij}$, $j=1,...,m_{hi}$ are correlated random variables ($Cor[\boldsymbol{Y}_{hia},\boldsymbol{Y}_{hib}]=\rho _{hi}^{2}$, $a\neq b$, $a,b\in \{1,...,m_{hi}\}$) and$$\boldsymbol{\Sigma }_{hi}=\boldsymbol{\Sigma }_{hi}(\nu _{hi},\boldsymbol{\beta })=\nu _{hi}m_{hi}\boldsymbol{\Delta }(\boldsymbol{\pi }_{hi}\left(
\boldsymbol{\beta }\right) ), \label{var2}$$but the distribution of $\widehat{\boldsymbol{Y}}_{hi}$ is not in principle used for the estimators. Distributions such as Dirichlet Multinomial, Random Clumped and $m$-inflated belong this family (see Alonso et al. (2017), Morel and Neerchal (2012) and Raim et al. (2015) for details). The weakest assumption is to consider that $\widehat{\boldsymbol{Y}}_{hi}$ has an unknown distribution, with $\boldsymbol{Y}_{hij}$, $j=1,...,m_{hi}$ being possibly correlated but with no specific pattern. It is worth of mentioning that $\nu _{hi}$ plays here a role of nuisance parameter and it is possible to consider a model with additional nuisance parameters, which are more complex than (\[var2\]) but simpler than the option of completely unknown distribution for $\widehat{\boldsymbol{Y}}_{hi}$ (see Morel and Koehler (1995) for details).
Taking into account weights for each cluster, the quasi weighted loglikelihood is defined as $$\mathcal{\ell (}\boldsymbol{\beta },w)={\sum\limits_{h=1}^{H}}{\sum\limits_{i=1}^{n_{h}}}w_{hi}\log \boldsymbol{\pi }_{hi}^{T}\left(
\boldsymbol{\beta }\right) \widehat{\boldsymbol{Y}}_{hi}. \label{llike}$$Then, the maximum quasi weighted likelihood estimator of $\boldsymbol{\beta }
$, say $\widehat{\boldsymbol{\beta }}_{P}$, is obtained by maximizing the quasi weighted loglikelihood, $\mathcal{\ell (}\boldsymbol{\beta },w)$, with respect to $\boldsymbol{\beta }$. The corresponding estimating equation is then given by $${\sum\limits_{h=1}^{H}}{\sum\limits_{i=1}^{n_{h}}}w_{hi}\frac{\partial
\boldsymbol{\pi }_{hi}^{T}\left( \boldsymbol{\beta }\right) }{\partial
\boldsymbol{\beta }}\mathrm{diag}^{-1}(\boldsymbol{\pi }_{hi}\left(
\boldsymbol{\beta }\right) )\left[ \widehat{\boldsymbol{Y}}_{hi}-m_{hi}\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \right] =\boldsymbol{0}_{d(k+1)}, \label{2.4}$$with $$\begin{aligned}
\frac{\partial \boldsymbol{\pi }_{hi}^{T}\left( \boldsymbol{\beta }\right) }{\partial \boldsymbol{\beta }}& =\mathcal{\boldsymbol{\Delta }}^{\ast }(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) )\otimes \boldsymbol{x}_{hi}, \quad \boldsymbol{\Delta }^{\ast }(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }
\right) ) =\left( \boldsymbol{I}_{d},\boldsymbol{0}_{d}\right) \boldsymbol{\Delta }(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) ) .\end{aligned}$$ The system of equations (\[2.4\]) can be written as $\boldsymbol{u}\left( \boldsymbol{\beta }\right) =\boldsymbol{0}_{d(k+1)}$, where $$\begin{aligned}
\boldsymbol{u}\left( \boldsymbol{\beta }\right) &
=\sum\limits_{h=1}^{H}\sum\limits_{i=1}^{n_{h}}\boldsymbol{u}_{hi}\left(
\boldsymbol{\beta },\boldsymbol{x}_{hi}\right) , \label{Un} \quad
\boldsymbol{u}\left( \boldsymbol{\beta },\boldsymbol{x}_{hi}\right) =w_{hi}\left[ \widehat{\boldsymbol{Y}}_{hi}^{\ast }-m_{hi}\boldsymbol{\pi }_{hi}^{\ast }\left( \boldsymbol{\beta }\right) \right] \otimes \boldsymbol{x}_{hi}, \end{aligned}$$with superscript $^{\ast }$ here, the vector (matrix) obtained by deletingthe last row from the initial vector (matrix) is denoted; thus $\boldsymbol{\pi }_{hi}^{\ast }\left( \boldsymbol{\beta }\right) =\left( \pi _{hi1}\left(
\boldsymbol{\beta }\right) ,...,\pi _{hid}\left( \boldsymbol{\beta }\right)
\right) ^{T}$ and $\widehat{\boldsymbol{Y}}_{hi}^{\ast }=\left( \widehat{Y}_{hi1}^{\ast
},...,\widehat{Y}_{hid}^{\ast }\right) ^{T}$. The derivation from (\[2.4\]) to (\[Un\]) is in Appendix and some additional details can be found in Morel (1989).
The Minimum Quasi Weighted Density Power Divergence estimators \[sec3\]
=======================================================================
Let $f_{\boldsymbol{\beta }}(\boldsymbol{y}_{hij}|\boldsymbol{x}_{hi})$ be the probability mass function of $\boldsymbol{Y}_{hij}|\boldsymbol{x}_{hi}$ as defined in the previous section, $g(\boldsymbol{y}_{hij}|\boldsymbol{x}_{hi})$ an unknown and true probability mass function of $\boldsymbol{Y}_{hij}|\boldsymbol{x}_{hi}$ and $\mathbb{Y}_{hi}$ the support. The DPD based on the probability mass functions of a single observation of the sample, between $f_{\boldsymbol{\beta }}(\boldsymbol{y}_{hij}|\boldsymbol{x}_{hi})$ and $g(\boldsymbol{y}_{hij}|\boldsymbol{x}_{hi})$, is given by, for $\lambda >0$, $$\begin{aligned}
d_{\lambda }\left( g(\boldsymbol{y}_{hij}|\boldsymbol{x}_{hi}),f_{\boldsymbol{\beta }}(\boldsymbol{y}_{hij}|\boldsymbol{x}_{hi})\right) &
=\int\nolimits_{\mathbb{Y}_{hi}}\left( f_{\boldsymbol{\beta }}^{\lambda +1}(\boldsymbol{y}|\boldsymbol{x}_{hi})-\frac{\lambda +1}{\lambda }f_{\boldsymbol{\beta }}^{\lambda }(\boldsymbol{y}|\boldsymbol{x}_{hi})g(\boldsymbol{y}|\boldsymbol{x}_{hi})+\frac{1}{\lambda }g^{\lambda +1}(\boldsymbol{y}|\boldsymbol{x}_{hi})\right) d\boldsymbol{y} \\
& =\int\nolimits_{\mathbb{Y}_{hi}}f_{\boldsymbol{\beta }}^{\lambda }(\boldsymbol{y}|\boldsymbol{x}_{hi})dF(\boldsymbol{y}|\boldsymbol{x}_{hi})-\frac{\lambda +1}{\lambda }\int\nolimits_{\mathbb{Y}_{hi}}f_{\boldsymbol{\beta }}^{\lambda }(\boldsymbol{y}|\boldsymbol{x}_{hi})dG(\boldsymbol{y}|\boldsymbol{x}_{hi})+K \\
& =d_{\lambda }^{\ast }\left( g(\boldsymbol{y}_{hij}|\boldsymbol{x}_{hi}),f_{\boldsymbol{\beta }}(\boldsymbol{y}_{hij}|\boldsymbol{x}_{hi})\right) +K,\end{aligned}$$where $K$ is a constant not depending on $\boldsymbol{\beta }$, $F(\boldsymbol{y}|\boldsymbol{x}_{hi})$ and $G(\boldsymbol{y}|\boldsymbol{x}_{hi})$ are the distribution functions corresponding to the densities $f(\boldsymbol{y}|\boldsymbol{x}_{hi})$ and $g(\boldsymbol{y}|\boldsymbol{x}_{hi})$, respectively, and$$d_{\lambda }^{\ast }\left( g(\boldsymbol{y}_{hij}|\boldsymbol{x}_{hi}),f_{\boldsymbol{\beta }}(\boldsymbol{y}_{hij}|\boldsymbol{x}_{hi})\right) =E[f_{\boldsymbol{\beta }}^{\lambda }(\boldsymbol{Y}_{hij}|\boldsymbol{x}_{hi})]-\frac{\lambda +1}{\lambda }\int\nolimits_{\mathbb{Y}_{hi}}f_{\boldsymbol{\beta }}^{\lambda }(\boldsymbol{y}|\boldsymbol{x}_{hi})dG(\boldsymbol{y}|\boldsymbol{x}_{hi})$$is the kernel of $d_{\lambda }\left( g(\boldsymbol{y}_{hij}|\boldsymbol{x}_{hi})\text{, }f_{\boldsymbol{\beta }}(\boldsymbol{y}_{hij}|\boldsymbol{x}_{hi})\right) $. In practice, since $G$ is unknown, it must be estimated from the sample, which is in this case a single individual, so$$d_{\lambda }^{\ast }\left( \widehat{g}(\boldsymbol{y}_{hij}|\boldsymbol{x}_{hi}),f_{\boldsymbol{\beta }}(\boldsymbol{y}_{hij}|\boldsymbol{x}_{hi})\right) =E[f_{\boldsymbol{\beta }}^{\lambda }(\boldsymbol{Y}_{hij}|\boldsymbol{x}_{hi})]-\frac{\lambda +1}{\lambda }f_{\boldsymbol{\beta }}^{\lambda }(\boldsymbol{y}_{hij}|\boldsymbol{x}_{hi}).$$Based on Ghosh and Basu (2013), the kernel of the ordinary DPD between the probability mass functions $\widehat{g}(\boldsymbol{y}_{hij}|\boldsymbol{x}_{hi})$ and $f_{\boldsymbol{\beta }}(\boldsymbol{y}_{hij}|\boldsymbol{x}_{hi})$ for the whole sample is defined as a total discrepancy given by $$d_{\lambda }^{\ast }(\widehat{g},f_{\boldsymbol{\beta }})=\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}\sum_{j=1}^{m_{hi}}\left( E\left[ f_{\boldsymbol{\beta }}^{\lambda }(\boldsymbol{Y}_{hij}|\boldsymbol{x}_{hi})\right] -\frac{\lambda
+1}{\lambda }f_{\boldsymbol{\beta }}^{\lambda }(\boldsymbol{y}_{hij}|\boldsymbol{x}_{hi})\right) ,\quad \text{for }\lambda >0.$$Since the support of $\boldsymbol{Y}_{hij}|\boldsymbol{x}_{hi}$, is the set of column vectors of identity matrix $\boldsymbol{I}_{d+1}$, it holds $$E\left[ f_{\boldsymbol{\beta }}^{\lambda }(\boldsymbol{Y}_{hij}|\boldsymbol{x}_{hi})\right] =E\left[ \boldsymbol{\pi }_{hi}^{\lambda ,T}\left(
\boldsymbol{\beta }\right) \boldsymbol{Y}_{hij}\right] =\sum_{l=1}^{d+1}\pi
_{hil}^{\lambda +1}\left( \boldsymbol{\beta }\right) .$$
In the current paper it is defined for the first time the kernel of the quasi weighted DPD between the probability mass functions $\widehat{g}(\boldsymbol{y}_{hij}|\boldsymbol{x}_{hi})$ and $f_{\boldsymbol{\beta }}(\boldsymbol{y}_{hij}|\boldsymbol{x}_{hi})$ in the whole sample as a weighted sum$$d_{\lambda }^{\ast }(\widehat{g},f_{\boldsymbol{\beta }},w)=\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}\sum_{j=1}^{m_{hi}}w_{hi}\left( E\left[ f_{\boldsymbol{\beta }}^{\lambda }(\boldsymbol{Y}_{hij}|\boldsymbol{x}_{hi})\right] -\frac{\lambda +1}{\lambda }f_{\boldsymbol{\beta }}^{\lambda }(\boldsymbol{y}_{hij}|\boldsymbol{x}_{hi})\right) ,\quad \text{for }\lambda >0,$$whose expression for the PLR model with complex sampling is$$d_{\lambda }^{\ast }(\widehat{g},f_{\boldsymbol{\beta }},w)=\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}w_{hi}\boldsymbol{\pi }_{hi}^{\lambda ,T}\left(
\boldsymbol{\beta }\right) \left( m_{hi}\boldsymbol{\pi }_{hi}\left(
\boldsymbol{\beta }\right) -\frac{\lambda +1}{\lambda }\widehat{\boldsymbol{y}}_{hi}\right) ,\quad \text{for }\lambda >0. \label{def}$$
Based on $d_{\lambda }^{\ast }(\widehat{g},f_{\boldsymbol{\beta }},w)$ the minimum quasi weighted DPD estimator is formally defined as follows.
The minimum quasi weighted DPD estimator, say $\widehat{\boldsymbol{\beta }}_{\lambda ,Q}$, of $\boldsymbol{\beta }$ is defined as$$\widehat{\boldsymbol{\beta }}_{\lambda ,Q}=\arg \min_{\boldsymbol{\beta }\in
\mathbb{R}
^{d(k+1)}}d_{\lambda }^{\ast }(\widehat{g},f_{\boldsymbol{\beta }},w).$$
At the particular choice $\lambda\rightarrow 0$, the DPD measure, defined as the limit of (\[def\]) coincides (in limit) with the weighted quasi-loglikelihood, $\mathcal{\ell (}\boldsymbol{\beta },w)$, given in (\[llike\]); thus the minimum quasi weighted DPD estimator of $\boldsymbol{\beta }$ at $\lambda =0$ coincides with the maximum weighted quasi likelihood estimator. With the same philosophy, the following result generalizes $\boldsymbol{u}(\boldsymbol{\beta },\boldsymbol{x}_{hi})$ given in (\[Un\]) which plays an important role for the derivation of the asymptotic distribution of $\widehat{\boldsymbol{\beta}}_{\lambda, Q}$.
\[th:u\] The minimum quasi weighted DPD estimate of $\boldsymbol{\beta }$, say $\widehat{\boldsymbol{\beta }}_{\lambda ,Q}$, can be obtained by solving the system of equations $\boldsymbol{u}_{\lambda }(\boldsymbol{\beta
})=\boldsymbol{0}_{d(k+1)}$, where $$\begin{aligned}
\boldsymbol{u}_{\lambda }(\boldsymbol{\beta })&
=\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}\boldsymbol{u}_{\lambda }(\boldsymbol{\beta
},\boldsymbol{x}_{hi}), \label{u0} \\
\boldsymbol{u}_{\lambda }(\boldsymbol{\beta },\boldsymbol{x}_{hi})& =\left[
w_{hi}\boldsymbol{\Delta }^{\ast }(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) )\mathrm{diag}^{\lambda -1}\{\boldsymbol{\pi }_{hi}\left(
\boldsymbol{\beta }\right) \}\{\widehat{\boldsymbol{y}}_{hi}-m_{hi}\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \}\right] \otimes
\boldsymbol{x}_{hi}. \label{u}\end{aligned}$$
Let $\boldsymbol{U}_{\lambda }(\boldsymbol{\beta },\boldsymbol{X})$ be a random variable generator of (\[u\]) associated with a generic random explanatory variable $\boldsymbol{X}$, with no stratum and cluster assignment. In what is to follow, $$\boldsymbol{U}_{\lambda }(\boldsymbol{\beta },\boldsymbol{x}_{hi})=\left[
w_{hi}\boldsymbol{\Delta }^{\ast }(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) )\mathrm{diag}^{\lambda -1}\{\boldsymbol{\pi }_{hi}\left(
\boldsymbol{\beta }\right) \}\{\widehat{\boldsymbol{Y}}_{hi}-m_{hi}\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \}\right] \otimes
\boldsymbol{x}_{hi}$$denotes $\boldsymbol{U}_{\lambda }(\boldsymbol{\beta },\boldsymbol{X})|\boldsymbol{X=x}_{hi}$. An important property is that the estimating equation given in Theorem \[th:u\] is unbiased. This property is very important to consider these estimators similar to the maximum quasi weighted likelihood ones in the construction of the asymptotic properties, which was not the case for other distance based estimators such as the one proposed in Castilla et al (2018).
Asymptotic distribution and estimates of the design effect
----------------------------------------------------------
The following results are generalized for the PLR model with complex sampling and random explanatory variables. Without any loss of generality it is assumed that $\widehat{\Pr }(\boldsymbol{X=x}_{hi})=\frac{1}{n}$ is estimated from the sample of strata in order to get assymptotic results for the fixed explanatory variables (as a particular case of random explanatory variables).
\[Th1\] Let $\widehat{\boldsymbol{\beta }}_{\lambda ,Q}$ the minimum quasi weighted DPD estimate of $\boldsymbol{\beta }$ in the PLR model ([2.1.0]{}) under a complex survey. Then we have$$\sqrt{n}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q}-\boldsymbol{\beta }_{0})\overset{\mathcal{L}}{\underset{n\mathcal{\rightarrow }\infty }{\longrightarrow }}\mathcal{N}\left( \boldsymbol{0}_{d(k+1)},\mathbf{\Psi}_{\lambda }^{-1}\left( \boldsymbol{\beta }_{0}\right) \mathbf{\Omega}_{\lambda
}\left( \boldsymbol{\beta }_{0}\right) \mathbf{\Psi}_{\lambda }^{-1}\left(
\boldsymbol{\beta }_{0}\right) \right) ,$$where $\boldsymbol{\beta }_{0}$ is the true parameter value and $$\begin{aligned}
\mathbf{\Omega}_{\lambda }\left( \boldsymbol{\beta }\right) & =\lim_{n\rightarrow
\infty }\mathbf{\Omega}_{n,\lambda }\left( \boldsymbol{\beta }\right)
=\lim_{n\rightarrow \infty }\frac{1}{n}\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}\mathbf{\Omega}_{hi,\lambda }\left( \boldsymbol{\beta },\boldsymbol{\Sigma }_{hi}\right) , \label{hs} \\
\mathbf{\Psi}_{\lambda }\left( \boldsymbol{\beta }\right) & =\lim_{n\rightarrow
\infty }\mathbf{\Psi}_{n,\lambda }\left( \boldsymbol{\beta }\right)
=\lim_{n\rightarrow \infty }\frac{1}{n}\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}\mathbf{\Psi}_{hi,\lambda }\left( \boldsymbol{\beta }\right) , \label{gs}\end{aligned}$$where$$\begin{aligned}
\mathbf{\Omega}_{hi,\lambda }\left( \boldsymbol{\beta },\boldsymbol{\Sigma }_{hi}\right) & =w_{hi}^{2}\boldsymbol{\Delta }^{\ast }(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) )\mathrm{diag}^{\lambda -1}\{\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \}\boldsymbol{\Sigma
}_{hi}\mathrm{diag}^{\lambda -1}\{\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \}\boldsymbol{\Delta }^{\ast T}(\boldsymbol{\pi }_{hi}\left(
\boldsymbol{\beta }\right) )\otimes \boldsymbol{x}_{hi}\boldsymbol{x}_{hi}^{T}, \label{hhi} \\
\boldsymbol{\Sigma }_{hi}& =\mathrm{Var}[\widehat{\boldsymbol{Y}}_{hi}],
\notag \\
\mathbf{\Psi}_{hi,\lambda }\left( \boldsymbol{\beta }\right) & =\left\{
\begin{array}{lc}
w_{hi}m_{hi}\boldsymbol{\Delta }^{\ast }(\boldsymbol{\pi }_{hi}\left(
\boldsymbol{\beta }\right) )\mathrm{diag}^{\lambda -1}\{\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \}\boldsymbol{\Delta }^{\ast T}(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) )\otimes \boldsymbol{x}_{hi}\boldsymbol{x}_{hi}^{T}, & \lambda >0 \\
w_{hi}m_{hi}\boldsymbol{\Delta }(\boldsymbol{\pi }_{hi}^{\ast }\left(
\boldsymbol{\beta }\right) )\otimes \boldsymbol{x}_{hi}\boldsymbol{x}_{hi}^{T}, & \lambda =0\end{array}\right. . \label{ghi}\end{aligned}$$
Notice that the expression of $\mathbf{\Psi}_{\lambda =0}\left( \boldsymbol{\beta }\right) $ is the same as the so called Fisher information matrix for multinomial sampling. In addition, if $\widehat{\boldsymbol{Y}}_{hi}$ has a multinomial sampling scheme $\mathbf{\Omega}_{\lambda =0}\left( \boldsymbol{\beta
}\right) =\mathbf{\Psi}_{\lambda =0}\left( \boldsymbol{\beta }\right) $ and $\mathbf{\Omega}_{\lambda =0}^{-1}\left( \boldsymbol{\beta }\right) \mathbf{\Psi}_{\lambda =0}\left( \boldsymbol{\beta }\right) \mathbf{\Omega}_{\lambda
=0}^{-1}\left( \boldsymbol{\beta }\right) $ is the inverse of the Fisher information matrix.
Consistency is usually considered as a minimal requirement for an inference procedure. The following result is useful as a tool for estimating $\mathbf{\Omega}_{\lambda }\left( \boldsymbol{\beta }\right) $ and $\mathbf{\Psi}_{\lambda }\left( \boldsymbol{\beta }\right) $ consistently plugging a consistent estimator into $\boldsymbol{\beta }$, and hence also $\mathbf{\Omega}_{\lambda }^{-1}\left( \boldsymbol{\beta }\right) \mathbf{\Psi}_{\lambda }\left( \boldsymbol{\beta }\right) \mathbf{\Omega}_{\lambda
}^{-1}\left( \boldsymbol{\beta }\right) $ again as a (double) plug-in estimator.
\[cor1\]The following ones are (weak) consistent estimators as $n$ goes to infinity:
a)
: $\widehat{\boldsymbol{\beta }}_{\lambda ,Q}$ is a consistent estimator of the true regression coefficient $\boldsymbol{\beta }_{0}$.
b)
: $\mathbf{\Psi}_{n,\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda
,Q})=\frac{1}{n}\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}\mathbf{\Psi}_{hi,\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})$ is a consistent estimator of $\mathbf{\Psi}_{\lambda }\left( \boldsymbol{\beta }_{0}\right) $.
c)
: $\mathbf{\Omega}_{n,\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda
,Q},\{\widehat{\boldsymbol{\Sigma }}_{hi}\}_{h=1,...,H;i=1,...,n_{h}})=\frac{1}{n}\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}\mathbf{\Omega}_{hi,\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda ,Q},\widehat{\boldsymbol{\Sigma }}_{hi})$ with$$\begin{aligned}
\mathbf{\Omega}_{hi,\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda ,Q},\widehat{\boldsymbol{\Sigma }}_{hi})& =w_{hi}^{2}\boldsymbol{\Delta }^{\ast }(\boldsymbol{\pi }_{hi}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q}))\mathrm{diag}^{\lambda -1}\{\boldsymbol{\pi }_{hi}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})\}\widehat{\boldsymbol{\Sigma }}_{hi} \notag \\
& \times \mathrm{diag}^{\lambda -1}\{\boldsymbol{\pi }_{hi}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})\}\boldsymbol{\Delta }^{\ast T}(\boldsymbol{\pi }_{hi}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q}))\otimes
\boldsymbol{x}_{hi}\boldsymbol{x}_{hi}^{T}, \label{Hhi}\end{aligned}$$is a consistent estimator of $\mathbf{\Omega}_{\lambda }\left( \boldsymbol{\beta }_{0}\right) $, whenever $\Sigma _{hi}=\mathrm{Var}[\widehat{\boldsymbol{Y}}_{hi}]$ is consistently estimated through $\widehat{\boldsymbol{\Sigma }}_{hi}$ for all $(h,i)\in \{1,...,H\}\times \{1,...,m_{hi}\}$.
The following two cases of $\mathbf{\Omega}_{n,\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda ,Q},\{\widehat{\boldsymbol{\Sigma }}_{hi}\}_{h=1,...,H;i=1,...,n_{h}})$ are taken into account for two cases of $\boldsymbol{\Sigma }_{hi}$ for which there exists a consistent estimator:
- if $\widehat{\boldsymbol{Y}}_{hi}$ has (an ordinary) multinomial sampling scheme then$$\mathbf{\Omega}_{n,\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})=\frac{1}{n}\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}\mathbf{\Omega}_{hi,\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})$$where $\mathbf{\Omega}_{hi,\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})$ is $\mathbf{\Omega}_{hi,\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda ,Q},\widehat{\boldsymbol{\Sigma }}_{hi})$ given in (\[Hhi\]) with $$\widehat{\boldsymbol{\Sigma }}_{hi}=\boldsymbol{\Sigma }_{hi}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})=m_{hi}\boldsymbol{\Delta }(\boldsymbol{\pi }_{hi}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q}));$$
- if $\widehat{\boldsymbol{Y}}_{hi}$ has an overdispersed multinomial sampling scheme then$$\mathbf{\Omega}_{n,\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda ,Q},\{\widetilde{\nu }_{hi}\}_{h=1,...,H;i=1,...,n_{h}})=\frac{1}{n}\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}\mathbf{\Omega}_{hi,\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda ,Q},\widetilde{\nu }_{hi}),$$where $\widetilde{\nu }_{hi}$ is a consistent estimator of the overdispersion parameter $\nu _{hi}$ which is established later in Corollary \[cor2\], and $\mathbf{\Omega}_{n,\lambda }(\widetilde{\nu }_{hi},\widehat{\boldsymbol{\beta }}_{\lambda ,Q})$ is $\mathbf{\Omega}_{hi,\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda ,Q},\widehat{\boldsymbol{\Sigma }}_{hi})$ given in (\[Hhi\]) with $$\widehat{\boldsymbol{\Sigma }}_{hi}=\boldsymbol{\Sigma }_{hi}(\widetilde{\nu
}_{hi},\widehat{\boldsymbol{\beta }}_{\lambda ,Q})=\widetilde{\nu }_{hi}m_{hi}\boldsymbol{\Delta }(\boldsymbol{\pi }_{hi}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q}));$$
If $\widehat{\boldsymbol{Y}}_{hi}$ has a multinomial sampling scheme, Theorem \[Th1\] can be similarly formulated taking the assumption that $\widehat{\Pr }(\boldsymbol{X=x}_{hi})=\frac{1}{m}$, where $m=\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}m_{hi}$, for each individual of all the clusters in the sample, and taking for the estimation equation $m$ summands rather than $n$, i.e. plugging ${\sum\limits_{j=1}^{m_{hi}}}\boldsymbol{y}_{hij}=\widehat{\boldsymbol{y}}_{hi}$ into (\[u0\]) and considering the system of equations $\boldsymbol{u}_{\lambda }(\boldsymbol{\beta })=\boldsymbol{0}_{d(k+1)}$, where$$\begin{aligned}
\boldsymbol{u}_{\lambda }(\boldsymbol{\beta })&
=\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}{\sum\limits_{j=1}^{m_{hi}}}\boldsymbol{v}_{j,\lambda }(\boldsymbol{\beta },\boldsymbol{x}_{hi}), \\
\boldsymbol{v}_{j,\lambda }(\boldsymbol{\beta },\boldsymbol{x}_{hi})& =\left[
w_{hi}\boldsymbol{\Delta }^{\ast }(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) )\mathrm{diag}^{\lambda -1}\{\boldsymbol{\pi }_{hi}\left(
\boldsymbol{\beta }\right) \}\{\boldsymbol{y}_{hij}-\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \}\right] \otimes \boldsymbol{x}_{hi}.\end{aligned}$$Hence, it holds$$\sqrt{m}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q}-\boldsymbol{\beta }_{0})\overset{\mathcal{L}}{\underset{m\mathcal{\rightarrow }\infty }{\longrightarrow }}\mathcal{N}\left( \boldsymbol{0}_{d(k+1)},\mathbf{\Psi}_{\lambda }^{-1}\left( \boldsymbol{\beta }_{0}\right) \mathbf{\Omega}_{\lambda
}\left( \boldsymbol{\beta }_{0}\right) \mathbf{\Psi}_{\lambda }^{-1}\left(
\boldsymbol{\beta }_{0}\right) \right) ,$$with$$\begin{aligned}
\mathbf{\Omega}_{\lambda }\left( \boldsymbol{\beta }\right) & =\lim_{m\rightarrow
\infty }\left\{
\begin{array}{lc}
\begin{array}{l}
\frac{1}{m}\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}{m}_{hi}w_{hi}^{2}\boldsymbol{\Delta }^{\ast }(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) )\mathrm{diag}^{\lambda -1}\{\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \}\boldsymbol{\Delta }(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) ) \\
\times \mathrm{diag}^{\lambda -1}\{\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \}\boldsymbol{\Delta }^{\ast T}(\boldsymbol{\pi }_{hi}\left(
\boldsymbol{\beta }\right) )\otimes \boldsymbol{x}_{hi}\boldsymbol{x}_{hi}^{T},\end{array}
& \lambda >0 \\
\frac{1}{m}\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}w_{hi}^{2}m_{hi}\boldsymbol{\Delta }(\boldsymbol{\pi }_{hi}^{\ast }\left( \boldsymbol{\beta }\right)
)\otimes \boldsymbol{x}_{hi}\boldsymbol{x}_{hi}^{T}, & \lambda =0\end{array}\right. \\
\mathbf{\Psi}_{\lambda }\left( \boldsymbol{\beta }\right) & =\lim_{m\rightarrow
\infty }\left\{
\begin{array}{lc}
\frac{1}{m}\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}w_{hi}m_{hi}\boldsymbol{\Delta }^{\ast }(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) )\mathrm{diag}^{\lambda -1}\{\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right)
\}\boldsymbol{\Delta }^{\ast T}(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) )\otimes \boldsymbol{x}_{hi}\boldsymbol{x}_{hi}^{T}, &
\lambda >0 \\
\frac{1}{m}\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}w_{hi}m_{hi}\boldsymbol{\Delta }(\boldsymbol{\pi }_{hi}^{\ast }\left( \boldsymbol{\beta }\right) )\otimes
\boldsymbol{x}_{hi}\boldsymbol{x}_{hi}^{T}, & \lambda =0\end{array}\right. .\end{aligned}$$The formal proof is omitted, but the derivation of the expressions is almost the same considering $\boldsymbol{U}_{\lambda }(\boldsymbol{\beta },\boldsymbol{x}_{hi})={\sum\limits_{j=1}^{m_{hi}}}\boldsymbol{V}_{j,\lambda }(\boldsymbol{\beta },\boldsymbol{x}_{hi})$, with $\boldsymbol{V}_{j,\lambda }(\boldsymbol{\beta },\boldsymbol{x}_{hi})$ i.i.d. random variables $j=1,...,m_{hi}$. This idea matches the philosophy of the asymptotic result developed in Castilla et al. (2019), where for $H=1$ and $w_{1i}=1$, $i=1,...,m_{1i}$.
The following result is useful for any sample of polytomous logistic regression with complex sample design, more general in comparison with Corollary \[cor1\], since it is not necessary to get any consistent estimators for $\boldsymbol{\Sigma }_{hi}$.
\[Th1b\]The estimator $\widehat{\mathbf{\Omega}}_{n,\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda
,Q}) =\frac{1}{n}\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}\widehat{\mathbf{\Omega}}_{hi,\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})$, with $$\begin{aligned}
\widehat{\mathbf{\Omega}}_{hi,\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda
,Q})& =\boldsymbol{U}_{\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda ,Q},\boldsymbol{x}_{hi})\boldsymbol{U}_{\lambda }^{T}(\widehat{\boldsymbol{\beta
}}_{\lambda ,Q},\boldsymbol{x}_{hi}) \\
& =\left[ w_{hi}^{2}\boldsymbol{\Delta }^{\ast }(\boldsymbol{\pi }_{hi}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q}))\mathrm{diag}^{\lambda -1}\{\boldsymbol{\pi }_{hi}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})\}\{\widehat{\boldsymbol{Y}}_{hi}-m_{hi}\boldsymbol{\pi }_{hi}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})\}\right. \\
& \left. \times \{\widehat{\boldsymbol{Y}}_{hi}-m_{hi}\boldsymbol{\pi }_{hi}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})\}^{T}\mathrm{diag}^{\lambda -1}\{\boldsymbol{\pi }_{hi}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})\}\boldsymbol{\Delta }^{\ast T}(\boldsymbol{\pi }_{hi}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q}))\right] \otimes \boldsymbol{x}_{hi}\boldsymbol{x}_{hi}^{T},\end{aligned}$$is consistent for $\mathbf{\Omega}_{\lambda }\left( \boldsymbol{\beta }\right)$ as $n$ goes to infinity.
\[cor2\]Let $\widehat{\boldsymbol{Y}}_{hi}$ be a random variable with overdisepersed multinomial sampling scheme with a common overdispersion parameter $\nu $ and $m_{hi}=\overline{m}$,$$\begin{aligned}
\boldsymbol{\Sigma }_{hi}& =\boldsymbol{\Sigma }_{hi}\left( \nu ,\boldsymbol{\beta }\right) =\nu \overline{m}\boldsymbol{\Delta }(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) ), \\
\nu & =1+\rho ^{2}(\overline{m}-1),\end{aligned}$$then, for $\nu $ and $\rho ^{2}$:
a)
: robust and consistent estimators based on the estimating equation are given respectively by$$\begin{aligned}
\widetilde{\nu }_{n,\lambda }^{E}=\widetilde{\nu }_{n,\lambda }^{E}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})& =\frac{1}{d(k+1)}\mathrm{trace}\left(
\mathbf{\Omega}_{n,\lambda }^{-1}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})\widehat{\mathbf{\Omega}}_{n,\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda
,Q})\right) , \label{vE} \\
\widetilde{\rho }_{n,\lambda }^{2,E}& =\frac{\widetilde{\nu }_{n,\lambda
}^{E}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})-1}{\overline{m}-1}, \notag\end{aligned}$$where the matrices of interest are as follows, the one associated with multinomial sampling$$\begin{aligned}
\mathbf{\Omega}_{n,\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})& =\frac{1}{n}\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}\mathbf{\Omega}_{hi,\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda ,Q}), \\
\mathbf{\Omega}_{hi,\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})& =\overline{m}w_{hi}^{2}\boldsymbol{\Delta }^{\ast }(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) )\mathrm{diag}^{\lambda -1}\{\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \}\boldsymbol{\Delta
}(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) ) \\
& \times \mathrm{diag}^{\lambda -1}\{\boldsymbol{\pi }_{hi}\left(
\boldsymbol{\beta }\right) \}\boldsymbol{\Delta }^{\ast T}(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) )\otimes \boldsymbol{x}_{hi}\boldsymbol{x}_{hi}^{T},\end{aligned}$$and Theorem \[Th1b\] for overdispersed multinomial sampling.
b)
: robust and consistent estimators based on the method of moments are given by$$\begin{aligned}
\widetilde{\nu }_{n,\lambda }^{M}& =\widetilde{\nu }_{n,\lambda }^{M}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})=\frac{1}{nd}\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}\sum_{j=1}^{d+1}\frac{(\widehat{Y}_{hij}-\overline{m}\pi
_{hij}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q}))^{2}}{\overline{m}\pi
_{hij}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})}, \label{vM} \\
\widetilde{\rho }_{n,\lambda }^{2,M}& =\frac{\widetilde{\nu }_{n,\lambda
}^{M}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})-1}{\overline{m}-1}. \notag\end{aligned}$$
The preceding results were established for $n=\sum\nolimits_{h=1}^{H}n_{h}$ tending to infinity which implies in practice that $H$ is fixed and there exists $\eta _{h}=\lim_{n\rightarrow \infty }\frac{n_{h}}{n}\in (0,1)$, $\sum\nolimits_{h=1}^{H}\eta _{h}=1$. This means that it is approriate to consider in the place of Theorem \[Th1\], the following one, assuming that $\widehat{\Pr }(\boldsymbol{X}_{h}\boldsymbol{=x}_{hi})=\frac{1}{n_{h}}$.
\[th2\] Let $\widehat{\boldsymbol{\beta }}_{\lambda ,Q}$ the quasi MDPDE of $\boldsymbol{\beta }$ in the PLR model (\[2.1.0\]) under a complex survey. Then we have$$\sqrt{n}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q}-\boldsymbol{\beta }_{0})\overset{\mathcal{L}}{\underset{n\mathcal{\rightarrow }\infty }{\longrightarrow }}\mathcal{N}\left( \boldsymbol{0}_{d(k+1)},\mathbf{\Psi}_{\lambda }^{-1}\left( \boldsymbol{\beta }_{0}\right) \mathbf{\Omega}_{\lambda
}\left( \boldsymbol{\beta }_{0}\right) \mathbf{\Psi}_{\lambda }^{-1}\left(
\boldsymbol{\beta }_{0}\right) \right) ,$$where $\boldsymbol{\beta }_{0}$ is the true parameter value and$$\begin{aligned}
\mathbf{\Omega}_{\lambda }\left( \boldsymbol{\beta }\right) & =\sum_{h=1}^{H}\eta
_{h}\mathbf{\Omega}_{\lambda }^{(h)}\left( \boldsymbol{\beta }\right)
=\sum_{h=1}^{H}\eta _{h}\lim_{n_{h}\rightarrow \infty }\frac{1}{n_{h}}\sum_{i=1}^{n_{h}}\mathbf{\Omega}_{i,\lambda }^{(h)}\left( \boldsymbol{\beta },\boldsymbol{\Sigma }_{hi}\right) , \\
\mathbf{\Psi}_{\lambda }\left( \boldsymbol{\beta }\right) & =\sum_{h=1}^{H}\eta
_{h}\mathbf{\Psi}_{\lambda }^{(h)}\left( \boldsymbol{\beta }\right)
=\sum_{h=1}^{H}\eta _{h}\lim_{n_{h}\rightarrow \infty }\frac{1}{n_{h}}\sum_{i=1}^{n_{h}}\mathbf{\Psi}_{i,\lambda }^{(h)}\left( \boldsymbol{\beta }\right) ,\end{aligned}$$where $\mathbf{\Omega}_{i,\lambda }^{(h)}\left( \boldsymbol{\beta },\boldsymbol{\Sigma }_{hi}\right) $ and $\mathbf{\Psi}_{i,\lambda }^{(h)}\left( \boldsymbol{\beta }\right) $ have the same expressions (not meaning) as $\mathbf{\Omega}_{hi,\lambda }\left( \boldsymbol{\beta },\boldsymbol{\Sigma }_{hi}\right) $and $\mathbf{\Psi}_{hi,\lambda }\left( \boldsymbol{\beta }\right) $respectively, given in (\[hhi\]), (\[ghi\]).
The idea of the previous theorem matches the philosophy of the asymptotic result developed in Castilla et al. (2018), where other family of divergences was considered and the derivations of the results were quite different. It is also appropriate to consider the following results in the place of Corollary \[cor1\], b and c.
\[cor1b\]The following ones are (weak) consistent estimators as $n_{h}$ goes to infinity, for every $h=1,...,H$:
a)
: $\mathbf{\Psi}_{\lambda }^{(h)}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})=\frac{1}{n_{h}}\sum_{i=1}^{n_{h}}\mathbf{\Psi}_{i,\lambda
}^{(h)}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})$ is a consistent estimator of $\mathbf{\Psi}_{\lambda }^{(h)}\left( \boldsymbol{\beta }_{0}\right) $, for every $h=1,...,H$.
b)
: $\mathbf{\Omega}_{\lambda }^{(h)}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q},\{\widehat{\boldsymbol{\Sigma }}_{hi}\}_{h=1,...,H:i=1,...,n_{h}})=\frac{1}{n_{h}}\sum_{i=1}^{n_{h}}\mathbf{\Omega}_{i,\lambda }^{(h)}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q},\widehat{\boldsymbol{\Sigma }}_{hi})$ is a consistent estimator of $\mathbf{\Omega}_{\lambda }^{(h)}\left( \boldsymbol{\beta }_{0}\right) $, for every $h=1,...,H$, whenever $\Sigma _{hi}=\mathrm{Var}[\widehat{\boldsymbol{Y}}_{hi}]$ is consistently estimated through $\widehat{\boldsymbol{\Sigma }}_{hi}$ for all $(h,i)\in \{1,...,H\}\times \{1,...,m_{hi}\}$.
The following result is useful for any sample of polytomous logistic regression with complex sample design, more general in comparison with Corollary \[cor1b\], since it is not neccessary to get any consistent estimators for $\boldsymbol{\Sigma }_{hi}$.
\[Th2b\]The estimator$$\begin{aligned}
\widehat{\mathbf{\Omega}}_{\lambda }^{(h)}(\widehat{\boldsymbol{\beta }}_{\lambda
,Q})& =\frac{1}{n_{h}}\sum_{i=1}^{n_{h}}\widehat{\mathbf{\Omega}}_{i,\lambda
}^{(h)}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q}), \\
\widehat{\mathbf{\Omega}}_{i,\lambda }^{(h)}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})& =\boldsymbol{U}_{\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda ,Q},\boldsymbol{x}_{hi})\boldsymbol{U}_{\lambda }^{T}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q},\boldsymbol{x}_{hi}) \\
& =\left[ w_{hi}^{2}\boldsymbol{\Delta }^{\ast }(\boldsymbol{\pi }_{hi}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q}))\mathrm{diag}^{\lambda -1}\{\boldsymbol{\pi }_{hi}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})\}\{\widehat{\boldsymbol{Y}}_{hi}-m_{hi}\boldsymbol{\pi }_{hi}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})\}\right. \\
& \left. \{\widehat{\boldsymbol{Y}}_{hi}-m_{hi}\boldsymbol{\pi }_{hi}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})\}^{T}\mathrm{diag}^{\lambda -1}\{\boldsymbol{\pi }_{hi}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})\}\boldsymbol{\Delta }^{\ast T}(\boldsymbol{\pi }_{hi}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q}))\right] \otimes \boldsymbol{x}_{hi}\boldsymbol{x}_{hi}^{T},\end{aligned}$$is consistent of $\mathbf{\Omega}_{\lambda }^{(h)}\left( \boldsymbol{\beta }\right) $, as $n_{h}$ goes to infinity, for every $h=1,...,H$.
\[cor3\]Let $\widehat{\boldsymbol{Y}}_{hi}$ be a random variable with overdisepersed multinomial sampling scheme with an overdispersion parameter $\nu _{h}$, specific for each stratum, and $m_{hi}=\overline{m}_{h}$,$$\begin{aligned}
\boldsymbol{\Sigma }_{hi}\left( \boldsymbol{\beta }\right) & =\nu _{h}\overline{m}_{h}\boldsymbol{\Delta }(\boldsymbol{\pi }_{hi}\left(
\boldsymbol{\beta }\right) ), \\
\nu _{h}& =1+\rho _{h}^{2}(\overline{m}_{h}-1),\end{aligned}$$then, for $\nu _{h}$ and $\rho _{h}^{2}$:
a)
: robust and consistent estimators based on the estimating equation are given respectively by$$\begin{aligned}
\widetilde{\nu }_{h,\lambda }^{E}(\widehat{\boldsymbol{\beta }}_{\lambda
,Q})& =\frac{1}{d(k+1)}\mathrm{trace}\left( \mathbf{\Omega}_{\lambda }^{(h),-1}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})\widehat{\mathbf{\Omega}}_{\lambda
}^{(h)}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})\right) , \\
\widetilde{\rho }_{h,\lambda }^{2,E}& =\frac{\widetilde{\nu }_{h,\lambda
}^{E}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})-1}{\overline{m}_{h}-1},\end{aligned}$$where the matrices of interest are the one associated with multinomial sampling for the $h$-th stratum, $$\begin{aligned}
\mathbf{\Omega}_{\lambda }^{(h)}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})& =\frac{1}{n_{h}}\sum_{i=1}^{n_{h}}\mathbf{\Omega}_{i,\lambda }^{(h)}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q}), \\
\mathbf{\Omega}_{i,\lambda }^{(h)}\left( \boldsymbol{\beta },\boldsymbol{\Sigma }_{hi}\right) & =\overline{m}_{h}w_{hi}^{2}\boldsymbol{\Delta }^{\ast }(\boldsymbol{\pi }_{hi}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q}))\mathrm{diag}^{\lambda -1}\{\boldsymbol{\pi }_{hi}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})\}\boldsymbol{\Delta }(\boldsymbol{\pi }_{hi}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})) \\
& \times \mathrm{diag}^{\lambda -1}\{\boldsymbol{\pi }_{hi}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})\}\boldsymbol{\Delta }^{\ast T}(\boldsymbol{\pi }_{hi}\widehat{\boldsymbol{\beta }}_{\lambda ,Q}))\otimes
\boldsymbol{x}_{hi}\boldsymbol{x}_{hi}^{T}\end{aligned}$$and Theorem \[Th2b\] for overdispersed multinomial sampling.
b)
: robust and consistent estimators based on the method of moments are given by$$\begin{aligned}
\widetilde{\nu }_{h,\lambda }^{M}(\widehat{\boldsymbol{\beta }}_{\lambda
,Q})& =\frac{1}{n_{h}d}\sum_{i=1}^{n_{h}}\sum_{j=1}^{d+1}\frac{(\widehat{Y}_{hij}-\overline{m}_{h}\pi _{hij}(\widehat{\boldsymbol{\beta }}_{\lambda
,Q}))^{2}}{\overline{m}_{h}\pi _{hij}(\widehat{\boldsymbol{\beta }}_{\lambda
,Q})}, \\
\widetilde{\rho }_{h,\lambda }^{2,M}& =\frac{\widetilde{\nu }_{h,\lambda
}^{M}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})-1}{\overline{m}_{h}-1}.\end{aligned}$$
Testing linear hypotheses for the PLR coefficients with complex survey design \[secWald\]
=========================================================================================
Based on the asymptotic distribution of the minimum quasi weighted DPD estimator, $\widehat{\boldsymbol{\beta }}_{\lambda ,Q},$ presented in Theorem \[th2\], we now introduce and study a family of Wald-type test statistics for testing $$H_{0}:\boldsymbol{M}^{T}\boldsymbol{\beta =l}\text{ against }H_{0}:\boldsymbol{M}^{T}\boldsymbol{\beta \neq l}\text{ } \label{A}$$where $\boldsymbol{M}$ is a $d(k+1)\times r$ full row-rank matrix with $r\leq d(k+1)$ and the right-hand-side $r-$vector consists of constants that in many situations are $\boldsymbol{l=0}_{r}.$ For solving the problem of testing given in (\[A\]) we define a family of Wald-type test statistics based on quasi MDPDE as follows.
Let $\widehat{\boldsymbol{\beta }}_{\lambda ,Q}$ the minimum quasi weighted DPD estimator of $\boldsymbol{\beta }$ in the PLR model (\[2.1.0\]) under a complex survey and we denote $$\begin{aligned}
\widehat{\mathbf{Q}}_{n,\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda
,Q})& =\frac{1}{n}\mathbf{\Psi}_{n,\lambda }^{-1}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})\widehat{\mathbf{\Omega}}_{n,\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})\mathbf{\Psi}_{n,\lambda }^{-1}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q}) \\
& =\left( n\mathbf{\Psi}_{n,\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda
,Q})\right) ^{-1}n\widehat{\mathbf{\Omega}}_{n,\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})\left( n\mathbf{\Psi}_{n,\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})\right) ^{-1},\end{aligned}$$where$$\begin{aligned}
n\widehat{\mathbf{\Omega}}_{n,\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda
,Q})& =\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}\left[ w_{hi}^{2}\boldsymbol{\Delta }^{\ast }(\boldsymbol{\pi }_{hi}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q}))\mathrm{diag}^{\lambda -1}\{\boldsymbol{\pi }_{hi}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})\}\{\widehat{\boldsymbol{Y}}_{hi}-m_{hi}\boldsymbol{\pi }_{hi}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})\}\right. \\
& \left. \times \{\widehat{\boldsymbol{Y}}_{hi}-m_{hi}\boldsymbol{\pi }_{hi}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})\}^{T}\mathrm{diag}^{\lambda -1}\{\boldsymbol{\pi }_{hi}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})\}\boldsymbol{\Delta }^{\ast T}(\boldsymbol{\pi }_{hi}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q}))\right. \otimes \boldsymbol{x}_{hi}\boldsymbol{x}_{hi}^{T}, \\
n\mathbf{\Psi}_{n,\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})&
=\left\{
\begin{array}{lc}
\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}w_{hi}m_{hi}\boldsymbol{\Delta }^{\ast }(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) )\mathrm{diag}^{\lambda -1}\{\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \}\boldsymbol{\Delta }^{\ast T}(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta
}\right) )\otimes \boldsymbol{x}_{hi}\boldsymbol{x}_{hi}^{T}, & \lambda >0
\\
\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}w_{hi}m_{hi}\boldsymbol{\Delta }(\boldsymbol{\pi }_{hi}^{\ast }\left( \boldsymbol{\beta }\right) )\otimes \boldsymbol{x}_{hi}\boldsymbol{x}_{hi}^{T}, & \lambda =0\end{array}\right. .\end{aligned}$$The family of Wald-type test statistics for testing the null hypothesis given in (\[A\]) is given by $$W_{n}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})=(\boldsymbol{M}^{T}\widehat{\boldsymbol{\beta }}_{\lambda ,Q}-\boldsymbol{l})^{T}\left(
\boldsymbol{M}^{T}\widehat{\mathbf{Q}}_{n,\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})\boldsymbol{M}\right) ^{-1}(\boldsymbol{M}^{T}\widehat{\boldsymbol{\beta }}_{\lambda ,Q}-\boldsymbol{l}) \label{B}$$
For $\lambda =0,$ $\widehat{\boldsymbol{\beta }}_{\lambda =0,Q},$ is the maximum quasi weighted likelihood estimator of $\boldsymbol{\beta }$, with estimating equations given in (\[2.4\]). It is not difficult to see that $\mathbf{Q}_{\lambda =0}(\widehat{\boldsymbol{\beta }}_{\lambda =0,Q})$ is the Fisher information matrix and $W_{n}(\widehat{\boldsymbol{\beta }}_{\lambda =0,Q})$ will be the classical Wald test statistic.
Under the null hypothesis given in (\[A\]) the asymptotic distribution of the Wald-type test statistics $W_{n}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})$, defined in (\[B\]), is chi-square with $r$ degrees of freedom.
The proof is immediate using the asymptotic distribution of the minimum quasi weighted DPD estimator, presented in Theorem \[th2\], and taking into account the consistence of the matrix $\mathbf{Q}_{\lambda }\left( \widehat{\boldsymbol{\beta }}_{\lambda ,Q}\right) $ presented in Corollary \[cor1b\]. Based on the previous theorem the null hypothesis given in (\[A\]) will be rejected if $$W_{n}\left( \widehat{\boldsymbol{\beta }}_{\lambda ,Q}\right) >\chi
_{r,\alpha }^{2}. \label{C}$$
Next we present a result in order to give an approximation of the power function for the test statisitcs given in (\[C\]). Let $\boldsymbol{\beta }^{\ast }$ $\in \Theta $ be such that $\boldsymbol{M}^{T}\boldsymbol{\beta }^{\ast }\neq \boldsymbol{c,}$ i.e., $\boldsymbol{\beta }^{\ast }$ does not belong to the null parameter space. Let us denote $$l_{\boldsymbol{\beta }_{1}}(\boldsymbol{\beta }_{2})=\left( \boldsymbol{M}^{T}\boldsymbol{\beta }_{1}-\boldsymbol{l}\right) ^{T}\left( \boldsymbol{M}^{T}\widehat{\mathbf{Q}}_{n,\lambda }\left( \boldsymbol{\beta }_{2}\right)
\boldsymbol{M}\right) ^{-1}\left( \boldsymbol{M}^{T}\boldsymbol{\beta }_{1}-\boldsymbol{l}\right) .$$Then we have the following result.
\[th:c\] Let $\boldsymbol{\beta }^{\ast }\in \Theta $ with $\boldsymbol{M}^{T}\boldsymbol{\beta }^{\ast }\neq \boldsymbol{l}$ the true value of the parameter so that $\widehat{\boldsymbol{\beta }}_{\lambda ,Q}\underset{n\rightarrow \infty }{\overset{P}{\longrightarrow }}\boldsymbol{\beta }^{\ast }$. The power function of the test statistic given in (\[C\]), at $\boldsymbol{\beta }^{\ast }$, is given by $$Po_{W_{n}\left( \widehat{\boldsymbol{\beta }}_{\lambda ,Q}\right) }\left(
\boldsymbol{\beta }^{\ast }\right) =1-\phi _{n}\left( \frac{1}{\sigma \left(
\boldsymbol{\beta }^{\ast }\right) }\left( \frac{\chi _{r,\alpha }^{2}}{\sqrt{n}}-\sqrt{n}l_{\boldsymbol{\beta }^{\ast }}(\boldsymbol{\beta }^{\ast
})\right) \right) \label{D}$$where $\phi _{n}\left( x\right) $ tends uniformly to the standard normal distribution $\phi \left( x\right) $ and $\sigma \left( \boldsymbol{\beta }^{\ast }\right) $ is given by $$\sigma \left( \boldsymbol{\beta }^{\ast }\right) ^{2}=\left. \frac{\partial
l_{\boldsymbol{\beta }}(\boldsymbol{\beta }^{\ast })}{\partial \boldsymbol{\beta }^{T}}\right\vert _{\boldsymbol{\beta }=\boldsymbol{\beta }^{\ast }}\widehat{\mathbf{Q}}_{n,\lambda }\left( \boldsymbol{\beta }^{\ast }\right)
\left. \frac{\partial l_{\boldsymbol{\beta }}(\boldsymbol{\beta }^{\ast })}{\partial \boldsymbol{\beta }}\right\vert _{\boldsymbol{\beta }=\boldsymbol{\beta }^{\ast }}.$$
It is clear that $$\lim_{n\rightarrow \infty }Po_{W_{n}\left( \widehat{\boldsymbol{\beta }}_{\lambda ,Q}\right) }\left( \boldsymbol{\beta }^{\ast }\right) =1$$for all $\alpha \in \left( 0,1\right) $. Therefore, the Wald-type test statistics are consistent in the sense of Fraser.
Based on the previous theorem we can obtain the sample size necessary to get a fix power $Po_{W_{n}\left( \widehat{\boldsymbol{\beta }}_{\lambda
,Q}\right) }\left( \boldsymbol{\beta }^{\ast }\right) =\pi _{0}.$ By (\[D\]) we must solve the equation $$1-\pi _{0}=\phi \left( \frac{1}{\sigma \left( \boldsymbol{\beta }^{\ast
}\right) }\left( \frac{\chi _{r,\alpha }^{2}}{\sqrt{n}}-\sqrt{n}l_{\boldsymbol{\beta }^{\ast }}(\boldsymbol{\beta }^{\ast })\right) \right)$$and we get that $n=\left[ n^{\ast }\right] +1$ with $$n^{\ast }=\frac{A+B+\sqrt{A(A+2B)}}{2l_{\boldsymbol{\beta }^{\ast }}(\boldsymbol{\beta }^{\ast })^{2}}$$ being $A=\sigma ^{2}\left( \boldsymbol{\beta }^{\ast }\right) \left( \phi
^{-1}\left( 1-\pi _{0}\right) \right) ^{2}\text{ and }B=\frac{1}{2}\chi
_{r,\alpha }^{2}l_{\boldsymbol{\beta }^{\ast }}(\boldsymbol{\beta }^{\ast }).$
We may also find an approximation of the power of the Wald-type tests given in (\[C\]) at an alternative close to the null hypothesis. Let $\boldsymbol{\beta }_{n}$ such that $\boldsymbol{M}^{T}\boldsymbol{\mathbf{\beta }_{n}\neq l}$ be a given alternative and let $\boldsymbol{\mathbf{\beta }}_{0}$ be the element such that $\boldsymbol{M}^{T}\boldsymbol{\mathbf{\beta }_{0}}=\boldsymbol{l}$ closest to $\boldsymbol{\beta }_{n}$ in the Euclidean distance sense. A first possibility to introduce contiguous alternative hypotheses is to consider a fixed $\boldsymbol{d}\in \mathbb{R}^{d(k+1)}$ and to permit $\boldsymbol{\beta }_{n}$ moving towards $\boldsymbol{\beta }_{0}$ as $n$ increases in the following way $$H_{1,n}:\boldsymbol{\beta}=\boldsymbol{\beta }_{n}, \quad \text{where} \quad \boldsymbol{\beta }_{n}=\boldsymbol{\beta }_{0}+n^{-1/2}\boldsymbol{d}. \label{E}$$A second approach is to relax the condition $\boldsymbol{M}^{T}\boldsymbol{\mathbf{\beta }_{0}}=\boldsymbol{l}$ defining the null hypothesis$.$ Let $\boldsymbol{\delta }\in \mathbb{R}^{r}$ and consider the following sequence, $\boldsymbol{\beta }_{n},$ of parameters moving towards $\boldsymbol{\beta }_{0}$ according to $$H_{1,n}^{\ast }:\boldsymbol{\beta}=\boldsymbol{\beta }_{n}, \quad \text{where} \quad \boldsymbol{M}^{T}\boldsymbol{\mathbf{\beta }_{n}}-\boldsymbol{l}=n^{-1/2}\boldsymbol{\delta }. \label{F}$$Note that $$\boldsymbol{M}^{T}\boldsymbol{\mathbf{\beta }_{n}}-l\boldsymbol{=M^{T}\boldsymbol{\beta }_{0}+M^{T}}n^{-1/2}\boldsymbol{\boldsymbol{d}}-\boldsymbol{l}=M\boldsymbol{^{T}}n^{-1/2}\boldsymbol{\boldsymbol{d}}.\label{G}$$Then the equivalence between the two hypotheses is given by $M\boldsymbol{^{T}\boldsymbol{d}}=\boldsymbol{\delta }.$
If we denote by $\chi_{r}^{2}(\Delta)$ the non central chi-square distribution with $r$ degrees of freedom and noncentrality parameter $\Delta, $ we can state the following theorem.
\[th:wald\] We have,
1. $W_{n}\left( \widehat{\boldsymbol{\beta }}_{\lambda ,Q}\right) \underset{n\longrightarrow \infty }{\overset{\mathcal{L}}{\longrightarrow }}\chi _{r}^{2}\left( \Delta _{1}\right) $under $H_{1,n}$ given in (\[E\]), with $$\Delta _{1}=\boldsymbol{d}^{T}\boldsymbol{M}\left[ M\boldsymbol{^{T}\widehat{\mathbf{Q}}_{n,\lambda }\left( \boldsymbol{\beta }_{0}\right) M}\right]
^{-1}M\boldsymbol{^{T}d.}$$
2. $W_{n}\left( \widehat{\boldsymbol{\beta }}_{\lambda ,Q}\right)\underset{n\longrightarrow \infty }{\overset{\mathcal{L}}{\longrightarrow }}\chi _{r}^{2}\left( \Delta _{2}\right) $ under $H_{1,n}^{\ast }$ given in (\[F\]), with $$\Delta _{2}=\boldsymbol{\delta }^{T}\left( M\boldsymbol{^{T}\widehat{\mathbf{Q}}_{n,\lambda }\left( \boldsymbol{\beta }_{0}\right) M}\right) ^{-1}\boldsymbol{\delta .}$$
Influence Function\[sec4\]
==========================
Influence function is a classical tool to measure robustness of an estimator (Hampel et al., 1968). However, the present set-up of complex survey is not as simple as the iid set-up; in fact the observations within a cluster of a stratum are iid but the observations in different cluster and stratum are independent non-homogeneous. So we need to modify the definition of the influence function accordingly. Recently, Ghosh and Basu (2013, 2016, 2018) have discussed the extended definition of the influence function for the independent but non-homogeneous observations; we will extend their approach to define the influence function in the present case of PLR model under complex design.
Influence Function of the minimum quasi weighted DPD estimator
--------------------------------------------------------------
We first need to define the statistical functional corresponding to the minimum quasi weighted DPD estimator as the minimizer of the DPD between the true and model densities. Assume the set-up and notations of Section 3 with the DPD kernel being given by $d_{\lambda }^{\ast }\left( g(\boldsymbol{y}_{hij}|\boldsymbol{x}_{hi}),f_{\boldsymbol{\beta }}(\boldsymbol{y}_{hij}|\boldsymbol{x}_{hi})\right)$ for individual densities $g(\boldsymbol{y}_{hij}|\boldsymbol{x}_{hi})$ and $f_{\boldsymbol{\beta }}(\boldsymbol{y}_{hij})$. Then, following Ghosh and Basu (2013), the minimum quasi weighted DPD estimator functional is to be defined by the minimizer of the total weighted DPD measure given by $$\begin{aligned}
d_\lambda(g, f_{\boldsymbol{\beta}}, w) &=& \sum_{h=1}^H\sum_{i=1}^{n_h}
d_{\lambda }^{\ast }\left( g(\boldsymbol{y}_{hij}|\boldsymbol{x}_{hi}),f_{\boldsymbol{\beta }}(\boldsymbol{y}_{hij}|\boldsymbol{x}_{hi})\right)
\nonumber\\
&=& \sum_{h=1}^{H}\sum_{i=1}^{n_{h}}w_{hi} \left(
m_{hi}\boldsymbol{\pi }_{hi}^{\lambda ,T}\left(\boldsymbol{\beta }\right)\boldsymbol{\pi }_{hi}\left(\boldsymbol{\beta }\right)
-\frac{\lambda +1}{\lambda }\int \boldsymbol{\pi }_{hi}^{\lambda ,T}\left(\boldsymbol{\beta }\right){\boldsymbol{y}}_{hi}
dG({\boldsymbol{y}}_{hi}|{\boldsymbol{x}}_{hi})\right),
\quad \text{for }\lambda >0. \nonumber\\
\label{def_func}\end{aligned}$$
We consider the PLR model with complex survey defined in (\[2.1.0\]). The minimum quasi weighted DPD estimator functional, ${\boldsymbol{T}}_{\lambda,Q}(\boldsymbol{G})$, of $\boldsymbol{\beta}$ at $\boldsymbol{G}=(G(\boldsymbol{y}_{hij}|\boldsymbol{x}_{hi}), ~ h=1, \ldots, H, i=1, \ldots, n_h)$ is defined as$${\boldsymbol{T}}_{\lambda,Q}(\boldsymbol{G}) = \arg \min_{\boldsymbol{\beta}\in \Theta }d_\lambda(g, f_{\boldsymbol{\beta}}, w),$$ where $d_\lambda(g, f_{\boldsymbol{\beta}}, w)$ is as defined above in (\[def\_func\]).
Note that, by the property of the DPD measure, it is immediate that the minimum quasi weighted DPD estimator functional ${\boldsymbol{T}}_{\lambda, Q}(\boldsymbol{G})$ is Fisher consistent at the assumed PLR model (\[2.1.0\]), i.e., ${\boldsymbol{T}}_{\lambda, Q}(\boldsymbol{\pi}(\boldsymbol{\beta})) =\boldsymbol{\beta}$ for all parameter values $\boldsymbol{\beta}$. Also, following the calculations for Theorem \[th:u\], one can see that the minimum quasi weighted DPD estimator functional ${\boldsymbol{T}}_{\lambda, Q}(\boldsymbol{G})$ can also be derived as a solution to the estimating equations $$\begin{aligned}
\boldsymbol{u}_{\lambda}^*\left(\boldsymbol{g}, \boldsymbol{\beta}\right)
& =\sum\limits_{h=1}^{H}\sum \limits_{i=1}^{n_{h}}
\boldsymbol{u}_{\lambda,hi}^*\left({g}_{hi}, \boldsymbol{\beta}\right)=\boldsymbol{0}_{d(k+1)}, \label{4.09} \end{aligned}$$ where ${g}_{hi}=g(\boldsymbol{y}_{hij}|\boldsymbol{x}_{hi})$ and $$\begin{aligned}
\boldsymbol{u}_{\lambda,hi}^*\left(g_{hi}, \boldsymbol{\beta}\right) &
=\left[w_{hi}m_{hi}\boldsymbol{\Delta }^{\ast }(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) )
\mathrm{diag}^{\lambda -1}\{\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \}
\{g_{hi}-\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \}\right] \otimes
\boldsymbol{x}_{hi}. \label{4.010}\end{aligned}$$ Note that, these estimating equations are unbiased at the model probability $\boldsymbol{g}=\boldsymbol{\pi}(\boldsymbol{\beta})$.
Now, in order to define the influence function of the minimum DPD functional ${\boldsymbol{T}}_{\lambda, Q}(\boldsymbol{g})$, we note that the functional itself depends on the sample sizes and cluster weights and so it IF will have the same dependence in analogue to the non-homogeneous case of Ghosh and Basu (2013, 2016, 2018). Also note that, the contamination can be in any one particular cluster within one stratum or simultaneously in many of them (or all). For simplicity, let us first assume that the contamination is only in one cluster probability ${g}_{h_0i_0}$ for some fix $h_0$ and $i_0$. Consider the contaminated probability vector ${g}_{h_0i_0, \epsilon} = (1-\epsilon) {g}_{h_0i_0} + \epsilon \delta_{\boldsymbol{t}}$, where $\epsilon$ is the contamination proportion and $ \delta_{\boldsymbol{t}}$ is the degenerate probability at the outlier point $\boldsymbol{t} =(t_1, \ldots, t_{d+1})^T \in \{0, 1\}^{d+1}$ with $\sum_{s=1}^{d+1}t_s=1$. Denote the corresponding contaminated full probability vector as $\boldsymbol{g}_{\epsilon}$ which is same as $\boldsymbol{g}$ except ${g}_{h_0i_0}$ being replaced by ${g}_{h_0i_0,\epsilon}$ and let the corresponding contaminated distribution vector be $\boldsymbol{G}_\epsilon$. Then, the corresponding influence function is defined as $$\mathcal{IF}(\boldsymbol{t}, {\boldsymbol{T}}_{\lambda, Q}, \boldsymbol{g})
= \lim_{\epsilon\rightarrow 0}
\frac{{\boldsymbol{T}}_{\lambda, Q}(\boldsymbol{G}_{\epsilon}) -{\boldsymbol{T}}_{\lambda, Q}(\boldsymbol{G})}{\epsilon}
= \frac{\partial}{\partial\epsilon}{\boldsymbol{T}}_{\lambda, Q}(\boldsymbol{G}_{\epsilon})\bigg|_{\epsilon=0}.$$
In order to calculate this influence function, we start with the estimating equation for ${\boldsymbol{T}}_{\lambda, Q}$. Note that, ${\boldsymbol{T}}_{\lambda, Q}(\boldsymbol{G}_{\epsilon})$ satisfies the equations $$\boldsymbol{u}_{\lambda}^*\left(\boldsymbol{g}_{\epsilon}, {\boldsymbol{T}}_{\lambda, Q}(\boldsymbol{G}_{\epsilon})\right)
=\boldsymbol{0}_{d(k+1)}.$$ Now, differentiating it with respect to $\epsilon$ art $\epsilon=0$ and simplifying, we get the required influence function as given by $$\begin{aligned}
\mathcal{IF}(\boldsymbol{t}, {\boldsymbol{T}}_{\lambda, Q}, \boldsymbol{g})=
\boldsymbol{\Psi}_{n,\lambda}^*(\boldsymbol{g}, \boldsymbol{\beta})^{-1}\frac{1}{n}
\left[\boldsymbol{u}_{\lambda,h_0i_0}^*\left(\delta_{\boldsymbol{t}}, \boldsymbol{\beta}\right)
- \boldsymbol{u}_{\lambda,h_0i_0}^*\left({g}_{h_0i_0}, \boldsymbol{\beta}\right)\right],\end{aligned}$$ where $$\begin{aligned}
\boldsymbol{\Psi}_{n,\lambda}^*(\boldsymbol{g}, \boldsymbol{\beta})= - \frac{1}{n} \sum\limits_{h=1}^{H}\sum \limits_{i=1}^{n_{h}}
\frac{\partial}{\partial\boldsymbol{\beta}}\boldsymbol{u}_{\lambda,hi}^*\left({g}_{hi}, \boldsymbol{\beta}\right).\end{aligned}$$ Note that, at the model probability $\boldsymbol{g}=\boldsymbol{\pi}(\boldsymbol{\beta})$, we have $\boldsymbol{u}_{\lambda,h_0i_0}^*\left(\boldsymbol{\pi}_{h_0i_0}(\boldsymbol{\beta}), \boldsymbol{\beta}\right)
= \boldsymbol{0}$ and $\boldsymbol{\Psi}_{n,\lambda}^*(\boldsymbol{\pi}(\boldsymbol{\beta}), \boldsymbol{\beta})=\boldsymbol{\Psi}_{n,\lambda}(\boldsymbol{\beta})$, as defined in Theorem \[Th1\]. Hence, the influence function of the proposed minimum quasi weighted DPD estimator at the model probability simplifies to $$\begin{aligned}
\mathcal{IF}(\boldsymbol{t}, {\boldsymbol{T}}_{\lambda, Q}, \boldsymbol{\pi}(\boldsymbol{\beta}))
=\boldsymbol{\Psi}_{n,\lambda}(\boldsymbol{\beta})^{-1}\frac{1}{n}
\boldsymbol{u}_{\lambda,h_0i_0}^*\left(\delta_{\boldsymbol{t}}, \boldsymbol{\beta}\right).\end{aligned}$$ Clearly, by the assumed form of the model probability $\boldsymbol{\pi}(\boldsymbol{\beta})$ this influence function is bounded for all $\lambda>0$ but unbounded at $\lambda=0$. So, the proposed minimum quasi weighted DPD estimators are expected to be robust for any $\lambda>0$ but the corresponding maximum quasi weighted likelihood estimator (at $\lambda=0$) is non-robust against data contamination.
Similarly, one can show that, if there is contamination in some of the clusters within some stratum indexed by ${h,i}\in \Gamma \subseteq \{h=1, \ldots, H; i=1, \ldots, n_h\}$ at the contamination points $\boldsymbol{t}_{hi}$, the corresponding influence function of the proposed minimum quasi weighted DPD estimator at $\boldsymbol{g}$ and the model $\pi(\boldsymbol{\beta})$ are respectively given by $$\begin{aligned}
\mathcal{IF}((\boldsymbol{t}: \{h, i\}\in \Gamma), {\boldsymbol{T}}_{\lambda, Q}, \boldsymbol{g})=
\boldsymbol{\Psi}_{n,\lambda}^*(\boldsymbol{g}, \boldsymbol{\beta})^{-1}\frac{1}{n}
\sum_{\{h,i\}\in \Gamma}\left[\boldsymbol{u}_{\lambda,hi}^*\left(\delta_{\boldsymbol{t}_{hi}}, \boldsymbol{\beta}\right)
- \boldsymbol{u}_{\lambda,hi}^*\left({g}_{hi}, \boldsymbol{\beta}\right)\right],\end{aligned}$$ and $$\begin{aligned}
\mathcal{IF}((\boldsymbol{t}: \{h, i\}\in \Gamma), {\boldsymbol{T}}_{\lambda, Q}, \boldsymbol{\pi}(\boldsymbol{\beta}))
=\boldsymbol{\Psi}_{n,\lambda}(\boldsymbol{\beta})^{-1}\frac{1}{n}
\sum_{\{h,i\}\in \Gamma}\boldsymbol{u}_{\lambda,hi}^*\left(\delta_{\boldsymbol{t}_{hi}}, \boldsymbol{\beta}\right).\end{aligned}$$ The boundedness and robustness implications for these influence functions are exactly the same as before.
Influence Function of the Wald-type Tests in PLRM
-------------------------------------------------
Let us now study the robustness of proposed Wald-type test through the influence function of the corresponding test statistics in (\[B\]). Considering the notation and set-up of Section 2 and 3, and define $$\mathbf{Q}_{\lambda }(\boldsymbol{\beta}) = \frac{1}{n}\mathbf{\Psi}_{\lambda}^{-1}({\boldsymbol{\beta}})
{\mathbf{\Omega}}_{\lambda }({\boldsymbol{\beta }})\mathbf{\Psi}_{\lambda }^{-1}({\boldsymbol{\beta }}).$$ Then the Wald-type test functional $W_\lambda(\boldsymbol{G})$ corresponding to (\[B\]) for testing the null hypothesis given in (\[A\]) at the true distribution vector $\boldsymbol{G}$ is defined as $$W_{\lambda}(\boldsymbol{G})=(\boldsymbol{M}^{T}{\boldsymbol{T}}_{\lambda, Q}(\boldsymbol{G})-\boldsymbol{l})^{T}
\left(\boldsymbol{M}^{T}{\mathbf{Q}}_{\lambda }({\boldsymbol{T}}_{\lambda, Q}(\boldsymbol{G}))\boldsymbol{M}\right) ^{-1}
(\boldsymbol{M}^{T}{\boldsymbol{T}}_{\lambda, Q}(\boldsymbol{G})-\boldsymbol{l}), \label{B1}$$ where ${\boldsymbol{T}}_{\lambda, Q}$ is the minimum quasi weighted DPD estimator functional as defined in the previous subsection. Also, let $\boldsymbol{\beta }_{0}$ denote the true null parameter value for the hypothesis in (\[A\]).
Let us now derive the influence function of the test functional $W_\lambda$. As before, here also the contamination can be any particular cluster and strata (a given $h_0, i_0$) combination or in many (or all) of them. The influence function of general Wald-type tests under such non-homogeneous set-up has been extensively studied in Basu et al. (2018). Here, we follow the general theory of Basu et al. (2018) to conclude that the first order influence functions of $W_\lambda$, defined as the first order derivative of its value at the contaminated distribution with respect to $\epsilon$ at $\epsilon=0$, in either case of contamination become identically zero at the null distribution $\boldsymbol{g}=\boldsymbol{\pi}(\boldsymbol{\beta}_0)$. Therefore, the first order influence function is not informative in this case of Wald-type tests, and we need to investigate the second order influence function of $W_\lambda$.
The second order influence function $IF^{(2)}$, which measures the second order approximation to the asymptotic bias due to infinitesimal contamination, is defined as the second order derivative of the value of $W_\lambda$ at the contaminated distribution with respect to the contamination proportion $\epsilon$ at $\epsilon=0$. Again, following Basu et al. (2018), we can derive these second order influence functions of the Wald-type tests in either case of contaminations; at the null distribution $\boldsymbol{g}=\boldsymbol{\pi}(\boldsymbol{\beta}_0)$, they simplifies to $$\begin{aligned}
&\mathcal{IF}^{(2)}(\boldsymbol{t}, W_{\lambda}, \boldsymbol{\pi}(\boldsymbol{\beta}_0))
=2\mathcal{IF}(\boldsymbol{t}, {\boldsymbol{T}}_{\lambda, Q}, \boldsymbol{\pi}(\boldsymbol{\beta}_0))^{T}
\boldsymbol{M}\left(\boldsymbol{M}^{T}{\mathbf{Q}}_{\lambda }(\boldsymbol{\beta}_0)\boldsymbol{M}\right) ^{-1}\boldsymbol{M}^{T}
\mathcal{IF}(\boldsymbol{t}, {\boldsymbol{T}}_{\lambda, Q}, \boldsymbol{\pi}(\boldsymbol{\beta}_0)),
\nonumber\\
&\mathcal{IF}^{(2)}((\boldsymbol{t}: \{h, i\}\in \Gamma), W_{\lambda}, \boldsymbol{\pi}(\boldsymbol{\beta}_0))\nonumber
\\
& =2\mathcal{IF}((\boldsymbol{t}: \{h, i\}\in \Gamma), {\boldsymbol{T}}_{\lambda, Q}, \boldsymbol{\pi}(\boldsymbol{\beta}_0))^{T}
\boldsymbol{M}\left(\boldsymbol{M}^{T}{\mathbf{Q}}_{\lambda }(\boldsymbol{\beta}_0)\boldsymbol{M}\right) ^{-1}\boldsymbol{M}^{T}
\mathcal{IF}((\boldsymbol{t}: \{h, i\}\in \Gamma), {\boldsymbol{T}}_{\lambda, Q}, \boldsymbol{\pi}(\boldsymbol{\beta}_0)),
\nonumber\end{aligned}$$for the two types of contamination as before in the case of estimator. Note that, the second order influence functions of the proposed Wald-type tests are a quadratic function of the corresponding influence functions of the minimum quasi weighted DPD estimator for any type of contamination. Therefore, the boundedness of the influence functions of minimum quasi weighted DPD estimator at $\lambda>0$ also indicates the boundedness of the influence functions of the Wald-type test functional $W_\lambda$ indicating their robustness against contamination in any cluster or stratum of the sample data.
Illustrative Example: BMI data set \[secEx\]
============================================
Let us consider an illustrative real-life dataset on BMI which was previously studied in Castilla et al. (2018). This data set, obtained from CANSIM Canada’s database and presented in Table \[table:CANSIM\_data\], shows the body mass indexes of population in Canada in the year 1994. Each person in the sample is divided by their body mass index category under the international standard: acceptable weight, overweight or obese. The data set consists on a stratified sample design with clusters nested on them, with the strata being three different age groups (20–34 years, 35–44 years and 45–64 years) and the genders (male or female) as the clusters. The qualitative explanatory variables are valid to distinguish the clusters within the strata. They are given by $\boldsymbol{x}_{h1}^T=(1,0)$, and $\boldsymbol{x}_{h2}^T=(0,1)$, $h=1,\dots,5$ for men and women, respectively.
2.8pt
$\begin{tabular}{llccc}
\hline
& & \multicolumn{2}{c}{Body mass index (BMI)} & \\ \cline{3-5}
Age group & Sex & Acceptable weight (18.5-24.9) &
Overweight (25.0-29.9) & Obese (30 or higher) \\ \hline
20-34 years & Men & \multicolumn{1}{c}{5438} & \multicolumn{1}{c}{4790} &
\multicolumn{1}{c}{1470} \\
& Women & \multicolumn{1}{c}{4910} & \multicolumn{1}{c}{2878} &
\multicolumn{1}{c}{802} \\
35-44 years & Men & \multicolumn{1}{c}{2458} & \multicolumn{1}{c}{3437} &
\multicolumn{1}{c}{1319} \\
& Women & \multicolumn{1}{c}{3100} & \multicolumn{1}{c}{1494} &
\multicolumn{1}{c}{1313} \\
45-64 years & Men$\^\*$ & \multicolumn{1}{c}{1968} & \multicolumn{1}{c}{3290} &
\multicolumn{1}{c}{1412} \\
& Women$\^\*$ & \multicolumn{1}{c}{1710} & \multicolumn{1}{c}{1481} &
\multicolumn{1}{c}{1078} \\ \hline
\end{tabular} $ \[table:CANSIM\_data\]
To illustrate the robustness of minimum quasi weighted DPD estimators, we contaminate the BMI data by permuting the categories overweight and obese in the Men with age range 45–64 years. After obtaining the correpsonding minimum quasi weighted DPD estimators estimates, we compute the mean absolute standardized deviations ($mssd$) between the estimated parameters and corresponding estimated probabilities obtained for the modified and original data, as given by $$masd(\widehat{\boldsymbol{\beta}}_{\lambda}^*,\widehat{\boldsymbol{\beta}}_{\lambda})=\frac{1}{4}\sum_{r=1}^{2}\sum_{s=1}^{2}\left|\frac{\widehat{\beta}^*_{\lambda,rs}-\widehat{\beta}_{\lambda,rs}}{\widehat{\beta}_{\lambda,rs}} \right| \quad \text{ and } \quad masd(\widehat{\boldsymbol{\pi}}^*_{\lambda},\widehat{\boldsymbol{\pi}}_{\lambda})=\frac{1}{6}\sum_{r=1}^{3}\sum_{s=1}^{2}\left|\frac{\pi_{rs}(\widehat{\boldsymbol{\beta}}_{\lambda}^*)-\pi_{rs}(\widehat{\boldsymbol{\beta}}_{\lambda})}{\pi_{rs}(\widehat{\boldsymbol{\beta}}_{\lambda})} \right|,$$ where, with superscript $^*$ we denote the contaminated case. Assuming common overdispersion parameter, absolute standardized deviation ($asd$) of the two versions of the intra-cluster correlation estimator are also computed as in Corollary \[cor2\]. Their values, as presented in Table \[table:CANSIM\_dataMSD\], clearly show that the minimum quasi weighted DPD estimators become more robust as $\lambda$ increases. Similar results are obtained when permuting the categories for Women instead of Men, as can also be seen in Table \[table:CANSIM\_dataMSD\].
[l rrr rrr]{} &\
& 0 & 0.2 & 0.4 & 0.6 & 0.8 & 1\
\
$masd$($\widehat{\boldsymbol{\beta}}^*_{\lambda}$,$\widehat{\boldsymbol{\beta}}_{\lambda}$) & 0.24396 & 0.23057 & 0.21731 & 0.20441 & 0.19187 & 0.17969\
$masd$($\widehat{\boldsymbol{\pi}}^*_{\lambda}$,$\widehat{\boldsymbol{\pi}}_{\lambda}$)& 0.10170 & 0.09700 & 0.09220 & 0.08750 & 0.08280 & 0.07810\
$asd$($\widehat{\rho}^{2*,E}_{\lambda},\widehat{\rho}^{2,E}_{\lambda}$)& 0.64785 & 0.58974 & 0.52955 & 0.46868 & 0.40858 & 0.35044\
$asd$($\widehat{\rho}^{2*,M}_{\lambda},\widehat{\rho}^{2,M}_{\lambda}$)& 1.57103 & 1.41605 & 1.25700 & 1.09724 & 0.94068 & 0.79037\
\
$masd$($\widehat{\boldsymbol{\beta}}^{*}_{\lambda}$,$\widehat{\boldsymbol{\beta}}_{\lambda}$) & 0.10516 & 0.09484 & 0.08533 & 0.07665 & 0.0687 & 0.06148\
$masd$($\widehat{\boldsymbol{\pi}}^{*}_{\lambda}$,$\widehat{\boldsymbol{\pi}}_{\lambda}$) & 0.03250 & 0.03030 & 0.02810 & 0.02600 & 0.0240 & 0.02210\
$asd$($\widehat{\rho}^{2*,E}_{\lambda},\widehat{\rho}^{2,E}_{\lambda}$)& 0.07038 & 0.05132 & 0.03358 & 0.01744 & 0.00321 & 0.00893\
$asd$($\widehat{\rho}^{2*,M}_{\lambda},\widehat{\rho}^{2,M}_{\lambda}$)& 0.03358 & 0.01169 & 0.05244 & 0.08819 & 0.11821 & 0.14206\
Monte Carlo Simulation Study\[sec5\]
====================================
Simulation Scheme\[sec:SimA\]
-----------------------------
We develop a complex design extension of the simulation scheme previously studied by Castilla et al. (2019), where a simple sample design was considered. $H$ strata consisting of $n_h=n$ clusters having $m_{hi}=m$ units each, are taken. Three overdispersed multinomial distributions: the random-clumped (RC), the $m$-inflated ($m$-I) and the Dirichlet Multinomial (DM) distributions (Alonso et al., 2017) having the same parameters $\boldsymbol{\pi}_{hi}\left(\boldsymbol{\beta}_{0}\right)$ and $\rho$, are then considered for $\widehat{\boldsymbol{Y}}_{hi}$; which are characterized by $$\begin{aligned}
\boldsymbol{E}[\widehat{\boldsymbol{Y}}_{hi}] & =m\boldsymbol{\pi}_{hi}\left( \boldsymbol{\beta}_{0}\right) \quad\text{and}\quad\boldsymbol{V}[\widehat{\boldsymbol{Y}}_{hi}]=\nu_{m}m\boldsymbol{\Delta}(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta}_{0}\right) ), \\
\nu_{m} & =1+\rho^{2}(m-1), ~~~~~~~i=1, \ldots, n; ~h=1,2.\end{aligned}$$ As in Castilla et al. (2019), we further assume that the outcome nominal variable $\boldsymbol{Y}$ has $d+1=3$ categories depending on $k=2$ explanatory variables through the PLR model probabilities $$\pi _{hir}\left( \boldsymbol{\beta }\right) =\left \{
\begin{array}{ll}
\dfrac{\exp \{ \boldsymbol{x}_{hi}^{T}\boldsymbol{\beta }_{r}\}}{1+{\sum_{l=1}^{d}}\exp \{ \boldsymbol{x}_{hi}^{T}\boldsymbol{\beta }_{l}\}}, &
r=1,2 \\
\dfrac{1}{1+{\sum_{l=1}^{d}}\exp \{ \boldsymbol{x}_{hi}^{T}\boldsymbol{\beta
}_{l}\}}, & r=3\end{array}\right. ,$$with $\boldsymbol{\beta }=(\beta_{01},\beta_{11},\beta_{21},\beta_{02},\beta_{12},\beta_{22})^T=(0,-0.9,0.1,0.6,-1.2,0.8)^T$ and $\boldsymbol{x}_{hi}\overset{iid}{\sim}\mathcal{N}(\boldsymbol{0},\boldsymbol{I})$ for all $i=1, \ldots,n$, $h=1,\dots,H$.
Considering different values of the intra-cluster correlation parameter ($\rho^2$), the number of clusters in each stratum ($n$) and the clusters sizes ($m$), we simulate data from different scenarios.
1. Scenario 1: $H=2$, $n\in\{10i\}_{i=4}^{15}$, $m=21$, $\rho^2=0.25$, RC distribution.
2. Scenario 1b: $H=2$, $n\in\{10i\}_{i=4}^{15}$, $m=21$, $\rho^2=0.50$, RC distribution.
3. Scenario 2: $H=2$, $n=60$, $m\in \{10i\}_{i=2}^{12}$, $\rho^2=0.25$, RC m-I and DM distributions.
4. Scenario 3: $H=2$, $n=60$, $m=21$, $\rho^2\in\{0.1i\}_{i=0}^9$, RC, m-I and DM distributions.
In order to study the robustness issue, these simulations are repeated under contaminated data having $7\%$ outliers. These outliers are generated by permuting the elements of the outcome variable, such that categories 1, 2, 3 are classified as categories 3, 1, 2 for the outlying observations.
Performance of minimum quasi weighted DPD estimators and $\protect\rho^2$ estimates
-----------------------------------------------------------------------------------
For the above scenarios, we compute the minimum quasi weighted DPD estimator of $\boldsymbol{\beta}$, for different tuning parameters $\lambda\in\{0,0.2,0.4,0.6,0.8\}$, and the corresponding estimate of $\rho^2$, both for the methods of moments and the estimating equations. The root of the mean square error (RMSE) are then computed based on $1000$ replications (Figures \[fig:S1\]–\[fig:S3\]).
While classical maximum quasi weighted likelihood estimator presents the best behaviour under pure data, minimum quasi weighted DPD estimators with $\lambda>0$ are much more robust. In particular, as $\lambda$ increases, the change on their behaviour is accentuated. This is independent to the sample size (Figures \[fig:S1\] and \[fig:S1b\]), the intra-cluster correlation and the distribution (Figure \[fig:S2\]) considered.
Estimator of $\rho^2$ by the method of moments seems less precise than the method of estimating equations, with independence of the tuning parameter chosen (Figures \[fig:S1\] and \[fig:S1b\]). Best estimators of $\rho^2$ by the estimating equations are obtained from minimum quasi weighted DPD estimators with $\lambda>0$, both for pure and contaminated data and for any of the distributions considered (Figure \[fig:S2\]). Error of the estimators of $\rho^2$ tends to be smaller with the DM and RC distributions in comparison with the mI distribution, as can be seen in Figure \[fig:hist\], where density plots based on $1,000$ replications are shown for $\rho^2=0.5, n=60, m=21$ and tuning parameter $\lambda=0.4$. These results on the design effect parameter are consistent with the previous work of Alonso et al. (2017).
![ Density plots with estimates obtained from observations of three distributions, DM, mI, RC, when $\protect\rho^2 = 0.5$ and $\protect\lambda=0.4$. $1,000$ samples.[]{data-label="fig:hist"}](figures/scenario_histograms.pdf)
Notice that $\boldsymbol{\beta}$ estimates are obtained through $nlm()$ procedure with tolerance $10^{-6}$ in the software R, used for the whole Monte Carlo study. In Figure \[fig:hist\] iterations needed to compute the corresponding minimum quasi weighted DPD estimators were also obtained, with not a significant difference among the different distributions.
![Pseudo minimum phi-divergence estimators: best choice of the tuning parameter, $m=21$ and $\protect\rho^2=0.5.$[]{data-label="fig:S_bar_phi"}](figures/phi_barplots_p05_lambda.pdf)
![Comparison among pseudo minimum phi−divergence and minimum quasi weighted DPD estimators[]{data-label="fig:S_rho_phi"}](figures/scenario_rho_phi_dpd_title.pdf "fig:") ![Comparison among pseudo minimum phi−divergence and minimum quasi weighted DPD estimators[]{data-label="fig:S_rho_phi"}](figures/scenario_nh_p025_phi_dpd.pdf "fig:")
Comparison of minimum quasi weighted DPD estimators with pseudo minimum phi-divergence estimators
-------------------------------------------------------------------------------------------------
In the previous Section we have illustrated how the quasi minimum quasi weighted DPD estimators present a much more robust behavior than the maximum quasi weighted likelihood estimator for the estimation of $\boldsymbol{\beta}$ when a moderate/large sample size was considered. In Castilla et al. (2018), another family of estimators, the pseudo minimum phi-divergence estimators, was defined in the PLR model with complex survey. In particular, pseudo minimum Cressie-Read divergence estimators for positive tuning parameter $\lambda$ were studied. Results of the simulation study suggested these estimators as an interesting alternative to maximum quasi weighted likelihood estimator in terms of efficiency when a small sample size and a large intra cluster correlation was considered. Nevertheless, the robustness of these estimators was not studied.
In this section we want to make a general empirical comparison between minimum quasi weighted DPD estimator and pseudo minimum phi-divergence estimators (with positive and negative tuning parameters in the Cressie-Read subfamily) of $\boldsymbol{\beta}$, in terms of efficiency and robustness. Although they were not studied in the cited paper of Castilla et al. (2018), phi-divergences with negative tuning parameter and, in particular, the Hellinger distance ($\lambda=-0.5$), are known to have good robustness properties in some models. See for example Beran (1977) Lindsay (1994) and Toma (2007). This seems to happen in the context of PLR model with complex survey too, when a moderate/large sample size is considered. Nevertheless, pseudo minimum phi-divergence estimators with positive tuning parameter present an important lack of robustness. This behaviour can be summarized in Figure \[fig:S\_bar\_phi\].
Let us now consider Scenario 3 of the previous section. A comparison between quasi minimum quasi weighted DPD estimators with $\lambda\in\{0.4,0.8\}$, classical maximum quasi weighted likelihood estimator, and pseudo minimum phi-divergence estimators with $\lambda\in\{-0.5,-0.3,0.66\}$ is made. Hellinger distance is shown to be, by far, the best choice when a contaminated scheme with low intra-cluster correlation is considered. With medium/high intra-correlation the behaviour turns to be the opposite. Minimum quasi weighted DPD estimators present a more stable behaviour, both in pure and contaminated schemes, with independence to the correlation parameter.
The same divergences are considered now in Scenario 1 for the estimation of the parameter $\rho^2$. The estimation of $\rho^2$ with pseudo minimum phi-divergence estimators is made by the method of Binder (Castilla et al., 2018) while the estimation of $\rho^2$ with minimum quasi weighted DPD estimators is made by the method of the estimating equations. Pseudo minimum phi-divergence estimators with negative tuning parameter are not competitive neither in the pure nor in the contaminated schemes. Pseudo minimum phi-divergence estimator with tuning parameter $\lambda=2/3$ is a good alternative to the minimum quasi weighted DPD estimators in contaminated data. This estimator showed also a good behaviour in the paper of Castilla et al. (2018).
![Estimated Levels (top) and Powers (bottom) under pure (left) and contaminated (right) data. Scenario 1.[]{data-label="fig:TESTS"}](figures/scenario_nh_p025_LEVEL.pdf "fig:") ![Estimated Levels (top) and Powers (bottom) under pure (left) and contaminated (right) data. Scenario 1.[]{data-label="fig:TESTS"}](figures/scenario_nh_p025_POWER.pdf "fig:")
![Scenario 1: RMSEs of minimum quasi weighted DPD estimators of $\boldsymbol{\protect\beta}$ and $\protect\rho^2$ by the equations method and the method of moments. Pure data (left) and contaminated data (right). RC distribution, $\protect\rho^2=0.25$. []{data-label="fig:S1"}](figures/scenario_nh_p025.pdf)
![Scenario 1b: RMSEs of minimum quasi weighted DPD estimators of $\boldsymbol{\protect\beta}$ and $\protect\rho^2$ by the equations method and the method of moments. Pure data (left) and contaminated data (right). RC distribution, $\protect\rho^2=0.50$.[]{data-label="fig:S1b"}](figures/scenario_nh_p050.pdf)
![Scenario 2: RMSEs of $\protect\rho^2$ by the equations method. Pure data (left) and contaminated data (right). RC (top), mI (middle) and DM (bottom) distribution, $\protect\rho^2=0.25, n=60$.[]{data-label="fig:S2"}](figures/scenario_m_p025.pdf)
![Scenario 3: RMSEs of minimum quasi weighted DPD estimators of $\boldsymbol{\protect\beta}$. Pure data (left) and contaminated data (right). RC (top), mI (middle) and DM (bottom) distribution, $n=60, m=21$.[]{data-label="fig:S3"}](figures/scenario_rho.pdf)
Performance of the Wald-type tests
----------------------------------
With the same model as in Section \[sec:SimA\] we now empirically study the robustness of the minimum quasi weighted DPD estimator based Wald-type tests for the PLR model. As in Castilla et al. (2018) we first study the level under the true null hypothesis $H_0:\beta_{02}=0.6$. For studying the power robustness, the true data generating parameter value is considered as $\beta_{02}=1.08$.
Under Scenario 1, and both for pure and contaminated data, we compute observed levels and powers (measured as the proportion of test statistics exceeding the corresponding chi-square critical value), as can be seen in Figure \[fig:TESTS\]. Under pure data, all levels present a very similar behaviour, while power attains their best value for classical maximum quasi weighted likelihood estimator. Under contamination, minimum quasi weighted DPD estimator based Wald-type tests for $\lambda>0$ present a more robust behaviour than maximum quasi weighted likelihood estimator, in concordance with previous results.
Concluding Remarks \[secCon\]
=============================
We have presented minimum quasi weighted DPD estimators, as a robust alternative to classical approaches, in the modeling of categorical responses with associated covariates through polytomous logistic regression models. A Wald-type family of tests based on these estimators is also presented, for the problem of testing linear hypotheses on regression coefficients. The robustness of both estimators and tests are theoretically justified in terms of the influence function, which is shown to be bounded for the new procedures. An extensive simulation study is also provided, to empirically illustrate their robustness along with a comparison with the pseudo minimum phi-divergence estimators of Castilla et al. (2018). The results clearly show how the proposed minimum quasi weighted DPD estimators seem to be the best choice for dealing with the robustness issue for a moderate sample size.
Appendix \[App\]
================
Derivation of (\[Un\]) from (\[2.4\]) {#derivation-of-un-from-2.4 .unnumbered}
-------------------------------------
From (\[2.4\]) it holds ${\sum\limits_{h=1}^{H}}{\sum\limits_{i=1}^{n_{h}}}\boldsymbol{u}_{hi}(\boldsymbol{\beta })=\boldsymbol{0}_{d(k+1)}$, with$$\boldsymbol{u}_{hi}(\boldsymbol{\beta })=w_{hi}\mathcal{\boldsymbol{\Delta }}^{\ast }(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) )\mathrm{diag}^{-1}(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) )\left[
\widehat{\boldsymbol{y}}_{hi}-m_{hi}\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \right] \otimes \boldsymbol{x}_{hi}.$$The first term of $\boldsymbol{u}_{hi}(\boldsymbol{\beta })$ is$
w_{hi}\boldsymbol{\Delta }^{\ast }(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) )\mathrm{diag}^{-1}\{\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \}\{\widehat{\boldsymbol{y}}_{hi}-m_{hi}\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \},
$ where$$\begin{aligned}
w_{hi}\boldsymbol{\Delta }^{\ast }(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) )\mathrm{diag}^{-1}\{\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \}& =\left( \boldsymbol{I}_{d},\boldsymbol{0}_{d}\right) \left[ \mathrm{diag}(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right)
)-\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \boldsymbol{\pi }_{hi}^{T}\left( \boldsymbol{\beta }\right) \right] \mathrm{diag}^{-1}\{\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \} \\
& =\left( \boldsymbol{I}_{d},\boldsymbol{0}_{d}\right) (\boldsymbol{I}_{d+1}-\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \boldsymbol{1}^{T}),\end{aligned}$$but $\boldsymbol{1}^{T}\{\widehat{\boldsymbol{y}}_{hi}-m_{hi}\boldsymbol{\pi
}_{hi}\left( \boldsymbol{\beta }\right) \}=0$ and hence$$\boldsymbol{u}_{hi}(\boldsymbol{\beta })=w_{hi}\left[ \widehat{\boldsymbol{y}}_{hi}-m_{hi}\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \right]
\otimes \boldsymbol{x}_{hi}.$$
Proof of Theorem \[th:u\] {#proof-of-theorem-thu .unnumbered}
--------------------------
The minimum quasi weighted DPD estimator of $\boldsymbol{\beta }$, is defined as $$\begin{aligned}
\widehat{\boldsymbol{\beta }}_{\lambda ,Q}& =\underset{\boldsymbol{\beta }\in \mathbb{R}^{d(k+1)}}{\arg \min }\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}w_{hi}\left\{ \left( m_{hi}\boldsymbol{\pi }_{hi}(\boldsymbol{\beta })-\frac{1+\lambda }{\lambda }\widehat{\boldsymbol{y}}_{hi}\right) ^{T}\boldsymbol{\pi }_{hi}^{\lambda }(\boldsymbol{\beta })\right\} \\
& =\underset{\boldsymbol{\beta }\in \mathbb{R}^{d(k+1)}}{\arg \min }\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}w_{hi}\left[ \sum_{l=1}^{d+1}\left\{ \left(
\pi _{hil}(\boldsymbol{\beta })-\frac{1+\lambda }{\lambda }\frac{\widehat{y}_{hil}}{m_{hi}}\right) \pi _{hil}^{\lambda }(\boldsymbol{\beta })\right\} \right]\end{aligned}$$which can also be obtained by solving the system of equations $\boldsymbol{u}_{\lambda }\left( \boldsymbol{\beta }\right) =\boldsymbol{0}_{d(k+1)}$ where $$\begin{aligned}
-\frac{1}{\lambda +1}\frac{\partial }{\partial \boldsymbol{\beta }}\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}w_{hi}\left\{ \left( m_{hi}\boldsymbol{\pi }_{hi}(\boldsymbol{\beta })-\frac{1+\lambda }{\lambda }\widehat{\boldsymbol{y}}_{hi}\right) ^{T}\boldsymbol{\pi }_{hi}^{\lambda }(\boldsymbol{\beta })\right\} & =\boldsymbol{0}_{d(k+1)}, \\
\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}w_{hi}m_{hi}\sum_{l=1}^{d+1}\left\{ \frac{\widehat{y}_{hil}}{m_{hi}}-\pi _{hil}(\boldsymbol{\beta })\right\} \pi
_{hil}^{\lambda -1}(\boldsymbol{\beta })\frac{\partial }{\partial
\boldsymbol{\beta }}\pi _{hil}(\boldsymbol{\beta })& =\boldsymbol{0}_{d(k+1)}, \\
\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}w_{hi}m_{hi}\left[ \frac{\partial }{\partial
\boldsymbol{\beta }}\boldsymbol{\pi }_{hi}^{T}\left( \boldsymbol{\beta }\right) \mathrm{diag}^{\lambda -1}\{\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \}\{\frac{\widehat{\boldsymbol{y}}_{hi}}{m_{hi}}-\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \}\right] & =\boldsymbol{0}_{d(k+1)}.\end{aligned}$$Now, taking into account that $$\frac{\partial \pi _{hil}(\boldsymbol{\beta })}{\partial \beta _{uv}}=x_{iu}\pi _{hil}(\boldsymbol{\beta })\left\{ \delta _{jv}-\pi _{hiv}(\boldsymbol{\beta })\right\} ,\qquad u=1,...,k,v=1,...,d,$$we get $$\begin{aligned}
\frac{\partial \boldsymbol{\pi }_{hi}^{T}(\boldsymbol{\beta })}{\partial
\boldsymbol{\beta }}& =\left( \boldsymbol{I}_{d},\boldsymbol{0}_{d}\right)
\boldsymbol{\Delta }(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right)
)\otimes \boldsymbol{x}_{i} =\boldsymbol{\Delta }^{\ast }(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) )\otimes \boldsymbol{x}_{i}, \label{der}\end{aligned}$$and hence the system of equations becomes $$\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}\left[ w_{hi}\boldsymbol{\Delta }^{\ast }(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) )\mathrm{diag}^{\lambda -1}\{\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \}\{\widehat{\boldsymbol{y}}_{hi}-m_{hi}\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \}\right] \otimes \boldsymbol{x}_{hi}=\boldsymbol{0}_{d(k+1)}.$$
Proof of Theorem \[Th1\] {#proof-of-theorem-th1 .unnumbered}
------------------------
By following formulas (3.3) and (3.4) of Ghosh and Basu (2013), it holds$$\begin{aligned}
\mathbf{\Omega}_{\lambda }\left( \boldsymbol{\beta }\right) & =\mathrm{Var}\left[
\boldsymbol{U}_{\lambda }(\boldsymbol{\beta },\boldsymbol{X})\right] , \quad
\mathbf{\Psi}_{\lambda }\left( \boldsymbol{\beta }\right) =\mathrm{E}\left[ -\frac{\partial }{\partial \boldsymbol{\beta }}\boldsymbol{U}_{\lambda }^{T}(\boldsymbol{\beta },\boldsymbol{X})\right] ,\end{aligned}$$which can be expressed as a limit of$$\begin{aligned}
\mathbf{\Omega}_{n,\lambda }\left( \boldsymbol{\beta }\right) & =\widehat{\mathrm{E}}[\mathrm{Var}[\boldsymbol{U}_{\lambda }(\boldsymbol{\beta },\boldsymbol{X})|\boldsymbol{X}]], \quad
\mathbf{\Psi}_{n,\lambda }\left( \boldsymbol{\beta }\right) =\widehat{\mathrm{E}}\left[ \mathrm{E}\left[ -\tfrac{\partial }{\partial \boldsymbol{\beta }}\boldsymbol{U}_{\lambda }^{T}(\boldsymbol{\beta },\boldsymbol{X})|\boldsymbol{X}\right] \right] ,\end{aligned}$$where the summands given in (\[hs\]) and (\[gs\]) are$$\begin{aligned}
\mathbf{\Omega}_{hi,\lambda }\left( \boldsymbol{\beta },\boldsymbol{\Sigma }_{hi}\right) & =\mathrm{Var}\left[ \boldsymbol{U}_{\lambda }(\boldsymbol{\beta },\boldsymbol{x}_{hi})\right] , \quad
\mathbf{\Psi}_{hi,\lambda }\left( \boldsymbol{\beta }\right) =\mathrm{E}\left[
-\frac{\partial }{\partial \boldsymbol{\beta }}\boldsymbol{U}_{\lambda }^{T}(\boldsymbol{\beta },\boldsymbol{x}_{hi})\right] .\end{aligned}$$With respect to $\mathbf{\Omega}_{\lambda }\left( \boldsymbol{\beta }\right) =\mathrm{Var}\left[ \boldsymbol{U}_{\lambda }(\boldsymbol{\beta },\boldsymbol{X})\right] =\mathrm{E}\left[ \mathrm{Var}\left[ \boldsymbol{U}_{\lambda }(\boldsymbol{\beta },\boldsymbol{X})|\boldsymbol{X}\right] \right] +\mathrm{Var}\left[ \mathrm{E}\left[ \boldsymbol{U}_{\lambda }(\boldsymbol{\beta },\boldsymbol{X})|\boldsymbol{X}\right] \right] $, the first summand estimated from the sample is$$\begin{aligned}
\widehat{\mathrm{E}}[\mathrm{Var}[\boldsymbol{U}_{\lambda }(\boldsymbol{\beta },\boldsymbol{X})|\boldsymbol{X}]]& =\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}\mathrm{Var}\left[ \boldsymbol{U}_{\lambda }(\boldsymbol{\beta },\boldsymbol{x}_{hi})\right] \widehat{\Pr }(\boldsymbol{X=x}_{hi}) =\frac{1}{n}\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}\mathrm{Var}\left[ \boldsymbol{U}_{\lambda }(\boldsymbol{\beta },\boldsymbol{x}_{hi})\right] ,\end{aligned}$$and the second one$$\begin{aligned}
\widehat{\mathrm{Var}}\left[ \mathrm{E}\left[ \boldsymbol{U}_{\lambda }(\boldsymbol{\beta },\boldsymbol{X})|\boldsymbol{X}\right] \right] & =\mathrm{Var}\left[ \sum_{h=1}^{H}\sum_{i=1}^{n_{h}}E\left[ \boldsymbol{U}_{\lambda }(\boldsymbol{\beta },\boldsymbol{x}_{hi})\right] \widehat{\Pr }(\boldsymbol{X=x}_{hi})\right] =\boldsymbol{0}_{d(k+1)},\end{aligned}$$for $\boldsymbol{\beta }=\boldsymbol{\beta }_{0}$, since $\mathrm{E}\left[
\boldsymbol{U}_{\lambda }(\boldsymbol{\beta }_{0},\boldsymbol{x}_{hi})\right]
=\boldsymbol{0}_{d(k+1)}$ from $\mathrm{E}[\widehat{\boldsymbol{Y}}_{hi}-m_{hi}\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }_{0}\right) ]=\boldsymbol{0}_{d(k+1)}$. On the other hand,$$\begin{aligned}
\mathrm{Var}\left[ \boldsymbol{U}_{\lambda }(\boldsymbol{\beta },\boldsymbol{x}_{hi})\right] & =\mathrm{Var}\left[ \left[ w_{hi}\boldsymbol{\Delta }^{\ast }(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) )\mathrm{diag}^{\lambda -1}\{\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right)
\}\{\widehat{\boldsymbol{Y}}_{hi}-m_{hi}\boldsymbol{\pi }_{hi}\left(
\boldsymbol{\beta }\right) \}\right] \otimes \boldsymbol{x}_{hi}\right] \\
& =\left( w_{hi}^{2}\boldsymbol{\Delta }^{\ast }(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) )\mathrm{diag}^{\lambda -1}\{\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \}\mathrm{Var}[\widehat{\boldsymbol{Y}}_{hi}]\mathrm{diag}^{\lambda -1}\{\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \}\boldsymbol{\Delta }^{\ast T}(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) )\right) \otimes
\boldsymbol{x}_{hi}\boldsymbol{x}_{hi}^{T}.\end{aligned}$$With respect to $\mathbf{\Psi}_{\lambda }\left( \boldsymbol{\beta }\right) =-\mathrm{E}\left[ \frac{\partial }{\partial \boldsymbol{\beta }}\boldsymbol{U}_{\lambda }^{T}(\boldsymbol{\beta },\boldsymbol{X})\right] $, estimated from the sample is given by$$\begin{aligned}
& \widehat{\mathrm{E}}\left[ \mathrm{E}\left[ -\tfrac{\partial }{\partial
\boldsymbol{\beta }}\boldsymbol{U}_{\lambda }^{T}(\boldsymbol{\beta },\boldsymbol{X})|\boldsymbol{X}\right] \right] \\
& =-\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}\mathrm{E}\left[ \frac{\partial }{\partial \boldsymbol{\beta }}\boldsymbol{U}_{\lambda }(\boldsymbol{\beta },\boldsymbol{x}_{hi})\right] \widehat{\Pr }(\boldsymbol{X=x}_{hi}) \\
& =-\frac{1}{n}\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}\left[ w_{hi}\frac{\partial }{\partial \boldsymbol{\beta }}\left[ \boldsymbol{\Delta }^{\ast }(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) )\mathrm{diag}^{\lambda -1}\{\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \}\right] \mathrm{E}[\widehat{\boldsymbol{Y}}_{hi}-m_{hi}\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) ]\right] ^{T}\otimes \boldsymbol{x}_{hi}^{T} \\
& +\frac{1}{n}\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}\left[ w_{hi}m_{hi}\boldsymbol{\Delta }^{\ast }(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) )\mathrm{diag}^{\lambda -1}\{\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \}\frac{\partial }{\partial \boldsymbol{\beta }}\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \right] ^{T}\otimes \boldsymbol{x}_{hi}^{T} \\
& =\frac{1}{n}\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}\left[ w_{hi}m_{hi}\boldsymbol{\Delta }^{\ast }(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) )\mathrm{diag}^{\lambda -1}\{\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \}\frac{\partial }{\partial \boldsymbol{\beta }}\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \right] ^{T}\otimes \boldsymbol{x}_{hi}^{T} \\
& =\frac{1}{n}\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}\left[ w_{hi}m_{hi}\boldsymbol{\Delta }^{\ast }(\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) )\mathrm{diag}^{\lambda -1}\{\boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \}\boldsymbol{\Delta }^{\ast T}(\boldsymbol{\pi }_{hi}\left(
\boldsymbol{\beta }\right) )\right] \otimes \boldsymbol{x}_{hi}\boldsymbol{x}_{hi}^{T},\end{aligned}$$where the third equality is again true for $\boldsymbol{\beta }=\boldsymbol{\beta }_{0}$. The derivations are omitted for $\lambda =0$, in which case similar ideas as $\lambda >0$ might be followed.
Proof of Theorem \[th:c\] {#proof-of-theorem-thc .unnumbered}
-------------------------
We have $$\begin{aligned}
Po_{W_{n}\left( \widehat{\boldsymbol{\beta }}_{\lambda ,Q}\right) }\left(
\boldsymbol{\beta }^{\ast }\right) & =\Pr \left( W_{n}\left( \widehat{\boldsymbol{\beta }}_{\lambda ,Q}\right) >\chi _{r,\alpha }^{2}\right) =\Pr
\left( n\left( l_{\widehat{\boldsymbol{\beta }}_{\lambda ,Q}}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})-l_{\boldsymbol{\beta }^{\ast }}(\boldsymbol{\beta }^{\ast })\right) >\chi _{r,\alpha }^{2}-nl_{\boldsymbol{\beta }^{\ast }}(\boldsymbol{\beta }^{\ast })\right) \\
& =\Pr \left( \sqrt{n}\left( l_{\widehat{\boldsymbol{\beta }}_{\lambda ,Q}}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})-l_{\boldsymbol{\beta }^{\ast }}(\boldsymbol{\beta }^{\ast })\right) >\frac{\chi _{r,\alpha }^{2}}{\sqrt{n}}-\sqrt{n}l_{\boldsymbol{\beta }^{\ast }}(\boldsymbol{\beta }^{\ast })\right) .\end{aligned}$$Now we are going to get the asymptotic distribution of the random variable $\sqrt{n}\left( l_{\widehat{\boldsymbol{\beta }}_{\lambda ,Q}}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})-l_{\boldsymbol{\beta }^{\ast }}(\boldsymbol{\beta }^{\ast })\right) $.
It is clear that $l_{\widehat{\boldsymbol{\beta }}_{\lambda ,Q}}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})$ and $l_{\widehat{\boldsymbol{\beta }}_{\lambda ,Q}}(\boldsymbol{\beta }^{\ast })$ have the same asymptotic distribution because $\widehat{\boldsymbol{\beta }}_{\lambda ,Q}\underset{n\rightarrow \infty }{\overset{P}{\longrightarrow }}\boldsymbol{\beta }^{\ast }.$ A first Taylor expansion of $l_{\widehat{\boldsymbol{\beta }}_{\lambda ,Q}}(\boldsymbol{\beta }^{\ast })$ at $\widehat{\boldsymbol{\beta }}_{\lambda ,Q}$ around $\boldsymbol{\beta }^{\ast }$ gives$$l_{\widehat{\boldsymbol{\beta }}_{\lambda ,Q}}(\boldsymbol{\beta }^{\ast
})-l_{\boldsymbol{\beta }^{\ast }}(\boldsymbol{\beta }^{\ast })=\left. \frac{\partial l_{\boldsymbol{\beta }}(\boldsymbol{\beta }^{\ast })}{\partial
\boldsymbol{\beta }^{T}}\right\vert _{\boldsymbol{\beta }=\boldsymbol{\beta }^{\ast }}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q}-\boldsymbol{\beta }^{\ast })+o_{p}\left( \left\Vert \widehat{\boldsymbol{\beta }}_{\lambda ,Q}-\boldsymbol{\beta }^{\ast }\right\Vert \right) .$$Therefore,$$\sqrt{n}\left( l_{\widehat{\boldsymbol{\beta }}_{\lambda ,Q}}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})-l_{\boldsymbol{\beta }^{\ast }}(\boldsymbol{\beta }^{\ast })\right) \underset{n\rightarrow \infty }{\overset{L}{\longrightarrow }}N\left( 0,\sigma ^{2}\left( \boldsymbol{\beta }^{\ast
}\right) \right)$$being $$\sigma ^{2}\left( \boldsymbol{\beta }^{\ast }\right) =\left. \frac{\partial
l_{\boldsymbol{\beta }}(\boldsymbol{\beta }^{\ast })}{\partial \boldsymbol{\beta }^{T}}\right\vert _{\boldsymbol{\beta }=\boldsymbol{\beta }^{\ast }}\widehat{\mathbf{Q}}_{n,\lambda }\left( \boldsymbol{\beta }^{\ast }\right)
\left. \frac{\partial l_{\boldsymbol{\beta }}(\boldsymbol{\beta }^{\ast })}{\partial \boldsymbol{\beta }}\right\vert _{\boldsymbol{\beta }=\boldsymbol{\beta }^{\ast }}.$$
Proof of Corollary \[cor1\] {#proof-of-corollary-cor1 .unnumbered}
---------------------------
It is based on the weak consistency of $\widehat{\boldsymbol{\beta }}_{\lambda ,Q}$ as well as in the continuity with respect to $\boldsymbol{\beta }$ of the elements of different matrices. Based on Theorem \[Th1\], it holds $\lim_{n\rightarrow \infty }\mathrm{E}[\widehat{\boldsymbol{\beta }}_{\lambda ,Q}]=\boldsymbol{0}_{d(k+1)}$ and $\lim_{n\rightarrow \infty }\mathrm{Var}[\widehat{\boldsymbol{\beta }}_{\lambda ,Q}]=\boldsymbol{O}_{d(k+1)\times d(k+1)}$ and hence $\widehat{\boldsymbol{\beta }}_{\lambda ,Q}\underset{n\rightarrow \infty }{\overset{P}{\longrightarrow }}\boldsymbol{\beta }_{0}$.
Proof of Theorem \[Th1b\] {#proof-of-theorem-th1b .unnumbered}
-------------------------
Let $$\begin{aligned}
\widehat{\mathbf{\Omega}}_{n,\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda
,Q})& =\widehat{\mathrm{E}}[\mathrm{Var}[\boldsymbol{U}_{\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda ,Q},\boldsymbol{X})|\boldsymbol{X}]] \\
& =\widehat{\mathrm{E}}[\mathrm{E}[\boldsymbol{U}_{\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda ,Q},\boldsymbol{X})\boldsymbol{U}_{\lambda
}^{T}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q},\boldsymbol{X})|\boldsymbol{X}]-\mathrm{E}[\boldsymbol{U}_{\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda ,Q},\boldsymbol{X})|\boldsymbol{X}]\mathrm{E}[\boldsymbol{U}_{\lambda }^{T}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q},\boldsymbol{X})|\boldsymbol{X}]]\end{aligned}$$be another possible consistent estimator of $\mathbf{\Omega}_{\lambda }\left(
\boldsymbol{\beta }\right) $, alternative to $\mathbf{\Omega}_{n,\lambda }\left(
\boldsymbol{\beta }\right) $. Since $$\widehat{\mathrm{E}}\left[ \mathrm{E}[\boldsymbol{U}_{\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda ,Q},\boldsymbol{X})|\boldsymbol{X}]\mathrm{E}[\boldsymbol{U}_{\lambda }^{T}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q},\boldsymbol{X})|\boldsymbol{X}]]\right] =\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}\mathrm{E}[\boldsymbol{U}_{\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda
,Q},\boldsymbol{x}_{hi})]\mathrm{E}[\boldsymbol{U}_{\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda ,Q},\boldsymbol{x}_{hi})]\widehat{\Pr }(\boldsymbol{X=x}_{hi}),$$and $\mathrm{E}[\boldsymbol{U}_{\lambda }(\widehat{\boldsymbol{\beta }}_{\lambda ,Q},\boldsymbol{X})|\boldsymbol{x}_{hi}]=0$, for being $\mathrm{E}[\widehat{\boldsymbol{Y}}_{hi}-m_{hi}\boldsymbol{\pi }_{hi}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})]=\boldsymbol{0}_{k+1}$, for all $(h,i)\in
\{1,...,H\}\times \{1,...,m_{hi}\}$ according to the PLR model with complex design, it holds$$\begin{aligned}
\widehat{\mathbf{\Omega}}_{n,\lambda }(\boldsymbol{\beta })& =\widehat{\mathrm{E}}[\mathrm{E}[\boldsymbol{U}_{\lambda }(\boldsymbol{\beta },\boldsymbol{X})\boldsymbol{U}_{\lambda }^{T}(\boldsymbol{\beta },\boldsymbol{X})|\boldsymbol{X}] =\frac{1}{n}\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}\boldsymbol{U}_{\lambda }(\boldsymbol{\beta },\boldsymbol{x}_{hi})\boldsymbol{U}_{\lambda }^{T}(\boldsymbol{\beta },\boldsymbol{x}_{hi})\end{aligned}$$is another possible consistent estimator of $\mathbf{\Omega}_{\lambda }\left(
\boldsymbol{\beta }\right) $, alternative to $\mathbf{\Omega}_{n,\lambda }\left(
\boldsymbol{\beta }\right) $.
Proof of Corollary \[cor2\] {#proof-of-corollary-cor2 .unnumbered}
----------------------------
Section a) is straightforward considering the expressions of matrix $\mathbf{G}_{n,\lambda }(\boldsymbol{\beta })$ and the related ones. For section b) we consider the vector$$\boldsymbol{Z}_{hi}^{\ast }\left( \boldsymbol{\beta }\right) =\sqrt{m}\Delta
^{-1/2}\left( \boldsymbol{\pi }_{hi}^{\ast }\left( \boldsymbol{\beta }\right) \right) (\tfrac{\widehat{\boldsymbol{Y}}_{hi}^{\ast }}{m}-\boldsymbol{\pi }_{hi}^{\ast }\left( \beta \right) ).$$Taking into account that $
\mathrm{Var}[\widehat{\boldsymbol{Y}}_{hi}^{\ast }]=m\nu \boldsymbol{\Delta }
\left( \boldsymbol{\pi }_{hi}^{\ast }\left( \beta \right) \right) ,
$ we have $$\mathrm{E}\left[ \boldsymbol{Z}_{hi}^{\ast }\left( \boldsymbol{\beta }\right) \right] =\boldsymbol{0}\text{ and }\mathrm{Var}\left[ \boldsymbol{Z}_{hi}^{\ast }\left( \boldsymbol{\beta }\right) \right] =\nu \boldsymbol{I}_{d},$$for $h=1,...,H$ and $i=1,...,n_{h}$. An unbiased estimator of $\mathrm{Var}\left[ \boldsymbol{Z}_{hi}^{\ast }\left( \boldsymbol{\beta }\right) \right] $ is $$\widehat{\mathrm{Var}}\left[ \boldsymbol{Z}_{hi}^{\ast }(\boldsymbol{\beta })\right] =\frac{1}{n}\sum\limits_{h=1}^{H}\sum\limits_{i=1}^{n_{h}}\boldsymbol{Z}_{hi}^{\ast }(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})\boldsymbol{Z}_{hi}^{\ast T}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q}),$$i.e. $\mathrm{E}\left[ \widehat{\mathrm{Var}}\left[ \boldsymbol{Z}_{hi}^{\ast }\left( \boldsymbol{\beta }\right) \right] \right] =\mathrm{Var}\left[ \boldsymbol{Z}_{hi}^{\ast }\left( \boldsymbol{\beta }\right) \right] $ for which the trace is$$\begin{aligned}
\mathrm{E}\left[ \mathrm{trace}\left( \widehat{\mathrm{Var}}\left[
\boldsymbol{Z}_{hi}^{\ast }(\boldsymbol{\beta })\right] \right) \right] & =\mathrm{trace}\left( \mathrm{Var}\left[ \boldsymbol{Z}_{hi}^{\ast }\left(
\boldsymbol{\beta }\right) \right] \right) , \\
\mathrm{E}\left[ \frac{1}{n}\sum\limits_{h=1}^{H}\sum\limits_{i=1}^{n_{h}}\boldsymbol{Z}_{hi}^{\ast }(\boldsymbol{\beta })\boldsymbol{Z}_{hi}^{\ast T}(\boldsymbol{\beta })\right] & =\nu d, \\
\mathrm{E}\left[ \frac{1}{nd}\sum\limits_{h=1}^{H}\sum\limits_{i=1}^{n_{h}}\boldsymbol{Z}_{hi}^{\ast }(\boldsymbol{\beta })\boldsymbol{Z}_{hi}^{\ast T}(\boldsymbol{\beta })\right] & =\nu .\end{aligned}$$Since $\widehat{\boldsymbol{\beta }}_{\lambda ,Q}$ is consistent of $\boldsymbol{\beta }$, it holds that $\frac{1}{nd}\sum\limits_{h=1}^{H}\sum\limits_{i=1}^{n_{h}}\boldsymbol{Z}_{hi}^{\ast }(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})\boldsymbol{Z}_{hi}^{\ast T}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})$ is consistent of $\nu $, but$$\frac{1}{nd}\sum\limits_{h=1}^{H}\sum\limits_{i=1}^{n_{h}}\boldsymbol{Z}_{hi}^{\ast }(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})\boldsymbol{Z}_{hi}^{\ast T}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})=\frac{1}{nd}\sum\limits_{h=1}^{H}\sum\limits_{i=1}^{n_{h}}(\widehat{\boldsymbol{Y}}_{hi}^{\ast }-m\boldsymbol{\pi }_{hi}^{\ast }(\widehat{\boldsymbol{\beta }}_{\lambda ,Q}))^{T}\frac{1}{m}\Delta ^{-1}\left( \boldsymbol{\pi }_{hi}^{\ast }\left( \boldsymbol{\beta }\right) \right) (\widehat{\boldsymbol{Y}}_{hi}^{\ast }-m\boldsymbol{\pi }_{hi}^{\ast }(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})),$$is in principle different in shape in comparison with (\[vM\]) with$$\sum_{h=1}^{H}\sum_{i=1}^{n_{h}}\sum_{j=1}^{d+1}\frac{(\widehat{Y}_{hij}-m\pi _{hij}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q}))^{2}}{m\pi
_{hij}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})}=\sum\limits_{h=1}^{H}\sum\limits_{i=1}^{n_{h}}(\widehat{\boldsymbol{Y}}_{hi}-m\boldsymbol{\pi }_{hi}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q}))^{T}\frac{1}{m}\Delta
^{-}\left( \boldsymbol{\pi }_{hi}^{\ast }\left( \boldsymbol{\beta }\right)
\right) (\widehat{\boldsymbol{Y}}_{hi}-m\boldsymbol{\pi }_{hi}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})),$$where $\Delta ^{-}\left( \boldsymbol{\pi }_{hi}^{\ast }\left( \boldsymbol{\beta }\right) \right) =\mathrm{diag}^{-1}\left( \boldsymbol{\pi }_{hi}\left( \boldsymbol{\beta }\right) \right) $. To assure that both expressions are equivalent we count with the result related to invariance of quadratic forms with generalized variances, Lemma 1a, of Moore (1977), from which is concluded that$$\begin{aligned}
& (\widehat{\boldsymbol{Y}}_{hi}-m\boldsymbol{\pi }_{hi}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q}))^{T}\frac{1}{m}\Delta ^{-}\left(
\boldsymbol{\pi }_{hi}^{\ast }\left( \boldsymbol{\beta }\right) \right) (\widehat{\boldsymbol{Y}}_{hi}-m\boldsymbol{\pi }_{hi}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})) \\
& =(\widehat{\boldsymbol{Y}}_{hi}^{\ast }-m\boldsymbol{\pi }_{hi}^{\ast }(\widehat{\boldsymbol{\beta }}_{\lambda ,Q}))^{T}\frac{1}{m}\Delta
^{-1}\left( \boldsymbol{\pi }_{hi}^{\ast }\left( \boldsymbol{\beta }\right)
\right) (\widehat{\boldsymbol{Y}}_{hi}^{\ast }-m\boldsymbol{\pi }_{hi}^{\ast
}(\widehat{\boldsymbol{\beta }}_{\lambda ,Q})).\end{aligned}$$
Proof of Theorem \[th2\] {#proof-of-theorem-th2 .unnumbered}
------------------------
The details are omitted for being very similar to Theorem \[Th1\] conditioning within each stratum, i.e.$$\begin{aligned}
\mathbf{\Omega}_{i,\lambda }^{(h)}\left( \boldsymbol{\beta },\boldsymbol{\Sigma }_{hi}\right) & =\widehat{\mathrm{E}}[\mathrm{Var}[\boldsymbol{U}_{\lambda }(\boldsymbol{\beta },\boldsymbol{X}_{h})|\boldsymbol{X}_{h}]], \quad
\mathbf{\Omega}_{\lambda }^{(h)}\left( \boldsymbol{\beta }\right) =\mathrm{Var}\left[ \boldsymbol{U}_{\lambda }(\boldsymbol{\beta },\boldsymbol{X}_{h})\right] , \\
\mathbf{\Psi}_{i,\lambda }^{(h)}\left( \boldsymbol{\beta }\right) & =-\widehat{\mathrm{E}}\left[ \mathrm{E}\left[ \tfrac{\partial }{\partial \boldsymbol{\beta }}\boldsymbol{U}_{\lambda }^{T}(\boldsymbol{\beta },\boldsymbol{X}_{h})|\boldsymbol{X}_{h}\right] \right] , \quad
\mathbf{\Psi}_{\lambda }^{(h)}\left( \boldsymbol{\beta }\right) =-\mathrm{E}\left[ \frac{\partial }{\partial \boldsymbol{\beta }}\boldsymbol{U}_{\lambda
}^{T}(\boldsymbol{\beta },\boldsymbol{X}_{h})\right] .\end{aligned}$$
Proof of Theorem \[th:wald\] {#proof-of-theorem-thwald .unnumbered}
----------------------------
We have $$\begin{aligned}
\boldsymbol{M}^{T}\widehat{\boldsymbol{\beta }}_{\lambda ,Q}-\boldsymbol{l}&
\boldsymbol{=M}^{T}\text{ }\boldsymbol{\mathbf{\beta }}_{n}-l\boldsymbol{+M}^{T}\left( \widehat{\boldsymbol{\beta }}_{\lambda ,Q}-\text{ }\boldsymbol{\mathbf{\beta }}_{n}\right) \\
& =\boldsymbol{M}^{T}\text{ }\boldsymbol{\mathbf{\beta }}_{0}+\boldsymbol{M^{T}}n^{-1/2}\boldsymbol{\boldsymbol{d}}-\boldsymbol{l+L^{T}\left( \widehat{\boldsymbol{\beta }}_{\lambda ,Q}-\text{ }\boldsymbol{\mathbf{\beta }}_{n}\right) } \\
& \boldsymbol{=M}\boldsymbol{^{T}}n^{-1/2}\boldsymbol{\boldsymbol{d+M}^{T}\left( \widehat{\boldsymbol{\beta }}_{\lambda ,Q}-\text{ }\boldsymbol{\mathbf{\beta }}_{n}\right) .}\end{aligned}$$Therefore,$$\boldsymbol{M}^{T}\text{ }\widehat{\boldsymbol{\beta }}_{\lambda ,Q}-\boldsymbol{l=M^{T}}n^{-1/2}\boldsymbol{\boldsymbol{d+M}^{T}\left( \widehat{\boldsymbol{\beta }}_{\lambda ,Q}-\text{ }\boldsymbol{\mathbf{\beta }}_{n}\right) .}$$We know, under $H_{1,n}$ that $$\sqrt{n}\left( \widehat{\boldsymbol{\beta }}_{\lambda ,Q}-\boldsymbol{\boldsymbol{\mathbf{\beta }}_{n}}\right) \underset{n\longrightarrow \infty }{\overset{\mathcal{L}}{\longrightarrow }}\mathcal{N}\left( \mathbf{0},\text{ }\mathbf{Q}_{\lambda }\boldsymbol{\left( \boldsymbol{\beta }_{0}\right) }\right)$$and $$\sqrt{n}\left( \boldsymbol{M}^{T}\text{ }\widehat{\boldsymbol{\beta }}_{\lambda ,Q}-\boldsymbol{c}\right) \underset{n\longrightarrow \infty }{\overset{\mathcal{L}}{\longrightarrow }}\mathcal{N}\left( M\boldsymbol{^{T}\boldsymbol{d}},\text{ }M\boldsymbol{^{T}\mathbf{\Psi}_{\lambda }\left(
\boldsymbol{\beta }_{0}\right) M}\right) .$$The Wald-type test statistics can be written as the quadratic form $
W_{n}\left( \widehat{\boldsymbol{\beta }}_{\lambda ,Q}\right) =\boldsymbol{Z}
^{T}\boldsymbol{Z}
$ with $$\boldsymbol{Z}=\sqrt{n}\left[ M\boldsymbol{^{T}\mathbf{\Psi}_{\lambda }\left(
\boldsymbol{\beta }_{0}\right) M}\right] ^{-1/2}\left( \boldsymbol{M}^{T}\text{ }\widehat{\boldsymbol{\beta }}_{\lambda ,Q}-\boldsymbol{l}\right)$$and $$\boldsymbol{Z}\underset{n\longrightarrow \infty }{\overset{\mathcal{L}}{\longrightarrow }}\mathcal{N}\left( \left[ M\boldsymbol{^{T}\mathbf{\Psi}_{\lambda }\left( \boldsymbol{\beta }_{0}\right) M}\right] ^{-1/2}\boldsymbol{M^{T}d,I}_{r\times r}\right) ,$$where $\boldsymbol{I}$ is the identity $r\times r$ matrix. The application of the previous gives i). The noncentrality parameter is $
\boldsymbol{d}^{T}\boldsymbol{M}\left[ M\boldsymbol{^{T}\mathbf{\Psi}_{\lambda
}\left( \boldsymbol{\beta }_{0}\right) M}\right] ^{-1}\boldsymbol{M^{T}d.}
$ Result ii) follows using relation (\[G\]).
[99]{} Agresti, A. (2002). *Categorical Data Analysis* (Second Edition). John Wiley & Sons.
Alonso-Revenga, J. M., Martín, N. and Pardo, L. (2017). New improved estimators for overdispersion in models with clustered multinomial data and unequal cluster sizes, *Statistics and Computing* **27**, 193-217.
Basu, A., Harris, I. R., Hjort, N. L. and Jones, M. C. (1998). Robust and efficient estimation by minimizing a density power divergence. *Biometrika*, **85**, 549–559
Basu, A. Ghosh, A. Mandal, N. Martin and L. Pardo (2017). A Wald-type test statistic for testing linear hypothesis in logistic regression models based on minimum density power divergence estimator. *Electon. J. Stat.*, **11**, 2741–2772
Basu, A. Ghosh, A., Martin and L. Pardo (2018). Robust Wald-type tests for non-homogeneous observations based on the minimum density power divergence estimator. *Metrika*, **81(5)**, 493–522.
Beran, Rudolf (1977). Minimum Hellinger Distance Estimates for Parametric Models. *Ann. Statist*. **5**, no. 3, 445–463.
Binder, D. A. (1983). On the variance of asymptotically normal estimators from complex surveys. *International Statistical Review, **51**, 279–292.*
Castilla, E., Martin, N. and Pardo, L. (2018). Minimum phi-divergence estimators for multinomial logistic regression with complex sample design. *Advances in Statistical Analysis*, **102**, 381–411.
Castilla, E., Ghosh, A., Martin, N. and Pardo, L. (2019). New robust statistical procedures for polytomous logistic regression models. *Biometrics*. **74**, 1282-1291.
Engel, J. (1988). Polytomous logistic regression. *Statistica Neerlandica,* **42,** 233–252.
Ghosh, A., and Basu, A. (2013). Robust estimation for independent non-homogeneous observations using density power divergence with applications to linear regression. *Electronic Journal of Statistics*, **7**, 2420–2456.
Ghosh, A. and Basu, A. (2015). Robust estimation for non-homogeneous data and the selection of the optimal tuning parameter: the density power divergence approach. *Journal of applied statitsics*, **42**, 2056–2072.
Ghosh, A., and Basu, A. (2016). Robust Estimation in Generalized Linear Models : The Density Power Divergence Approach. *TEST*, **25**, 269–290.
Ghosh, A., and Basu, A. (2018). Robust Bounded Influence Tests for Independent but Non-Homogeneous Observations. *Statistica Sinica*, **28**, 1133–1155.
Gupta, A. K., Kasturiratna, D., Nguyen, T. and Pardo, L. (2006). A new family of BAN estimators for polytomous logistic regression models based on density power divergence measures. Statistical Methods & Applications, **15**, 159–176.
Gupta, A. K.; Nguyen, T.; Pardo, L. (2008). Residuals for polytomous logistic regression models based on density power divergences test statistics. *Statistics*, **42**, 495–514.
Hampel, F. R., Ronchetti, E., Rousseeuw, P. J., and Stahel W.(1986). Robust Statistics: The Approach Based on Influence Functions. *New York, USA: John Wiley & Sons*.
Huber, P. J. (1983). Minimax aspects of bounded-influence regression (with discussion). *Journal of the American Statistical Association*, **69**, 383-393.
Lesaffre, E.and Albert, A. (1989). Multiple-group logistic regression diagnostic. *Applied Statistics,* **38,** 425–440.
Lindsay, Bruce G. (1994). Efficiency Versus Robustness: The Case for Minimum Hellinger Distance and Related Methods. *Ann. Statist*. **22**, no. 2, 1081–1114.
Liu, I. and Agresti, A. (2005). The analysis of ordered categorical data: an overview and a survey of recent developments. With discussion and a rejoinder by the authors. *Test*, **14**, 1–73.
McCullagh, P. (1980). Regression models for ordinary data. *Journal of the Royal Statistical Society*-Series B, **42,** 109–142.
Morel, G. (1989). Logistic regression under Complex Survey Designs. *Survey Methodology, **15**, 203–223.*
Morel, J. G. and Koehler, K. J. (1995). A One‐Step Gauss‐Newton Estimator for Modelling Categorical Data with Extraneous Variation. *Journal of the Royal Statistical Society: Series C*, **44(2)**, 187–200.
Morel, G. and Neerchal, N. K. (2012). Overdispersion Models in SAS. SAS Institute.
Pardo, L. (2005). *Statistical Inference Based on Divergence Measures. Statistics: Texbooks and Monographs.* Chapman & Hall/CRC, New York.
Raim, A. M. , Neerchal, N. K. and Morel, J. G. (2015). Modeling overdispersion in R. Technical Report HPCI-2015-1 UMBCH High Performance Computing Facility, University of Maryland, Baltimore Country.
Roberts, G., Rao, J.N.K. and Kumer, S. (1987). Logistic Regression Analysis of Sample Survey Data, *Biometrika*, **74**, 1–12.
Toma, A. (2007). Minimum Hellinger distance estimators for some multivariate models: influence functions and breakdown point results. *C. R. Acad. Sci. Paris*, Ser. I 345, 353–358.
Warwick, J. and Jones, M. C. (2005). Choosing a robustness tuning parameter. *Journal of Statistical Computation and Simulation*, **75**, 581–588.
Wedderburn, R. W. M. (1974). Quasi-likelihood functions, generalized linear models, and the Gauss-Newton method. *Biometrika*, **61**, 439–447.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
The family of superconducting fullerides (NH$_3$)$_x$NaK$_2$C$_{60}$ shows an anomalous correlation between $T_c$ and lattice parameter. To better understand the origin of this anomaly we have studied a representative $x=0.75$ compound using SQUID magnetometry and $\mu$SR spectroscopy.
The lower critical field $H_{c1}$, measured by the trapped magnetization method, is less than 1 G, a very small value as compared with that of other fullerides. Muon spin depolarization in the superconducting phase shows also quite small local field inhomogeneities, of the order of those arising from nuclear dipolar fields. On the other hand, the 40 T value for $H_{c2}$, as extracted from magnetometry data, is comparable to that of other fullerides. We show that these observations cannot be rationalized within the framework of the Ginzburg-Landau theory of superconductivity. Instead, the anomalous magnetic properties could be interpreted taking into account the role played by polaronic instabilities in this material.
author:
- 'M. Riccò'
- 'T. Shiroka'
- 'E. Zannoni'
- 'F. Barbieri'
- 'C. Bucci'
- 'F. Bolzoni'
bibliography:
- 'ricco.bib'
title: '$\mu$SR and SQUID Investigation of Superconductivity in (NH$_3$)$_{0.75}$NaK$_2$C$_{60}$'
---
Introduction {#sec:intro}
============
In fullerene based superconductors ammonia intercalates as a neutral molecule, without interacting with the host electronic system and preserving the superconducting properties of the material. Since it acts simply as a molecular spacer, a change in lattice parameters and an increase of unit cell volume are often observed[@Zhou93]. As a consequence, the $t_{1u}$ conduction band of the fullerene compound narrows and its density of states at the Fermi level increases, thus determining an increment in the superconducting transition temperature. An example of this mechanism has been reported for Na$_2$CsC$_{60}$[@Zhou93] which gives (NH$_3$)$_4$Na$_2$CsC$_{60}$ after ammoniation, with a conspicuous increase in transition temperature from 10.5 to 29.6 K.
The intercalation of ammonia can also induce a metal-to-insulator transition as e.g. in NH$_3$K$_3$C$_{60}$,[@Iwasa96; @Kitano02] where the superconductivity of the ammoniated compound can be restored only after the application of external pressure.[@Zhou95]
We will deal here with a family of fullerides such as (NH$_3$)$_x$NaK$_2$C$_{60}$ and (NH$_3$)$_x$NaRb$_2$C$_{60}$, whose precursors NaK$_2$C$_{60}$ and NaRb$_2$C$_{60}$ cannot exist as a single phase in normal conditions, but become stable only as ammoniated compounds.[@Shimoda96] X-ray diffraction in these systems shows that the NH$_3$-Na groups occupy the large octahedral sites[@Shimoda96] with a consequent off-centering of the Na$^+$ ions. Since their discovery, these compounds have revealed puzzling features concerning the relation between the superconducting transition temperature both with the lattice parameter as well as with the density of states at the Fermi level. Indeed, the progressive removal of NH$_3$, accomplished by pumping on the sample above room temperature, results in a *decrease* of the lattice parameters accompanied by an *increase* of the superconducting transition temperature, a trend opposite to that observed in (NH$_3$)$_4$Na$_2$CsC$_{60}$. In addition, we have recently shown[@Ricco01] that in a series of (NH$_3$)$_x$NaK$_2$C$_{60}$ compounds (with $0.5<x<0.8$) the Pauli-Landau spin susceptibility yields a *lower* density of states at the Fermi level in compounds having a *higher* $T_c$, in contrast with BCS or Migdal-Eliashberg predictions.
It is therefore interesting to ask whether these anomalies are related to essential differences in the nature of superconductivity in [(NH$_3$)$_{0.75}$NaK$_2$C$_{60}$]{} with respect to most common superconducting fullerides. A possible non-conventional nature of the superconductivity in this system would manifest itself in the values of the critical parameters of its superconducting phase: the London penetration depth $\lambda$ and the lower and upper critical fields, $H_{c1}$ and $H_{c2}$ respectively. Such fundamental parameters have never so far been measured and we will tackle the problem by providing their values and discussing the implications.
From the available data in the literature, the other so-called “normal” fullerene based superconductors appear to be extreme type-II superconductors[@Holczer91; @Sparn92] characterized by $\kappa = \lambda/\xi \gg 1$, where $\kappa$ is the Ginzburg-Landau parameter. Our results will be first analyzed in this framework and, in case of discrepancies, alternative suggestions will be offered.
When $\lambda \gg \xi$, both of these fundamental lengths can be easily extracted from measurements of the lower ($H_{c1}$) and the upper ($H_{c2}$) critical magnetic fields or, more precisely, from their extrapolated values at zero temperature. Roughly speaking, $H_{c2}$, the field at which the transition from the superconducting to the normal state occurs, corresponds to the field at which one quantum of magnetic flux $\Phi_0=hc/2e \simeq 2 \times 10^{-7}$ G cm$^2$ extends over the coherence area of an electron pair, so that: $$\label{eq:hc2}
H_{c2}(0) = \frac{\Phi_0}{2\pi\xi^2} ,$$ from which the Ginzburg-Landau coherence length $\xi$ can be determined. On the other hand, the knowledge of $H_{c1}$, the field at which the magnetic flux starts to penetrate the sample, allows the computation of the penetration depth $\lambda$ by using the well known equation (valid for $\kappa \gg 1$): $$\label{eq:hc1}
H_{c1}(0) = \frac{\Phi_0}{4\pi\lambda^2}\ln\kappa .$$
In addition to independent measurements of $H_{c1}$ and $H_{c2}$, as obtained by standard SQUID magnetometry, we will provide also the value for $\lambda$, which yields a stringent check of the validity of eq. (\[eq:hc2\]) and (\[eq:hc1\]). Since $\lambda$ represents the transverse extension of the vortices in the Abrikosov intermediate phase, its value can be also determined from the local magnetic field distribution. It is well known that Muon Spin Rotation ($\mu$SR) gives reasonable estimates of $\lambda$, even for irregular flux line lattices. Details on SQUID and $\mu$SR measurements are given in the next experimental section.
Experimental {#sec:experiment}
============
The samples were prepared following the procedures outlined in Ref. . Stoichiometric amounts of alkali metals (Aldrich, 99.95%) and C$_{60}$ (Southern Chem. 99.5%) were dissolved in anhydrous ammonia (Aldrich, 99.99+%) at 230 K. After the reaction had taken place, the temperature was slowly increased until the ammonia was completely evaporated. The successive pumping at 120$^\circ$C for 30 minutes yielded the compound [(NH$_3$)$_{0.75}$NaK$_2$C$_{60}$]{}; the sample was then annealed at 100$^\circ$C for 10 days. Its transition temperature was $T_c = 12$ K and it showed a 20% superconducting fraction, indicative of bulk superconductivity. The ammonia concentration $x=0.75$ was determined from the superconducting transition temperature by interpolating the $T_c-x$ data reported in Ref. .
Evaluation of the granulometry of the samples was performed using a scanning electron microscope (SEM), whose micrographs indicate an average size of the particles $d \sim 2$ $\mu$m.
DC magnetometry measurements were performed with a Quantum Design SQUID magnetometer equipped with a home built Helmholtz cube which surrounded the whole instrument body, thus allowing a reduction of the residual field on the sample to less than 2 mG. The SQUID superconducting magnet was cooled from RT to liquid He temperature in zero external field. For a good thermal contact even at low temperatures the sample was sealed in a long quartz tube under 1 mbar of He atmosphere. The sample was suspended in the middle of the tube, whose length was so chosen as to have always a tube portion face the SQUID coils, even when the sample had to move in and out of them during the magnetic moment measurement. This expedient allowed an accurate subtraction of the quartz diamagnetic contribution.
$\mu$SR, which measures the spin precession of implanted muons, is very sensitive to local magnetic fields and therefore it constitutes a valuable technique for our purposes. Indeed, when the applied transverse field exceeds $H_{c1}$, the field distribution of the flux line lattice will damp the muon spin precession signal, hence the penetration depth can be readily extracted[@MacFarlane98] from the damping rate.
In common metals and superconductors all the implanted muons usually sit interstitially and, being screened by conduction electrons, they will not form any paramagnetic bound state (muonium). Hence, their precession frequency in an external magnetic field remains that of a free particle (diamagnetic muon). In C$_{60}$ based superconductors, besides this major component there is an additional part of implanted muons (typically 10–20%) which will form endohedral muonium (located inside a C$_{60}$ molecule), whose precession frequency is much higher than that of diamagnetic muon. Here we are interested only in the majority of muons that come at rest in the *fcc* lattice interstices (the precise location is not well known) and precess as diamagnetic muons, i.e. with a gyromagnetic ratio $\gamma_{\mu}=13.55$ kHz/G.
When a superconductor is in the intermediate state, i.e. $\mu_0H_{c1}<B<\mu_0H_{c2}$, the muon precession signal will be damped by the inhomogeneous magnetic field distribution of the vortices. The expected damping profile (or the corresponding lineshape in Fourier space) in the case of a triangular flux-line lattice has been computed[@Brandt88] and, for single crystals, also successfully measured. Unlike single crystals, polycrystalline or powder materials exhibit a smeared out magnetic field distribution, with the consequence that the $\mu$SR line will assume a Gaussian shape ($\sigma_{\text{sc}}\sim0.1$–0.6 $\mu$s$^{-1}$ for fullerides) below $T_c$.[@MacFarlane98] The established relation between the $\mu$SR damping rate $\sigma_{\text{sc}(0)}$ and the internal field rms deviation $\Delta B$ is given by $\Delta B = \sigma_{\text{sc}}(0)/2\pi\gamma_{\mu}$. Once $\Delta B$ is known from an experiment that measures $\sigma_{\text{sc}}(0)$ in the appropriate intermediate field range $\mu_0H_{c1}<B<\mu_0H_{c2}$, the penetration depth $\lambda$ is given by:[@Brandt88] $$\label{eq:lamdB}
\lambda \approx 3.71\times 10^{-3}\cdot\left[\frac{\Phi_0^2}{(\Delta B)^2}\right]^{1/4}.$$
$\mu$SR experiments were performed on the EMU spectrometer of the ISIS Facility (Rutherford Laboratory, UK). The pulsed nature of the muon beam sets an upper frequency cut-off that prevents the direct observation of high muonium frequencies. Nevertheless, the interesting fraction of diamagnetic muons will precess well within the pass-band of the spectrometer and therefore could be readily measured. In our case, the sample ($\sim 500$ mg) was pressed inside an air tight aluminum cell equipped with a thin (75 $\mu$m) Kapton window. A pure silver foil was put both directly behind the sample and around the cell window to make it easy to subtract the signal coming from the sample holder.
SQUID Measurements {#sec:squid}
==================
Lower critical field
--------------------
In spite of many previous measurements of the lower critical field by SQUID magnetometry in fullerides,[@Sparn92; @Politis92a; @Politis92b; @Holczer91; @Buntar92; @Buntar95a; @Buntar95b; @Buntar96; @Buntar97; @Buntar98] its precise determination is still controversial due to considerable experimental difficulties. The simplest way to measure $H_{c1}$ consists in observing the field at which the $M=M(H)$ curve starts to deviate from linearity. Unfortunately data on fullerides never show a good linearity and this brings to an overestimate of the $H_{c1}$ value. Alternative methods based on Bean’s critical state model[@Bean62] give quite different values.[@Buntar95a] Recently it was shown[@Buntar96] that the measurement of the trapped magnetization in the intermediate phase yields a more reliable determination of $H_{c1}$. The procedure is as follows: first the sample is cooled in zero field from above $T_c$ and its initial magnetic moment ($M_1$) is measured (ideally it should be zero, but a residue always exists). Then a magnetic field $H$ is applied for at least 30 s and, after switching it off, the final moment ($M_2$) is measured. When the applied field exceeds a threshold value $H_{\text{thr}}$ the magnetic flux is trapped inside the sample. This trapped magnetization, given by the difference $M_2-M_1$, is then plotted against the applied field $H$, as shown in Figure \[fig:hc1\_trapped\] for measurements on [(NH$_3$)$_{0.75}$NaK$_2$C$_{60}$]{} performed at 7 K. The increase of the trapped magnetization above $H_{\text{thr}}$, which follows a linear behavior at all the investigated temperatures, allows a much more precise determination of $H_{c1}$ than alternative procedures. The fitted value for trapping onset (0.51 G for the linear fit shown in Figure \[fig:hc1\_trapped\]) is related to the lower critical field value by $H_{c1} = H_{\text{thr}}/(1-n)$, where $n$ is the demagnetization factor. Preliminary measurements on powdered samples gave essentially the same results, although the trapped magnetization values were rather scattered around the (same) fit line (see Figure \[fig:hc1\_trapped\]). The use of (weakly compressed) pellets remarkably reduced the spread without an appreciable change in $H_{\text{thr}}$, confirming that the demagnetization factor of a set of independent spheres ($n=1/3$), correctly adopted for powders,[@Buntar96] is appropriate also for pellets.
![\[fig:hc1\_trapped\]The trapped magnetization $M_2-M_1$ as a function of the applied field $H$ in [(NH$_3$)$_{0.75}$NaK$_2$C$_{60}$]{}. The lower critical field $H_{c1}$ is determined from the onset of the magnetization which starts is trapped at $H=H_{\text{thr}}$.](fig1.eps){width="45.00000%"}
The results of several $H_{c1}$ measurements at different temperatures are illustrated in Figure \[fig:hc1\_T\]. The zero temperature value $H_{c1}(0)=0.87$ G was then extrapolated from a parabolic fit (BCS weak coupling gave a similar result and experimental errors do not allow to distinguish between the different behaviors).
![\[fig:hc1\_T\]Measured $H_{c1}$ values as a function of temperature. The extrapolation of the parabolic fit at $T=0$ yields $H_{c1}(0)=0.87(9)$ G.](fig2.eps){width="45.00000%"}
The striking feature about the $H_{c1}(0)$ is its low value with respect to that of other fullerides, shown in Table \[tab:fullerides\] for a comparison. Even an unphysical linear fit of the data would at most yield an extrapolated value $H_{c1}(0)\sim 1.2$ which still is one order of magnitude smaller than $H_{c1}(0)$ for the quoted fullerides. The obvious suspicion that the powdered nature of the sample affects the measured values was addressed in a previous experiment, reported in Ref. , where a powdered RbCs$_2$C$_{60}$ sample and a single crystal were compared and found to have the same lower critical field value. Hence, the measured low $H_{c1}$ value given here will be considered an intrinsic feature of the compound.
Compound $H_{c1}(T=0)$ \[G\] Ref.
--------------------------------------- --------------------- -----------
Rb$_3$C$_{60}$ $\sim 50$
Rb$_3$C$_{60}$ 13\*
RbCs$_2$C$_{60}$ $\sim 80$
RbCs$_2$C$_{60}$ 16\*
K$_3$C$_{60}$ 12\*
[(NH$_3$)$_{0.75}$NaK$_2$C$_{60}$]{} 0.87(9)\* this work
: \[tab:fullerides\]Comparison of lower critical fields $H_{c1}$ for several alkali metal doped fullerides (powders denoted by asterisk).
Upper critical field
--------------------
The upper critical field $H_{c2}$ can be determined from the temperature dependence of the field cooled magnetization.[@Buntar95a] In this case the sample is cooled in an externally applied field $H$ (0.5 T $<H<$ 5.5 T) starting from $T>T_c$ and its magnetic moment is measured. The value of the external field corresponds to $H_{c2}$ when the temperature equals the relative $T_c$. Figure \[fig:measHc2\] shows an example of such a measurement in a 3 T magnetic field.
![\[fig:measHc2\]Upper graph: Field cooled magnetization of [(NH$_3$)$_{0.75}$NaK$_2$C$_{60}$]{} at $H=3$ T: the fitted curve accounts for paramagnetic impurities and Pauli contributions. Lower graph: After the subtraction of the above mentioned contributions, $T_c$, at $H_{\text{appl}}=H_{c2}$, is determined from the intersection of the two linear fits.](fig3.eps){width="45.00000%"}
The data above $T_c$ were fitted to a Curie behavior coming from paramagnetic impurities, to which one must add a temperature independent component resulting from Pauli, Landau and core contributions to susceptibility (upper part of the figure).[@Ricco01] After subtraction of all these contributions the curve shown in the lower part of Figure \[fig:measHc2\] is obtained. $T_c$ was estimated as the temperature at which the linear interpolation of the data in the superconducting state intersects the normal state baseline. The observed superconducting transition temperature $T_c(H)$ decreases on increasing $H_{\text{appl}}$. In Figure \[fig:TvsHc2\] we report the dependence of the critical temperature on $H_{\text{appl}}$.
![\[fig:TvsHc2\]Critical temperature dependence on $H_{\text{appl}}$. The upper critical field $H_{c2}(0)=40 \pm 2.5$ T is found from the extrapolated slope $\partial H_{\text{appl}}/\partial T_c$ using formula (\[eq:whh\]).](fig4.eps){width="45.00000%"}
Unfortunately, the maximum field available in our conventional SQUID magnetometer ($H_{\text{max}}= 5.5$ T) does not allow us to investigate the full range of $T_c=T_c(H_{\text{appl}})$ dependence so that we have to resort to extrapolation. The extrapolated field value at zero temperature (relevant for $\xi$ calculation) is usually extracted from the slope of the observed linear behavior using the Werthamer-Helfand-Hohenberg (WHH) formula:[@Werthamer66] $$\label{eq:whh}
H_{c2}(0) = 0.69\;T_c\cdot\left.\frac{\partial H_{c2}}{\partial T}\right|_{T=T_c}.$$
The value for the derivative is $5.0 \pm 0.3$ T/K and from eq. (\[eq:whh\]) we find $H_{c2}(0)=40\pm 2$ T. If we use eq. \[eq:hc2\], we can extract a coherence length $\xi=2.86(8)$ nm. Both of these values, although different, do not appear to be inconsistent with those of other C$_{60}$ based superconductors.[@Baenitz95]
Similarly to the measurement of $H_{c1}$ it is important to examine possible factors that could affect the estimated value for $H_{c2}$:\
a) According to analogous measurements on K$_3$C$_{60}$[@Boebinger92] the parabolic $H_{c2}(T)$ dependence predicted by the WHH theory[@Werthamer66] was not observed: the use of such theory to extract $H_{c2}(0)$ was shown to produce an underestimated value. By guessing from Ref. a plausible enhancement factor of 2 in our datum, we point out that eq. (\[eq:hc2\]) would predict a still smaller $\xi$ ($\sim 2$ nm), which has an important role in the discussion that will follow.\
b) It is known that sample granularity enhances the measured $H_{c2}(0)$ values due to the onset of 0-D fluctuations, as detected in conductivity measurements in K$_3$C$_{60}$.[@Hou94] This effect is expected to become dramatic when the grain size $d$ becomes comparable to $\xi$. Indeed, also conventional superconductors like Al, in a suitable fine granular form, can have an upper critical field nearly two orders of magnitude larger than that of bulk samples.[@Abeles67] The granularity of the material we have investigated, however, involves an average particle size of $\sim 2$ $\mu$m, which is more than three orders of magnitude larger than any estimate for $\xi$, thus allowing us to definitely exclude any appreciable $H_{c2}$ enhancement effect due to granularity.
In conclusion we can state that the coherence length we find from $H_{c2}$ measurements is accurate and, due to its inverse square-root dependence from $H_{c2}$, even large uncertainties in the determination of the latter would not appreciably affect the $\xi$ value.
Muon Spin Rotation Measurements {#sec:musr}
===============================
As briefly described in Sec. \[sec:experiment\], $\mu$SR damping rate measurements yield a reasonable value for the penetration depth $\lambda$, even for an irregular flux-line configuration, typical for powders or polycrystalline samples. In our experiment the sample was field-cooled in an external transverse field of 50 G from above $T_c$ and the muon precession histograms measured at fixed $T$. The presence of two precessing signals at all temperatures indicated that a fraction of the total muons came at rest in the sample holder while the other in the sample. The two components could be easily singled out thanks to their differences not only in amplitude, but mainly in their precession frequency (diamagnetic shift due to superconducting material) and damping rate (internal field second moment due to flux-line lattice). In accounting for the whole expected signal, a missing fraction was detected: it is well known that this fraction is entirely due to the formation of endohedral muonium — i.e. muonium atoms at rest within the C$_{60}$ cage — while the formation of muonium adduct radicals is inhibited in C$_{60}^{n-}$ compounds. This fraction is “missing” because, at the applied field of 50 G, its characteristic frequency is too high to be observable at the ISIS facility. In passing, it must be pointed out that this missing muonium fraction is not interesting for the present experiment, where we look for the information yielded by the unbound muons in the material’s interstitial sites.
The signal due to the sample shows a Gaussian decay with a decay rate $\sigma$, which was fitted using the function $\sin(\omega t+\phi)\exp(-\sigma^2t^2/2)$. Figure \[fig:sigmaT\] shows the fitted values for $\sigma$ in the temperature range 4 $<T<$ 15 K. The temperature independent residual value as observed above $T_c$ is due to the magnetic field distribution of the randomly oriented nuclear dipoles ($^1$H, $^{14}$N, $^{23}$Na, $^{39,40,41}$K, $^{13}$C).
![\[fig:sigmaT\]$\mu$SR Gaussian decay rate as a function of temperature in a 50 G transverse field. The fit of data below $T_c$ yields $\sigma_{\text{sc}}$, from which the field distribution inside the sample is found. The penetration depth $\lambda$ is then calculated using eq. (\[eq:lamdB\]).](fig5.eps){width="45.00000%"}
The additional broadening below $T_c$, due to the flux-line lattice formation, is smaller than the nuclear term and much smaller than that found in similar fullerides such as Rb$_3$C$_{60}$ and K$_3$C$_{60}$.[@MacFarlane98] The superconducting contribution and the normal state nuclear dipole broadening add in quadrature, hence the former ($\sigma_{\text{sc}}$) can be extracted. The temperature dependence of $\sigma_{\text{sc}}$ was fitted to the phenomenological temperature dependence: $\sigma_{\text{sc}}(T)=\sigma_{\text{sc}}(0)[1-(T/T_c)^{\alpha}]$, which yields $\sigma_{\text{sc}}(0)= 5.5\cdot10^{-2}$ $\mu s^{-1}$ and $\alpha=2.55$. The same experiment was repeated applying a 100 G transverse field; in this case no significant increase in $\sigma_{\text{sc}}$ was observed. The value for the London penetration depth $\lambda$ can be extracted from eq. (\[eq:lamdB\]) which gives $\lambda=1.40$ $\mu$m. This value is more than two times larger than that of other superconducting fullerides. In thinking of effects that could enhance the measured $\lambda$ a possible candidate is the granulometry of the sample, when the particles size is not much larger than $\lambda$. But this is just our case, where being $d \sim \lambda$, we do expect strong surface effects which tend to increase the measured $\lambda$ value with respect to that of a bulk sample. Although we have no means to implement a quantitative correction to the measured value, we can definitely state that it represents an upper limit for the real penetration depth, which certainly cannot be larger than 1.4 $\mu$m.
Discussion {#sec:discuss}
==========
In the previous section we described the independent measurements of $H_{c1}$, $H_{c2}$, and $\lambda$ and anticipated that they would be used within the framework of the GL-Abrikosov equations (\[eq:hc2\]) and (\[eq:hc1\]). In principle, the knowledge of $H_{c1}$ and $H_{c2}$ values is sufficient to extract the coherence length $\xi$ and the penetration depth $\lambda$. The additional experimental value of $\lambda$ from $\mu$SR measurements helps in checking the internal consistency. From eq. (\[eq:hc2\]) the experimental value of $H_{c2} = 40$ T yields $\xi = 2.86(8)$ nm; with this value and the experimental result for $H_{c1}= 0.87$ G eq. (\[eq:hc1\]) yields $\lambda = 3.82$ $\mu$m. If only $H_{c1}$ and $H_{c2}$ were measured, the values for $\xi$ and $\lambda$ (although the latter seems considerably larger than the value found in other superconducting fullerides) would appear acceptable for an extreme type-II superconductor.
The picture appears inconsistent, however, since we must account for the *measured* value of $\lambda = 1.4$ $\mu$m, which is nearly three times smaller than that predicted by eq. (\[eq:hc2\]) and (\[eq:hc1\]). Such a discrepancy is by no means trivial because, in order to weaken it, one has to make unphysical assumptions on the other measured parameters, i.e. the two critical fields. Indeed, as can be seen from eq. (\[eq:hc2\]) and (\[eq:hc1\]), even quite a small variation in $\lambda$ will determine orders of magnitude variations in the critical fields.
As an additional check for the observed discrepancy we have also used some approximate analytical solutions of the GL equations. Specifically, we compute the muon spin depolarization rate $\sigma_{\text{sc}}$, which can easily be compared with the experimental value. The calculation requires a numerical solution of the GL equations which was performed using the efficient algorithm developed in Ref. . This fast iterative variational procedure, which minimizes the free energy density $f$, gives as an output the local magnetic field distribution $B=B(x,y)$. The second moment of the local fields, $\Delta B$, yields the damping rate ($\sigma_{\text{sc}}$) of the $\mu$SR.
![\[fig:sigma\_vs\_xi\]$\mu$SR signal decay rate $\sigma_{\text{sc}}$ as a function of coherence length $\xi$ as obtained from the numerical solution of the GL equations. The different curves correspond to several values of applied magnetic field; the lower critical field value was fixed at $H_{c1}=0.87$ G.](fig6.eps){width="45.00000%"}
The results of this calculation are shown in Fig. \[fig:sigma\_vs\_xi\] which reports the dependence of $\sigma_{\text{sc}}$ on the coherence length $\xi$ as predicted by the GL equations for a fixed $H_{c1}=0.87$ G. The different curves refer to different values of the applied field in the range 5 G $< B < 100$ G. It appears evident that the maximum predicted value for the muon decay rate at the applied field of 50 G is $9\cdot10^{-3}$ $\mu$s$^{-1}$, much smaller than the measured value of $55\cdot10^{-3}$ $\mu$s$^{-1}$ and, in any case, too small to be measured. This confirms that the $H_{c1}$, $H_{c2}$ and $\lambda$ values measured in [(NH$_3$)$_{0.75}$NaK$_2$C$_{60}$]{} cannot be representative of a superconducting system described by the Ginzburg-Landau theory. In addition, the fact that a series of samples with different ammonia content have critical temperatures inversely proportional to the density of states at the Fermi level, as recalled in the Introduction, is consistent with this conclusion.
Even though the phonon-mediated superconductivity in fullerides is well established, we suggest that polaron-based theories of superconductivity can correctly describe some of these systems, such as those considered in this work. Unlike the BCS or Migdal-Eliashberg alternatives, the polaronic approach does not require the phonon energy scale to be much smaller than that of electrons. In fullerides, indeed, typical phonon energies involved in the superconducting coupling are or the order of 0.15 eV, due to the intra-molecular $H_{\text{g}}$ modes of C$_{60}$.[@Gunnarsson97] According to the early predictions of N.L. Bulaevskii and co-workers[@Bulaevskii84] for the superconducting properties of systems with local pairs, critical fields and critical length values are expected to be significantly different from those of ordinary BCS superconductors; in particular, the lower critical field $H_{c1}$ is expected to be much smaller and the penetration depth $\lambda$ much larger than the respective BCS counterparts. More recently, a polaron-based analysis has been formulated, with the inclusion of non adiabatic effects in the Migdal-Eliashberg theory, in order to better predict transition temperatures and photoelectron spectra of fullerides.[@Alexandrov96] The presence of charge instabilities or fluctuations (charge disproportion, charge density waves etc.) predicted in polaronic systems manifests itself, indirectly, in superconducting fullerides in the following circumstances: a) the NMR detection of a spin gap in the Na$_2$CsC$_{60}$ superconductor as due to the presence of Jahn-Teller distorted C$_{60}^{(2,4)-}$ originating from a dynamic charge disproportion;[@Brouet00] and b) the existence of a non-magnetic insulating phase in (NH$_3$)$_2$NaK$_2$C$_{60}$,[@Riccounpub] interpreted in terms of a possible (dynamic) charge disproportion. Polaronic instabilities are in general expected at high values of the electron-phonon coupling constant ($\lambda_c\sim
1.5$–2).[@Paci02] In the case of fullerides, though, the coupling values suggested by different experimental techniques (even though still under debate[@Gunnarsson01]) settle in a $0.5<\lambda_{c}<1.2$ range. The closeness of the upper limit of this range to that of the previously mentioned polaronic superconductivity boundary could suggest a possible role played by polarons in these systems.
Conclusions {#sec:conclusions}
===========
In this work we have shown that the magnetic properties of the fullerene based superconductor [(NH$_3$)$_{0.75}$NaK$_2$C$_{60}$]{} are quite different from those of other fullerides, namely it displays a very small $H_{c1}$ associated with very large $H_{c2}$ and $\lambda$ values. These results cannot be explained within the framework of the Ginzburg-Landau (GL) theory. It is suggested that the system considered here represents a border-line case, in which the nature of superconducting coupling begins to switch from a phonon-mediated to a polaronic character.
Indeed, electron-phonon interaction could be sufficiently large in general to make it possible for some systems, like (NH$_3$)$_x$NaK$_2$C$_{60}$, to exhibit an enhancement of the coupling which, in turn, favours the development of polaronic instabilities. In the present situation, specific quantitative predictions for measurable superconducting quantities are needed in order to understand the experimental results and clarify the essential nature of superconductivity in fullerides.
We thank Prof. E. H. Brandt who kindly provided us with the simulation code for muon spin depolarization calculations and Dr G. Salviati with Dr N. Armani who performed the SEM granulometry measurements.\
We acknowledge the financial support by the EU Improvement of Human Potential (IHP) programme, which sponsored part of our $\mu$SR experiments at ISIS facility (UK).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We construct a universal code for stationary and memoryless classical-quantum channel as a quantum version of the universal coding by Csiszár and Körner. Our code is constructed by the combination of irreducible representation, the decoder introduced through quantum information spectrum, and the packing lemma.'
author:
- Masahito Hayashi
date: 'Received:'
title: '[Universal coding for classical-quantum channel]{}'
---
Introduction
============
The channel coding theorem for a stationary and memoryless[^1] (classical-)quantum channel has been established by combining the direct part shown by Holevo [@Holevo-QCTh] and Schumacher-Westmoreland [@Schumacher-Westmoreland] with the (weak) converse (impossible) part which goes back to 1970’s works by Holevo[@Holevo-bounds; @Holevo-bounds2]. Its strong converse part has been shown by Ogawa and Nagaoka[@Oga-Nag:channel] and Winter[@Winter]. This theorem is a fundamental element of quantum information theory[@H-book]. After their achievement, Ogawa and Nagaoka [@ON07] and Hayashi and Nagaoka[@Hay-Nag] constructed other codes attaining the capacity. However, since the existing codes depend on the form of the channel, they are not robust against the disagreement between the sender’s frame and receiver’s frame. In the classical system, Csiszár and Körner [@CK] constructed a universal channel coding, whose construction does not depend on the channel and depends only on the mutual information and the ‘type’ of the input system, i.e., the empirical distribution of code words, whose precise explanation will be explained in Section \[s3\]. Such a universal code for the quantum case was also constructed for variable-length source coding[@Ly; @Da] and fixed-length source coding[@CK].
Concerning the quantum system, Jozsa et al. [@JH] constructed a universal fixed-length source coding, which depends only on the compression rate and attains the minimum compression rate. Hayashi [@Expo-s] discussed the exponential decreasing rate of its decoding error. Further, Hayashi and Matsumoto [@HayaMa] constructed a universal variable-length source coding in the quantum system. However, any universal coding for classical-quantum channel was not constructed. In fact, the universal coding is required when the receiver cannot synchronize his frame with the sender’s frame.
In the present paper, we construct a universal coding for a classical-quantum channel, which attains the quantum mutual information and depends only on the coding rate and the ‘type’ of the input system. In the proposed construction, the following three methods play essential roles. One is the decoder given by the proof of the information spectrum method. In the information spectrum method, the decoder is constructed by the square root measurement of the projectors given by the quantum analogue of the likelihood ratio between the signal state and the mixture state[@Hay-Nag; @Verdu-Han].
The second method is the irreducible decomposition of the dual representation of the special unitary group and the permutation group. The method of irreducible decomposition provides the universal protocols in quantum setting[@JH; @Haya97; @H2001; @MaHa07; @HayaMa; @KeylW; @Sanov]. However, even in the classical case, the universal channel coding requires the conditional type as well as the type[@CK]. In the present paper, we introduce a quantum analogue of the conditional type, which is the most essential part of the present paper.
The third method is the packing lemma, which yields a suitable combination of the signal states independent of the form of the channel in the classical case[@CK]. This method plays the same role in the present paper.
The remainder of the present paper is organized as follows. In section \[s2\], we give the notation herein and the main result including the existence of a universal coding for classical-quantum channel. In this section, we presented the exponential decreasing rate of the error probability of the presented universal code. In section \[s3\], the notation for group representation theory is presented and a quantum analogue of conditional type is introduced. In section \[s4\], we give a code that well works universally. In section \[s5\], the exponential decreasing rate mentioned in section \[s2\] is proven by using the property given in section \[s3\].
Main Result
===========
In the classical-quantum channel, we focus on the set of input alphabets ${\cal X}:=\{1, \ldots, k\}$ and the representation space ${\cal H}$ of the output system, whose dimension is $d$. Then, a classical-quantum channel is given as the map from ${\cal X}$ to the set of densities on ${\cal H}$ with the form $i \mapsto W(i)$. The $n$-th discrete memoryless extension is given as the map from ${\cal X}^n$ to the set of densities on the $n$-th tensor product system ${\cal H}^{\otimes n}$. That is, this extension maps the input sequence $\vec{i}=(i_1, \ldots, i_n)$ to the state $W_n(\vec{i}_n):=W(i_1)\otimes \cdots \otimes W(i_n)$. Sending the message $\{1, \ldots, M_n\}$ requires an encoder and a decoder. The encoder is given as a map $\varphi_n$ from the set of messages $\{1, \ldots, M_n\}$ to the set of alphabets ${\cal X}^n$, and the decoder is given by a POVM $Y^n=\{Y_i^n\}_{i=1}^{M_n}$. Thus, the triplet $\Phi_n:=(M_n, \varphi_n, Y)$ is called a code. Its performance is evaluated by the size $|\Phi_n|:=M_n$ and the average error probability given by $$\begin{aligned}
\varepsilon [\Phi_n,W]:=\frac{1}{M_n} \sum_{i=1}^{M_n} \Tr W_n (\varphi_n(i)) (I-Y_i^n).\end{aligned}$$ As mentioned in the following main theorem, there exists an asymptotically optimal code that depends only on the coding rate.
For any distribution $\vec{p}=\{p_i\}_{i=1}^k$ on the set of input alphabets ${\cal X}:=\{1, \ldots, k\}$ and any real number $R$, there is a sequence of codes $\{\Phi_n\}_{n=1}^{\infty}$ such that $$\begin{aligned}
\lim_{n \to \infty}\frac{-1}{n}\log \varepsilon [\Phi_n,W]
& \ge \max_{0 \le t \le 1} \frac{\phi_{W,\vec{p}}(t)-tR}{1+t} \\
\lim_{n \to \infty}\frac{1}{n}\log |\Phi_n| &= R\end{aligned}$$ for any classical-quantum channel $W$, where $\phi_{W,\vec{p}}(t)$ is given by $$\begin{aligned}
\phi_{W,\vec{p}}(t):=- (1-t) \log \Tr (\sum_{i=1}^k p_i W(i)^{1-t})^{\frac{1}{1-t}} .\end{aligned}$$ Note that the code $\{\Phi_n\}_{n=1}^{\infty}$ does not depend on the channel $W$, and depends only on the distribution $\vec{p}$ and the coding rate $R$.
The derivative of $\phi_{W,\vec{p}}(t)$ is given as $$\begin{aligned}
\phi_{W,\vec{p}}'(0)&=I(p,W):=\sum_{i=1}^k p_i \Tr W(i)(\log W_i- \log W_{\vec{p}}) \\
W_{\vec{p}}&:=\sum_{i=1}^k p_i W(i).\end{aligned}$$ When the transmission rate $R$ is smaller than the mutual information $I(\vec{p},W)$, $$\begin{aligned}
\max_{0 \le t \le 1} \frac{\phi_{W,\vec{p}}(t)-tR}{1+t} >0\end{aligned}$$ because there exists a parameter $t \in (0,1)$ such that $\phi_{W,\vec{p}}(t) -tR >0$. That is, the average error probability $\varepsilon [\Phi_n,W]$ goes to zero.
Group representation theory
===========================
In this section, we focus on the dual representation on the $n$-fold tensor product space by the the special unitary group $SU(d)$ and the $n$-th symmetric group $S_n$[^2]. For this purpose, we focus on the Young diagram and the ‘type’. The former is a key concept in group representation theory and the latter is that in information theory[@CK]. When the vector of integers $\vec{n}=(n_1,n_2, \ldots, n_d)$ satisfies the condition $n_1 \ge n_2 \ge \ldots \ge n_d\ge 0$ and $\sum_{i=1}^d n_i=n$, the vector $\vec{n}$ is called the Young diagram (frame) with size $n$ and depth $d$, the set of which is denoted as $Y_n^d$. When the vector of integers $\vec{n} $ satisfies the condition $n_i \ge 0$ and $\sum_{i=1}^d n_i=n$, the vector $\vec{p}=\frac{\vec{n}}{n}$ is called the ‘type’ with size $n$, the set of which is denoted as $T_n^d$. Further, for $\vec{p} \in T_n^d$, the subset of ${\cal X}^n$ is defined as: $$\begin{aligned}
T_{\vec{p}}:= \{\vec{x} \in {\cal X}^n| \hbox{The empirical distribution of } \vec{x}
\hbox{ is equal to }\vec{p}\}.\end{aligned}$$ The numbers of these sets are evaluated as follows: $$\begin{aligned}
|Y_n^d| &\le |T_n^d|\le (n+1)^{d-1}\Label{9}\\
(n+1)^{-d} e^{n H(\vec{p})} &\le |T_{\vec{p}}| ,\Label{7}\end{aligned}$$ where $H(\vec{p}):= -\sum_{i=1}^d p_i \log p_i$[@CK]. Using the Young diagram, the irreducible decomposition of the above representation can be characterized as follows: $$\begin{aligned}
{\cal H}^{\otimes n}=\bigoplus_{\vec{n}\in Y_n^d} {\cal U}_{\vec{n}} \otimes {\cal V}_{\vec{n}},\end{aligned}$$ where ${\cal U}_{\vec{n}}$ is the irreducible representation space of $SU(d)$ characterized by $\vec{n}$, and ${\cal V}_{\vec{n}}$ is the irreducible representation space of $n$-th symmetric group $S_n$ characterized by $\vec{n}$. Here, the representation of the $n$-th symmetric group $S_n$ is denoted as $V: s \in S_n \mapsto V_s$. For $\vec{n}\in Y_n^d$, the dimension of ${\cal U}_{\vec{n}}$ is evaluated by $$\begin{aligned}
\dim {\cal U}_{\vec{n}} \le n^{\frac{d(d-1)}{2}} \Label{11}.\end{aligned}$$ Then, denoting the projection to the subspace ${\cal U}_{\vec{n}} \otimes {\cal V}_{\vec{n}}$ as $I_{\vec{n}}$, we define the following. $$\begin{aligned}
\rho_{\vec{n}}&:= \frac{1}{\dim {\cal U}_{\vec{n}} \otimes {\cal V}_{\vec{n}}} I_{\vec{n}}\\
\rho_{U,n}&:= \sum_{\vec{n}\in Y_n^d} \frac{1}{|Y_n^d|} \rho_{\vec{n}}.\Label{15}\end{aligned}$$ Any state $\rho$ and any Young diagram $\vec{n}\in Y_n^d$ satisfy the following: $$\begin{aligned}
\dim {\cal U}_{\vec{n}} \rho_{\vec{n}}
\ge I_{\vec{n}} \rho^{\otimes n} I_{\vec{n}}.\end{aligned}$$ Thus, (\[9\]), (\[11\]), and (\[15\]) yield the inequality $$\begin{aligned}
n^{\frac{d(d-1)}{2}} |Y_n^d| \rho_{U,n} \ge \rho^{\otimes n}.\Label{ineq-2}\end{aligned}$$
Next, we focus on two systems ${\cal X}$ and ${\cal Y}=\{1,\ldots, l\}$. When the distribution of ${\cal X}$ is given by a probability distribution $\vec{p}=(p_1, \ldots, p_d)$ on $\{1, \ldots, d\}$, and the conditional distribution on ${\cal Y}$ with the condition on ${\cal X}$ is given by $\vec{V}$, we denote the joint distribution on ${\cal X} \times {\cal Y}$ by $\vec{p} \vec{V}$ and the distribution on ${\cal Y}$ by $\vec{p} \cdot \vec{V}$. When the empirical distribution of $\vec{x}\in {\cal X}^n$ is $(\frac{n_1}{n},\ldots, \frac{n_d}{n})$, the sequence of types $\vec{V}=(\vec{v}_1, \ldots, \vec{v}_d)\in
T_{n_1}^{l}\times \cdots \times T_{n_d}^{l}$ is called a conditional type for $\vec{x}$[@CK]. We denote the set of conditional types for $\vec{x}$ by $V(\vec{x},{\cal Y})$. For any conditional type $V$ for $\vec{x}$, we define the subset of ${\cal Y}^n$: $$\begin{aligned}
T_{\vec{V}}(\vec{x})
:= \left\{\vec{y} \in {\cal Y}^n\left|
\begin{array}{l}
\hbox{The empirical distribution of } \\
((x_1,y_1), \ldots, (x_n,y_n))
\hbox{ is equal to }
\vec{p} \vec{V}.
\end{array}
\right.\right\},\end{aligned}$$ where $\vec{p}$ is the empirical distribution of $\vec{x}$.
We define the state $\rho_{\vec{x}}$ for $\vec{x} \in {\cal X}^n$. For this purpose, we consider a special element $\vec{x}'=(\underbrace{1,\ldots,1}_{m_1},\underbrace{2,\ldots,2}_{m_2},\ldots,
\underbrace{k,\ldots, k}_{m_k})$. The state $\rho_{\vec{x}'}$ is defined as $\rho_{\vec{x}'}:= \rho_{U,m_1} \otimes \rho_{U,m_2} \otimes \cdots \otimes \rho_{U,m_k}$. For a general element $\vec{x} \in {\cal X}^n$, we choose a permutation $s\in S_n$ such that $\vec{x}=s\vec{x}'$. Then, we define the state $\rho_{\vec{x}}$ is defined as $\rho_{\vec{x}}:=U_s \rho_{\vec{x}'}U_s^\dagger$, where $U_s$ is the unitary representation of $S_n$. This state plays a similar role as the conditional type in the classical case. Using the inequality (\[ineq-2\]), we have $$\begin{aligned}
n^{\frac{kd(d-1)}{2}} |Y_n^d|^k
\rho_{\vec{x}} \ge W_n(\vec{x}).
\Label{ineq-1}\end{aligned}$$ For $\vec{n}_1\in Y_{m_1}^d, \vec{n}_2\in Y_{m_2}^d, \ldots, \vec{n}_k\in Y_{m_k}^d$, the density $\rho_{\vec{n}_1}\otimes \rho_{\vec{n}_2}\otimes \cdots \otimes
\rho_{\vec{n}_k}$ is commutative with the projector $I_{\vec{n}}$ for $\vec{n} \in Y_n^d$. This fact implies that the density $\rho_{\vec{x}}$ is commutative with the density $\rho_{U,n}$. This property is essential for the construction of the proposed decoder.
Construction of code
====================
According to Csiszár and Körner[@CK], the proposed code is constructed as follows.
For a positive number $\delta>0$, a type $\vec{p} \in T_n^d$, and a real positive number $R< H(\vec{p})$, there exist $M_n:=e^{n(R-\delta)}$ distinct elements ${\cal M}_n:=\{\vec{x}_1,\ldots, \vec{x}_{M_n}\} \subset T_{\vec{p}}$ such that their empirical distributions are $\vec{p}$ and $$\begin{aligned}
|T_{\vec{V}}(\vec{x}) \cap ({\cal M}_n\setminus \{\vec{x}\})|
\le |T_{\vec{V}}(\vec{x})|
e^{-n(H(\vec{p})-R)} $$ for $\vec{x} \in {\cal M}_n \subset T_{\vec{p}}$ and $\vec{V} \in V(\vec{x},{\cal X})$.
This lemma can be shown by substituting the identical map into $\hat{V}$ in Lemma 5.1 in Csiszár and Körner[@CK]. Since Csiszár and Körner proved Lemma 5.1 using the random coding method, we can replace $\delta$ by $\frac{1}{\sqrt{n}}$. That is, there exist $M_n:= e^{n R- \sqrt{n}}$ distinct elements ${\cal M}_n:=\{\vec{x}_1,\ldots, \vec{x}_{M_n}\} \subset T_{\vec{p}}$ such that their empirical distributions are $\vec{p}$ and $$\begin{aligned}
|T_{\vec{V}}(\vec{x}) \cap ({\cal M}_n\setminus \{\vec{x}\})|
\le |T_{\vec{V}}(\vec{x})|
e^{-n(H(\vec{p})-R)} \Label{20}\end{aligned}$$ for $\vec{x} \in {\cal M}_n \subset T_{\vec{p}}$ and $\vec{V} \in V(\vec{x},{\cal X})$. Now, we transform the property (\[20\]) to a more useful form.
Using the encoder ${\cal M}_n$, we can define the distribution $P_{{\cal M}_n}$ as $$\begin{aligned}
p_{{\cal M}_n}(\vec{x})=
\left\{
\begin{array}{cc}
\frac{1}{|{\cal M}_n|} & \vec{x} \in {\cal M}_n \\
0 & \vec{x} \notin {\cal M}_n.
\end{array}
\right.\end{aligned}$$ For any $\vec{x} \in {\cal X}^n$, we define the invariant subgroup $S_{\vec{x}}\subset S_n$: $$\begin{aligned}
S_{\vec{x}}: = \{s \in S_n | s(\vec{x})=\vec{x} \}.\end{aligned}$$ Since $\vec{x}' \in T_{\vec{p}}$ implies that $$\begin{aligned}
\vec{p}^{n}(\vec{x}')=e^{-nH(\vec{p})} ,\end{aligned}$$ any element $\vec{x}' \in T_{\vec{V}}(\vec{x}) \cap {\cal M}_n \subset T_{\vec{p}}$ satisfies $$\begin{aligned}
&\sum_{s \in S_{\vec{x}}}
\frac{1}{|S_{\vec{x}}|}
p_{{\cal M}_n}\circ s (\vec{x}')
=
\frac{|T_{\vec{V}}(\vec{x}) \cap {\cal M}_n|}{|T_{\vec{V}}(\vec{x})|}
\cdot
\frac{1}{|{\cal M}_n|}
=
\frac{|T_{\vec{V}}(\vec{x}) \cap ({\cal M}_n\setminus \{\vec{x}\})|}{|T_{\vec{V}}(\vec{x})||{\cal M}_n|} \nonumber \\
\le &
e^{-nH(\vec{p})}
e^{\sqrt{n} }
=\vec{p}^{n}(\vec{x}') e^{\sqrt{n} } \Label{8}\end{aligned}$$ when the conditional type $\vec{V}$ is not identical. Relation (\[8\]) holds for any $\vec{x}'(\neq \vec{x}) \in {\cal M}_n $ because there exists a conditional type $\vec{V}$ such that $\vec{x}' \in T_{\vec{V}}(\vec{x})$ and $\vec{V}$ is not identical.
Next, for any $\vec{x} \in {\cal X}^n$ and any real number $C_n$, we define the projection $$\begin{aligned}
P(\vec{x}):=
\{\rho_{\vec{x}}- C_n \rho_{U,n}\ge 0\},\end{aligned}$$ where $\{X\ge 0\}$ presents the projection $\sum_{i:x_i \ge 0}E_i$ for a Hermitian matrix $X$ with the diagonalization $X=\sum_ix_i E_i$. Remember that the density $\rho_{\vec{x}}$ is commutative with the other density $\rho_{U,n}$. Using the projection $P(\vec{x})$, we define the decoder: $$\begin{aligned}
Y_{\vec{x}'}:=
\sqrt{\sum_{\vec{x} \in {\cal M}_n}P(\vec{x})}^{-1}
P(\vec{x}')
\sqrt{\sum_{\vec{x} \in {\cal M}_n}P(\vec{x})}^{-1}.\end{aligned}$$ In the following, the above-constructed code $(e^{nR-\sqrt{n}}, {\cal M}_n,\{Y_{\vec{x}}\}_{\vec{x}\in {\cal M}_n})$ is denoted by $\Phi_{U,n}(\vec{p},R)$.
Exponential evaluation
======================
Hayashi and Nagaoka[@Hay-Nag] showed that $$\begin{aligned}
I- Y_{\vec{x}'} \le 2 (I- P(\vec{x}'))+ 4 \sum_{\vec{x} (\neq \vec{x}')\in {\cal M}_n}P(\vec{x}).\end{aligned}$$ Then, the average error probability of $\Phi_{U,n}(\vec{p},R)$ is evaluated by $$\begin{aligned}
& \frac{1}{|{\cal M}_n|}
\sum_{\vec{x}' \in {\cal M}_n}
\Tr W_n(\vec{x}')(I- Y_{\vec{x}'}) \nonumber \\
\le &
\frac{2}{|{\cal M}_n|}
\sum_{\vec{x}' \in {\cal M}_n} \Tr W_n(\vec{x}') (I- P(\vec{x}'))
+
\frac{4}{|{\cal M}_n|}
\sum_{\vec{x}' \in {\cal M}_n}
\Tr W_n(\vec{x}')
\sum_{\vec{x} (\neq \vec{x}')\in {\cal M}_n}P(\vec{x})\nonumber \\
=&
\frac{2}{|{\cal M}_n|}
\sum_{\vec{x} \in {\cal M}_n} \Tr W_n(\vec{x}) (I- P(\vec{x}))
\nonumber \\
&+
4
\Tr
\left[
\sum_{\vec{x} \in {\cal M}_n}P(\vec{x})
\left(\frac{1}{|{\cal M}_n|}
\sum_{\vec{x}' (\neq \vec{x}) \in {\cal M}_n} W_n(\vec{x}')\right)
\right]\Label{37}.\end{aligned}$$
Since the density $\rho_{\vec{x}}$ is commutative with the density $\rho_{U,n}$, we have $$\begin{aligned}
(I- P(\vec{x}))
=
\{\rho_{\vec{x}}- C_n \rho_{U,n}< 0\}
\le
\rho_{\vec{x}}^{-t} C_n ^t \rho_{U,n} ^t \Label{ineq-4}\end{aligned}$$ for $0 \le t \le 1$. Since the density $\rho_{\vec{x}}$ is commutative with the density $W_n(\vec{x})$, $W_n(\vec{x})\rho_{\vec{x}}^{-t} $ is a Hermite matrix and (\[ineq-1\]) implies that $$\begin{aligned}
W_n(\vec{x})\rho_{\vec{x}}^{-t}
\le
n^{\frac{ktd(d-1)}{2}} |Y_n^d|^{kt}
W_n(\vec{x})^{1-t}.\Label{ineq-5}\end{aligned}$$ Using (\[ineq-4\]) and (\[ineq-5\]), we have $$\begin{aligned}
& \Tr W_n(\vec{x}) (I- P(\vec{x}))
\le
\Tr
W_n(\vec{x})\rho_{\vec{x}}^{-t}
\rho_{U,n} ^t C_n ^t \nonumber \\
\le &
n^{\frac{ktd(d-1)}{2}} |Y_n^d|^{kt} C_n ^t
\Tr W_n(\vec{x})^{1-t}
\rho_{U,n} ^t .\Label{l11}\end{aligned}$$ Since the quantity $\Tr W_n(\vec{x}) (I- P(\vec{x}))$ is invariant for the action of the permutation and the relation (\[7\]) implies that $$\begin{aligned}
\vec{p}^{n}(\vec{x})= e^{-n H(\vec{p})}
\ge \frac{(n+1)^{-d}}{|T_{\vec{p}}|} \Label{l12}\end{aligned}$$ for $\vec{x} \in T_{\vec{p}}$, we obtain $$\begin{aligned}
& \Tr W_n(\vec{x}) (I- P(\vec{x}))
=
\frac{1}{|T_{\vec{p}}|}
\sum_{\vec{x}'\in T_{\vec{p}}}
\Tr W_n(\vec{x}') (I- P(\vec{x}')) \nonumber \\
\le &
(n+1)^d \sum_{\vec{x}'\in {\cal X}^n
}
\vec{p}^{n}(\vec{x}')
\Tr W_n(\vec{x}') (I- P(\vec{x}'))\Label{l5} \\
\le &
(n+1)^d n^{\frac{ktd(d-1)}{2}} |Y_n^d|^{kt} C_n ^t
\Tr
(\sum_{\vec{x}'\in {\cal X}^n}
\vec{p}^{n}(\vec{x}')
W_n(\vec{x}')^{1-t})
\rho_{U,n} ^t \Label{l6} \\
\le &
(n+1)^{d+\frac{ktd(d-1)}{2}} |Y_n^d|^{kt} C_n ^t
\max_{\sigma}
\Tr
\left[\sum_{x\in {\cal X}} \vec{p}(x) W_n(x)^{1-t}\right]^{\otimes n}
\sigma^t \nonumber \\
\le &
(n+1)^{d+\frac{ktd(d-1)}{2}} |Y_n^d|^{kt} C_n ^t
\left( \Tr
\left( \left[\sum_{x\in {\cal X}} \vec{p}(x) W_n(x)^{1-t}\right]^{\otimes n}\right)
^{\frac{1}{1-t}}\right)^{1-t} \Label{l1}\\
= &
(n+1)^{d+\frac{ktd(d-1)}{2}} |Y_n^d|^{kt} C_n ^t
\left(\Tr
\left(\sum_{x\in {\cal X}} \vec{p}(x) W_n(x)^{1-t}\right)^{\frac{1}{1-t}}
\right)^{n(1-t)} \nonumber \\
=& (n+1)^{d+\frac{ktd(d-1)}{2}} |Y_n^d|^{kt} C_n ^t
e^{- n \phi_{W,\vec{p}}(t)},\Label{49}\end{aligned}$$ where (\[l5\]), (\[l6\]), and (\[l1\]) follow from (\[l12\]), (\[l11\]), and Lemma \[lem1\] in Appendix, respectively.
Next, we evaluate the second term of (\[37\]) using the invariant property of $S_{\vec{x}}$: $$\begin{aligned}
& \Tr
\left[
P(\vec{x})
\left(\frac{1}{|{\cal M}_n|}
\sum_{\vec{x}' (\neq \vec{x}) \in {\cal M}_n} W_n(\vec{x}')\right)
\right]\nonumber \\
=&
\Tr
\left[
P(\vec{x})
\sum_{\vec{x}' (\neq \vec{x}) \in {\cal M}_n}
p_{{\cal M}_n}(\vec{x}')
W_n(\vec{x}')
\right] \nonumber\\
= &
\Tr
\left[
P(\vec{x})
\sum_{s \in S_{\vec{x}}}
\frac{1}{|S_{\vec{x}}|}
\sum_{\vec{x}' (\neq \vec{x}) \in {\cal M}_n}
p_{{\cal M}_n}(\vec{x}')
V_{s}
W_n(\vec{x}')V_{s}^{*}
\right] \nonumber\\
= &
\Tr
\left[
P(\vec{x})
\sum_{\vec{x}' (\neq \vec{x}) \in {\cal M}_n}
\sum_{s \in S_{\vec{x}}}
\frac{1}{|S_{\vec{x}}|}
p_{{\cal M}_n}\circ s^{-1}(\vec{x}')
W_n(\vec{x}')
\right] \nonumber\\
\le &
\Tr
\left[
P(\vec{x})
\sum_{\vec{x}' (\neq \vec{x}) \in {\cal M}_n}
\vec{p}^{n}(\vec{x}') e^{\sqrt{n}}
W_n(\vec{x}')
\right] \Label{l13} \\
= &
e^{\sqrt{n}}
\Tr
\left[
P(\vec{x})
W_{\vec{p}}^{\otimes n}
\right] \nonumber\\
\le &
e^{\sqrt{n}}
\Tr
\left[
P(\vec{x})
n^{\frac{d(d-1)}{2}} |Y^d_n|
\rho_{U,n}
\right] \Label{l14} \\
\le &
e^{\sqrt{n} }
\Tr
\left[
P(\vec{x})
n^{\frac{d(d-1)}{2}} |Y^d_n|
C_n^{-1} \rho_{\vec{x}}
\right] \Label{l15} \\
\le &
e^{\sqrt{n} }
\Tr
\left[
n^{\frac{d(d-1)}{2}}|Y^d_n|
C_n^{-1} \rho_{\vec{x}}
\right]
=
e^{\sqrt{n} }
n^{\frac{d(d-1)}{2}}|Y^d_n|
C_n^{-1},\Label{59}\end{aligned}$$ where (\[l13\]), (\[l14\]), and (\[l15\]) follow from (\[8\]), (\[ineq-2\]), and the inequality $P(\vec{x})(\rho_{U,n}- C_n^{-1} \rho_{\vec{x}}) \le 0$.
For any $t \in (0,1)$ and $R>0$, we choose $|{\cal M}_n|:=e^{nR-\sqrt{n}}$, $C_n:=e^{n(R+r(t))}$, and $r(t):=\frac{\phi_{W,\vec{p}}(t)-tR }{1+t}$. Since $r(t)=\phi_{W,\vec{p}}(t)-t(R+r(t))$, from (\[37\]), (\[49\]) and (\[59\]), the exponential decreasing rate of the average error probability is evaluated as $$\begin{aligned}
\lim_{n\to \infty}
\frac{-1}{n}\log \varepsilon (\Phi_{U,n}(\vec{p},R),W)
\ge \min\{\phi_{W,\vec{p}}(t)-t(R+r(t)),r(t)\} =
\frac{\phi_{W,\vec{p}}(t)-tR }{1+t}.\end{aligned}$$ That is, when we choose $t_0:=\argmax_{t \in (0,1)}\frac{\phi_{W,\vec{p}}(t)-tR }{1+t}$, $|{\cal M}_n|:=e^{nR-\sqrt{n}}$, and $C_n:=e^{n(R+r(t_0))}$, we obtain $$\begin{aligned}
\lim_{n\to \infty}
\frac{-1}{n}\log \varepsilon (\Phi_{U,n}(\vec{p},R),W)
\ge \max_{t \in (0,1)}
\frac{\phi_{W,\vec{p}}(t)-tR }{1+t}\end{aligned}$$ for any channel $W$. Therefore, we obtain Theorem \[thm1\].
Discussion
==========
We have constructed a universal code attaining the quantum mutual information based on the combination of information spectrum method, group representation theory, and the packing lemma. The presented code well works because any tensor product state $\rho^{\otimes n}$ is close to the state $\rho_{U,n}$. Indeed, Krattenthaler and Slater [@KS] demonstrated the existence of the state $\sigma_n$ such that $\frac{1}{n}D(\rho^{\otimes n}\|\sigma_n)\to n$ for any state $\rho$ in the qubit system as a quantum analogue of Clarke and Barron’s result[@CB]. Its $d$-dimensional extension is discussed in another paper[@prep].
Further, Hayashi [@Expo-c] derived an exponential decreasing rate of error probability in classical-quantum channel, which is $
\max_{t:0\le t\le 1}
-(\log \sum_i p_i \Tr [W(i)^{1-t} W_p^t]) -tR$. Since $$\begin{aligned}
& e^{-\frac{\phi_{W,\vec{p}}(t)-t(R+r(t))}{1+t}}
=
e^{-(\phi_{W,\vec{p}}(t)-t(R+r(t)))}
=
e^{t(R+r(t))}
\max_\sigma
\Tr (\sum_i p_i W(i)^{1-t})\sigma^t \\
\ge &
e^{tR}
\Tr (\sum_i p_i W(i)^{1-t})(\sum_i p_i W(i))^t
=
e^{-
(-(\log \sum_i p_i \Tr [W(i)^{1-t} W_{\vec{p}}^t]) -tR)},\end{aligned}$$ we obtain $$\begin{aligned}
\max_{t:0\le t\le 1}
-(\log \sum_i p_i \Tr [W(i)^{1-t} W_{\vec{p}}^t]) -tR
\ge
\max_{t:0\le t\le 1}
\frac{\phi_{W,\vec{p}}(t)-tR}{1+t}.\end{aligned}$$ That is, the obtained exponential decreasing rate is smaller than that of Hayashi[@Expo-c]. However, according to Csiszár and Körner [@CK], the exponential decreasing rate of the universal coding is the same as the optimal exponential decreasing rate in the classical case when the rate is close to the capacity. Hence, if a more sophisticated evaluation is applied, a better exponential decreasing rate can be expected. Such an evaluation is left as a future problem.
Acknowledgment {#acknowledgment .unnumbered}
==============
This research was partially supported by a Grant-in-Aid for Scientific Research on Priority Area ‘Deepening and Expansion of Statistical Mechanical Informatics (DEX-SMI)’, No. 18079014 and a MEXT Grant-in-Aid for Young Scientists (A) No. 20686026.
Maximization
============
The following lemma is used for the derivation in Section \[s5\].
When $X$ is a positive semi-definite, we have $$\begin{aligned}
\max_{\sigma} \Tr X \sigma^t= (\Tr X^{\frac{1}{1-t}})^{1-t}\Label{l3}\end{aligned}$$ for $0 \le t \le 1$, where $\sigma$ is a density matrix.
First, we prove $$\begin{aligned}
\max_{q_i\ge 0:\sum_i q_i=1}
\Tr X \sum_{i}q_i^t |i\rangle \langle i|=
\left(\sum_{i}
\langle i|X |i\rangle^{\frac{1}{1-t}}
\right)^{1-t}\Label{l4}\end{aligned}$$ by the Lagrange multiplier method. Let $\lambda$ be the Lagrange multiplier. Then, $$\begin{aligned}
0=
\sum_i (\langle i|X |i\rangle t q_i^{t-1}+ \lambda )\delta q_i\end{aligned}$$ Thus, $$\begin{aligned}
0=\langle i|X |i\rangle t q_i^{t-1}+ \lambda.\end{aligned}$$ That is, $$\begin{aligned}
-\frac{t}{\lambda}\langle i|X |i\rangle = q_i^{1-t} .\end{aligned}$$ Then, when the maximizing $q_i$ has the form $C \langle i|X |i\rangle^{\frac{1}{1-t}}$ with the normalizing constant $C$, the constant $C$ has the form $C=\frac{1}{\sum_j \langle j|X |j\rangle^{\frac{1}{1-t}}}$. Substituting $\frac{\langle i|X |i\rangle^{\frac{1}{1-t}}}{\sum_j \langle j|X |j\rangle^{\frac{1}{1-t}}}$ into $q_i$, we obtain (\[l4\]).
Since $$\begin{aligned}
\left(\sum_{i}
\langle i|X |i\rangle^{\frac{1}{1-t}}
\right)^{1-t}
=
\Tr X \left(\sum_{i}
\langle i|\frac{1}{\Tr X}X |i\rangle^{\frac{1}{1-t}}
\right)^{1-t},\end{aligned}$$ the maximum $\max_{\sigma} \Tr X \sigma^t$ is given when we choose the basis $\{|i\rangle \}$ maximizing $\sum_{i}
\langle i|\frac{1}{\Tr X}X |i\rangle^{\frac{1}{1-t}}$. Since the function $x \mapsto x^{\frac{1}{1-t}}$ is a convex function, $
\langle i|\frac{1}{\Tr X}X |i\rangle^{\frac{1}{1-t}}
\le
\langle i|(\frac{1}{\Tr X}X)^{\frac{1}{1-t}}
|i\rangle$. Therefore, $$\begin{aligned}
\left(\sum_{i}
\langle i|X |i\rangle^{\frac{1}{1-t}}
\right)^{1-t}
\le
(\Tr X^{\frac{1}{1-t}})^{1-t}.\end{aligned}$$ The equality holds when we choose the basis $\{|i\rangle\}$ as the eigenvectors of $X$. Therefore, we obtain (\[l3\]).
[99]{} A.S. Holevo, “The capacity of the quantum channel with general signal states,” [*IEEE Trans. Inform. Theory*]{}, vol.44, 269–273, 1998. B. Schumacher and M.D. Westmoreland, “Sending classical information via noisy quantum channels,” [*Phys. Rev. A*]{}, vol.56, 131–138, 1997. A.S. Holevo, “Bounds for the quantity of information transmitted by a quantum communication channel,” [*Probl. Inform. Transm.*]{}, vol.9, 177–183, 1973. A.S. Holevo, “On the capacity of quantum communication channel,” [*Probl. Inform. Transm.*]{}, vol. 15, no. 4, pp. 247–253, 1979. T. Ogawa and H. Nagaoka, “Strong Converse to the Quantum Channel Coding Theorem," [*IEEE Trans. Inform. Theory*]{}, vol.45, 2486-2489, 1999. A. Winter, “Coding theorem and strong converse for quantum channels,” [*IEEE Trans. Inform. Theory*]{}, vol.45, 2481-2485, 1999. M. Hayashi, [*Quantum Information: An Introduction*]{}, (Springer, Berlin, 2006).
T. Ogawa and H. Nagaoka, “Making Good Codes for Classical-Quantum Channel Coding via Quantum Hypothesis Testing,” [*IEEE Trans. Inform. Theory*]{}, vol.53, 2261 - 2266, (2007).
M. Hayashi and H. Nagaoka: “General formulas for capacity of classical-quantum channels,” [*IEEE Trans. Infor. Theory*]{}, [**49**]{}, 1753-1768 (2003).
I. Csiszár and J. Körner, [*Information Theory: Coding Theorems for Discrete Memoryless Systems*]{}, (Academic Press, 1981).
T. J. Lynch, “Sequence time coding for data compression,” [*Proc. IEEE*]{}, [**54**]{}, 1490-1491, (1966).
L. D. Davisson, “Comments on ‘Sequence time coding for data compression’,” [*Proc. IEEE*]{}, [**54**]{}, 2010, (1966).
R. Jozsa, M. Horodecki, P. Horodecki and R. Horodecki, “Universal Quantum Information Compression,” [*Phys. Rev. Lett.*]{}, [**81**]{}, 1714 (1998); quant-ph/9805017 (1998). M. Hayashi, “Exponents of quantum fixed-length pure state source coding,” [*Phys. Rev. A*]{}, [**66**]{}, 032321 (2002).
M. Hayashi and K. Matsumoto, “Quantum universal variable-length source coding,” [*Phys. Rev. A*]{} [**66**]{}, 022311 (2002).
S. Verdú and T.S. Han, “A general formula for channel capacity,” [*IEEE Trans. Inform. Theory*]{}, [**40**]{}, 1147–1157 (1994).
M. Hayashi: “Asymptotics of quantum relative entropy from a representation theoretical viewpoint,” [*J. Phys. A: Math. and Gen.*]{}, [**34**]{}, 3413-3419 (2001).
M. Keyl and R. F. Werner, “Estimating the spectrum of a density operator,” [*Phys. Rev. A*]{}, [**64**]{}, 052311 (2001).
M. Hayashi, “Optimal sequence of POVMs in the sense of Stein’s lemma in quantum hypothesis,” [*J. Phys. A: Math. and Gen.*]{}, [**35**]{}, 10759-10773 (2002).
I. Bjelaković, J.-D. Deuschel, T. Kruger, R. Seiler, R. Siegmund-Schultze, and A. Szko[ł]{}a, “A Quantum Version of Sanov’s Theorem,” [*Comm. Math. Phys.*]{}, [**260**]{}, 659-671 (2005).
K. Matsumoto, and M. Hayashi, “Universal distortion-free entanglement concentration,” [*Physical Review A*]{}, [**75**]{}, 062338 (2007). M. Christandl, “The Structure of Bipartite Quantum States - Insights from Group Theory and Cryptography,” PhD thesis, February 2006, University of Cambridge, quant-ph/0604183.
C. Krattenthaler and P. Slater “Asymptotic Redundancies for Universal Quantum Coding,” [*IEEE Trans. Inform. Theory*]{}, [**46**]{}, 801-819 (2000).
B. S. Clarke and A. R. Barron, “Information-theoretic asymptotics of Bayes methods,” [*IEEE Trans. Inform. Theory*]{}, [**36**]{}, 453–471 (1990).
M. Hayashi, Universal approximation of multi-copy states and universal quantum lossless data compression, arXiv:0806.1091.
M. Hayashi, “Error exponent in asymmetric quantum hypothesis testing and its application to classical-quantum channel coding,” [*Phys. Rev. A*]{}, [**76**]{}, 062301 (2007).
[^1]: Throughout the paper, a stationary memoryless channel without using entangled input states is simply referred to as a stationary memoryless channel.
[^2]: Christandl[@Christandl] contains a good survey of representation theory for quantum information.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The neutron lifetime is determined for now most accurately by two methods –the so-called “beam method” and the method of neutron storage in a trap. The goal of this research is to obtain the neutron lifetime with a higher precision by eliminating the traditional systematic errors. A new feature is extracting the neutron lifetime only from the set of electron counting rates with their errors. The newly received neutron lifetime value is $$\tau_{NT}=883.31\pm0.02(stat.)\pm0.02(sys.) \quad \mbox{s.}$$ The obtained neutron lifetime is the weighted average of two lifetimes, defined for different subsets of neutrons providing the well-known effect of the asymmetry in the neutron $\beta$-decay. Simple calculations predict the values of these lifetimes and show their correspondence to the known asymmetry parameters of neutron decay. The arithmetic mean (central) value for the two newly introduced neutron lifetimes was determined. The resulting weighted value $\tau_{NT}$ is in a good agreement with the lifetimes in the first and second methods in limits of their double errors, but exceeds them significantly in precision. In addition, estimates of the new lifetimes, which are called here as $L$-neutron lifetime and $R$-neutron lifetime, are fulfilled. The estimation results correspond to the known parameters of the neutron decay asymmetry.'
author:
- |
V.V. Vasiliev\
e-mail [ basil\[email protected]]{}
date: 'Received: date / Accepted: date'
title: 'Direct determination of neutron lifetime in $\beta$-decay'
---
Introduction {#intro .unnumbered}
============
Measuring the neutron lifetime is a whole epoch in the history of nuclear physics. This era began in the late 1940s and consists of two periods. In the first period, the results for the neutron lifetime exceeded 900 seconds. The so-called beam method measured a counting rate of protons or electrons from the decay of neutrons $n \to p+e^-+\tilde \nu $ in a neutron beam of a nuclear reactor [@SS59], [@Chr72], [@Brn80]. For the second period, typical neutron lifetimes were less than 900 seconds. During this period, another method of measuring the neutron lifetime becomes the leading one, namely the method of storing ultracold neutrons in a trap until beta-decay. In 2018, this method yielded the neutron lifetime equal to $881.5 \pm 0.9$ s [@Srb2018]. The most accurate result of the beam method is $\tau_n=887.7 \pm2.2$ s [@Yue2013]. There were cases of recalculating the results of the experiments and shifting it from the upper range of values to the lower one. Herewith, the shift in the estimates of the lifetime significantly exceeded the indicated experimental errors. In the case of [@SS59] in 1959, the result was $1013 \pm26$ seconds and was reduced in 1978 to the value of $877 \pm 8$ seconds [@BS78] while repeating the basic scheme of the experiment. In the experiment [@Brn80] the result of 1980 was $937 \pm 18$ seconds, but 16 years later, the authors published their result [@Brn96] as equal to $889.2 \pm 4.8$ s. The averaging out all the results for the whole mentioned measurement period without any restriction leads to an average neutron lifetime of about 900 s.
The beam experiments use the differential decay equation: $$\frac{dN_d}{dt}=\frac{1}{\tau_n} \cdot \varepsilon \times N, \eqno (1)$$ where $\frac{dN_d}{dt}$ is the counting rate of the decay electron (or proton) detector, $\tau_n$ is the neutron lifetime, $N$ is the number of neutrons at time $t$ in some region of the neutron beam, $\varepsilon$ is the total efficiency. The total efficiency includes the efficiency of neutron detection by a neutron detector, the efficiency of collecting electrons (or protons) from the decay region to the electron (proton) detector, and the efficiency of their registration by the detector. The most difficult problem is the exact experimental determination of the number $N_d=(\varepsilon \times N)$ on the right side of equation (1), i.e. the number of neutrons whose decay is in the field of view for the electron (or proton) detector. The value $(\varepsilon \times N)$ includes all sources of systematic errors of the beam method.
A new method for the beam experiments has been developed in some years. This method excludes necessities to measure precisely the number of neutrons in the beam, and eliminates the need to determine the absolute values of the efficiency components.
Variation method of neutron decay scale tuning {#sect1:VARM}
===============================================
The method proposed by the author [@VV1] uses a step-wise variation of the initial number of neutrons passing through a region controlled by a detector of electrons. The method is based on a system of differential equations of type (1) for $k$ steps of neutron numbers: $$R_d(i)=\frac{1}{\tau_n} \times N_d(i), \eqno (2)$$ where $R_d(i)$ is the electron count rate from the detector, $N_d(i)$ is a number of neutrons seen by the electron detector at the $i$-th variation step, $i=1, 2, 3,\cdots k$. The problem of measuring the number of neutrons is not posed here at all. At each $i$-th stage of the neutron number variation the count rate $R_d(i)$ of the electron detector is measured and, after multiple iterations, the count rate error $\sigma_d(i)$ at each stage is determined. As a result, an array $(R_d(i),\sigma_d(i))$ with $k$ lines is formed. To simplify the notation, it is worth eliminating the indices $d$ and $n$ in (2). The next step is to represent an unknown set of neutron numbers $N_i$ by members of an arithmetic progression $N_i\approx \frac{1}{\mu}\times m_i$ with $\frac{1}{\mu}$ as a decimal common difference. The common difference of the required arithmetic progression is the step of the neutron number scale that describes the distribution of counting rates with their errors. The integer $m_i$ is the number of the scale division corresponding to the neutron number $N_i$. The parameter $\mu$, the inverse of the scale step, is called a scale factor or $\mu$-factor. Then the system of differential equations of decay has the following form: $$\tau \times (R_i \pm \sigma_i)\approx \frac{1}{\mu} \times m_i, \eqno (3)$$ where $i=1, 2, 3,\cdots k$, $k$ is the full number of variation steps.
The goal is set as follows. It is necessary to choose an optimal scale step to describe in the best way the measured data array of count rates by a certain sample of members of the obtained arithmetic progression of neutron numbers. The $\frac{1}{\mu}$ scale step uniquely identifies the set of integers $m_i$ - the scale division numbers corresponding to the array of pairs $(R_i,\sigma_i)$ for a given value of $\tau$–trial lifetime. The estimate $\aleph_i$ of neutron number $N_i$ is $$\aleph_i =
round \left[\frac{round \left[\mu \cdot \tau \cdot R_i,0 \right]}{\mu}, p \right] \mbox{.} \eqno (4)$$ The operator $round[C, p]$ rounds to the nearest number with $p$ significant digits. The operator $round[C, 0]$ means rounding $C$ to the nearest integer. The following error functional is constructed for the required range of $\tau$ for different $p$: $$F_{\mu,p}(\tau)=\sum_{i=1}^k \frac{(R_i-\frac{1}{\tau} \cdot
\aleph_i(p,\mu,\tau))^2}{\sigma^2_i} \mbox{.} \eqno (5)$$ The scale factor $\mu_0$ is to be selected for the best approximation by the estimate $\aleph_i(p,\mu_0,\tau)$ for any $p$. The neutron lifetime $\tau_0$ is determined from the equation $$\frac{dF_{\mu_0,p}(\tau)}{d\tau}=0 \mbox{.} \eqno (6)$$
![Description of count rates with integer neutron numbers.[]{data-label="fig1"}](fig1_lines)
Fig. \[fig1\] illustrates the method in the case of integer neutron numbers. The figure shows the best correspondence of count rates to the set of neutron numbers 1, 2, 3, 5 by the straight line L1. The path AB-BC-CD-DE-EF describes the action of the operator (4) in the case of the count rate $R2$ for the line L2. At the counting rate $R3$ the operator implements the trajectory GH-HK-KD-DE-EF. The description of the line L2 in integers leads to an increase in the error functional by the deviations FA and GF. Hence a deviation from the optimal line (from L1 to L2) leads to an increase in the approximation error while processing in the same neutron number scale. The main requirement for obtaining an accurate result is a high accuracy of counting rate measurements.
Experimental data {#sect2:ExpData}
=================
The experimental data of background measurement in the last ITEP experiment on the magnetic storage of ultra-cold neutrons (UCN) was used. The background consisted of electrons generated in the vacuum chamber of the magnetic trap by the decay of neutrons. Electrons were transported from the trap to the UCN detector [@VV2]. The proportional gas chamber of the UCN detector operated as an electron detector of high efficiency. The vertical and horizontal channels of the reactor with open shutters were sources of thermal, intermediate and fast neutrons of the neutron background. The set of those neutrons penetrating through the walls of the trap was the generator of decay electrons.
To measure the electron background, a separate long experiment was performed. The low pressure gas detector was specially optimized for counting electrons emanating from the magnetic trap. A special absorber was installed into the trap chamber for eliminating the ultracold neutrons. A virtually complete storage cycle for electrons collected in the trap from background neutrons was carried out. The electron counts in the intervals with the magnetic shutter on and in the drain intervals with the magnetic shutter off were measured. The counting of the electron background from the neutron flux through the magnetic trap was an analogue of the beam experiment. The count of electrons flowing to the detector from the magnetic trap changed cyclically with the changes in the set of simultaneously operating neutron channels of the reactor. The data on background measurements in the readout intervals of the outgoing electrons were processed and shown in the order of growth in Fig. \[fig2\]–Fig. \[fig4\].
![Electron count rate vs number. Number 1-71.[]{data-label="fig2"}](fig2_rates_a)
![Electron count rate vs number. Number 72-145.[]{data-label="fig3"}](fig3_rates_b)
![Electron count rate vs number. Number 147-151.[]{data-label="fig4"}](fig4_rates_c)
The full series of 152 count rates is divided into two series. There are seventy-one values (“series 71”, S-71) in the first series (Fig. \[fig2\]). The most accurate among these values is $6.6667\cdot10^{-3} \pm 6\cdot10^{-7}$ $s^{-1}$. In addition, eighty-one values formed the second series (“series 81”, S-81) Fig. \[fig3\] –Fig. \[fig4\]. Out of 81 measurements, for reasons of scale, two points are not shown: $$\mbox{point number} \quad 146: R = 1.355 \cdot10^{-2} \pm 6 \cdot 10^{-5} \quad s^{-1}$$ and $$\mbox{point number} \quad 152: R = 3.738 \cdot10^{-2} \pm 4 \cdot10^{-4}\quad s^{-1} \mbox{.}$$ There are several steps of the background electrons in Fig. \[fig2\] –Fig.\[fig4\]. Fig. \[fig4\] shows the count of electrons flowing out of the magnetic trap after opening the magnetic shutter. All results for different reading-out intervals were received over a period of more than 100 days. The paper [@VV2] described the scheme and description of the experimental set-up in more details.
Decay asymmetry and neutron lifetime {#sect3:DAs}
====================================
The well-known phenomenon of electron-spin asymmetry of the neutron beta decay $n \to p+e^{-}+\tilde \nu$ reveals itself in the fact that neutron decays with an electron emitting in the direction of the neutron spin occur less frequently than neutron decays with an electron emitting against the neutron spin direction. The neutron decay probability [@Jacks] modified for decay electrons emitting is $$dW(E_e,\Theta_e)=W_0dE_ed\Theta_e\left(1+A\cdot\frac{v_e}{c}
\cdot\cos\theta_e\right) \mbox{,} \eqno (7)$$ where $A$ is the correlation coefficient of the electron emission with the direction of the neutron spin, $\theta_e$ is an electron emission angle relative to the direction of the neutron spin, $\Theta_e$ is a solid angle of electron emission, $\frac{v_e}{c}$ is an electron helicity, and $W_0$ is a constant. From numerous experiments [@PDG] it is known that the coefficient $A = - 0.1173 \pm 0.0013$. Thus, neutron decays with electron emission in the direction of the neutron spin and against the spin differ qualitatively and quantitatively. The decay asymmetry in case of a transversely polarized neutron beam is the relative difference of the electrons emitted in the direction of the neutron spin and against the direction of the neutron spin. However, the fact that even with a completely depolarized ensemble of neutrons, the asymmetry of decay leads to two different frequencies of electron generation in the decay region remains unnoticed. Nevertheless, the asymmetry of neutron decay is a phenomenon of existence of two $\beta$-decay constants, i.e. two reduced decay frequencies defined at different subsets of neutrons. The total set $T$ of neutrons is a sum of two subsets. Those are the subset of $L$-neutrons with decays by an electron ejection against the direction of the neutron spin ($L$-channel) and the subset of $R$-neutrons, decaying with the emission of an electron in the direction of the neutron spin ($R$-channel). Hence $T=L\cup R$ and the number of neutrons $N_T$ in the total set $T$ is the sum of $L$-neutrons ($L$-subset) and $R$-neutrons ($R$-subset): $N_T=N_L+N_R$. Without loss of generality the decay constants for these neutron sets are equal to $$\lambda_S=\frac{1}{N_{S}}\cdot\frac{dN_{S}}{dt}\mbox{,} \eqno (8)$$ where $S=T, L, R$. Differentiation of the sum $N_T$ in the expression (8) for $S=T$ with the subsequent insularity of the partial constants for $S=L$ and $S=R$ leads to the expression: $$\lambda_T=\lambda_L \cdot W_L+ \lambda_R \cdot W_R \mbox{.} \eqno (9)$$ Here the total decay constant $\lambda_T$ has the form of a weighted average of the partial constants $\lambda_L$ and $\lambda_R$. The weights $W_L$ and $W_R$, when use the parallel decay rule [@Blatt p. 344] as $\frac{N_R}{N_L}=\frac{\lambda_R}{\lambda_L}$ are the following relations $$W_L=\frac{N_L}{N_L + N_R}=\frac{\lambda_L}{\lambda_L + \lambda_R} \mbox{,} \eqno (10.1)$$ $$W_R=\frac{N_R}{N_L + N_R}=\frac{\lambda_R}{\lambda_L + \lambda_R} \mbox{.} \eqno (10.2)$$ The lifetime of neutrons in every $S$-set is $\tau_{NS} = \frac {1} {\lambda_S}$ and the total lifetime on the full set $T$ is equal to $\tau_{NT} = \frac {1} {\lambda_T}$. Therefore, a record for the total lifetime follows from (9) in the form of a weighted average value, namely: $$\tau_{NT}=\tau_{NL} \cdot W_{\tau L}+\tau_{NR} \cdot W_{\tau R}\mbox{,} \eqno (11)$$ where the weights for the partial lifetimes receive the following forms: $$W_{\tau L} = \frac{\tau^2_{NR}}{\tau^2_{NL}+\tau^2_{NR}}\mbox{,} \eqno (12.1)$$ $$W_{\tau R} = \frac{\tau^2_{NL}}{\tau^2_{NL}+\tau^2_{NR}}\mbox{.} \eqno (12.2)$$ The introduced partial decay constants receive the following dependence on the asymmetry parameter $\Delta$ $$\lambda_L=\lambda_0\cdot (1+\Delta) \mbox{,} \eqno (13.1)$$ $$\lambda_R=\lambda_0\cdot (1-\Delta) \mbox{,} \eqno (13.2)$$ where $\lambda_0$ is the arithmetic mean of the introduced constants, $\Delta=A\cdot\frac{\bar v_e}{c}$ , where $\frac{\bar v_e}{c}$ is the average helicity of electrons emitted during neutron decay. The average helicity of electrons can be calculated from the electronic neutron decay spectrum, for example, from the results of [@Robsn]. The partial lifetimes are $$\tau_{NL}=\tau_{NCenter} \cdot (1-\Delta) \mbox{,} \eqno (14.1)$$ $$\tau_{NR}= \tau_{NCenter} \cdot (1+\Delta) \mbox{,} \eqno (14.2)$$ where $\tau_{NCenter}$ is a central neutron life time , i.e. the central point between values of $\tau_{NL}$ and $\tau_{NR}$ . The weighted average decay constant $\lambda_T$ is related to the average decay constant $\lambda_0$ as follows: $$\lambda_T=\lambda_0 \cdot (1+\Delta^2) \mbox{.} \eqno (15)$$ The total neutron lifetime $\tau_{NT}$ vs $\Delta$ is $$\tau_{NT}=
\tau_{NCenter} \cdot \left(1-\frac{2 \cdot \Delta^2}{1+\Delta^2}\right) \mbox{.} \eqno (16)$$ Therefore, the dependence of the error functional (5) on the trial lifetime is symmetrical with respect to the point $\tau_{NCenter}$. Then the weighted average lifetime according to (16) is to the left of the center of symmetry. The goal of this study, therefore, turned out to be two-fold. Firstly, it is a direct determination of an experimental value for the observed neutron lifetime, i.e. the weighted average neutron lifetime. On the other hand, it became necessary to determine the value of the “central” neutron lifetime. In the case of a convincing correspondence between a determined displacement value and a calculated value (16), it is advisable to introduce new physical quantities into the physical dictionary – the lifetimes of $L$-neutrons and $R$-neutrons and to estimate their numerical values using formulas (14.1) and (14.2).
Direct determination of the neutron lifetime from the experimental data {#sect4:DDED}
=======================================================================
The interval of the trial lifetime from 860 seconds to 940 seconds is quite informative for solving the given problem. This interval includes the results of measurements of the lifetime of the last three decades. To determine the neutron lifetime, it is necessary to find the minimum value of the scale-factor $\mu$, providing the condition of “reduction-to-one”. This condition means that the minimum point of the error functional is the closest to the unity for any data series in the considered range of the lifetime. The “reduction-to-one” is applied to the data shown in Fig. \[fig2\]–Fig. \[fig4\] for various values of the scale step in the indicated range of the trial neutron lifetime $\tau$. The results of the data processing are illustrated by Fig. \[fig5\]–Fig. \[fig7\].
![ Error functional dependence on trial lifetime for S-152.[]{data-label="fig5"}](fig5_sigma_152)
Fig. \[fig5\] shows the result of the reduction-to-one for the S-152 series with $p=3$. The closest to the unity is the value of the functional minimum equal to 1.074, corresponding to the lifetime of 883.305 s for the scale-factor $\mu =72$. At the same scale-factor $\mu$, minimums of the functional of the second, third and fourth orders of accuracy reached a good agreement. The value for the neutron lifetime is $\tau_{NT}=(529983\pm10)\cdot\frac{1}{6}\cdot10^{-2}$ s after using the half-width of the parabola of the third order functional at height $\chi^2=1.2$.
![ Error functional dependence on trial lifetime for S-71 and S-81.[]{data-label="fig6"}](fig6_sigma_71_81)
Fig. \[fig6\] shows the results of the reduction-to-one for the S-71 (1) and S-81 (2) series separately. The dependence of the average weighted functional over two series on the lifetime gives the result for the neutron lifetime $\tau_{NT}=(264992\pm7)\cdot\frac{1}{3}\cdot10^{-2}$ s. Thus, the complete result for the neutron lifetime is $\tau_{NT}=883.31\pm0.02$ s, 95%, $CL$.
All these facts prove the following:
a\) both independent data series S-71 and S-81 are qualitatively homogeneous and describe electrons from the decay of neutrons without admixture of an additional background;
b\) the results of independent data series S-71 and S-81 are compatible within statistical errors, reasonably processed together and do not require an additional introduction of a systematic error to describe the consolidated result.
Therefore, the applied method has reached its goal of eliminating known systematic errors for higher accuracy in the experimental determination of the neutron lifetime.
Additional studies were made to prove the existence of a common center of symmetry of the error functional on the trial lifetime interval from 895 to 905 seconds for all orders of accuracy $p = 0, 1, 2, \cdots $.
![Center of symmetry of the error functional for S-152 at $\mu=72$[]{data-label="fig7"}](fig7_symmetry)
Fig. \[fig7\] shows results for the scale-factor $\mu=72$ by dependencies (1)-(3) of error functional $F_{\mu,p}(\tau)$: (1) integers ($p = 0$); (2) $p=1$; (3) $p=2$. The coincidence of the minimums of all three orders of the functional on the $\tau$-scale at the point of 899.975 s, with symmetry of the left and right wings of the functional shows that this point is the center of symmetry and confirms the result of [@VV2]. This value can serve as an estimate of the so-called “central” neutron lifetime. The symmetry of the coordinates of the jumps of the functional (3) with respect to this $\tau_{NCenter}=899.975$ s point also confirms its central position. The dependence of the functional on the trial lifetime allows determining the value of the central neutron lifetime with a completely moderate error $\tau_{NCenter}=(179995\pm 4)\cdot 5 \cdot 10^{-3}$ s. It is important to emphasize that the error functional in the range of the trial lifetime of 750-1050 s has only one specified center of symmetry for all accuracy level $p$.
Upper limit of a systematic error {#sect5:SYS}
=================================
As a source of systematic error, a hypothetical difference between the electrons flowing through the closed magnetic shutter and the electrons flowing onto the detector through the open magnetic shutter may appear. If only the electrons of the first type are presented in the first S-71 series, then in the second S-81 series there is a fraction of electrons of the second type providing the maximum values of the counting rate. The result indicated in Fig. \[fig6\] proves that the independently obtained lifetime values in these series are mutually compatible within the statistical error. In the case of functional (1), the result is $883.29 \pm 0.03$ s, and in the case of functional (2), the result is $883.33 \pm 0.03$ s. The errors indicated here are equal to the half-width of the parabolas at the level $\chi^2=1.3$. It means that the displacement is within the limits of the statistical errors at 95$\% CL$. Calculation of the half-sum of these lifetimes and its error gives $\tau_{NT}=883.31\pm 0.02$ s, and comparison with the result shown in Fig. \[fig5\] confirms complete coincidence. Moreover, the minimums of the functional for S-152, $\mu=72$ at $p=2$, $p=3$ and $p=4$ coincide in those limits altogether.
Nevertheless, the fact that the minimums at the accuracy $p=4$ shift from 883.29 s for S-71 to 883.33 s for S-81 gives a reason to interpret it due to some unaccounted factors and to introduce on this basis a systematic error equal to the half of the bias value. Thus, the estimate of the systematic error is 0.02 s, and the result for the neutron lifetime as a weighted average is $\tau_{NT}=883.31 \pm 0.02(stat.)\pm 0.02(sys.)$ s, $95\% CL$. In the same format, the central neutron lifetime has got the following value $\tau_{NCenter}=899.98 \pm 0.02(stat.) \pm 0.02(sys.)$ s.
Discussion of the result {#sect6:DISR}
========================
In details the previous two results in the introduction are $$887.7 \pm 1.2 (stat.) \pm 1.9 (syst.) \quad \mbox{s}$$ for the beam method, and $$881.5 \pm 0.7 (stat.) \pm 0.6 (syst.) \quad \mbox{s}$$ for the storage method. Taking into account limits within double total errors for these two results, it is easy to note the coincidence of the result of this work with the lower limit for the first result (883.3 s) and with the upper limit for the second result (883.3 s) . Therefore, a good agreement between these three results, including the presented one, is evident.
It also means that there are no grounds for any conclusions about the so-called “neutron anomaly” as a possible interpretation of the difference between the two mentioned results of measuring the neutron lifetime.
A side and unexpected result of the present research is the above evidence of the difference between the two obtained neutron lifetimes, the first of them is the weighted average value and the second is the “central” lifetime of neutron. Within the error limits (0.02 s), the obtained value of the weighted average neutron lifetime in a more convenient form is $$\tau_{NT}=900 \cdot \frac{53}{54} \quad \mbox{s,}$$ while the central lifetime is $\tau_{NCenter}\approx 900$ s within the same error. The value of the relative shift $$\delta=\frac{\tau_{NCenter} - \tau_{NT}}{\tau_{NCenter}+\tau_{NT}}$$ is a characteristic of the displacement of the weighted average lifetime relative to the central lifetime. Then the estimate of the relative shift is $$\delta=\frac{1}{107} \mbox{.}$$ The weighted average lifetime is connected with the central lifetime as $$\tau_{NT}=\tau_{NCenter} \cdot \left(1-\frac{2 \cdot \delta}{1+\delta}\right) \mbox{.}
\eqno (17)$$ From (16) $\delta$ is interpreted as $\delta=\Delta^2$. The obtained value for $\Delta=\frac{1}{\sqrt{107}}$ is in a good agreement with $A = - 0.1173 \pm 0.001$ and $\frac{\bar v_e}{c}=0.824$. This value for the average helicity is confirmed by the spectrum of electrons from the neutron decay [@Robsn].
Thus, the displacement of the weighted average neutron lifetime relative to the so-called central lifetime is a consequence of the electron-spin asymmetry of neutron decay. The main purpose of this research is the direct determination of the observed neutron lifetime. Nevertheless, it is easy to predict from the displacement the numerical values for the lifetime of $L$-neutrons $\tau_{NL}$ (14.1) and the lifetime of $R$-neutrons $\tau_{NR}$ (14.2). They are the following: $$\tau_{NL}=(2 \cdot 3 \cdot 5)^2 \cdot \left (1-\frac{1}{\sqrt{107}} \right) \quad \mbox{s,}$$ $$\tau_{NR}=(2 \cdot 3 \cdot 5)^2 \cdot \left (1+\frac{1}{\sqrt{107}} \right) \quad \mbox{s.}$$ The errors of the indicated values are estimated not exceeding 0.08 s. A direct determination of these parameters of the neutron $\beta$-decay will make it possible in the future to indicate their values more accurately.
Now it is worth noticing that the weighted average of the mentioned above values of $\tau_{NL}$ and $\tau_{NR}$ is in a good agreement with the neutron lifetime obtained in this research from the experimental data. This can be easily verified using numerical expressions for the lifetime weights of (12.1) and (12.2), namely: $$W_{\tau L}=\frac{1}{2}+\frac{\sqrt{107}}{108} \mbox{,}$$ $$W_{\tau R}=\frac{1}{2}-\frac{\sqrt{107}}{108} \mbox{.}$$ The calculation result by formula (11) gives 883.33 s that coincides with the experimental result of this work within $\pm 0.01$ s.
Conclusion {#sect7:CONCL}
==========
Using the proposed method, the neutron lifetime is determined with the accuracy of 0.02 s. The system of differential equations of decay with a step-wise variation of the number of neutrons turned out to be a sufficient tool for determining the neutron lifetime by the modified least-squares method. Traditional sources of systematic errors inherent in the beam method of measuring the neutron lifetime are successfully excluded. As a result, the accuracy of the neutron lifetime is significantly improved. The observed (weighted average) neutron lifetime obtained in this work is expressed in primes as $$\tau_{NT}=2 \cdot 5^2 \cdot \frac{53}{3} \quad \mbox{s}$$ within no more than 0.03 s (taking into account the introduced systematic error).
In addition to the main result, estimates of the effect of splitting the neutron lifetime are obtained. It is indicated that neutron decay can be described by two lifetimes: $L$-neutron lifetime and $R$-neutron lifetime, differing in the sign of the scalar product of the electron momentum and the neutron spin. The estimation results correspond to the known parameters of the electron-spin asymmetry of the neutron decay. This fact provides the grounds for introducing new quantities - $L$-neutron lifetime and the $R$-neutron lifetime into the neutron $\beta$ -decay physics. The obtained interval of neutron lifetimes from $\tau_{NL} = 813$ s to $\tau_{NR} = 987$ s gives an explanation of the range of experimental values obtained over the 70-year history of of the neutron lifetime measurements. This method opens up the prospect of reducing the neutron lifetime errors to thousandths of a second.
The author is deeply grateful to V.V. Vladimirsky for his personal support. The author is grateful to the staff of the UCN-ITEP group - V. F. Belkin, A. A. Belonozhenko, N. I. Kozlov, E. N. Mospan, A. Yu. Karpov, I. B. Rozhnin, A.M. Salomatin , V.G. Frankovskaya, and also and especially G.M. Kukavadze and S.P. Boroblev for the creation and operation of the KENTAVR set-up at the ITEP HWR reactor.
[99]{} Sosnovskii A.N., Spivak P.E., Prokof’ ev Yu. A., et al. Measurement of the Neutron Half-life. JETP, vol. 9, Num. 4, p. 717, (1959). Christensen C.J., Nielsen A., Bahnsen A. , et al. Free-Neutron Beta-Decay Half-Life. Phys. Rev., D5, Num. 7, p. 1628, (1972). Byrne J., Morse J. , Smith K. F., et al. A new measurement of the neutron lifetime. Physics Letters B, Vol. 92, rel: 3-4, p. 274-278, (1980). Serebrov A.P. , Kolomensky E.A., Fomin A.K. , et al. Neutron lifetime measurements with a large gravitational trap for ultracold neutrons. Physical Review C, vol. 97, rel: 5, Num. 055503, (2018). Yue T. Y, Dewey M. S. , Gilliam D. M., et al. Determination of the Neutron Lifetime. Phys. Rev. Lett., 111, 222501, (2013). Bondarenko L.N. , Kurguzov V.V. , Prokof’ev Yu. A. , et al. Measurement of the neutron half-life period, JETP Letters, vol.28, rel. 5, pp. 329-333, (1978). Byrne J. , Dawber P.G., Habeck C.G. , et al. A revised value for the neutron lifetime measured using Penning trap, Europhysical Letters, 33 (3), pp.187-192, (1996). Vasiliev V.V., A method for measuring radioactive decay constant. Bulletin of inventions Num. 36, 2009. RF patent of invention Num. 2377599, G01T 1/30. Vasiliev V.V., Neutron Lifetime and the Background Structure in a Neutron Magnetic Trap, JETP Letters, vol.77, Num. 5, pp. 207-2011, (2003). Jackson J.D., Treiman S.B. and H.D. Wyld, Jr., Possible Tests of Time Reversal Invariance in Beta-Decay, Physical Review, vol.106, Num. 3, p.517-521, (1957). K. Nakamura et al. (Particle Data Group), J. Phys. G 37, 075021, p.10 , (2010). Blatt J.M., Weisskopf V.F., Theoretical nuclear physics, John Wiley & Sons, New York (1952). Robson V.V., The Radioactive Decay of the Neutron, Physical Review, vol.83, rel. 2, p.349-358 (1951).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Classification algorithms face difficulties when one or more classes have limited training data. We are particularly interested in classification trees, due to their interpretability and flexibility. When data are limited in one or more of the classes, the estimated decision boundaries are often irregularly shaped due to the limited sample size, leading to poor generalization error. We propose a novel approach that penalizes the Surface-to-Volume Ratio (SVR) of the decision set, obtaining a new class of SVR-Tree algorithms. We develop a simple and computationally efficient implementation while proving estimation and feature selection consistency for SVR-Tree. SVR-Tree is compared with multiple algorithms that are designed to deal with imbalance through real data applications.'
author:
- |
Yichen Zhu[^1]\
Department of Statistical Science, Duke University\
and\
David B. Dunson[^2]\
\
Department of Statistical Science, Duke University
bibliography:
- 'IBbib.bib'
title: '**Classification Trees for Imbalanced and Sparse Data: Surface-to-Volume Regularization**'
---
1
[1]{}
0
[1]{}
[**Classification Trees for Imbalanced and Sparse Data: Surface-to-Volume Regularization**]{}
[*Keywords:*]{} CART, Categorical data, Decision boundary, Feature Selection, Shape penalization, Variable selection
Introduction
============
We are interested in the common setting in which one has a set of training data $\mathscr{D}_n = \{(X_i, Y_i)\}_{i=1}^n$, with $X_i \in \Omega \subset \mathbb{R}^d$ a vector of features and $Y_i \in \{0,\ldots,J-1\}$ a class label. The goal is to estimate a classifier $f: \Omega \to \{0,\ldots,J-1\}$, which outputs a class label given an input feature vector. This involves partitioning the feature space $\Omega$ into subsets having different class labels, as illustrated in Figure 1. Classification trees are particularly popular to their combined flexibility and interpretability. Competitors like deep neural networks (DNN; @schmidhuber2015deep) often have higher accuracy in prediction, but are essentially uninterpretable black boxes.
Our focus is on the case in which $n_j = \sum_{k=1}^n \bm{1}_{\{Y_i = j\}}$ is small, for one or more $j \in \{0,\ldots,J-1\}$. For simplicity in exposition, we assume $J=2$, so that there are only two classes. Without loss of generality, suppose that $j=0$ is the majority class, $j=1$ is the minority class, and call set $\{x\in\Omega: f(x)=1\}$ the decision set (of the minority class). Hence, $n_1$ is relatively small compared to $n_0$ by this convention.
To illustrate the problems that can arise from this imbalance, we consider a toy example. Let the sample space of $X$ be $\Omega = [0,1]\times[0,1]\subset \mathbb{R}^2$. Let the conditional distribution of $X$ given $Y$ be $X|Y=1 \sim U([0,0.75]\times[0.25,0.75])$, $X|Y=0 \sim U(\Omega)$. We generate a training data set with $5$ minority samples and $200$ majority samples. The estimated decision sets of unpruned CART and optimally pruned CART [@breiman1984classification] are shown in Figure (a) and (b), respectively. Minority samples are given weight 32 while majority samples are given weight 1 to naively address the imbalance. Both decision sets are inaccurate, with unpruned CART overfitting and pruned CART also poor.
Imbalanced data problems have drawn substantial interest; see [@haixiang2017learning], [@he2008learning], [@krawczyk2016learning] and [@fernandez2017insight] for reviews. Some early work relied on random under- or over-sampling, which is essentially equivalent to modifying the weights and cannot address the key problem. [@chawla2002smote] proposed SMOTE, which instead creates synthetic samples. For each minority class sample, they create synthetic samples along the line segments that join each minority class sample with its k nearest neighbors in the minority class. Building on this idea, many other synthetic sampling methods have been proposed, including ADASYN [@he2008adasyn], Borderline-SMOTE [@han2005borderline], SPIDER [@stefanowski2008selective], safe-level-SMOTE [@bunkhumpornpat2009safe] and WGAN-Based sampling [@wang2019wgan]. These synthetic sampling methods have been demonstrated to be relatively effective.
However, current understanding of synthetic sampling is inadequate. [@chawla2002smote] motivates SMOTE as designed to “create large and less specific decision regions”, “rather than smaller and more specific regions”. Later papers fail to improve upon this heuristic justification. Practically, the advantage of synthetic sampling versus random over-sampling diminishes as the dimension of the feature space increases. In general, for each minority sample, we require at least $d$ synthetic samples to fully describe its neighborhood. This is often infeasible due to the sample size of the majority class and to computational complexity. Hence, it is typical to fix the number of synthetic samples regardless of the dimension of the feature space [@chawla2002smote], which may fail to “create large and less specific decision regions” when the dimension is high.
Motivated by these issues, we propose to directly penalize the Surface-to-Volume Ratio (SVR) of the decision set. A primary issue with imbalanced data is estimating a decision set consisting of small neighborhoods around each minority class sample. By penalizing SVR we favor regularly shaped decision sets much less subject to such over-fitting. With this motivation, we propose a new class of SVR-Tree algorithms.
The rest of the paper is organized as follows. Section 2 describes our methodology. Section 3 provides theory on estimation and feature selection consistency. Section 4 presents numerical studies for real datasets. Section 5 contains a discussion, and proofs are included in an Appendix.
[0.32]{} ![Decision sets for different methods. Red crosses denote minority class samples, while blue points denote majority class samples. Rectangles with dashed frames denote the support of minority class samples, while rectangles filled with green color denote the minority class decision set.[]{data-label=""}](toy_CART_unpruned.png "fig:"){width="\textwidth"}
[0.32]{} ![Decision sets for different methods. Red crosses denote minority class samples, while blue points denote majority class samples. Rectangles with dashed frames denote the support of minority class samples, while rectangles filled with green color denote the minority class decision set.[]{data-label=""}](toy_CART_pruned.png "fig:"){width="\textwidth"}
[0.32]{} ![Decision sets for different methods. Red crosses denote minority class samples, while blue points denote majority class samples. Rectangles with dashed frames denote the support of minority class samples, while rectangles filled with green color denote the minority class decision set.[]{data-label=""}](toy_SVRTree.png "fig:"){width="\textwidth"}
Methodology
===========
We first introduce the definition of surface-to-volume ratio (SVR) and tree impurity, and then define SVR-Tree as the minimizer of a weighted average of tree impurity and SVR. We then state the algorithm to estimate this tree from training data $\mathscr{D}_n = \{(X_i, Y_i)\}_{i=1}^n$. We assume readers have basic knowledge of tree-based classifiers like CART [@breiman1984classification] and C4.5 [@quinlan2014c4]. In the rest of the paper, the word “tree” refers specifically to classification trees that specify a class label associated with each leaf node.
Notations
---------
Training data are denoted as $\mathscr{D}_n = \{(X_i, Y_i)\}_{i=1}^n$, with $X_i \in \Omega \subset \mathbb{R}^d$ a vector of features and $Y_i \in \{0,1\}$ a class label. Uppercase letters $(X, Y)$ denote random variables, while lowercase $x, y$ denote specific values. We denote the $j$th feature of $X$ as $X[j]$. The true distribution of $(X_i, Y_i)$ is denoted by $\mathbb{P}$, while the empirical distribution, which assigns mass $1/n$ to each training data point, is denoted by $\mathbb{P}_n$. For a constant $\alpha>1$, let $\mathbb{P}_\alpha$ modify $\mathbb{P}$ to up weight minority class samples by $\alpha$. That is, for any subset $A\subset \Omega$, $$\mathbb{P}_\alpha(A\times\{1\}) = \frac{\alpha\mathbb{P}(A\times\{1\})}{\mathbb{P}(\Omega\times\{0\})+\alpha\mathbb{P}(\Omega\times\{1\})}, \;\; \;\mathbb{P}_\alpha(A\times\{0\}) = \frac{\mathbb{P}(A\times\{0\})}{\mathbb{P}(\Omega\times\{0\})+\alpha\mathbb{P}(\Omega\times\{1\})}.$$ Similarly, let $\mathbb{P}_{n,\alpha}$ be the weighted version of $\mathbb{P}_n$. To avoid complex subscripts, with some abuse of notation, we also use $\mathbb{P}, \mathbb{P}_\alpha, \mathbb{P}_n, \mathbb{P}_{n,\alpha}$ to denote the corresponding marginal probability measures on $X$. For example, $$\mathbb{P}_\alpha(A) \triangleq \mathbb{P}_\alpha(A\times\{0,1\}) = \frac{\mathbb{P}(A\times\{0\})+\alpha\mathbb{P}(A\times\{1\})}{\mathbb{P}(\Omega\times\{0\})+\alpha\mathbb{P}(\Omega\times\{1\})}.$$ Whether they refer to the joint distribution of $(X, Y)$ or the marginal distribution of $X$ should be clear from the context. When discussing the probability of certain events that include $n$ random variables $\{(X_i, Y_i)\}_{i=1}^n$, we simply use $\mathbb{P}$ to represent the probability measure in the $n$-product space. For example, for $\mu>0$, $\mathbb{P}\left((\sum_{i=1}^n X_i)/n >\mu\right)$ denotes the probability of the event $\{\{(X_i, Y_i)\}_{i=1}: (\sum_{i=1}^n X_i)/n > \mu\}$. We use $\mathbb{E}$ for expectations over $\mathbb{P}$ and $\mathbb{E}_n$ for expectations over $\mathbb{P}_n$.
Surface-to-Volume Ratio
-----------------------
For all $d \in \mathbb{N}$, define a $d$-dimensional measure space as $(\mathbb{R}^d, \mathscr{B}, \mu)$, where $\mathscr{B}$ is the collection of Borel sets, and $\mu$ is Lebesgue measure. For any measurable closed set $A \in \mathscr{B}$, we define its volume as Lebesgue measure of set $A$: $V(A) = \mu(A)$. We define the surface as the $d-1$ dimensional Lebesgue measure of the boundary of $A$: $S(A) = \mu_{d-1}(\partial A).$ For any set $A$ with $0<\mu(A)<\infty$, the surface-to-volume ratio (SVR) can be obtained as $r(A) = \frac{S(A)}{V(A)}.$ For sets with the same volume, the $d$-dimensional ball has the smallest SVR, while sets having multiple disconnected subsets and/or irregular boundaries have relatively high SVR.
#### Surface-to-Volume Ratio of a Classification Tree
For training data $\mathscr{D}_n$, we define a closed bounded sample space $\Omega\subset \mathbb{R}^d$ such that the support of $X$ is a subset of $\Omega$. A classification tree $T$ divides the sample space $\Omega \subset \mathbb{R}^d$ into two disjoint subsets $\Omega_0$, $\Omega_1$, where $\Omega_0 \cup \Omega_1= \Omega$. The tree $T$ predicts a new sample $X^*$ belongs to class 0 if $X^* \in \Omega_0$, and class 1 if $X^* \in \Omega_1$. The surface-to-volume ratio of a classification tree is defined as the surface-to-volume-ratio of set $\Omega_1$: $r(T) = r(\Omega_1).$
Impurity Function and Tree Impurity
-----------------------------------
A classification tree partitions the sample space into multiple leaf nodes, assigning one class label to each leaf node. The tree should be built to maximize homogeneity of the training sample class labels within nodes. This motivates the following definition.
An impurity function $\phi(\cdot, \cdot)$ is defined on the set of pairs $(p_0, p_1)$ satisfying $p_0\ge0, p_1\ge0$, $p_0+p_1=1$ with the properties (i) $\phi(\cdot, \cdot)$ achieves its maximum only at $(1/2, 1/2)$; (ii) $\phi(\cdot, \cdot)$ achieves its minimum only at $(1, 0), (0,1)$; (iii) $\phi(\cdot, \cdot)$ is symmetrical in $(p_0, p_1)$, i.e., $\phi(p_0, p_1) = \phi(p_1, p_0)$.
Let $p_0$ and $p_1$ represent the probabilities of belonging to the majority and minority class, respectively, within some branch of the tree. Ideally, splits of the tree are chosen so that, after splitting, $p_0$ and $p_1$ move closer to 0 or 1 and further from 1/2. Many different tree building algorithms use impurity to measure the quality of a split; for example, CART uses Gini impurity and C4.5 uses entropy, which is effectively a type of impurity measure.
When data are imbalanced, it is important to modify the definition of impurity to account for the fact that $p_1$ is much smaller than $p_0$. With this in mind, we propose the following weighted impurity function.
\[weighted impurity\] Letting $\phi(\cdot, \cdot)$ be an impurity function, a weighted impurity function with weight $\alpha$ for the minority class is defined as $$\varphi_\alpha(p_0, p_1) = \phi\Big(\frac{p_0}{p_0+\alpha p_1}, \frac{\alpha p_1}{p_0 + \alpha p_1}\Big).$$
In the remainder of the paper, the term ‘impurity function’ refers to $\varphi_\alpha$ in Definition \[weighted impurity\] with $\alpha\ge1$ and the function $\phi(\cdot,\cdot)$ refers to the Gini impurity. That is, $$\varphi_\alpha(p_0, p_1) = 1 - \Big(\frac{p_0}{p_0+\alpha p_1}\Big)^2 - \Big(\frac{\alpha p_1}{p_0 + \alpha p_1}\Big)^2.$$
Let $A_1, A_2, \ldots, A_m$ be the leaf nodes of a classification tree $T$ and let $z_j\in\{0,1\}$ be the predictive class label for node $A_j$, for $j = 1,\ldots,m$. Let $\mathbb{P}$ be a probability measure over $\Omega\times\{0,1\}$. Then the impurity of leaf node $A_j$ is $I_\alpha(A_j, \mathbb{P}) = \varphi_\alpha (\mathbb{P}(Y=0|X\in A_j), \mathbb{P}(Y=1|X\in A_j))$. The impurity of node $A_j$ measures the class homogeneity in $A_j$, but does not depend on the predictive class label $z_j$. Let $\tilde{z}_j =\mathbbm{1}_{\{\mathbb{P}(Y=1|X\in A_j)\ge1/(1+\alpha)\}}$ denote the dominant class label in $A_j$ under weight $\alpha$. We define a signed impurity taking into account $z_j$ as $$\tilde{I}_\alpha(A_j, \mathbb{P}) = \mathbbm{1}_{\{z_j=\tilde{z}_j\}} I_\alpha(A_j, \mathbb{P}) + \mathbbm{1}_{\{z_j\ne\tilde{z}_j\}} (1-I_\alpha(A_j, \mathbb{P})).$$ If the predictive class label $z_j$ matches the dominant class label $\tilde{z}_j$ in node $A_j$, the signed impurity of node $A_j$ is equal to the impurity of node $A_j$ and is no greater than $1/2$. Otherwise, an extra penalty is applied. Taking a weighted average of the signed impurities across the leaf nodes, one obtains the tree impurity and signed tree impurity.
\[tree impurity\] Let $T$ be a tree and $A_1, A_2, \ldots A_m$ be all the leaf nodes of this tree. Up-weighting the minority class by $\alpha$, the tree impurity of $T$ is $$I_\alpha(T, \mathbb{P}) = \sum_{j=1}^m \mathbb{P}_\alpha(X\in A_j) I_\alpha(A_j, \mathbb{P}),$$ where $\mathbb{P}_\alpha(X\in A_j)$ is the weighted version of $\mathbb{P}$ defined in section 2.1.
\[signed tree impurity\] Under the notation of Definition \[tree impurity\], the signed tree impurity is $\tilde{I}_\alpha(T, \mathbb{P}) = \sum_{j=1}^m \mathbb{P}_\alpha(X\in A_j) \tilde{I}_\alpha(A_j, \mathbb{P}).$
SVR-Tree Classifiers
--------------------
The SVR-Tree classifier is the minimizer of the weighted average of signed tree impurity and surface-to-volume ratio. Letting $\mathscr{T}$ be the collection of possible trees, the SVR-Tree classifier is formally defined as $$\label{Tree Impu}
\hat{T} = \operatorname*{argmin}_{T\in\mathscr{T}}[\tilde{I}_\alpha(T,\mathbb{P}_n) + \lambda_n r(T)],$$ where $\lambda_n$ is a penalty. The unknown probability measure $\mathbb{P}$ is replaced with the empirical measure $\mathbb{P}_n$ that assigns mass $1/n$ to each training sample $(X_i, Y_i), 1\le i\le n$. Unfortunately, without restricting the space of trees $\mathscr{T}$, optimization problem (\[Tree Impu\]) is intractable. In the following subsection, we introduce an iterative greedy search algorithm that limits the size of $\mathscr{T}$ in each step to efficiently obtain a tree having provably good performance.
The SVR-Tree Algorithm
----------------------
The SVR-Tree Algorithm is designed to find a nearly optimal SVR-Tree classifier. SVR-Tree proceeds in a greedy manner. We begin with the root node. At each step, we operate on one leaf node of the current tree, partitioning it into two new leaf nodes by finding the solution of (\[Tree Impu\]). The node to partition at each step is uniquely specified by a breadth-first searching order. After partitioning, the tree is updated and the node to partition in the next step will be specified. The process stops when further splitting of leaf nodes either does not improve the loss or a prespecified maximum number of leaf nodes is achieved.
We first describe how to split a current leaf node to improve the solution to (\[Tree Impu\]). Suppose the current tree is $T$ and the node to partition is $A$, with $n'$ training samples. For each feature $j$, sort all samples in $A$ by increasing order of the $j$th feature as $X[j]_{j_1}, X[j]_{j_2}, \ldots, X[j]_{j_n'}$. We only allow partitions to occur at $(X[j]_{j_i}+ X[j]_{j_{i+1}})/2, \;1\le i\le n'-1,\; 1\le j\le d$, the midpoint of two adjacent values of each feature. The total number of different partitions is no more than $(n'-1)d$. After each such split of $A$, we keep all other leaf nodes unchanged while allowing all $4$ different class label assignments at the two new daughter nodes of $A$. The current set of trees $\mathscr{T}$ to choose from in optimizing (\[Tree Impu\]) include the initial $T$ and all the partitioned trees described above. The cardinality of $\mathscr{T}$ is no more than $1 + 4(n'-1)d$, a linear order of $n'$. We compute the risk for all $T\in\mathscr{T}$ to find the minimizer. If the initial $T$ is the risk minimizer, we do not make any partition in this step and mark the node $A$ as “complete”. Any node marked as complete will no longer be partitioned in subsequent steps.
It remains to specify the ‘breadth-first’ searching order determining which node to partition in each step. Let depth of a node be the number of edges to the root node, which has depth 0. Breadth-first algorithms explore all nodes at the current depth before moving on (@cormen2009introduction, chapter 22.2). To keep track of changes in the tree, we use a queue[^3]. We begin our algorithm with a queue where the only entity is the root node. At each step, we remove the node at the front terminal of the queue, and partition at this node as described in the previous paragraph. If a partitioned tree is accepted as the risk minimizer over the current set $\mathscr{F}$, we enqueue two new leaf nodes; otherwise, the unpartitioned tree is the risk minimizer over the current set $\mathscr{F}$, so we don’t enqueue any new node. The nodes in the front of the queue have the lowest depth. Therefore, our algorithm naturally follows a breadth-first searching order. We preset the maximal number of leaf nodes as $\bar{a}_n$. The process is stopped when either the queue is empty, in which case all the leaf nodes are marked as complete, or the number of leaf nodes is $\bar{a}_n$.
Our SVR-Tree algorithm has a coarse to fine tree building style, tending to first partition the sample space into larger pieces belonging to two different classes, followed by modifications to the surface of the decision set to decrease tree impurity and SVR. The steps are sketched in Algorithm 1, where feature selection steps are marked as optional. A more detailed and rigorous version is in the last section of supplemental material.
Input training data $\{(X_i, Y_i)\}_{i=1}^n$, impurity function $f(\cdot)$, weight for minority class $\alpha$, SVR penalty parameter $\lambda_n$, and maximal number of leaf nodes $\bar{a}_n\in\mathbb{N}$. Let ${\mathrm}{node\_queue}$ be a queue of only root node
#### Optional Step for Feature Selection
It is likely that some features will not help to predict class labels. These features should be excluded from our estimated tree. Under some mild conditions, a split on a redundant feature has minimal impact on the tree impurity compared to a partition in a non-redundant feature. Thus feature selection can be achieved by thresholding. Suppose we are partitioning node $A$ into two new leaf nodes $A_1, A_2$. Then the (unsigned) **tree impurity decrease** after this partition is defined as: $$\Delta I_\alpha(T, \mathbb{P}_n) = \mathbb{P}_{n,\alpha}(A) [I_\alpha(A, \mathbb{P}_n) - I_\alpha(A_1,\mathbb{P})\mathbb{P}_{n,\alpha}(A_1|A) - I_\alpha(A_2,\mathbb{P})\mathbb{P}_{n,\alpha}(A_2|A)].$$ Let $J_0\subset \{1, 2, \ldots d\}$ be the indices of features that have been partitioned in previous tree building steps. Given that we are partitioning on node $A$, let $\Delta I_0$ be the maximal tree impurity decrease over all partitions in feature $X[j], j\in J_0$. Then a partition in a new feature $X[j'], j'\not\in J_0$, with tree impurity decrease $\Delta I_\alpha(T, \mathbb{P}_n)$, is accepted if $$\label{impu decr}
\Delta I_\alpha(T, \mathbb{P}_n) \ge \Delta I_0 + c_0\lambda_n,$$ where $c_0$ is a constant independent of the training data. By equation (\[impu decr\]), a partition on a new feature is accepted if its tree impurity decrease is greater than $\Delta I_0$, the tree impurity decrease over all previously partitioned features, plus a threshold term $c_0\lambda_n$. Theoretical support for this thresholding approach is provided in Section 3.
Computational Complexity
------------------------
Recall the number of training samples is $n$ and the number of features is $d$. Denote the depth of the estimated tree as $h$ and let the maximal number of leaf nodes be $\bar{a}_n = O(\sqrt{n})$.[^4] The storage complexity is trivial and is the same as usual decision trees, i.e., $O(dn)$. We analyze the time complexity of Algorithm 1 in this section. We first introduce the approach to compute surface-to-volume ratio, then discuss the time complexity of building the tree.
#### Computing SVR
Suppose in some intermediate state of building the SVR Tree, the current tree has $m$ leaf nodes $R_1, R_2, \cdots R_m$. We already know the surface and volume of the current tree. Now we need to perform a partition at node $R_1$, which has $n'$ samples. Suppose we partition at $X[j] = x_j$ and obtain two child nodes. The volume of the tree after this partition can be computed by adding the volume of the minority class child node(s), which takes $O(d)$ time. The major concern lies in the computation of the surface area. If both child nodes are in the majority class or the minority class, the surface after partitioning is the surface that node $R_1$ is labeled as the majority class or the minority class. It takes $O(d^2)$ time to compute the surface of $R_1$, and $O(md)$ time to compute all the overlapping surface between $R_1$ and $R_j, 2\le j\le m$. Therefore, the time complexity to compute the surface of the tree after partitioning is $O(d^2+md)$.
If one child node belongs to the minority class and the other belongs to the majority class, the problem becomes more complicated. Let $S_{01,j}(x_j)$ be the surface of the partitioned tree if $R_1$ is partitioned at $X[j]=x_j$ and the left child is labeled as $0$ and the right child is labeled as $1$. $S_{10,j}(x_j)$ is similarly defined. Both $S_{01,j}(x_j)$ and $S_{10,j}(x_j)$ are piecewise constant functions whose change points can only exist at borders of $R_j, 2\le j\le m$. Therefore, we first compute the analytical forms of $S_{01,j}(x_j)$ and $S_{10,j}(x_j)$ for all $1\le j\le m$. This requires us to find all the borders of $R_j, 2\le j\le m$, to compute all the overlapping surface between $R_1$ and $R_j, 2\le j\le m$ and to compute the surface of $R_1$ itself. The process of computing the analytical forms of $S_{01}(x_j)$ and $S_{10}(x_j)$ for all $1\le j\le m$, takes $O(d^2+md)$ time. The cost of evaluating $S_{01,j}(x_j)$ and $S_{10,j}(x_j)$ at a specific value $x_j$ can takes as much as $O(m)$ time; but if the samples are pre-sorted for all features, it only takes $O(dm+dn')$ time to evaluate $S_{01}(x_j)$ and $S_{10}(x_j)$, $\forall 1\le j\le m$, at all the possible partition locations of $R_1$. Therefore, it takes $O(d^2+md+n'd)$ time to compute surface area for all the possible partition locations and class label assignments. Similarly, the costs of computing volume for all possible partition locations is $O(n'd)$. Thus for all the possible partition locations and class label assignments at $R_1$, the total cost of computing SVR is $O(d^2+md+n'd)$.
#### Time Complexity of Algorithm 1
Before working on any partitions, we first need to sort the whole dataset for all features, taking $O(dn\log n)$ time. For a node with $n'$ samples and $m$ leaf nodes in the current tree, there are $dn'$ possible partition locations. It takes $O(dn')$ time to compute signed impurity for all possible class label assignments and all partition locations; $O(d^2+md+n'd)$ time to compute surface-to-volume ratio for all possible class label assignments and all partition locations; $O(dn')$ time to find the best partition when all the signed impurities and SVR are already computed. The overall time complexity of finding the best partition at this node is $O(d^2+md+n'd)$. Let $a_n$ be the number of leaf nodes of the tree output by Algorithm 1, with $a_n\le \bar{a}_n = O(\sqrt{n})$. The time complexity is $$O(dn\log n) + \sum_{m=1}^{ a_n} O(d^2+md+n'd) \le O(dn\log n + \sqrt{n}d^2 + nhd),$$ where we use the fact that $\sum_{m=1}^{ a_n} n' \le nh$. This shows the efficiency of Algorithm 1.
Theoretical Properties
======================
In this section, we will discuss the consistency of our classification tree obtained from Algorithm 1 in two aspects: the first is estimation consistency – as sample size goes to infinity, a generalized distance between SVR-Tree and the oracle classifier converges to zero; the other is feature selection consistency – for redundant features that are conditionally independent of $Y$ given the other features, the probability of SVR-Tree excluding these features converges to one.
Estimation Consistency
----------------------
We define a generalized metric on the space of all nonrandom classifiers, introduce classifier risk, and define our notions of classifier consistency.
For any weight $\alpha$ and classifier $f:\Omega \to \{0,1\}$, we define the risk as $$R(f) = \mathbb{E} \left[\alpha \max(Y-f(X),0) + \max(f(X)-Y,0)\right],$$ where the expectation is taken over probability measure $\mathbb{P}$ for random variables $(X,Y)$.
Let $\mathscr{F}$ be the collection of all measurable functions from $\Omega$ to $\{0,1\}$. Then the oracle classifier and minimal risk are defined as $f^* = \arg\min_{f\in\mathscr{F}} R(f)$ and $R^* = R(f^*),$ respectively. Without loss of generality, we assume $\mathbb{P}(\mathbb{E}(Y|X)=1/(1+\alpha))=0$. Then the oracle classifier is unique almost surely. We now define the consistency of a sequence of classifiers based on L1 distance from the oracle classifier.
A sequence of classifiers $f_n$ is consistent if $ \lim_{n\to\infty} \mathbb{E} |f_n - f^*| = 0.$
Denote the tree obtained from Algorithm 1 as $\hat{T}_n$ and the classifier$:\Omega\to\{0,1\}$ associated with $\hat{T}_n$ as $\hat{f}_n$. Theorem \[consistency\] shows $\hat{f}_n$ is consistent under mild conditions.
\[consistency\] Let $\mathbb{P}$ be absolutely continuous with respect to Lebesgue measure on $\mathbb{R}^d$. Assume $\lim_{n\to\infty}\bar{a}_n = \infty,$ $\lim_{n\to\infty}\lambda_n = 0,$ and $\lim_{n\to\infty}\frac{\bar{a}_n d\log n}{n} = 0.$ Let $\alpha$ be a constant no smaller than one. Then the classifier $\hat{f}_n$ obtained from Algorithm 1 is consistent.
Feature Selection Consistency
-----------------------------
Let $X[-j]$ denote the set of all features except $X[j]$. We say $X[j]$ is redundant if conditionally on $X[-j]$, $Y$ is independent of $X[j]$. We denote $X'$ as the collection of all non-redundant features, and $X''$ as the collection of all redundant features. Under two conditions on the distribution of $Y, X'$ and $X''$, we can show if $\lambda_n$ goes to zero slower than $n^{-1/2}$, the probability of $\hat{T}_n$ excluding all redundant features goes to one. We assume there are $q \;(1\le q < d)$ non-redundant features denoted as $X[j_1], X[j_2], \cdots X[j_q]$. The redundant features are denoted as $X[j_{q+1]}, X[j_{q+2}], \cdots X[j_d]$.
Before stating these two conditions, we need to discuss how a partition can decrease the tree impurity. Suppose we are partitioning on node $A$ at feature $X[j]$, resulting in two new leaf nodes: $A_1 = \{X\in A: X[j] \le x_j\}$ and $A_2 = \{X\in A: X[j] > x_j\}$. Then the weighted conditional expectation of $Y$ in node $A_1, A_2$ is $$\label{p1alpha fs}
p_{h, \alpha} = \frac{\alpha \mathbb{E}(Y|X\in A_h)}{1-\mathbb{E}(Y|X\in A_h) + \alpha\mathbb{E}(Y|X\in A_h)},\quad h=1,2.$$ Thus the impurity of $A_1, A_2$ is $I_\alpha(A_1, \mathbb{P}) = 1 - p_{1, \alpha}^2 - (1-p_{1, \alpha})^2 = 2p_{1, \alpha}(1-p_{1,\alpha}),$ and $I_\alpha(A_2, \mathbb{P}) = 2p_{2,\alpha}(1-p_{2,\alpha}).$ Similarly, denoting the weighted conditional expectation of $Y$ in node $A$ as $p_\alpha$, the impurity of $A$ is $I_\alpha(A,\mathbb{P}) = 2p_\alpha(1-p_\alpha)$. The impurity decrease on node $A$ is $$\label{delta Ialpha}
\Delta I_\alpha(A, \mathbb{P}) = 2p_\alpha(1-p_\alpha) - 2p_{1,\alpha}(1-p_{1,\alpha})\mathbb{P}_\alpha(A_1|A) - 2p_{2,\alpha}(1-p_{2,\alpha})\mathbb{P}_\alpha(A_2|A),$$ Noting $A_1, A_2$ is determined by $x_j$, $\Delta I_\alpha(A, \mathbb{P})$ can be considered as a function of $x_j$. Denoting $\Delta I_\alpha(A,\mathbb{P}) = h_{A,j}(x_j)$, then the maximal impurity decrease at feature $X[j]$ is $$\label{Mj}
M_{A,j} = \max_{x_j\in(x_{j,1},x_{j,2})} h_{A,j}(x_j).$$ The quality of $M_{A,j}$ in reducing the impurity of node $A$ is measured relative to the oracle impurity decrease at node $A$. Suppose we partition node $A$ into measurable sets $A_1$ and $A_2$ satisfying $\mathbb{P}_\alpha(A_1| A) = V_1,$ $A_1\cup A_2=\Omega,$ $A_1\cap A_2 = \emptyset,$ for all $X_1\in A_1,$ $X_2\in A_2,$ and $\mathbb{E}(Y|X_1)\le\mathbb{E}(Y|X_2),$ where $V_1\in[0,1]$ and $A_1, A_2$ are not required to be hyper rectangles. By definition, the set $A_1$ corresponds to the $V_1$ proportion of $A$ having the smallest $\mathbb{E}(Y|X)$ values, while the set $A_2$ corresponds to the $1-V_1$ proportion of $A$ having the largest $\mathbb{E}(Y|X)$ values. Similar to equation (\[p1alpha fs\]) and (\[delta Ialpha\]), we can define $p_{1,\alpha}, p_{2,\alpha}$ and the impurity decrease $\Delta I_\alpha(A,\mathbb{P})$. Given $V_1\in [0,1]$, the impurity decrease $\Delta I_\alpha(A,\mathbb{P})$ is unique, which is also the maximal impurity decrease for all measurable $A_1, A_2$ satisfying $\mathbb{P}_\alpha(A_1| A) = V_1,\; \mathbb{P}_\alpha(A_2|A) = 1-V_1$. Thus we can denote the impurity decrease as $\Delta I_\alpha(A,\mathbb{P}) = h^*_A(V_1)$. The oracle impurity decrease is the maximal value of $h^*_A(V_1)$: $M_{A}^* = \max_{V_1\in[0,1]} h^*_A(V_1).$ In general, the larger impurity decrease will tend to correspond to non-redundant features; this will particularly be the case under Conditions 1-2.
\[condition 1\] There exists $c_1\in [0,1]$, such that for all $A\subset\Omega$, $\sup_{1\le j\le q} M_{A,j} \ge c_1 M_A^*.$
Condition \[condition 1\] relates the impurity decrease in non-redundant features to the oracle impurity decrease. The strength of the condition is dependent on the value of $c_1$. If $c_1=0$, the condition does not impose any restrictions; if $c_1 =1$, partitions in non-redundant features can fully explain the oracle impurity decrease. Here we give some examples of models with different $c_1$ values.
Let $\Omega = [0,1]^d$ and the marginal distribution of $\mathbb{P}_\alpha$ be the uniform distribution on $\Omega$. Suppose $\mathbb{E}(Y|X)= \phi(a^T X + b),$ where $\phi(\cdot)$ is a monotonically increasing function, $a\in\mathbb{R}^d$ and $b\in\mathbb{R}$. Let $A_{y1}, A_{y2}$ be measurable sets having $A_{y1}\cap A_{y2}=\emptyset$ and $A_{y1}\cup A_{y2} = \Omega$, which achieve the oracle impurity decrease $M_A^*$, and let $\min\{\mathbb{P}_\alpha(A_{y1}), \mathbb{P}_\alpha(A_{y2})\}=h$. Then the constant $c_1$ in Condition 1 satisfies $c_1 \ge (\sqrt{2}-1)h + o(h)$ if $h \le 1/d!$ and $c_1 \ge F_{d-1}(F_d^{-1}(h)-h) - F_{d-1}(F_d^{-1}(h)-2h)$ if $h>1/d!$, where $F_d(z) = \frac{1}{d!} \sum_{k=0}^{\lfloor z\rfloor}(-1)^k {d\choose k} (z-k)^{d-1}$ is the cumulative distribution function for the Irwin-Hall distribution with parameter $d$.
Let $\Omega = [0,1]^2$ and $\mathbb{P}_X$ be the uniform distribution on $\Omega$. Let $A_1 = [0,1/2]\times[0,1/2],$ $A_2 = [1/2,1]\times[0,1/2],$ $A_3 = [0,1/2]\times[1/2,1],$ $A_4 = [1/2,1]\times[1/2,1],$ and $\mathbb{E}(Y|X) = 1$ if $X \in A_1\cup A_4$ and $0$ if $X\in A_2\cup A_3.$ Further assume $\alpha=1$. Then $c_1=0$.
\[condition 2\] The weighted probability measure $\mathbb{P}_\alpha$ has density $\rho_\alpha(X', X'')$ in $\Omega$. Moreover, $\rho_\alpha(X', X'') = \rho_{1,\alpha}(X') \rho_{2,\alpha}(X'') + \rho_{3,\alpha}(X', X''),$ where for all $(X', X'')\in \Omega$, $\rho_{3,\alpha}(X', X'') \le c_2 \rho_\alpha(X', X''),$ with $c_2\in [0,1]$ a constant.
Condition \[condition 2\] asserts that the joint density of $X', X''$ can be decomposed into an independent component plus a dependent component, where the dependent component is dominated by the independent component up to a constant. This condition controls the dependence between $X'$ and $X''$. Since given $X'$, $Y$ is independent of $X''$, this condition will also control the dependence between $Y$ and $X''$. The strength of the condition depends on the constant $c_2$. When $c_2=1$, the condition imposes no restrictions; when $c_2=0$, the condition asserts $X'$ and $X''$ are completely independent.
Using Conditions \[condition 1\]-\[condition 2\], we establish feature selection consistency in Theorem \[feature selection\].
\[feature selection\] If the optional steps in Algorithm 1 are enabled, condition \[condition 1\], \[condition 2\] are satisfied with $c_1>(1-c_2)/[c_2(2-c_2)]$ and $\lambda_n\ge c_3 n^{-(1/2-\beta)}$ for some constant $c_3>0$, $\beta\in (0, 1/2)$, we have $$\lim_{n\to\infty}\mathbb{P}(\hat{T}_n \;\text{does not have partitions in}\; X[j_{q+1}], X[j_{q+2}], \cdots X[j_d]) = 1.$$
In Theorem \[feature selection\] conditions 1 and 2 are complementary. If $c_2$ is smaller (i.e., condition 2 is stronger), $c_1$ can be smaller (i.e., condition 1 is weaker). The opposite also holds. The following two Corollaries cover special cases.
\[feature selection monotonic\] If the optional steps in Algorithm 1 are enabled, condition \[condition 1\] is satisfied with $c_1=1$ (that is, in each hyper-rectangle, the maximal impurity decrease at non-redundant features is equal to the oracle impurity decrease) and $\lambda_n\ge c_3 n^{-(1/2-\beta)}$ for some constant $c_3>0$, $\beta\in (0, 1/2)$, we have $$\lim_{n\to\infty}\mathbb{P}(\hat{T}_n \;\text{does not have partitions in}\; X[j_{q+1]}, X[j_{q+2}], \cdots X[j_d]) = 1.$$
\[feature selection independent\] If the optional steps in Algorithm 1 are enabled, $X'$ and $X''$ are independent and $\lambda_n\ge c_3 n^{-(1/2-\beta)}$ for some constant $c_3>0$, $\beta\in (0, 1/2)$, we have $$\lim_{n\to\infty}\mathbb{P}(\hat{T}_n \;\text{does not have partitions in}\; X[j_{q+1]}, X[j_{q+2}], \cdots X[j_d]) = 1.$$
Corollary \[feature selection independent\] is a direct result of theorem \[feature selection\] with $c_2=0$ and an arbitrary value of $c_1$.
Numerical Studies
=================
We compare SVR Tree with popular imbalanced classification methods on real datasets, adding redundant features to these datasets in evaluating feature selection. A confusion matrix (Table \[confusion matrix\]) is often used to assess classification performance. A common criteria for accuracy is $(TP + TN)/(TP + TN + FP + FN).$ When $1$s are relatively rare, it is often important to give higher priority to true positives, which is accomplished using the true positive rate (recall) $TPR = \frac{TP}{TP+FN}$ and precision $\frac{TP}{TP+FP}.$ To combine these, the F-measure is often used: $(1+\beta^2)TPR\times Precision/(\beta^2 TPR + Precision)$.
[ccc|c|]{} & &\
& & 1 & 0\
& & True Positive (TP) & False Positive (FP)\
& & False Negative (FN) & True Negative (TN)\
Datasets
--------
We test our method on 5 datasets from the UCI Machine Learning Repository [@Dua:2019], varying in size, number of features and level of imbalance.
#### Glass dataset
<https://archive.ics.uci.edu/ml/datasets/Glass+Identification> consists of 213 samples and 9 features. The objective is to classify samples into one of seven types of glass. We choose class 7 (headlamps) as the minority class and class 1-6 as the majority class, yielding 29 minority class samples.
#### Vehicle dataset
[@siebert1987vehicle] <https://archive.ics.uci.edu/ml/datasets/Statlog+(Vehicle+Silhouettes)> consists of 846 samples and 18 features. The aim is to classify a silhouette into one of four types of vehicles. As in [@he2008adasyn], we choose class “Van” as the minority class and the other three types of vehicles as the majority class, resulting in 199 minority class samples.
#### Abalone dataset
<https://archive.ics.uci.edu/ml/datasets/Abalone> aims to predict the age of abalone by physical measurements. As in [@he2008adasyn], we let class “18” be the minority class and class “9” be the majority class. This yields 731 samples in total, among which 42 samples belong to the minority class. We also remove the discrete feature “sex”, which gives us 9 features.
#### Satimage dataset
<https://archive.ics.uci.edu/ml/datasets/Statlog+(Landsat+Satellite)> consists of a training set and a test set. We have 6435 samples and 36 features. As in [@chawla2002smote], we choose class ‘4’’ as the minority class and collapsed all other classes into a single majority class, resulting in 626 minority class samples.
#### Wine dataset
[@cortez2009modeling] <https://archive.ics.uci.edu/ml/datasets/Wine+Quality> collects information on wine quality. We focus on the red wine subset, which has 1599 samples and 11 features. We let the minority class be samples having quality $\ge 7$, while the majority class has quality $\le 6$. This generates 217 minority class samples.
Experimental Setup {#setup}
------------------
We test the performance of SVR-Tree, SVR-Tree with feature selection, CART [@breiman1984classification] with duplicated oversampling, CART with SMOTE [@chawla2002smote], CART with Borderline-SMOTE [@han2005borderline] and CART with ADASYN [@he2008adasyn] on all five datasets. Features are linearly transformed so that samples lie in $[0,1]^d$.
For each method and dataset, we run the algorithm 50 times. For each run, we randomly choose 2/3 samples as training and 1/3 as testing. The average values of TP, TN, FP, FN on testing sets are used to compute the accuracy, TPR, precision and F-measure. The specific settings for each method are discussed below.
SVR-Tree with and without feature selection are described in Algorithm 1. The weight for the minority class, $\alpha$, is set to be the largest integer that makes the total weight of the minority class no greater than the total weight of the majority class; If $\alpha$ is greater than $15$, it is truncated to be $15$. The maximal number of leaf nodes $\bar{a}_n$ is $2\sqrt{n}$. The penalty parameter $\lambda_n$ for SVR is chosen from a geometric sequence in $[2^0, 2^{10}]\times10^{-3}\times n^{-1/3}$; the parameter with highest F-measure on 50 runs is selected. For SVR-Tree with feature selection, the constant $c_0$ in equation (\[impu decr\]) is fixed to $c_0=4$; In practice, the results are insensitive to $c_0$.
For the other methods, we first over sample the minority class samples, such that the number of minority samples are multiplied by $\alpha$ after oversampling. We than build the CART tree on the over sampled dataset and prune it. The pruning parameter of CART is selected to maximize the F-measure. By the algorithm proposed by [@breiman1984classification], the range from which to choose the pruning parameter will be available after we build the tree and does not need to be specified ahead of time.
For duplicated oversampling, we sample each minority class sample $\alpha-1$ times; For SMOTE, the number of nearest samples is set as $k=5$; For Borderline-SMOTE, we adopt the Borderline-SMOTE1 of [@han2005borderline], with the number of nearest samples $k=5$. For both SMOTE and Borderline-SMOTE, if $\alpha-1\ge k$, some nearest neighbors may be used multiple times to generate synthetic samples. For ADASYN, denote the number of majority class samples as $n_0$ and the number of minority class samples as $n_1$, then the parameter $\beta$ is set to be $\beta = \alpha n_1 / n_0$.
Results
-------
The average values of accuracy, precision, TPR, F-measure and number of selected features across the 50 runs are shown in Table \[results\]. In the column “Method”, SVR = SVR-Tree, SVR-Select = SVR-Tree with feature selection, Duplicate = CART with duplicated oversampling, SMOTE = CART with SMOTE, and BSMOTE = CART with Borderline-SMOTE and ADASYN = CART with ADASYN. For each dataset and evaluation measure, the method that ranks first is highlighted in bold. The number of wins for each performance measure are also calculated. The SVR methods perform the best overall, with SVR-Select having particularly good performance. Furthermore, SVR-Select chooses the fewest number of features for each dataset, so has a good balance of accuracy and parsimony.
Data set Method Accuracy Precision TPR F-measure $n(\mathrm{features})$
---------- ------------ ------------ ------------ ------------ ------------ ------------------------
SVR 0.9583 0.8161 0.8956 0.8540 4.98
SVR-Select **0.9683** **0.8668** 0.9067 **0.8863** **1.0**
Duplicate 0.9646 0.8489 0.9000 0.8737 **1.0**
SMOTE 0.9624 0.8295 0.9111 0.8684 1.02
BSMOTE 0.9602 0.8160 **0.9133** 0.8619 **1.0**
ADASYN 0.9498 0.7713 0.8978 0.8297 5.72
SVR **0.9368** **0.8659** 0.8652 0.8655 14.76
SVR-Select 0.9355 0.8544 0.8748 0.8645 **5.7**
Duplicate 0.9362 0.8377 **0.9039** **0.8696** 11.64
SMOTE 0.9317 0.8385 0.8791 0.8583 10.98
BSMOTE 0.9309 0.8417 0.8697 0.8554 13.96
ADASYN 0.9341 0.8423 0.8855 0.8634 11.58
SVR 0.9212 0.3251 0.3457 0.3351 6.88
SVR-Select **0.9244** **0.3431** 0.3457 **0.3444** **5.6**
Duplicate 0.9184 0.2956 0.3043 0.2999 6.94
SMOTE 0.8974 0.2578 **0.4186** 0.3191 6.94
BSMOTE 0.8960 0.2479 0.3986 0.3057 6.96
ADASYN 0.8937 0.2480 **0.4186** 0.3114 6.92
SVR **0.9036** **0.5032** 0.6969 0.5844 34.5
SVR-Select 0.9029 0.5008 0.7020 **0.5845** **28.12**
Duplicate 0.9032 0.5017 0.6553 0.5683 34.58
SMOTE 0.8895 0.4578 **0.7364** 0.5646 29.2
BSMOTE 0.8945 0.4720 0.71 0.5671 31.38
ADASYN 0.8946 0.4711 0.6831 0.5576 34.72
SVR 0.8476 0.4564 0.6433 0.5340 10.96
SVR-Select 0.8481 0.4578 **0.6475** **0.5363** **10.54**
Duplicate **0.8553** **0.4744** 0.6103 0.5338 11.0
SMOTE 0.8513 0.4647 0.6311 0.5353 11.0
BSMOTE 0.8503 0.4608 0.6047 0.5231 11.0
ADASYN 0.8477 0.4554 0.6228 0.5261 11.0
SVR **2** **2** 0 0
SVR-Select **2** **2** 1 **4**
Duplicate 1 1 1 1
SMOTE 0 0 **1.5** 0
BSMOTE 0 0 1 0
ADASYN 0 0 0.5 0
: Results of Numerical Studies on five real world datasets.[]{data-label="results"}
Additional Experiments for Feature Selection
--------------------------------------------
For the Wine and Abalone datasets, we generate 10 additional uninformative features independently from $U([0,1])$. We reran the analyses as above and results are shown in Table \[results feature selection\], where $n(\mathrm{features}_o)$ denotes the average number of original features selected by the method and $n(\mathrm{features}_r)$ denotes the average number of artificially generated features selected by the method.
SVR-Select performs well when there are a considerable number of redundant features. SVR-Select selects significantly more original features than artificially generated features, suggesting effectiveness in feature selection. For all other methods, the relative difference between number of original features and number of artificially generated features is much smaller. In addition, nearly all methods select fewer of the original features when compared with results in Table \[results\].
Data set Method Accuracy Precision TPR F-measure $n(\mathrm{features}_o)$ $n(\mathrm{features}_r)$
---------- ------------ ------------ ------------ ------------ ------------ -------------------------- --------------------------
SVR **0.8277** **0.4242** 0.5486 0.4785 10.72 8.58
SVR-Select 0.8115 0.3910 **0.6975** **0.5011** **3.96** **0.26**
Duplicate 0.8186 0.3963 0.6439 0.4906 9.58 7.32
SMOTE 0.8085 0.3838 0.6789 0.4904 8.8 5.74
BSMOTE 0.8136 0.3883 0.6492 0.4859 7.6 4.24
ADASYN 0.8073 0.3816 0.6767 0.4880 8.6 6.44
SVR 0.8779 0.2173 0.4329 0.2894 5.34 8.0
SVR-Select **0.9106** **0.2788** 0.3500 **0.3104** 2.42 0.7
Duplicate 0.7730 0.1472 **0.6157** 0.2376 **1.0** **0.0**
SMOTE 0.8471 0.1706 0.43 0.2443 3.32 4.14
BSMOTE 0.8761 0.2028 0.3943 0.2678 3.7 3.7
ADASYN 0.8383 0.1640 0.4429 0.2393 3.48 4.12
: Additional Numerical Study for Feature Selection.[]{data-label="results feature selection"}
Discussion
==========
A major challenge in analyzing imbalanced data is small sample size in the minority class leading to overfitting. It is natural to consider using regularization to address this problem. Regularization of classification trees is an old idea; for example, [@breiman1984classification] proposed to penalize the number of leaf notes in the tree. Other classification trees like C4.5 [@quinlan2014c4] and Ripper [@cohen1995fast] also prune the overgrown tree. However, the number of leaf nodes may not be a good measure of complexity of a classification tree. Recently, [@hahn2020bayesian] adds a Bayesian prior to an ensemble of trees, which functions as indirect regularization. Following the idea of directly regularizing the shape of the decision set and complexity of the decision boundary, we instead penalize the surface-to-volume ratio. To our knowledge, this is a new idea in the statistical literature on classification.
SVR-Tree can be trivially generalized to the multi-class case and balanced datasets. For multiple classes, we can apply SVR to one or more minority classes and take the sum of these ratios as regularization. For balanced datasets, we can either compute the SVR ratio of all classes, or we can simply regularize the surface of the decision boundary. The principle behind these generalizations is to regularize the complexity of the decision boundaries and shapes of the decision sets.
[**SUPPLEMENTARY MATERIAL**]{}
[**Proofs and a Detailed Algorithm:**]{} Supplemental Material for “Classification Trees for Imbalanced and Sparse Data: Surface-to-Volume Regularization”.\
[**Codes:**]{} <https://github.com/YichenZhuDuke/Classification-Tree-with->\
[Surface-to-Volume-ratio-Regularization.git](Surface-to-Volume-ratio-Regularization.git).
Proofs
======
We prove Theorem 1 and 2 here. Proofs for Corollary 1, $c_1$ bounds in examples 1-2 and lemmas in Appendix are in the supplemental material. Without loss of generality, we assume $\Omega = [0,1]^d$ in this section.
Proof of Theorem \[consistency\]
--------------------------------
The proof builds on [@nobel1996histogram], [@gyorfi2006distribution], [@tsybakov2004optimal] and [@scornet2015consistency]. We first establish a sufficient condition for consistency, showing a classification tree whose signed impurity converges to an oracle bound is consistent. We then break the difference between signed impurity of $\hat T_n$ and the oracle bound into two parts. The first is estimation error, which goes to zero if the number of leaves $a_n$ increases slower than $n$; the second is approximation error, which goes to zero if $\mathbb{E}(Y|X\in A)$ goes to a constant within each leaf node and penalty $\lambda_n$ goes to zero.
Denote $p(x) = \mathbb{E}(Y=1|X=x)$, and its weighted version $p_\alpha(x) = [\alpha p(x)][1-p(x)+\alpha p(x)]$. Define the oracle lower bound for signed impurity as $I_\alpha^* = \int_\Omega 2p_\alpha(x)(1-p_\alpha(x)) d\mathbb{P}_\alpha(x).$ The following lemma shows if the signed impurity of a classification tree converges to $T^*$ as $n\to\infty$, the classifier associated with the tree is consistent.
\[lem: signed impu consistency\] Let $T_n$ be a sequence of classification trees, let $f_n: \Omega\to \{0,1\}$ be the classifier associated with $T_n$. $T_n$ is consistent if $\tilde{I}_\alpha(T_n, \mathbb{P}) \to I^*_\alpha$ in probability.
We then decompose the difference between signed impurity of $\hat{T}_n$ and the oracle bound into estimation error and approximation error.
\[lem:decompose\] Let $\hat{T}_n$ be a classification tree trained from data $\mathscr{D}_n$, $A_1, A_2, \ldots A_m$ be all the leaf nodes of $\hat{T}_n$. Define the set of classifiers $\mathscr{T}_n$ as: $$\mathscr{T}_n = \{T: T\text{'s associated classifier }f:\Omega\to\{0,1\}\text{ is constant on all }A_j, \;\; 1\le j \le m\}.$$ We have $$\label{decompose}
\tilde{I}_\alpha(\hat{T}_n,\mathbb{P}) - I^*_\alpha \le 2 \sup_{T\in\mathscr{T}_n}|\tilde{I}_\alpha(T, \mathbb{P}) - \tilde{I}_\alpha(T,\mathbb{P}_n)| + \inf_{T\in\mathscr{T}_n}|\tilde{I}_\alpha(T, \mathbb{P})+\lambda_n r(T) - I^*_\alpha|.$$
The first term on the right hand side of equation (\[decompose\]) is the “estimation error”, which measures the difference between functions evaluated under the empirical and true distributions. The second term is “approximation error”, which measures the ability of $\mathscr{T}_n$ to approximate the oracle prediction function. The next two lemmas show both terms go to zero in probability.
\[estimation error\] If $\frac{\bar{a}_n d\log n}{n} = o(1),$ we have $\sup_{T\in\mathscr{T}_n}|\tilde{I}_\alpha(T, \mathbb{P}) - \tilde{I}_\alpha(T,\mathbb{P}_n)| \to 0$ in probability.
\[approximation error\] As $n\to\infty$, if $\lambda_n\to 0$ and $\bar{a}_n\to\infty$, $\inf_{T\in\mathscr{T}_n} |\tilde{I}_\alpha(T,\mathbb{P}) + \lambda_n r(T) - I_\alpha^*| \to 0$ in probability.
Combining lemma \[lem: signed impu consistency\], \[lem:decompose\], \[estimation error\], \[approximation error\], we finish the proof.
Proof of Theorem \[feature selection\]
--------------------------------------
The proof of theorem \[feature selection\] mainly consists of two parts. The first works on the true distribution $\mathbb{P}$, proving that under $\mathbb{P}$, the partition with highest impurity decrease is always in non-redundant features; The second works on the randomness brought by $\mathbb{P}_n$, proving that with high probability, the randomness of the tree impurity decrease can be controlled by the threshold $c_0\lambda_n$. Combining these two parts with theorem \[consistency\], we can show the probability of rejecting all redundant features in the tree building procedure goes to zero. We now focus on the first part.
\[theoretical exclude\] Under the conditions of theorem \[feature selection\], for all hyperrectangles $A\subset \Omega$, we have $\sup_{1\le j\le q} M_{A,j} \ge \sup_{q+1\le l\le d} M_{A,l},$ where $M_{A,j}$ is the maximal impurity decrease of feature $j$ at node $A$ defined in equation (\[Mj\]).
We now work on the second part, proving the randomness in tree impurity decrease can be bounded by $c_0\lambda_n$ with high probability.
\[noise thre\] Suppose $\lambda_n \ge n^{-1/2+\beta}$ and the density of $X$ exists, then for all hyperrectangles $A\subset \Omega$, let $A_1, A_2$ be the two child nodes of $A$ obtained by an arbitrary partition, and let $n'$ be the number of samples in $A$. To highlight the dependence of impurity decrease and the child node $A_1, A_2$, denote the impurity decrease on node $A$ after partitioning $A$ into $A_1, A_2$, under measure $\mathbb{P}$ and $\mathbb{P}_n$, as $\Delta I_\alpha(A,A_1,\mathbb{P})$ and $\Delta I_\alpha(A,A_1,\mathbb{P}_n)$, respectively. Then we have $$\lim_{n\to\infty}\mathbb{P}\left(\sup_{A_1} |\Delta I_\alpha(A, A_1,\mathbb{P}) - \Delta I_\alpha(A, A_1,\mathbb{P}_n)|>\frac{c_0\lambda_n n}{2n'}\right) = 0.$$
We are now prepared to prove Theorem \[feature selection\].
Let $T_k^*$ be the theoretical tree with $k$ leaf nodes defined in Algorithm 2 of supplemental material, and let $\hat{T}_{n,k}$ be the tree consisting of the first $k$ partitions of tree $\hat{T}_n$. The proof of theorem \[feature selection\] consists of two steps. In the first step, we show there exists $k_0\in\mathbb{N}$, such that as $n$ goes to infinity, the probability of $\hat{T}_{n,k_0}$ including all non-redundant features and excluding all redundant features goes to one. In the second step, we show if all the non-redundant features are already included in the tree, the probability of including any new redundant features in all the following partition procedures goes to zero.
We begin with the first step. By lemma \[theoretical exclude\], the theoretical tree $T_k^*$ will not include any redundant features. We now show all non-redundant features are included in $T_k^*$ provided $k$ is large enough. Let $f_k^*$ be the classifier associated with tree $T_k^*$. Let $\psi_k(x)$ be the leaf node of $T_k^*$ that contains $x$, and let $e_k(x)$ be the average of $\mathbb{E}(Y|X)$ in each leaf node $e_k(x) \triangleq \mathbb{E}(Y|X\in\psi_k(x)).$ By Lemma 1 in [@scornet2015consistency], we have $e_k(x)$ converges to $\mathbb{E}(Y|X)$ almost surely. If there exists a non-redundant feature $X[j]$, such that $X[j]$ is excluded in partitions of $T_k^*,\; \forall k\in\mathbb{N}$, we have $\mathbb{E}(e_k(X)|X[-j]) = e_k(X),\;\forall k\in \mathbb{N}$, hence $\mathbb{E}(Y|X[-j]) = \mathbb{E}(Y|X)$. This contradicts the fact that $X[j]$ is a non-redundant feature. Therefore, for all non-redundant features $X[j]$, there exists some $k_j\in\mathbb{N}$, such that feature $X[j]$ is included in partitions of $T_k^*$ for $k\ge k_j$. Since the number of features is finite, there exists $k_0\in\mathbb{N}$, such that $T_{k_0}^*$ includes all non-redundant features. By Technical Lemma 4 in the supplemental material, as $n$ goes to infinity, with probability tending to one, $\hat T_{n,k_0}$ includes all non-redundant features while excluding all redundant features.
It remains to show no redundant features will be included after the $k_0$th partition. For the $k'$th ($k_0<k'\le \bar{a}_n$) partition, which partitions node $A$ into $A_1, A_2$, denote $\Delta I_{\alpha,n}(T,\mathbb{P}_n)$ and $\Delta I_{\alpha,r}(T,\mathbb{P}_n)$ as the maximal tree impurity decrease measured by $\mathbb{P}_n$ when the partition is in a non-redundant feature and redundant feature, respectively. Then by Algorithm 1, the partition at a redundant feature will be rejected if $\Delta I_{\alpha,r}(T,\mathbb{P}_n) \le \Delta I_{\alpha,n}(T,\mathbb{P}_n) + c_0\lambda_n.$ Similarly, denote $\Delta I_{\alpha,n}(A,\mathbb{P}_n)$ and $\Delta I_{\alpha,r}(A,\mathbb{P}_n)$ as the maximal impurity decrease of node $A$ measured by $\mathbb{P}_n$ when the partition is in a non-redundant feature and redundant feature, respectively. Further denote $\Delta I_{\alpha,n}(A,\mathbb{P})$ and $\Delta I_{\alpha,r}(A,\mathbb{P})$ as the maximal impurity decrease of node $A$ when the probability measure is the true probability $\mathbb{P}_n$. By lemma \[theoretical exclude\], for the impurity decrease on node $A$ measured under true probability $\mathbb{P}$, we have $\Delta I_{\alpha,n}(A,\mathbb{P}) \le \Delta I_{\alpha,r}(A,\mathbb{P}).$ By lemma \[noise thre\], we have with probability greater than $1-\sigma(n)$ $$|\Delta I_{\alpha,n}(A,\mathbb{P}_n) - \Delta I_{\alpha,n}(A,\mathbb{P})|\le \frac{c_0\lambda_n n}{2n'\alpha},$$ $$|\Delta I_{\alpha,r}(A,\mathbb{P}_n) - \Delta I_{\alpha,r}(A,\mathbb{P})|\le \frac{c_0\lambda_n n}{2n'\alpha}.$$ Therefore, we have with probability greater than $1-\sigma(n)$ $$\begin{aligned}
|\Delta I_{\alpha,r}(T,\mathbb{P}_n)-\Delta I_{\alpha,n}(T,\mathbb{P}_n)|
\le & \mathbb{P}_{n,\alpha}(A) |\Delta I_{\alpha,r}(A,\mathbb{P}_n) -\Delta I_{\alpha,n}(A,\mathbb{P}_n)| \\
\le & \alpha \frac{n'}{n} \Big(0 + \frac{c_0\lambda_n n}{2n'\alpha} + \frac{c_0\lambda_n n}{2n'\alpha}\Big) \\
\le & c_0\lambda_n.\end{aligned}$$ Therefore, for all $k'$ ($k_0<k'\le \bar{a}_n$), the probability of partitioning at a redundant feature is smaller than $\sigma(n)$. Since $\bar{a}_n\le n$ and $\lim_{n\to \infty} n\sigma(n) = 0$, with probability tending to one, $\hat{T}_n$ will not partition at redundant features.
[^1]: Yichen Zhu is Ph.D. student, Department of Statistical Science, Duke University, Durham, NC 27708.
[^2]: David B. Dunson is art and science professor, Department of Statistical Science, Duke University, Durham, NC 27708. The research was partially supported by grant N000141712844 of the United States Office of Naval Research.
[^3]: A queue is a dynamic set in which the elements are kept in order and the principal operations on the set are the insertion of elements to the **tail**, known as enqueue, and deletion of elements from the **head**, known as dequeue. See chapter 10.1 of [@cormen2009introduction] for details.
[^4]: In our experiments, we set $\bar{a}_n = 2\sqrt{n}$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Platooning is a way to significantly reduce fuel consumption of trucks. Vehicles that drive at close inter-vehicle distance assisted by automatic controllers experience substantially lower air-drag. In this paper, we deal with the problem of coordinating the formation and the breakup of platoons in a fuel-optimal way. We formulate an optimization problem which accounts for routing, speed-dependent fuel consumption, and platooning decisions. An algorithm to obtain an approximate solution to the problem is presented. It first determines the shortest path for each truck. Then, possible platoon configurations are identified. For a certain platoon configuration the optimal speed profile is the solution to a convex program. The algorithm is illustrated by a realistic example.'
author:
- 'Sebastian van de Hoef, Karl H. Johansson and Dimos V. Dimarogonas [^1]'
bibliography:
- 'citations.bib'
title: |
Fuel-Optimal Centralized Coordination of Truck Platooning\
Based on Shortest Paths
---
Introduction
============
Truck-platooning means that several trucks drive with small inter-vehicle distance. The distance is maintained by appropriate, automatic control of the vehicle speeds. Earlier research on truck-platooning has mostly been motivated by the possible increase in traffic throughput [@frameworks_overview_article; @path_overview_conference; @dolphin_framework_conference; @Halle_car_platoons_multiagent_auto21; @Hall_platoon_sorting]. The potential for fuel saving has renewed the interest in the topic[@assad_thesis]. When the spacing between the vehicles is sufficiently small, the follower vehicles experience reduced aerodynamic drag [@air_drag_reduction_platooning; @Bonnet2000]. This, in turn, leads to reduced fuel consumption. Advances in related technologies have made the commercial availability of trucks that can platoon likely in near future.
Much of the research on truck-platooning focuses on the control of the inter-vehicle distances [@string_stability; @barooah_platooning]. Unless trucks have the same route and start at the same time, platoons have to be created during their journeys. If we use platooning to reduce fuel consumption and only have a small part of vehicles equipped with a platooning system, the formation of the platoons needs to be coordinated over large regions, which is the subject of this work. Related problems arise for example in air traffic management [@air_traffic_management], convoy movement [@convoy_movement_opt], and other areas [@pinedo2012scheduling; @uav_beard; @Bullofrazzoli_vehicle_routing]. Related research on coordination of truck platooning is considered in [@jeff_distributed_coordination] where a heuristic scheme based on distributed controllers at road intersections is proposed. The complexity of an off-line optimization algorithm is discussed in [@jeff_complexity]. Schemes where platoons are formed on the on-ramps as in [@Hall_platoon_sorting] require huge investments in infrastructure. The authors of [@datamining_platooning] consider that trucks wait to form platoons and use data-mining techniques. [@its_energy_project] computes the global fuel saving opportunities based on a simulation study but does not explicitly detail the means of coordination.
The goal of this paper is to introduce a framework for a fuel-optimal coordination where the trucks adapt their speeds in order to form platoons during their journeys while driving. The influence of speed on the fuel consumption is explicitly considered and trucks are guaranteed to arrive at their destinations by their arrival deadlines.
We consider a number of transport assignments, each with a start position and time and an arrival position and deadline. In Section \[sec:problem\_statement\] we introduce notation and formulate an optimization problem. The route of a truck is modeled as a path in a graph. The average speed of the trucks is constrained to take into account legal speed limits, performance of the truck, and traffic. Decision variables are the paths taken from start to destination and the speed on each edge. We optimize the total fuel consumption for all transport assignments while taking into account both reduced fuel consumption due to platooning as well as speed-dependent fuel consumption. Platooning is considered to take place when two or more trucks traverse an edge at the same time with the same speed. We illustrate why this optimization problem is difficult to solve. In Section \[sec:approx\_sol\], we derive an approximate solution to the problem. The approximate solution is based on the shortest path for each transport assignment, which can be efficiently computed with available algorithms. The remaining selection of edge traversal speeds is solved by identifying platoon configurations taking into account the route and the timing constraints. For a certain platoon configuration, the optimal speeds are computed as the solution to a convex, continuous optimization problem, which can be efficiently solved. We describe how the existence of a solution to the convex problem can be checked for a platoon configuration. By searching through the possible platoon configurations, we obtain optimal speed profiles for the trucks. We show that each two trucks form a platoon at most once during a journey, which can significantly reduce the discrete search space. We demonstrate the algorithm on a realistic scenario.
Problem Statement {#sec:problem_statement}
=================
We consider a road network modeled as a weighted, directed graph ${\mathcal{G}} = ({\mathcal{N}},{\mathcal{E}},W)$, where ${\mathcal{N}}$ is a set of nodes, ${\mathcal{E}} \in {\mathcal{N}} \times {\mathcal{N}}$ is a set of edges, and $W: {\mathcal{E}} \rightarrow \mathbb{R}^+$ are positive edge weights. Each edge corresponds to a road segment and the nodes correspond to the intersections. The edge-weights model the length of the road segments.
We have $K$ transport assignments, each characterized by a start node $n_k^S \in {\mathcal{N}}$ and a destination node $n_k^D \in {\mathcal{N}}$. There is a start-time $t_k^S$ at which the transport starts from $n_k^S$ and a deadline $t_k^D$ by which the transport needs to reach $n_k^D$. We consider that each transport assignment is assigned to one truck. The truck carrying out transport assignment $k$ has index $k$, i.e., the same index as the transport assignment.
The speed of a truck is positive and smaller than an allowed maximum speed $v_{\max}$. We assume that the speed over each edge is constant. So, if the speed of a truck over edge $(n_1,n_2)$ is $v$ and the truck starts traversing the edge at time $t$, it arrives at $t + W((n_1,n_2))/v$ at node $n_2$.
A path of ${\mathcal{G}}$ is a sequence of nodes $(n_k[1],n_k[2],\dots)$, so that $(n_k[i],n_k[i+1]) \in {\mathcal{E}}$ for $i = 1,\dots,(|n_k|-1)$, where $|n_k|$ denotes the length of the sequence, i.e., the number of nodes in the path. A path connecting $n_k^S$ and $n_k^D$ has $n_k[1] = n_k^S$ and $n_k[|n_k|] = n_k^D$.
We say that a set of edges $\tilde{{\mathcal{E}}}$ forms a path if and only if there exists a path $\tilde{n}$ so that $\tilde{{\mathcal{E}}} = \{(\tilde{n}[i],\tilde{n}[i+1]): i = 1,\dots,(|\tilde{n}|-1)\}$. Let $\tilde{n}$ be a path. We say that $\tilde{n}_s$ is a sub-path of $\tilde{n}$ if and only if $\tilde{n}_s$ is a path and there exists $j$ such that for every $i = 1,\dots,(|\tilde{n}_s|-1)$ it holds that $(\tilde{n}_s[i],\tilde{n}_s[i+1]) = (\tilde{n}[j + i],\tilde{n}[j+i+1])$. We denote the sequence of edges on the path of transport assignment $k$ as $e_k[i] = (n_k[i],n_k[i+1])$ for $i = 1,\dots,(|n_k|-1)$.
For a path $n_k$, we can collect the edge traversal speeds in a sequence $v_k$ of length $|e_k|$. The $i$th element $v_k[i]$ is the edge traversal speed on edge $e_k[i]$.
The arrival time $t_k[i]$ of a truck $k$ at the $i$th node on its path is the sum of the start time and the edge traversal times up to this point. Given the sequences $n_k$ and $v_k$, we calculate $t_k$ recursively as $$\begin{split}
t_k[1] &= t_k^S\\
t_k[i] &= t_k[i-1] + \frac{W(e_k[i-1])}{v_k[i-1]},
\; i \in \{2,\dots,|n_k|\}. \label{eq:t_k_def}
\end{split}$$ We denote the set of paths as ${\mathcal{R}} = \{n_1,\dots,n_K\}$ and the set of speed sequences as ${\mathcal{V}} = \{v_1,\dots,v_K\}$.
Fuel Consumption Model {#sec:fuel_model}
----------------------
We model the fuel consumption per distance traveled as a function of speed $f(v)$. A simple model of fuel consumption per distance is a second order polynomial in the speed: $
f(v) = F_r + F_a v^2,
$ where $F_r, F_a$ are positive constants [@assad_thesis]. $F_r$ accounts for forces mainly caused by rolling resistance and $F_a v^2$ accounts for aerodynamic forces. For the sake of simplicity, we neglect other factors influencing the fuel consumption such as selected gear, road grade, wind, etc. Many of these influences could be added without fundamentally changing the problem formulation, but they would complicate notation.
When platooning as a follower, i.e., not the first vehicle in the platoon, we assume that the aerodynamic coefficient is reduced to $\eta F_a$ with $0 < \eta \leq 1$, i.e., $
f(v,\eta) = F_r + \eta F_a v^2.
$ We neglect the small fuel savings of the platoon leader, i.e., the first vehicle in the platoon. Note that $\eta = 1$ corresponds to the fuel consumption without platooning. Recent experiments have shown that, typically, $\eta \approx 0.6$ [@Bonnet2000]. We say that a group of trucks platoon over edge $(n_1,n_2)$, if the start times of the edge traversal and the speeds on the edge are the same for all trucks in the group. One of the trucks is assigned to be the platoon leader. The coefficients $F_r$ and $F_a$ are assumed to be the same for all trucks. This implies that $\eta$ for the followers does not depend on the platoon leader or on how the trucks are ordered in the platoon. We appoint the truck in the platoon with the highest index the role of the platoon leader.
The complete fuel expense is the sum of the fuel expenses of each truck on each edge on its path: $
f_c({\mathcal{R}},{\mathcal{V}}) = \sum\limits_{k = 1}^{K} \sum\limits_{i = 1}^{|e_k|} W\big(e_k[i]\big) f\big(v_k[i],\eta_l(k,i)\big),
$ with $\eta_l(k,i) = \eta$ if $\exists\, k_l,i_l: k_l > k, e_{k_l}[i_l] = e_k[i], t_{k_l}[i_l] = t_k[i], v_{k_l}[i_l] = v_k[i]$ and $\eta_l(k,i) = 1$ otherwise. ${\mathcal{R}}$, ${\mathcal{V}}$ is the set of paths and the set of speed sequences respectively. The function $\eta_l$ is equal to $\eta$, if a truck with higher index $k_l$ traverses the same edge $e_{k_l}[i_l] = e_k[i]$ at the same time $t_{k_l}[i_l] = t_k[i]$ with the same speed $v_{k_l}[i_l] = v_k[i]$, and 1 otherwise. Hence, $\eta_l = \eta$ if the current truck gets the role of a platoon follower and then experiences reduced air-drag on this edge, and $\eta =1$ otherwise.
Problem Statement {#problem-statement}
-----------------
We are ready to state the optimization problem. For each transport assignment $k$ we want to find a path in ${\mathcal{G}}$ from $n_k^S$ to $n_k^D$ denoted by $n_k$ and a speed $v_k$ for each edge on the path so that the truck arrives in time, does not violate the speed constraints, and that the total fuel expense $f_c$ is minimized.
\[prob:basic\_problem\]
$$\begin{aligned}
&\min\limits_{ {\mathcal{R}},{\mathcal{V}}} f_c({\mathcal{R}},{\mathcal{V}}) \label{eq:opt_prob_obj}\\
&\;\;\;{\mathrm{ s.t. }}\nonumber\\
&{\mathrm{for}} \; k \in \{1,\dots,K\}\nonumber\\
&\;\; 0 \leq v \leq v_{\max},\; v \in v_k \label{eq:opt_prob_v_con}\\
&\;\; n_k[1] = n_k^S,\; n_k[|n_k|] = n_k^D && \label{eq:opt_prob_n_1_con}\\
&\;\; e_k[j] \in {\mathcal{E}},\; j \in \{1,\dots,|e_k|\} \label{eq:opt_prob_n_path_con}\\
&\;\; t_k[1] = t_k^S && \label{eq:opt_prob_t_s}\\
&\;\; t_k[|n_k|] \leq t_k^D && \label{eq:opt_prob_t_d_con}\end{aligned}$$
The constraints ensure that the speed $v$ on each edge is positive and below the maximum speed $v_{\max}$. The constraints ensure that the paths connect the start and the destination of each transport assignment. The constraints ensure that each sequence of node $n_1,\dots,n_k$ describes a path in ${\mathcal{G}}$, i.e., that between any two successive nodes in $n_k$ there is an edge in ${\mathcal{E}}$ connecting these nodes. The constraints ensure that each truck starts at $t_k^S$. The constraints ensure that each truck arrives in time with $t_k$ given by . It would not be difficult to make $v_{\max}$, $F_r$, and $F_a$ dependent on the edge. We omitted this, however, in order to keep the notation simple. In general, Problem \[prob:basic\_problem\] is not straightforward to solve, since small changes in the constraints can entirely change the routes of the trucks when there are several possible routes of similar length.
Approximate Solution Based on Shortest Paths {#sec:approx_sol}
============================================
The problem stated in Section \[sec:problem\_statement\] is hard to solve exactly [@jeff_distributed_coordination; @jeff_complexity]. Therefore, we propose here an approximate solution which addresses routing and speed selection in two successive steps.
In the first step, we determine the paths ${\mathcal{R}}$ [@shortest_path_overview]. We choose $n_k$ as the shortest paths with respect to the weights $W$ between $n_k^S$ and $n_k^D$. We assume that the shortest path for any pair of nodes in ${\mathcal{N}}$ is unique.
Having computed the path for each truck, the problem remains to select the speeds on each edge of the paths such that $f_c$ is minimized. Without the fuel saving effect of platooning, this would just be the slowest speed for each truck with which it arrives before the deadline $t_k^D$. Taking platooning into account, it might, however, be beneficial to choose higher speeds on some edges, in order to form platoons. There is a non-trivial trade-off between higher fuel consumption due to increased speed and reduced fuel consumption due to platooning. Notice that the speed-independent part of the fuel consumption $
\sum\limits_{k = 1}^{K} \sum\limits_{i = 1}^{|e_k|} W(e_k[i]) F_r
$ is fixed, if the paths are fixed, and hence it does not have to be considered when selecting the speeds.
We formulate the approximate problem with ${\mathcal{R}}$ ($e_k$) as the shortest routes:
\[prob:shortest\_path\_problem\]
$$\begin{aligned}
&\min\limits_{{\mathcal{V}}} \sum\limits_{k = 1}^{K} \sum\limits_{i = 1}^{|e_k|} W\big(e_k[i]\big) F_a v_k[i]^2\eta_l(k,i) \label{eq:opt_prob_obj_path}\\
&\;\;\;{\mathrm{s.t.}}\nonumber\\
&{\mathrm{for}}\; k \in \{1,\dots,K\}\nonumber\\
&\;\; 0 \leq v \leq v_{\max},\; v \in v_k \label{eq:opt_prob_v_con_path}\\
&\;\; t_k[1] = t_k^S \label{eq:opt_prob_t_s_path}\\
&\;\; t_k[|t_k|] \leq t_k^D && \label{eq:opt_prob_t_d_con_path}\end{aligned}$$
We cannot directly apply standard continuous optimization techniques to Problem \[prob:shortest\_path\_problem\] since $f_c({\mathcal{R}},{\mathcal{V}})$ is not continuous in ${\mathcal{V}}$. Instead, we propose explicitly considering which trucks platoon together over which edges as a discrete decision, which we call *platoon configuration*. Based on this decision, we can formulate a continuous, convex optimization problem with linear constraints, whose solution gives ${\mathcal{V}}$ in a way that $f_c({\mathcal{R}},{\mathcal{V}})$ is minimal for this particular platoon configuration. In order to find the optimal platoon configuration, we solve a convex program for each feasible platoon configuration and select the best one.
Optimal Speed for a Given Platoon Configuration
-----------------------------------------------
In this section, we will describe how to model the platoon configuration and how to formulate the convex program whose optimal solution are the optimal velocities ${\mathcal{V}}$ for a given platoon configuration. A program like this can be efficiently solved. In Section \[sec:platooning\_posibilities\], we will describe how to search through the possible platoon configurations in order to determine the optimal one.
Firstly, we specify in mathematical terms what we mean by *platoon configuration*. We assign each transport assignment $k$ on each edge of its path $n_k$ a *predecessor*. If the truck is not part of a platoon or if it is a platoon leader on this particular edge, the predecessor is the truck itself. If it is a follower in a platoon, the predecessor is the truck in the platoon with the smallest index higher than $k$. We collect the predecessors for truck $k$ in a sequence $l_k$.
We reformulate the problem in terms of the traversal times $T_k[i] = W(e_k[i])/v_k[i]$ as opposed to the traversal speeds $v_k$ since this allows us to express all constraints as linear constraints. The choice of $l_k$ determines $\eta_l$ to $\eta_l(k,i) = \eta$ if $l_k[i] \neq k$. We will assume that $l_k$ is chosen so that $\eta_l(k,i) = 1$ if $l_k[i] = k$. We can write the objective function as $$F_a \sum\limits_{k=1}^{K} \sum\limits_{i=1}^{|e_k|} \eta(k,i) \frac{W(e_k[i])^3}{T_k[i]^2}, \label{eq:red_objective_f}$$ where we have only kept the speed dependent part of $f_c$ and substituted $v_k[i]^2 = W(e_k[i])^2/T_k[i]^2$. We can also remove $F_a$ from the objective function since it is a constant positive scaling factor in .
Constraint is reformulated in terms of $T_k$ as $
0 \leq W(e_k[i])/v_{\max} \leq T_k[i].
$ Constraints and turn into $
t_k^S + \sum\limits_{i=1}^{|e_k|} T_k[i] \leq t_k^D.
$ Finally, we have to introduce a number of constraints that ensure that the trucks do platoon as specified by $l_k$. Therefore, we have to add a constraint for each truck $k \in \{1,\dots,K\}$ on each link $j \in \{1,\dots,|e_k|\}$ on its path, ensuring that it arrives at the same time at the beginning of the edge as its predecessor: $t_k^S + \sum\limits_{i = 1}^{j-1} T_k[i] = t_{l_k[j]}^S + \sum\limits_{i = 1}^{j_l-1} T_{l_k[j]}[i]$, and that it traverses the edge at the same speed: $T_k[j] = T_{l_k[j]}[j_l]$, where we define $j_l: n_k[j] = n_{l_k[j]}[j_l]$ as the index of node $n_k[j]$ in the path of the predecessor of truck $k$ on edge $e_k[j]$. We get the following optimization problem for a given platoon configuration $l_1,\dots,l_K$.
\[prop:convex\_cont\_problem\]
$$\begin{aligned}
&\min\limits_{ \{T_1,\dots,T_K\}} \sum\limits_{k=1}^{K} \sum\limits_{i=1}^{|e_k|} \eta(k,i) \frac{W(e_k[i])^3}{T_k[i]^2} \\
&\;\;\;\mathrm{s.t.}\nonumber\\
&\mathrm{for} \; k \in \{1,\dots,K\}\nonumber\\
&\;\; \frac{W(e_k[i])}{v_{\max}} \leq T_k[i], \; i \in \{1,\dots,|e_k|\} \label{eq:cont_opt_v_max_con}\\
&\;\; t_k^S + \sum\limits_{i=1}^{|e_k|} T_k[i] \leq t_k^D \label{eq:cont_opt_t_d_con}\\
&\;\; t_k^S + \sum\limits_{i = 1}^{j-1} T_k[i] = t_{l_k[j]}^S + \sum\limits_{i = 1}^{j_l-1} T_{l_k[j]}[i],
\; j \in \{1,\dots,|e_k|\} \label{eq:cont_opt_plat_arr_con}\\
&\;\; T_k[j] = T_{l_k[j]}[j_l],
\; j \in \{1,\dots,|e_k|\}, \label{eq:cont_opt_plat_trav_con}\end{aligned}$$
with $j_l: n_k[j] = n_{l_k[j]}[j_l]$
We can verify that we have a convex, twice-differentiable objective function on the domain of positive traversal times $T_k[i] > 0: k \in \{1,\dots,K\},\; i \in \{1,\dots,|T_k|\}$ and constraints that are linear in the decision variables $\{T_1,\dots,T_K\}$. We provide an example to illustrate the formulation and solution of Problem \[prop:convex\_cont\_problem\].
Figure \[fig:example\_graph\_cont\_problem\] depicts a small road network and the paths of two trucks. We choose for the platoon configuration that the two trucks platoon from node 3 to 5, which means $l_1=(1,2,2,1,1)$ and $l_2=(2,2,2,2,2)$.
Figure \[fig:two\_truck\_cont\_opt\] shows the solution to this problem. We can see that the two trucks do meet at node 3 and platoon until node 5. Then they split up and have different velocities due to their different deadlines. An interesting observation is that both trucks attain their maximum speed, corresponding to minimum slope in Figure \[fig:two\_truck\_cont\_opt\], while platooning whereas lower speeds are optimal when they drive on their own. This is due to the reduced air drag that can render beneficial to choose high speed while driving in a platoon and low speed otherwise.
Evaluation of Platooning Possibilities {#sec:platooning_posibilities}
--------------------------------------
In order to find the optimal platoon configuration, we propose solving Problem \[prop:convex\_cont\_problem\] for different platoon configurations. We first describe how to enumerate all platoon configurations so that Problem \[prop:convex\_cont\_problem\] has a feasible solution. We then establish a property of optimal platoon configurations which further limits the discrete search space. Finally, we provide examples to illustrate the procedure.
In the first step, we determine which paths share edges. Two trucks can only platoon on common edges on their paths. We define the set of shared edges between two paths $n_{k_1}$, $n_{k_2}$ as $
n_{k_1} \cap_{\mathcal{E}} n_{k_2} = \{(n_{k_1}[i],n_{k_1}[i+1]): \exists j: (n_{k_1}[i],n_{k_1}[i+1]) = (n_{k_2}[j],n_{k_2}[j+1]) \}.
$ The following lemma is useful when determining the possible platoon configurations.
\[prop:path\_intesection\] If the paths $n_{k_1}$ and $n_{k_2}$ are unique shortest paths in ${\mathcal{G}}$, then either $n_{k_1} \cap_{\mathcal{E}} n_{k_2}$ is empty or the path formed by $n_{k_1} \cap_{\mathcal{E}} n_{k_2}$ also forms a shortest path between two nodes of ${\mathcal{G}}$.
It is well known that every sub-path of a shortest path is a shortest path [@introduction_to_algorithms]. But then the path between any two nodes of $n_{k_1}$ as well as between any two nodes of $n_{k_2}$ is unique. If $n_{k_1} \cap_{\mathcal{E}} n_{k_2}$ does not form a shortest path there exists a sub-path in $n_{k_1}$ as well as in $n_{k_2}$ connecting the same nodes which is not a sub-path of the path formed by $n_{k_1} \cap_{\mathcal{E}} n_{k_2}$. Then there are different sub-paths in $n_{k_1}$ and $n_{k_2}$ connecting the same nodes. This, however, contradicts the assumption that the shortest path between two nodes is unique.
If the paths of two trucks share common edges, we can check if it is possible that the two trucks form a platoon according to the constraints on the maximum speed and the before-deadline arrival .
Therefore, we calculate the earliest and latest arrival time at each node of the paths ${\mathcal{R}}$. The earliest arrival times for the path of truck $k$, denoted by $\underline{t}_k$, can be calculated forward the same way $t_k$ is calculated in with $v_k[1] = v_k[2] = \dots = v_k[|v_k|] = v_{\max}$: $\underline{t}_k[1] = t_k^S$ and $\underline{t}_k[i] = t_k[i-1] + W(e_k[i-1])/v_{\max}$ for $i \in \{2,\dots,|n_k|\}$.
In a similar way, we can calculate the latest arrival time $\bar{t}_k$ at each node backward. We have $t_k[1] = t_k^S$. Since there is no immediate constraint on how large $T_k[1]$ can be, the latest arrival time at nodes $n_k[2], n_k[3], \dots$ is only constrained by and . The latest arrival time $n_k[|n_k|]$ is $\bar{t}_k[|n_k|] = t_k^D$. Then, the latest time the truck can leave at $n_k[|n_k|-1]$ in order to arrive at $n_k[|n_k|]$ before $t_k^D$ is given by $\bar{t}_k[|n_k|] - W(e_k[|n_k|-1])/v_{\max}$. In this way we can work backwards until $n_k[2]$. So, we define $\bar{t}_k$ recursively as $\bar{t}_k[|n_k|] = t_k^D$, $\bar{t}_k[1] = t_k^S$ and $\bar{t}_k[i-1] = \bar{t}_k[i] - W(e_k[i-1])/v_{\max}$ for $i \in \{3,\dots,|n_k|\}$.
If there is a solution to Problem \[prob:shortest\_path\_problem\], we have $\bar{t}_k[i] \geq \underline{t}_k[i]$, and since all elements of $v_k$ can take any value in the interval $[0,v_{\max}]$, the truck can reach any $t_k[i] \in \big[\underline{t}_k[i], \bar{t}_k[i]\big]$ for a particular node $n_k[i]$. Due to not every sequence $t_k[i]$ which fulfills $t_k[i] \in \big[\underline{t}_k[i], \bar{t}_k[i]\big]$ has to be feasible. Figure \[fig:graph\_timining\_constraints\] illustrates the calculation of the sequences $\underline{t}_k$ and $\bar{t}_k$. The following result can now be stated.
\[prop:possible\_l\_per\_link\] Assume that $k_2 > k_1$ and that there exists a solution to Problem \[prop:convex\_cont\_problem\] without platoon followers, i.e., $l_k[\tilde{j}] = k$ for all $\tilde{j} \in \{1,\dots,|l_k|\}$ and all $k \in \{1,\dots,K\}$. Then, there exists a solution to Problem \[prop:convex\_cont\_problem\] with $k_2 = l_{k_1}[j]$ with $j \in \{1,\dots,|l_{k_1}|\}$ if and only if $\big[\underline{t}_{k_1}[j], \bar{t}_{k_1}[j]\big] \cap \big[\underline{t}_{k_2}[j_l], \bar{t}_{k_2}[j_l]\big] \neq \emptyset$ where $j_l: n_{k_1}[j] = n_{l_{k_1}[j]}[j_l]$.
Constraints \[eq:cont\_opt\_plat\_arr\_con,eq:cont\_opt\_plat\_trav\_con\] enforce $t_{k_1}[j] = t_{k_2}[j_l]$ and $t_{k_1}[j+1] = t_{k_2}[j_l+1]$, with $j$ and $j_l$ as in the theorem. On the other hand, we have $t_{k_1}[j] \in \big[\underline{t}_{k_1}[j], \bar{t}_{k_1}[j]\big]$ and $t_{k_2}[j_l] \in \big[\underline{t}_{k_2}[j_l], \bar{t}_{k_2}[j_l]\big]$. So, if $\big[\underline{t}_{k_1}[j], \bar{t}_{k_1}[j]\big] \cap \big[\underline{t}_{k_2}[j_l], \bar{t}_{k_2}[j_l]\big] = \emptyset$, we cannot find a solution to Problem \[prop:convex\_cont\_problem\] with $t_{k_1}[j] = t_{k_2}[j_l]$. This proves necessity.
Assume $l_k[p] = k$ for $(k,p) \neq (k_1,j),(k_1,j+1)$. If $\big[\underline{t}_{k_1}[j], \bar{t}_{k_1}[j]\big] \cap \big[\underline{t}_{k_2}[j_l], \bar{t}_{k_2}[j_l]\big] \neq \emptyset \wedge \big[\underline{t}_{k_1}[j+1], \bar{t}_{k_1}[j+1]\big] \cap \big[\underline{t}_{k_2}[j_l+1], \bar{t}_{k_2}[j_l+1]\big] \neq \emptyset$, then there exists $T_{k_1}[1], \dots, T_{k_1}[j-1]$ and $T_{k_2}[1], \dots, T_{k_2}[j_l-1]$ such that ${\mathrm{max}}(\underline{t}_{k_1},\underline{t}_{k_2}) \leq t_{k_1}^S + \sum\limits_{p = 1}^{j-1} T_{k_1}[p] = t_{k_2}^S + \sum\limits_{p = 1}^{j_l-1} T_{k_2}[p] \leq {\mathrm{min}}(\bar{t}_{k_1},\bar{t}_{k_2})$. Assign now $T_{k}[p] = W(e_{k}[p])/v_{\max}$ for $(k,p) \neq (k_1,1),\dots,(k_1,i-1),(k_2,1),\dots,(k_2,j-1)$. Since the solution for $l_k[i] = k$ exists, these $T_1,\dots,T_K$ fulfill constraints \[eq:cont\_opt\_v\_max\_con,eq:cont\_opt\_t\_d\_con,eq:cont\_opt\_plat\_arr\_con,eq:cont\_opt\_plat\_trav\_con\]. This proves sufficiency.
Theorem \[prop:possible\_l\_per\_link\] allows us to easily identify possible values for the elements of the sequences $l_k$, $k \in \{1,\dots,K\}$. However, Problem \[prop:convex\_cont\_problem\] does not have a solution for each combination of these values. Therefore, we describe next how to check feasibility for a particular platoon configuration. We do this by iterating over all paths of the trucks and the nodes therein. We calculate the intersection of the feasible arrival time at this node between a truck and its predecessor on the next edge on the truck’s path. In each iteration, we set $\big[\underline{t}_{k}[j], \bar{t}_{k}[j]\big]^+ = \big[\underline{t}_{l_k[j]}[j_l], \bar{t}_{l_k[j]}[j_l]\big]^+ = \big[\underline{t}_{k}[j], \bar{t}_{k}[j]\big] \cap \big[\underline{t}_{l_k[i]}[j_l], \bar{t}_{l_k[i]}[j_l]\big]$, where the superscript “+” indicates the values after the iteration and $j_l: n_k[j] = n_{l_k[j]}[j_l]$ is the index of node $n_k[j]$ in the path of the predecessor of truck $k$ on edge $e_k[j]$. Then, we propagate constraint for the trucks associated with transport assignments $k$ and $l_k[j]$. In this way, we prune the intervals created by $\underline{t}$, $\bar{t}$ until there either is a node where $\big[\underline{t}_{k}[j], \bar{t}_{k}[j]\big] \cap \big[\underline{t}_{l_k[j]}[j_l], \bar{t}_{l_k[j]}[j_l]\big]$ is empty or there is no change after an iteration over all nodes. In the first case, the platoon configuration is not feasible. In the second case, there is a solution to Problem \[prop:convex\_cont\_problem\]. Furthermore, the sequences $t_k, k \in \{1,\dots,K\}$ of the solution lie in the intervals with lower bound $\underline{t}_k$ and upper bound $\bar{t}_k$.
The above procedure will give us all platoon configurations for which Problem \[prop:convex\_cont\_problem\] has a solution. Recall that we are only interested in platoon configurations where the predecessor of a truck on an edge is the truck in the platoon with next largest index. However, we want to find the solution to Problem \[prob:shortest\_path\_problem\]. To this end, we can show that in an optimal platoon configuration two trucks will meet and split up at most once.
\[prop:only\_one\_split\] If the solution to Problem \[prop:convex\_cont\_problem\] with platoon configuration $l_1,\dots,l_k$ is the unique solution to Problem \[prob:shortest\_path\_problem\], then $l_1,\dots,l_k$ have the property that for $k_l \neq k$ there exists for each $k \in \{1,\dots,K\}$ at most one pair $(j_1,j_2): j_1 \leq j_2$ such that $k_l \neq l_k[j_1-1], k_l = l_k[j_1]$ and $k_l = l_k[j_2], k_l \neq l_k[j_2+1]$.
Let $\tilde{n}$ be the path formed by $n_{k_1} \cap_{\mathcal{E}} n_{k_2}$ and let $\tilde{e}$ be the corresponding sequence of edges. We know from Lemma \[prop:path\_intesection\] that $\tilde{n}$ will indeed be a path. Assume there is a unique optimal solution with $v_{k_1}$, $v_{k_2}$ to Problem \[prob:shortest\_path\_problem\] where the above statement does not hold, i.e., assume there exists $j_1,j_2: j_2 > j_1$ so that trucks $k_1$, $k_2$ platoon on $(\tilde{n}[j_1-1],\tilde{n}[j_1])$ and on $(\tilde{n}[j_2],\tilde{n}[j_2+1])$ but not on the edges between $\tilde{n}[j_1]$ and $\tilde{n}[j_2]$. Then, $k_1$, $k_2$ could also platoon on the edges between $\tilde{n}[j_1]$ and $\tilde{n}[j_2]$. Denote the speed profile of truck $k_1$, $k_2$ on these edges as $v_{k_1}$, $v_{k_2}$ respectively. Assume without loss of generality $\sum\limits_{m} W(\tilde{e}[j_1-1+m])v_{k_1}[m]^2 \leq \sum\limits_{m} W(\tilde{e}[j_1-1+m]) v_{k_2}[m]^2$. We assign truck $k_2$ on these nodes the speed profile of truck $k_1$ which implies they platoon on these edges. For each edge $m$ there are two cases: $k_2$ was part of another platoon. Then the change in the contribution of this edge to the objective function is $\eta(v_{k_1}[m]^2 - v_{k_2}[m]^2)$. If $k_2$ was not part of a platoon then the change for this edge in the objective function is $\eta v_{k_1}[m]^2 - v_{k_2}[m]^2 \leq \eta(v_{k_1}[m]^2 - v_{k_2}^2)$. So, the change in the objective function from the old to the new speed profile is upper bounded by $\eta \sum\limits_{m} W(\tilde{e}[j_1-1+m])(v_{k_1}[m]^2 - v_{k_2}[m]^2) \leq 0$ which contradicts the assumption that $v_{k_1}$, $v_{k_2}$ are part of a unique optimal solution.
In the next example, we illustrate how our method is applied to a realistic scenario.
We have four transport assignments. The road network ${\mathcal{G}}$ is shown in Figure \[fig:companion\_problem\_map\]. It is a handmade abstraction of an area in the east of Hungary and the west of Romania. The weights are in kilometers.
We set $v_{\max} = 90$ and $\eta = 0.6$. The start times are set to $t_1^S = 0.2,t_2^S=-0.3,t_3^S=0.75,t_4^s=-0.05$. The arrival deadlines are set to the arrival time when traveling the whole path with a speed of $80$. This creates 340 valid platoon configurations. The result of the optimization is shown in Figure \[fig:companion\_problem\_result\_3d\]. The best configuration turns out to be $l_1 = (1, 1, 2, 2, 2, 2, 1)$, $l_2 = (2, 2, 2, 2, 2, 3, 2, 2, 2)$, $l_3 = (3, 3, 4, 4)$, $l_4 = (4, 4, 4)$. It is interesting to note that trucks 1 and 2 (blue and green) do not platoon on the edges $(2,3)$ and $(16,18)$, even though they could according to the constraints. The value of the objective function is 5% lower than the solution where no coordination takes place, i.e., the air drag coefficient is in average reduced by 5%. If more favorable starting times are chosen, the reduction can be much bigger.
Conclusion
==========
In this paper, we have investigated the problem of how to plan the speed of trucks by exploiting the possibility of forming platoons. We ensured that speed limitations and arrival deadlines are not exceeded. We proposed an approximate solution, where each truck takes its shortest path, yielding a number of convex optimization problems. We demonstrated our solution on a realistic scenario.
We made a number of simplifications that would need to be considered in a real world deployment. Some of them can be easily tackled, such as different speed levels per edge and fuel consumption depending on the edge. Other factors, such as unknown traffic conditions and mandatory rest times for the driver, might be more difficult to handle and will be subject to future work. Furthermore, we will investigate how to decrease complexity. The complexity of the algorithm can be problematic in scenarios where many trucks can platoon.
[^1]: The authors are with the ACCESS Linnaeus Center and the School of Electrical Engineering, KTH Royal Institute of Technology, SE-100 44, Stockholm, Sweden. [{shvdh, kallej, dimos}@kth.se]{}. This work was supported by the COMPANION EU project.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Using methods of A. Grigor’yan and L. Saloff-Coste we prove that on a manifold with a conical end the heat kernel has a Gaussian bound. This result is applied to asymptotically conical Kähler manifolds. It is a result of the author and R. Goto that a crepant resolution $\pi:Y\rightarrow X$ of a Ricci-flat Kähler cone $X$ admits a Ricci-flat Kähler metric asymptotic to the cone metric in every Kähler class. We prove the sharp rate of convergence of the metric to the cone metric. For compact Kähler classes this is the same as for the Ricci-flat ALE metrics of P. Kronheimer and D. Joyce.'
address: 'Department of Mathematics, Massachusetts Institute of Technology, 77 Massachusetts Avenue, Cambridge, MA 02139-4307'
author:
- Craig van Coevering
date: 'December 20, 2009'
title: 'Regularity of asymptotically conical Ricci-flat Kähler metrics'
---
Introduction
============
This article considers Riemannian manifolds with a conical end. A metric cone is a manifold $(C(S),g)$ with $C(S)={\mathbb{R}}_{>0} \times S$ and $g=dr^2 +r^2 g_S$. We consider manifolds with an end for which the metric is approximated by a cone metric. This can be considered a generalization of ALE manifolds. But in this case the compact manifold $(S,g_S)$ is arbitrary. We will also consider Kähler manifolds with an end approximated by a Kähler cone. If $(C(S),g)$ is Kähler, then $(S,g_S)$ is a Sasaki manifold. So in this case $(S,g_S)$ is far from arbitrary, and such manifolds have been studied extensively [@BG3]. In particular we will consider Ricci-flat asymptotically conical Kähler manifolds.
Ricci-flat Kähler manifolds with an asymptotically conical end, and in particular resolutions of a Ricci-flat Kähler cone $(C(S),g)$, have been of particular interest recently due to their relevance to the AdS/CFT correspondence. See [@MS2] for the construction of many explicit examples which are resolutions of Ricci-flat Kähler cones. See also [@MS3; @MS4] for more on the relevance of these manifolds to AdS/CFT.
Asymptotically conical Ricci-flat Kähler manifolds have been extensively studied by solving the Monge-Ampére equation in various cases. There are the existence results of G. Tian and S.-T. Yau [@TY2], and independently S. Bando and R. Kobayashi [@BK2], on Ricci-flat Kähler metrics on quasi-projective manifolds $X\setminus D$. And there are the results of D. Joyce on the existence of Ricci-flat ALE metrics [@Joy1]. The author [@vC3] and R. Goto [@Got] have proved existence results on resolutions of Ricci-flat Kähler cones. This article shows what solutions to the Monge-Ampére equation behave essentially as in the ALE case due to D. Joyce. In particular, we have the following for resolutions of cones.
\[thm:main\] Let $\pi: \hat{X}\rightarrow X=C(S)\cup\{o\}$ be a crepant resolution of a Ricci-flat Kähler cone $(X,{\bar{g}})$ of complex dimension $m$. Then in every Kähler class in $H_c^2(\hat{X},{\mathbb{R}})\subset H^2(\hat{X},{\mathbb{R}})$ there is a unique Ricci-flat Kähler metric $g$ asymptotic to ${\bar{g}}$ as follows. There exists an $R>0$ such that for each $k\geq 0$ $$\label{eq:conv-cpt}
|\nabla^k(\pi_* g-{\bar{g}})| =O(r^{-2m-k})\quad\text{on }\{x\in C(S): r(x)>R\},$$ where $\nabla$ and the point-wise norm are with respect to ${\bar{g}}$.
Furthermore, in every Kähler class in $H^2(\hat{X},{\mathbb{R}})\setminus H_c^2(\hat{X},{\mathbb{R}})$ there is a Ricci-flat metric $g$ asymptotic to ${\bar{g}}$. In this case there exists an $R>0$ such that for each $k\geq 0$ $$\label{eq:conv-noncpt}
|\nabla^k(\pi_* g-{\bar{g}})| =O(r^{-2-k})\quad\text{on }\{x\in C(S): r(x)>R\}.$$
Both rates of convergence in (\[eq:conv-cpt\]) and (\[eq:conv-noncpt\]) are sharp. The contribution of this article is in proving the sharp convergence in (\[eq:conv-cpt\]). The author proved a weaker version in [@vC3] where the exponent in (\[eq:conv-cpt\]) is $-2m+\delta,\ \delta>0$. This was also proved by R. Goto [@Got] along with a clever proof of the existence for non-compact Kähler classes.
The first part of this article considers the more general case of arbitrary real manifolds with a conical end. Using methods of A. Grigor’yan and L. Saloff-Coste we prove that every such manifold satisfies the parabolic Harnack inequality, and as a consequence we have a Gaussian bound on the heat kernel. This follows from the scale-invariant Poincaré inequality, and the volume doubling condition, which is proved using a discretization technique. We use this to prove some results on the Laplacian on weighted Hölder spaces that will be needed later.
In the second part we consider asymptotically conical Kähler manifolds. The main result is a version of the Calabi conjecture for asymptotically conical Kähler manifolds. The proof due to D. Joyce [@Joy1] for ALE Kähler manifolds goes through as is using the Sobolev inequality and the results on the Laplacian on weighted Hölder space in this context. The result is that solutions to the Monge-Ampére equation on asymptotically conical Kähler manifolds behave as on ALE Kähler manifolds.
The final section gives an overview of some examples of Ricci-flat asymptotically conical manifolds given by Theorem \[thm:main\]. One is considering cones over Sasaki-Einstein manifolds which have crepant resolutions. Examples can be found from hypersurface singularities. Many such examples of Sasaki-Einstein manifolds are known (cf. [@BG2; @BGJ; @BG3]). We also briefly discuss the toric case. The existence problem of Sasaki-Einstein metrics in this case is solved [@FOW]. Therefore it is an easy source of examples.
Theorem \[thm:main\] does not settle the existence problem of asymptotically conical Ricci-flat Kähler metrics. More generally, one can consider crepant resolutions $\pi:\hat{X}\rightarrow X$ of more general affine varieties $X$ such that $\hat{X}$ is a quasi-projective variety with an end which is diffeomorphic to a cone, but not holomorphically.
Acknowledgements {#acknowledgements .unnumbered}
----------------
I would like to thank Gilles Carron for referring me to the articles [@Min] of V. Minerbe and [@GriSal-Cos] of A. Grigor’yan and L. Saloff-Coste which allowed me to prove the heat kernel bounds.
Notation {#notation .unnumbered}
--------
We will use ‘$\subset$’ to denote set inclusion, proper and otherwise. The geodesic ball centered at $x\in M$ of radius $r$ will be denoted by $B(x,r)$, and we will use $V(x,r)$ to denote its volume with respect to the Riemannian measure.
Asymptotically conical manifolds and
====================================
Asymptotically conical manifolds
--------------------------------
We cover some analysis on asymptotically conical manifolds. This will be used in the sequel in the proof of the Calabi conjecture. But it may be of independent interest.
Let $(X,g)$ be a complete Riemannian manifold. Using the metric $g$ and Levi-Civita connection $\nabla$ we have the Sobolev spaces $L_k^q(X)$, the $C^k$-spaces $C^k(X)$, and the Hölder spaces $C^{k,\alpha}(X)$. But we will also need *weighted Sobolev spaces* and *weighted Hölder spaces*.
Let $o\in X$ be a fixed point and $d(o,x)$ the Riemannian distance to $x\in X$. Then we define a weight function $\rho(x)=(1+d(o,x)^2)^{1/2}$.
Let $q\geq 1$, $\beta\in{\mathbb{R}}$, and $k$ a nonnegative integer. We define the *weighted Sobolev space* $L_{k,\beta}^q$ to be the set of functions $f$ on $X$ which are locally integrable and whose weak derivatives up to order $k$ are locally integrable and for which the norm $$\|f\|_{L_{k,\beta}^q}:=\left(\sum_{j=0}^k \int_X |\rho^{j-\beta}\nabla^j f|^q \rho^{-n}d\mu_g \right)^{1/q}$$ is finite. Then $L_{k,\beta}^q$ is a Banach space with this norm.
For $\beta\in{\mathbb{R}}$ and $k$ a nonnegative integer we define $C_\beta^k(X)$ to be the space of continuous functions $f$ with $k$ continuous derivatives for which the norm $$\|f\|_{C_\beta^k} := \sum_{j=0}^k \underset{X}{\sup}|\rho^{j-\beta}\nabla^j f|$$ is finite. Then $C_\beta^k(X)$ is a Banach space with this norm.
Let ${\operatorname{inj}}(x)$ be the injectivity radius at $x\in X$, and $d(x,y)$ the distance between $x,y\in X$. Then for $\alpha,\gamma\in{\mathbb{R}}$ and $T$ a tensor field define $$[T]_{\alpha,\gamma} :=\sup_{\substack{x\neq y\\ d(x,y)<{\operatorname{inj}}(x)}}\left[\min(\rho(x),\rho(y))^{-\gamma}\cdot\frac{|T(x)-T(y)|}{d(x,y)^\alpha}\right],$$ where $|T(x)-T(y)|$ is defined by parallel translation along the unique geodesic between $x$ and $y$.
For $\alpha\in(0,1)$ define the *weighted Hölder space* $C_\beta^{k,\alpha}(X)$ to be the set of $f\in C_\beta^k(X)$ for which the norm $$\|f\|_{C_\beta^{k,\alpha}} :=\|f\|_{C_\beta^k} +[\nabla^k f]_{\alpha,\beta-k-\alpha}$$ is finite. Then $C_\beta^{k,\alpha}(X)$ is a Banach space with this norm.
Define $C_\beta^\infty (X)$ to be the intersection of the $C_\beta^k(X)$, for $k\geq 0$.
It will be convenient to use a different weight function $\rho(x)$ on the manifolds we consider but it will define equivalent norms. See [@Ch-SCh-B] for an introduction to the theory of weighted Hölder spaces.
\[defn:cone\] Let $(S,g_S)$ be a compact Riemannian manifold. The cone over $(S,g_S)$ is the Riemannian manifold $(C(S),g)$ with $C(S)={\mathbb{R}}_{>0} \times S$ and $g=dr^2 +r^2 g_S$ where $r$ usual coordinate on $R_{>0}$.
We will sometimes consider $(C(S),g)$ with the apex $o\in C(S)$ at $r=0$. This is singular at $o\in C(S)$ unless $S=S^{n-1}$ is the sphere with the round metric.
\[defn:AC\] Let $(C(S),g_0)$ be a metric cone. Then $(X,g)$ is asymptotically conical of order $(\delta, k+\alpha)$, if there is compact subset $K\subset X$, a compact neighborhood $o\in K_0 \subset C(S)$, and diffeomorphism $\phi: X\setminus K \rightarrow C(S)\setminus K_0$ so that $$|\phi_* g -g_0| \in C_{\delta} ^{k,\alpha}\quad\text{on } C(S)\setminus K_0.$$ We will abbreviate this by ${\operatorname{AC}}(\delta, a+\alpha)$, and always $\delta<0$.
Suppose $K_0\subset D_{r_0} =\{(r,s)\in C(S): r< r_0\}$, and assume $r_0 \geq 2$. Then a smooth extension of $\phi^* r :X\setminus\phi^{-1}(D_{r_0}) \rightarrow[2,\infty)$ to $\rho: X\rightarrow [1,\infty)$ is a *radius function* of the ${\operatorname{AC}}$ manifold $(X,g)$.
It will be convenient sometimes to define weighted Hölder and Sobolev spaces using the radius function as the weight function $\rho$. In other cases we will use $\tilde{\rho(x)}=(1+d(o,x)^2)^{1/2}$ for $o\in K$. This will mostly be a matter of convenience. If $(X,g)$ is ${\operatorname{AC}}(\delta,0), \delta>0,$ it is not difficult to check that $c^{-1}\tilde{\rho}\leq\rho\leq c\tilde{\rho}$ for $c>0$. Thus the weighted norms are equivalent. We will denote such a relation between functions by $\rho\sim\tilde{\rho}$.
The Sobolev inequality
----------------------
We give a proof of the Sobolev inequality on asymptotically conical manifolds. Recall that metrics $g$ and $\tilde{g}$ on $X$ are *quasi-isometric* if there is a $c>0$ so that $$\label{eq:quasi-isom}
c^{-1}g_x (W,W)\leq \tilde{g}_x (W,W)\leq cg_x (W,W),\quad\forall x\in X\text{ and }\forall W\in T_x X.$$
\[thm:Sobolev\] Let $(X,g)$ be AC of order $(\delta, 0), \delta<0,$ or merely quasi-isometric to an AC manifold. Then there is a constant $C>0$ so that we have the *Sobolev inequality* $$\label{eq:Sobolev}
\|f\|_{n/(n-1)} \leq C\|\nabla f\|_1 ,\quad\forall f\in C_0 ^\infty (X).$$
And easy argument with the Hölder inequality gives the following.
\[cor:Sobolev\] For any real $p, 1\leq p<n$ we have $$\|f\|_{np/(n-p)} \leq C\|\nabla f\|_p ,\quad\forall f\in C_0 ^\infty (X).$$
Note that (\[eq:Sobolev\]), and most of what follows in this section, is stable under quasi-isometries. So all that is essential in Theorem \[thm:Sobolev\] is that the end of $(X,g)$ is quasi-isometric to a cone.
We prove Theorem \[thm:Sobolev\] with a discretization procedure used in [@GriSal-Cos] to prove Poincaré inequalities and generalized in [@Min] to prove more general Poincaré-Sobolev inequalities. A proof of (\[eq:Sobolev\]) is given in [@TY2], but our result is more general and the proof is simpler. Let $\mu$ denote the Riemannian measure on $(X,g)$.
\[defn:good-cov\] Let $A\subset A^{\#}$ be subsets of $X$. A family $\mathcal{U}=(U_i ,U^*_i ,U^{\#}_i)_{i\in I}$ consisting of subsets of $X$ having finite measure is said to be a good covering of $A$ in $A^{\#}$ if the following is true:
$A\subset\cup_i U_i \subset\cup_i U_i^{\#} \subset A^{\#}$;\[defn:gc-i\]
$\forall i\in I, U_i \subset U_i^* \subset U_i^{\#}$;\[defn:gc-ii\]
$\exists Q_1 ,\forall i_0 \in I, {\operatorname{Card}}\{i\in I :U_{i_0}^{\#} \cap U_i^{\#} \neq\emptyset\}\leq Q_1$;\[defn:gc-iii\]
For every $(i,j)\in I^2$ satisfying ${\bar{U}}_i \cap{\bar{U}}_j \neq\emptyset$, there is an element $k(i,j)$ such that $U_i \cup U_j \subset U^*_{k(i,j)}$;\[defn:gc-iv\]
There exists a $Q_2$ such that for every $(i,j)\in I^2$ with ${\bar{U}}_i \cap{\bar{U}}_j \neq\emptyset$ we have $\mu(U^*_{k(i,j)})\leq Q_2 \min(\mu(U_i),\mu(U_j))$.\[defn:gc-v\]
Given a Borel set $U$ with finite $\mu$-measure and a $\mu$-integrable function $f$, denote by $f_U$ the mean value of $f$ on $U$: $$f_U =\frac{1}{\mu(U)}\int_U f d\mu.$$
Given any good covering we have an associated weighted graph $(\mathcal{G},m)$ as follows.
\[defn:weig-graph\] Let $\mathcal{U}=(U_i ,U^*_i ,U^{\#}_i)_{i\in I}$ be a good covering of $A$ in $A^{\#}$. The associated weighted graph $(\mathcal{G},m)$ has vertices $\mathcal{V}=I$ and edges $\mathcal{E}=\{\{i,j\}\subset\mathcal{V}: i\neq j, {\bar{U}}_i \cap{\bar{U}}_j \neq\emptyset\}$.
Measures, both denoted $m$, are defined on $\mathcal{V}$ and $\mathcal{E}$ as follows:
$\forall i\in\mathcal{V}, m(i)=\mu(U_i)$;
$\forall \{i,j\}\in\mathcal{E}, m(i,j)=\max(m(i),m(j)).$
We will patch together Sobolev inequalities on the subsets $(U_i ,U^*_i ,U^{\#}_i)$ of a good covering using discrete Sobolev inequalities on the associated weighted graph $(\mathcal{G},m)$.
\[defn:Sobolev-cont\] Let $1\leq p<\infty$, and suppose $p<\nu\leq\infty$. We say that a good covering $\mathcal{U}$ satisfies a continuous $L^p$ Sobolev inequality of order $\nu$ if there exists a constant $S_c$ such that for every $i\in I$ $$\left(\int_{U_i} |f-f_{U_i}|^{\frac{p\nu}{\nu-p}}d\mu\right)^{\frac{\nu-p}{\nu}} \leq S_c \int_{U^*_i}|\nabla f|^p d\mu, \quad\forall f\in C^{\infty}(U^*_i),$$ and $$\left(\int_{U^*_i} |f-f_{U_i}|^{\frac{p\nu}{\nu-p}}d\mu\right)^{\frac{\nu-p}{\nu}} \leq S_c \int_{U^{\#}_i}|\nabla f|^p d\mu, \quad\forall f\in C^{\infty}(U^{\#}_i).$$
\[defn:Sobolev-dis-Dir\] Given $1\leq p<\nu\leq\infty$, we say that the weighted graph $(\mathcal{G},m)$ satisfies a discrete $L^p$ Sobolev-Dirichlet inequality of order $\nu$ if there exists a constant $S_d$ such that for every $f\in L^p(\mathcal{V},m)$ $$\left(\sum_{i\in\mathcal{V}} |f(i)|^{\frac{p\nu}{\nu-p}} m(i)\right)^{\frac{\nu-p}{\nu}}\leq S_d \sum_{{i,j}\in\mathcal{E}} |f(i)-f(j)|^p m(i,j).$$
Suppose $(\mathcal{G},m)$ is a finite weighted graph. Then for $f\in{\mathbb{R}}^\nu$, denote by $m(f)$ the mean of $f$ $$m(f)=\frac{1}{m(\mathcal{V})}\sum_{i\in\mathcal{V}} f(i).$$
\[defn:Sobolev-dis-Neu\] Given $1\leq p<\nu\leq\infty$, we say that a finite weighted graph $(\mathcal{G},m)$ satisfies a discrete $L^p$ Sobolev-Neumann inequality of order $\nu$ if there exists a constant $S_d$ such that for every $f\in{\mathbb{R}}^{N}$, $N=|\mathcal{V}|$, $$\left(\sum_{i\in\mathcal{V}} |f(i)-m(f)|^{\frac{p\nu}{\nu-p}} m(i)\right)^{\frac{\nu-p}{\nu}}\leq S_d \sum_{{i,j}\in\mathcal{E}} |f(i)-f(j)|^p m(i,j).$$
Note that the $L^p$ Sobolev inequality of order $\nu=\infty$ is the $L^p$ Poincaré inequality.
The following theorem is crucial in the proof of the Sobolev inequality on $(X,g)$. We also state the “Neumann” analogue of this theorem since we will use it in a proof of a Poincaré inequality on $(X,g)$.
\[thm:cov-Sobolev-Dir\] Suppose $1\leq p<\nu\leq\infty$. If a good covering $\mathcal{U}$ of $A$ in $A^{\#}$ satisfies the continuous $L^p$ Sobolev inequality of order $\nu$ (Definition \[defn:Sobolev-cont\]) and the discrete $L^p$ Sobolev-Dirichlet inequality of order $\infty$ (Definition \[defn:Sobolev-dis-Dir\]), then the following Sobolev-Dirichlet inequality is true: $$\left(\int_A |f|^{\frac{p\nu}{\nu-p}} d\mu\right)^{\frac{\nu-p}{p}}\leq S\int_{A^{\#}}|\nabla f|^p d\mu,\quad\forall f\in C_c^{\infty}(A).$$ Furthermore, one can choose $S=S_c Q_1 2^{p-1+\frac{p}{\nu}}(1+S_d Q_2(2^p Q_1^2)^{\frac{\nu}{\nu-p}})^{\frac{\nu-p}{\nu}}$.
\[thm:cov-Sobolev-Neu\] Suppose $1\leq p<\nu\leq\infty$. If a finite good covering $\mathcal{U}$ of $A$ in $A^{\#}$ satisfies the continuous $L^p$ Sobolev inequality of order $\nu$ (Definition \[defn:Sobolev-cont\]) and the discrete $L^p$ Sobolev-Neumann inequality of order $\infty$ (Definition \[defn:Sobolev-dis-Neu\]), then the following Sobolev-Neumann inequality is true: $$\left(\int_A |f-f_A|^{\frac{p\nu}{\nu-p}} d\mu\right)^{\frac{\nu-p}{p}}\leq S\int_{A^{\#}}|\nabla f|^p d\mu,\quad\forall f\in C^{\infty}(A^{\#}).$$ Furthermore, one can choose $S=S_c Q_1 2^{2p-1+\frac{p}{\nu}}(1+S_d Q_2(2^p Q_1^2)^{\frac{\nu}{\nu-p}})^{\frac{\nu-p}{\nu}}$.
We will need prove discrete Sobolev inequalities to apply these theorems. The discrete Sobolev-Dirichlet inequality in Definition \[defn:Sobolev-dis-Dir\] follows from an isoperimetric inequality on the graph $(\mathcal{G},m)$.
Given a graph $\mathcal{G}$, we define the boundary $\partial\Omega$ of a subset $\Omega\subset\mathcal{V}$ as $$\partial\Omega:=\{\{i,j\}\in\mathcal{E}:\{i,j\}\cap\Omega\neq\emptyset\text{ and } \{i,j\}\cap(\mathcal{V}\setminus\Omega)\neq\emptyset\}.$$
The following result is completely analogous to the situation with continuous Sobolev inequalities.
\[prop:disc-isop\] Let $(\mathcal{G},m)$ be an infinite weighted graph and fix $1<\nu\leq\infty$. Then the $L^1$ Sobolev inequality of order $\nu$ $$\left(\sum_{i\in\mathcal{V}} |f(i)|^{\frac{\nu}{\nu-1}} m(i)\right)^{\frac{\nu-1}{\nu}}\leq C\sum_{\{i,j\}\in\mathcal{E}} |f(i)-f(j)|m(i,j),\quad\forall f\in L^1(\mathcal{V},m),$$ is equivalent to the isoperimetric inequality of order $\nu$ $$m(\Omega)^{\frac{\nu-1}{\nu}} \leq C m(\partial\Omega),\quad\forall\Omega\subset\mathcal{V}\text{ with } m(\Omega)<\infty.$$
In the following section we will need a good bound on the constant in the discrete Sobolev-Neumann inequality. Suppose $(\mathcal{G},m)$ is a finite weighted graph with $|\mathcal{V}|=N$. The *spectral gap* $\lambda(\mathcal{G},m)$ is defined by $$\label{eq:spec-gap}
\lambda(\mathcal{G},m)=\inf\left\{\frac{\sum_{\{i,j\}\in\mathcal{E}}|f(i)-f(j)|^2 m(i,j)}{\sum_{i\in\mathcal{V}}|f(i)-m(f)|^2 m(i)}: f\in{\mathbb{R}}^N \setminus\{0\} \right\}.$$ Clearly, if $S_d$ is the minimum constant so that the discrete Sobolev-Neumann $L^2$ inequality of order $\infty$ holds, then $S_d= 1/\lambda(\mathcal{G},m)$.
The *Cheeger constant* $h(\mathcal{G},m)$ is defined by $$h(\mathcal{G},m)=\inf\left\{\frac{\sum_{\{i,j\}\in\mathcal{E}}|f(i)-f(j)| m(i,j)}{\inf_c \sum_{i\in\mathcal{V}}|f(i)-c| m(i)}: f\in{\mathbb{R}}^N \right\}.$$ It is well-known (cf. [@Sal-Cos1]) that the Cheeger constant is equivalent to the isoperimetric inequality $$\label{eq:disc-isop-Neu}
h(\mathcal{G},m)=\inf\left\{\frac{m(\partial U)}{m(U)}: U\subset\mathcal{V}, 0<m(U)\leq\frac{1}{2}m(\mathcal{V})\right\}.$$ And we can compare the constants $h(\mathcal{G},m)$ and $\lambda(\mathcal{G},m)$ by $$\label{eq:spect-Cheeger}
\frac{h^2(\mathcal{G},m)}{8m_0}\leq\lambda(\mathcal{G},m)\leq h(\mathcal{G},m),$$ where $$m_0 :=\underset{i\in\mathcal{V}}{\max}\left\{\frac{1}{m(i)}\sum_{j\in\mathcal{V}}m(i,j)\right\}.$$
We state the Euclidean scale-invariant Sobolev-Poincaré inequalities. They will be used in the proof of Theorem \[thm:Sobolev\] and also in the proof of the scale invariant Poincaré inequality on $(X,g)$. Let $B_r =B(o,r)\subset{\mathbb{R}}^n$ be the ball of radius $r>0$. For $1\leq p\leq n$ and $n\leq\nu\leq\infty$ there exists a constant $S_{p,\nu}$, only depending on $p,\nu$, so that $$\label{eq:Sobolev-loc}
\left(\int_{B_r}|f-f_r|^{\frac{p\nu}{\nu-p}} d\mu\right)^{\frac{\nu-p}{p\nu}} \leq S_{p,\nu} r^{1-\frac{n}{\nu}}\left(\int_{B_r}|\nabla f|^p d\mu\right)^{\frac{1}{p}},\quad\forall f\in C^{\infty}(B_r).$$
We may assume that the end of $(X,g)$ has the conical metric. That is, if $\rho$ is a radius function, then there is an $r_0 \geq 2$ so that $\phi: X\setminus K_{r_0}\rightarrow C(S)\setminus{\bar{D}}_{r_0}$, with $K_{r_0} =\{x\in X: \rho(x)\leq r_0 \}$ and ${\bar{D}}_{r_0} =\{x\in C(S): r(x)\leq r_0 \}$, is an isometry.
\[lem:Sobolev-anal\] Fix $R \geq r_0 ,\kappa >1$ and consider the annulus $A=A(R,\kappa R)= D_{\kappa R} \setminus D_{R}$. Then if we let $A_\delta$ be the $\delta R$-neighborhood of $A$, with $0<\delta<1$ sufficiently small, there is a constant $C>0$ independent of $R$ so that $$\label{eq:Sobolev-anal}
\left(\int_A |f-f_A|^{\frac{n}{n-1}} d\mu\right)^{\frac{n-1}{n}} \leq C \int_{A_\delta}|\nabla f|d\mu,\quad\forall f\in C^{\infty}(A_\delta).$$
Set $s=\delta R$, and let $\{x_i\}_{i\in I}$ be a maximal subset of $A$ such that the distance between any two of its elements is at least $s$. Set $V_i =B(x_i ,s)$ and $V_i^* =V_i^{\#}=B(x_i,3s)$. Then $(V_i,V_i^*,V_i^{\#})$ is a finite good covering of $A$ in $A_\delta$. Conditions (\[defn:gc-iii\]) and (\[defn:gc-v\]) are satisfied with $Q_1 ={\operatorname{Card}}(I)$ and $Q_2 =\frac{\max(\mu(V_i^*) :i\in I)}{\min(\mu(V_i):i\in I)}$. And in (\[defn:gc-iv\]) we may take $k(i,j)=i$.
We will apply Theorem \[thm:cov-Sobolev-Neu\]. The exponential map gives coordinates $\phi_i :B_{s} \rightarrow V_i$ and $\phi_i^* :B_{3s}\rightarrow V_i^*$. Then if $g_0$ is the flat metric on $B_{3s}$, by finiteness there is a $c>0$ so that $c^{-1}g_0 \leq (\phi_i^*)^* g\leq cg_0$ for all $i\in I$.
We will need the following which is easy to prove using the Hölder inequality. Let $(U,\lambda)$ be a finite measure space, then $$\label{eq:aver-min}
\|f-f_{U,\lambda}\|_{L^q(U,\lambda)} \leq 2\underset{c\in{\mathbb{R}}}{\inf}\|f-c\|_{L^q(U,\lambda)}.$$ Then (\[eq:Sobolev-loc\]) with $p=1$ and $\nu=n$, (\[eq:aver-min\]), and the uniform bound on $g$ imply that for $V=V_i$ or $V_i^*$ $$\left(\int_V |f-f_{V,\mu}|^{\frac{n}{n-1}} d\mu\right)^{\frac{n-1}{n}} \leq 2 S_{1,n} c^n \int_V |\nabla f|d\nu.$$ Thus the covering $(V_i,V_i^*,V_i^{\#})$ satisfies the continuous $L^1$ Sobolev inequality of order $\nu=n$. It remains to show the discrete $L^1$ Sobolev-Neumann inequality. But this follows because $\mathcal{G}$ is a finite connected graph, and any two norms on a finite dimensional vector space are equivalent. This proves (\[eq:Sobolev-anal\]). Then one can see that the same $C$ can be used in (\[eq:Sobolev-anal\]) for any $R\geq r_0$ by considering the Euler action $\psi_a :C(S)\rightarrow C(S), a>0,$ which acts by homotheties. We have $\psi_a :A(R,\kappa R)\rightarrow A(aR,a\kappa R)=A(R',\kappa R')$. And if $A'_\delta$ denotes the $\delta R'$-neighborhood of $A(R',\kappa R')$, then $\psi_a :A_\delta \rightarrow A'_\delta$ is a homothetic diffeomorphism. Since (\[eq:Sobolev-anal\]) is invariant under homotheties, the lemma follows.
Let $R>0$ and $\kappa>1$ be as above and define $R_i =\kappa^i R$. We define a good covering of $(X,g)$. Define $A_i =A(R_{i-1},R_i)$ for $i\geq 1$, and $A_0 =D_R =\{x\in X: \rho(x)<R\}$. We let $A^*_i$ be the union of all the $A_j$ whose closure intersects ${\bar{A}}_i$. And similarly, $A^{\#}_i$ is the union of all the $A_j$ whose closure intersects ${\bar{A}}^*_i$.
Lemma \[lem:Sobolev-anal\] shows that the continuous $L^1$ Sobolev inequality of order $n$ holds for the pairs $(A_i,A^*_i)$ with $i\geq 1$, and for $(A^*_i,A^{\#}_i)$ with $i\geq 2$. Since $A_0 =D_R$ is pre-compact one can show that (\[eq:Sobolev-anal\]) holds with $A=A_0$ using the same argument in Lemma \[lem:Sobolev-anal\]. Thus the continuous $L^1$ Sobolev inequality holds for $(A_0,A^*_0)$, and the remaining cases are identical.
Proposition \[prop:disc-isop\] shows that the discrete $L^1$ Sobolev-Dirichlet inequality, of order $\infty$, holds if $$\label{eq:disc-isop}
\frac{m(\Omega)}{m(\partial\Omega)}\leq C,\quad\forall\Omega\subset\mathcal{V}\text{ with } m(\Omega)<\infty.$$ Suppose $\Omega\subset\{0,\ldots,j\}$ with $j\in\Omega$. Then $$\begin{split}
m(\Omega) & \leq\sum_{i=0}^j m(i) \\
& =\mu(A_0) +\sum_{i=1}^j \mu(A(\kappa^{i-1}R,\kappa^i R))\\
& =\mu(A_0)+ \mu(A_1)\sum_{i=1}^j \kappa^{n(i-1)}\\
& =\mu(A_0) +\mu(A_1)\frac{\kappa^{nj} -1}{\kappa^n -1}
\end{split}$$ And $$\mu(\partial\Omega)\geq m(j+1)=\mu(A_1)\kappa^{nj}.$$ The isoperimetric inequality (\[eq:disc-isop\]) holds with $C=\frac{\mu(A_0)}{\mu(A_1)} +\frac{1}{\kappa^n -1}$. We then apply Theorem \[thm:cov-Sobolev-Dir\], and the prove is complete.
Let $V(x,r)$ be the volume of the geodesic ball $B(x,r)$ of radius $r$ centered at $x\in X$. It is well known (cf. [@Sal-Cos2]) that (\[eq:Sobolev\]) implies the following volume growth condition.
\[cor:Sobolev-vol\] There is a constant $c>0$, depending only on $n$ and $C$ in (\[eq:Sobolev\]), so that $V(x,r)\geq cr^{n}$.
\[rem:Sob\] The above arguments can easily be adapted to prove the Sobolev inequality (\[eq:Sobolev\]) on an ${\operatorname{AC}}$ manifold with multiple, but finitely many, ends.
Gaussian bound on the heat kernel
---------------------------------
Recall that the *heat kernel* $h(t,x,y)$ is a smooth function on ${\mathbb{R}}_{>0}\times X\times X$ symmetric in $x,y$ which is the fundamental solution to the heat diffusion equation, $(\partial_t +\Delta_x)h=0$, with $\underset{t\rightarrow 0^+}{\lim}h(t,x,y)=\delta_x$. In this section we prove that $h(t,x,y)$ satisfies a Gaussian bound. First we need a definition.
\[defn:Poin\] Let $U\subset U'$ be relatively compact open subsets of $(X,g)$. The *Poincaré constant* of the pair $(U,U')$ is the smallest positive number $\Lambda(U,U')$ such that $$\int_U |f-f_U|^2 d\mu\leq\Lambda(U,U')\int_{U'} |\nabla f|^2 d\mu, \quad\forall f\in C^{\infty}(U').$$ We say that $(X,g)$ satisfies a *Poincaré inequality* with *parameter* $0<\delta\leq 1$ if there is a constant $C_P >0$ so that for any ball $B(x,r)$ $$\label{eq:Poin}
\Lambda(B(x,\delta r),B(x,r))\leq C_P r^2.$$
\[thm:para-Harn\] For a complete manifold $(X,g)$ the following are equivalent.
$(X,g)$ satisfies the *volume doubling* condition. That is,\[thm:item:vd-Poin\] there exists a constant $C_D >0$ so that for any ball $B(x,r)$ $$\label{eq:vol-doub}
V(x,2r)\leq C_D V(x,r).$$ And $(X,g)$ satisfies a Poincaré inequality with parameter $0<\delta\leq 1$.
The *heat kernel* $h(t,x,y)$ of $(X,g)$ satisfies the two-sided Gaussian bound\[thm:item:Gauss\] $$\label{eq:Gauss}
\frac{c_1 \exp(-C_1 d^2(x,y)/t)}{V(x,\sqrt{t})}\leq h(t,x,y)\leq\frac{C_2 \exp(-c_2 d^2(x,y)/ t)}{V(x,\sqrt{t})}.$$
We also have the following time derivative estimates on $h(t,x,y)$ if the equivalent conditions in the theorem hold. For any integer $k$, $$|\partial_t^k h(t,x,y)|\leq \frac{A_k \exp(-c_2 d^2(x,y)/t)}{t^k V(x,\sqrt{t})},$$ and there is an $\epsilon>0$ so that $$\label{eq:Gaus-der}
|\partial_t^k (t,x,y)-\partial_t^k h(t,x,z)|\leq \frac{d(y,z)^\epsilon \exp(-c_2 d(x,y)/t)}{t^{k+\frac{\epsilon}{2}} V(x,\sqrt{t})},$$ for $d(y,z)\leq\sqrt{t}$.
It is known [@Jer] that if a Poincaré inequality (\[eq:Poin\]) holds for $0<\delta<1$ then it holds for all $\delta\in (0,1]$.
A. Grigor’yan [@Gri] and L. Soloff-Coste [@Sal-Cos] proved that (\[thm:item:vd-Poin\]) is equivalent to a parabolic Harnack inequality. The equivalence of (\[thm:item:Gauss\]) and the parabolic Harnack inequality goes back to [@FabStro]. See also [@Sal-Cos2] for proofs of both of these equivalences.
\[rem:vd-Poin-quasi\] It is not difficult to check that both the volume doubling condition and the existence of a Poincaré inequality in Theorem \[thm:para-Harn\].\[thm:item:vd-Poin\] are invariant under quasi-isometry. This will be used to simplify the proofs below.
The main result of this section is the following.
\[thm:Poin-con\] Suppose $(X,g)$ is ${\operatorname{AC}}$ or merely quasi-isometric to an ${\operatorname{AC}}$ metric. Then the equivalent conditions of Theorem \[thm:para-Harn\] hold on $(X,g)$.
We will prove Theorem \[thm:Poin-con\] by showing that $(X,g)$ satisfies the volume doubling (\[eq:vol-doub\]) and the Poincaré inequality (\[eq:Poin\]). We will need some definitions for the proof of Theorem \[thm:Poin-con\].
Fix $o\in X$ and a parameter $0<\epsilon\leq 1$ (the *remote parameter*).
We say that a ball $B(x,r)$ is *remote* if $r\leq\epsilon\frac{1}{2}d(o,x)$.
We say that a ball $B(o,r)$ is *anchored*.
\[lem:v-d\] Let $(X,g)$ be ${\operatorname{AC}}$. Then $(X,g)$ satisfies the volume doubling condition (\[eq:vol-doub\]).
Per Remark \[rem:vd-Poin-quasi\] we may assume $(X,g)$ is conical outside a compact set. We may assume that $\rho$ is a radius function and there is an $r_0 \geq 2$ so that $\phi: X\setminus D_{r_0}\rightarrow C(S)\setminus K_{r_0}$, with $D_{r_0} =\{x\in X: \rho(x)\leq r_0 \}$ and $K_{r_0} =\{x\in C(S): r(x)\leq r_0 \}$, is an isometry.
We first prove that volume doubling is satisfied for remote balls. Choose $\delta<{\operatorname{inj}}(X,g)$ and $R>r_0$ so that $R-\delta >r_0$. Let $\{x_i \}_{i\in I}$ be a maximal set of points in ${\bar{D}}_R$ such that the distance between any two is at least $\delta$. For each $i\in I$ let $\mathcal{B}_i \subset T B(x_i,\delta)$ be the radius $\delta$ disk bundle. And fix an isomorphism $\beta_i :B_\delta \times B(x_i,\delta)\cong\mathcal{B}_i$ linear and preserving distances on the fibers. Then define maps $\psi_i : B_\delta \times B(x_i,\delta) \rightarrow X$ by $\psi_i (w,x)=\exp_x (\beta_i(w,x))$. Let $g_i$ be the restriction of of $\psi_i^*g$ to the fibers on $B_\delta \times B(x_i,\delta)$. By an easy compactness argument it is easy to see that if $g_0$ is the flat metric on $B_\delta$, then $c^{-1}g_0 \leq g_i \leq cg_0$ for $c>1$ uniformly in $B(x_i,\delta)$ for each $i\in I$. If $x\in{\bar{D}}_R$, then the above arguments show that $g$ in the chart $\exp_x :B_\delta \rightarrow B(x,\delta)$ satisfies $c^{-1}g_0 \leq g \leq cg_0$. Then one can show that the volume doubling condition holds for balls $B(x,r)$ with $x\in{\bar{D}}_R$ and $r<\frac{1}{2}\delta$, i.e. there exists a $C>0$ so that $V(x,2r)\leq C V(x,r)$.
Now let $x\in X$ with $\rho(x)>R$. Then by applying the Euler action $\psi_a :C(S)\rightarrow C(S)$ to the above charts, we have the chart $\exp_x :B_{\frac{\rho(x)}{R}\delta} \rightarrow B(x,\frac{\rho(x)}{R}\delta)$ in which $c^{-1}g_0 \leq g \leq cg_0$. Thus we have volume doubling for balls $B(x,r)$ with $r<\frac{\rho(x)}{2R}\delta$. Since $\rho\sim d(o,-)$, volume doubling holds for remote balls if we chose a small enough remote parameter $\epsilon>0$.
There is a constant $C_o >0$ so that $V(o,r)\leq C_o r^n$. And by Corollary \[cor:Sobolev\] there is a constant $C>0$ so that $V(x,r)\geq Cr^n$ for any $x\in X$.
Choose $\epsilon$ to be the remote parameter from above. Set $d(o,x)=\ell$. We consider three cases.
*Case 1:* If $r\leq\frac{1}{2}\epsilon\ell$, then the ball $B(x,r)$ is remote and volume doubling holds.
*Case 2:* If $r\geq\frac{3}{2} \ell$, then we have $$V(x,2r)\leq V(o,\frac{8}{3}r)\leq C_o \frac{8^n}{3^n}r^n \leq\frac{8^n C_o}{3^n C} V(x,r).$$
*Case 3:* If $\frac{1}{2}\epsilon\ell\leq r\leq\frac{3}{2} \ell$, then $$V(x,2r)\leq V(o,8\ell)\leq C_o 8^n \ell^n \leq \frac{C_o}{C}\frac{16^n}{\epsilon^n}V(x,r).$$
The proof of the following is straight forward.
\[lem:rem-anch\] Given $o\in X$ suppose the Poincaré inequality holds for all remote and anchored balls with parameter $0<\delta_0 \leq 1$ and constant $C_P >0$. That is, for any remote or anchored ball $B(x,r)$ $$\Lambda(B(x,\delta_0 r),B(x,r))\leq C_P r^2.$$ Then the Poincaré inequality holds for any ball with parameter $\delta=\epsilon\delta^2_0 /8$ and constant $C_P>0$. That is, for any ball $B(x,r)$ $$\Lambda(B(x,\delta r),B(x,r))\leq C_P r^2.$$
We will use the same notation used in the proof of Lemma \[lem:v-d\]. From the proof of Lemma \[lem:v-d\] we have a remote parameter $\epsilon>0$ so that for $r\leq\epsilon\frac{1}{2}d(o,x)$ in the chart $\exp_x :B_\delta \rightarrow B(x,r)$ the metric $g$ satisfies $c^{-1}g_0 \leq g \leq cg_0$. Then by the local Poincaré inequality, (\[eq:Sobolev-loc\]) with $p=2, \nu=\infty$, there is a $C>0$ so that $$\int_{B(x,r)}|f-f_r|^{2} d\mu \leq\int_{B(x,r)}|f-f_{B_r ,dx}|^{2} d\mu \leq Cr^2 \int_{B(x,r)}|\nabla f|^2 d\mu, \quad\forall f\in C^{\infty}(B(x,r)),$$ for any remote ball $B(x,r)$, where $f_r$ is the average with respect to the $g$ volume $d\mu_g$.
Thus by Lemma \[lem:rem-anch\] it remains to prove the Poincaré inequality for anchored balls. As in the proof of Lemma \[lem:v-d\] we may assume the end $X\setminus{\bar{D}}_{r_0}$ of $(X,g)$ is conical. Fix $R>r_0$ and $\kappa>1$. And choose $\delta>0$ so that $s=\delta R<\frac{1}{3}{\operatorname{inj}}(g)$. Let $A=A(R,\kappa R)$ and $A_\delta$ the $\delta R$-neighborhood of $A$. Let $\{x_i \}_{i\in I}$ be a maximal subset of $A$ so that the distance between any two elements is at least $s$. Let $V_i =B(x_i ,s)$ and $V_i^* =V_i^{\#} =B(x_i ,3s)$. Then $(V_i,V_i^*,V_i^{\#})_{i\in I}$ is a finite good cover of $A$ in $A_\delta$. By uniformly bounding $g$ as above, there is a constant $S_c$ so that $\Lambda(V_i,V_i^*)\leq S_c$ and $\Lambda(V_i^*,V_i^{\#})\leq S_c$. In other words, the covering satisfies the continuous $L^2$ Sobolev inequality of order $\nu=\infty$. The associated graph $(\mathcal{G},m)$ if finite and connected, thus there is a $S_d >0$ so that the discrete $L^2$ Sobolev-Neumann inequality of order $\infty$ holds. Theorem \[thm:cov-Sobolev-Neu\] gives a constant $S>0$ so that $\Lambda(A,A_\delta)\leq S$. By considering the homothetic action $\psi_a :A(R,\kappa R)_\delta \rightarrow A(R',\kappa R')_\delta$ with $R'=aR$ we see there is a constant $C>0$ independent of $R$ so that $$\label{eq:cont-Poin}
\Lambda(A,A_\delta)\leq CR^2, \quad\forall R>r_0.$$
Let $R>r_0$ and $\kappa>1$ be as above. Choose $\delta>0$ so that $R-r_0 >\delta R$, and set $R_i =\kappa^i R$. By increasing $r_0$ if necessary, we may assume there is an $r_1 >0$ so that $D_{R_0}\subset B(o,r_1)\subset D_{R_1}$. We define a covering with the following sets $$\begin{gathered}
A_0 =D_{R_0} \text{ and } A_i =A(R_{i-1},R_i)\text{ for }i\geq 1,\\
A^*_i =A_{i-1}\cup A_i \cup A_{i+1},\\
A^{\#}_i =A^*_{i-1}\cup A^*_i \cup A^*_{i+1},\\\end{gathered}$$ where we assume $A_i =\emptyset$ for $i<0$. Choose the least $\ell\geq 1$ so that $B(o,r)\subset D_{R_{\ell}}$. Then $\mathcal{A}=\{(A_i,A^*_i,A^{\#}_i )\}_{i=0}^\ell$ is a good covering of $B(o,r)$ in $D_{R_{\ell +2}}\subset B(o,\kappa^4 r)$. If we set ${\bar{r}}=R_\ell$, then from (\[eq:cont-Poin\]) we have $\Lambda(A_i,A^*_i)\leq C{\bar{r}}^2$ for $i\geq 1$ and $\Lambda(A^*_i,A^{\#}_i)\leq C{\bar{r}}^2$ for $i\geq 2$.
It remains to prove the discrete Sobolev-Neumann inequality. We will show that there is a constant $c>0$ so that the spectral gap (\[eq:spec-gap\]) satisfies $c<\lambda(\mathcal{G},m)$. And by (\[eq:disc-isop-Neu\]) and (\[eq:spect-Cheeger\]) it suffices to show there is a $C>0$ independent of $\ell$ such that $$m(U)\leq Cm(\partial U),\quad\forall U\subset\mathcal{V}\text{ such that }m(U)\leq\frac{1}{2}m(\mathcal{V}).$$ Let $j:=\max\{i\in\mathcal{V}: (i,i+1)\in\partial(U)\}$. Then $$\begin{split}
m([0,j]) & =\sum_{i=0}^j m(i) \\
& =\mu(A_0) +\sum_{i=1}^j \mu(A(\kappa^{i-1}R,\kappa^i R))\\
& =\mu(A_0)+ \mu(A_1)\sum_{i=1}^j \kappa^{n(i-1)}\\
& =\mu(A_0) +\mu(A_1)\frac{\kappa^{nj} -1}{\kappa^n -1}
\end{split}$$ Set $C=\frac{\mu(A_0)}{\mu(A_1)} +\frac{1}{\kappa^n -1}$. Since either $U$ or $\mathcal{V}\setminus U$ is contained in $[0,j]$, $$\begin{gathered}
m(U)\leq\min\{m(U),m(\mathcal{V}\setminus U)\}\leq m([0,j]) \\
\leq C\mu(A_1)\kappa^{nj}=Cm(j+1)\leq Cm(\partial U).\end{gathered}$$ By Theorem \[thm:cov-Sobolev-Neu\] we have $\Lambda(B(o,r),D_{R_\ell})\leq C{\bar{r}}^2$, for $r\geq r_1$ where $C$ is independent of ${\bar{r}}$. From (\[eq:loc-Poin\]) there is a $C_1>0$ so that $\Lambda(B(o,r),B(o,r))\leq C_1 r^2$ for $r\leq r_1$. The proof is completed by observing that $d(o,-)\sim\rho$ on $X\setminus D_{R_0}$.
Unlike Theorem \[thm:Sobolev\], Theorem \[thm:Poin-con\] does not generalize to ${\operatorname{AC}}$ manifolds with more than one end. For example, it is known that the Poincaré inequality does not hold on a connected sum of Euclidean spaces ${\mathbb{R}}^n \#{\mathbb{R}}^n$. See [@GriSal-Cos] for more information on this.
Laplacian on Asymptotically conical manifolds {#subsect:Lap}
---------------------------------------------
We will need some properties of the Laplacian on an asymptotically conical manifold. These will follow from some good bounds on the Green’s function which follow from Theorem \[thm:para-Harn\]. First we state an elementary lemma whose proof is an easy exercise .
\[lem:lap\] Let $(X,g)$ be ${\operatorname{AC}}(\delta,k+\alpha), k\geq 1, 0<\alpha<1.$ Suppose $u\in C^2_{\beta}(X)$ and $v\in C^2_{\gamma}(X)$ where $\beta,\gamma\in{\mathbb{R}}$ with $\beta+\gamma <2-n$. Then $$\int_X u\Delta v d\mu =\int_X v\Delta u d\mu.$$ If $\rho$ is a radius function of $(X,g)$, then $\Delta(\rho^{2-n}) \in C^{k-1,\alpha}_{\delta-n}(X)$. And if $\Omega:={\operatorname{Vol}}(S,g_S)$, where $S$ is the link in the conical end, then $$\int_X \Delta(\rho^{2-n}) d\mu =(n-2)\Omega.$$
Recall, the Green’s function satisfies $$\Delta_y G(x,y)=\delta_x(y), \quad\forall x\in X.$$ Equivalently $G(x,y)$ satisfies $$\int_X G(x,y)\Delta f(y)d\mu(y) =f(x),$$ and $$\Delta_x \int_X G(x,y)f(y)d\mu(y)=f(x),$$ for any smooth compactly supported function $f$ on $X$.
If $(X,g)$ satisfies Theorem \[thm:para-Harn\], then $(X,g)$ admits a positive symmetric Green’s function if and only if $\int^\infty V(x,\sqrt{t})^{-1}dt <\infty$. And in this case $$G(x,y) =\int_0^\infty h(t,x,y) dt.$$
If $n>2$, then from Corollary \[cor:Sobolev\] and integrating (\[eq:Gauss\]) we have for some $C>0$ depending only on $g$ $$\label{eq:Green-bound}
\begin{split}
0< G(x,y) & \leq c\int_{d(x,y)^2}^\infty \frac{dt}{V(x,\sqrt{t})} \\
& \leq Cd(x,y)^{2-n},
\end{split}$$ for all $x,y\in X$, where the second inequality uses Corollary \[cor:Sobolev\].
Similarly, by integrating (\[eq:Gaus-der\]) for $x\neq y$ and $d(y,z)\leq d(x,y)/2$ we have $$\label{eq:Green-bound-dif}
\begin{split}
\frac{|G(x,y)-G(x,z)|}{d(y,z)^\epsilon} & \leq c \int_{d(x,y)^2}^\infty \frac{dt}{t^\frac{\epsilon}{2} V(x,\sqrt{t})} \\
& \leq C d(x,y)^{2-n-\epsilon},
\end{split}$$ where again we have used Corollary \[cor:Sobolev\] in the second inequality.
If $f\in C^{k,\alpha}_{\beta}(X)$ with $k\geq 0,\alpha\in (0,1)$ and $\beta<-2$, then (\[eq:Green-bound\]) and standard regularity arguments show that $u(x)=\int_X G(x,y)f(y)d\mu(y)$ is locally in $C^{k+2,\alpha}(X)$ and $\Delta u=f$. We will extend these arguments to prove the following.
\[thm:lap-weight\] Suppose $(X,g)$ is ${\operatorname{AC}}(\delta,\ell+\alpha),\delta<-\epsilon, \ell\geq 0, \alpha\in (0,1)$ of dimension $n>2$. Let $k\leq\ell$, then we have the following.
Suppose $-n<\beta<-2$. There exists a $C>0$ so that for each $f\in C^{k,\alpha}_\beta (X)$ there is a unique $u\in C^{k+2,\alpha}_{\beta+2}(X)$ with $\Delta u=f$ which satisfies $\|u\|_{C^{k+2,\alpha}_{\beta+2}}\leq C\|f\|_{C^{k,\alpha}_\beta}$.
Suppose $-n-\epsilon <\beta<-n$. There exist $C_1 ,C_2 >0$ such that for each $f\in C^{k,\alpha}_\beta (X)$ there is a unique $u\in C^{k+2,\alpha}_{2-n}(X)$ with $\Delta u=f$. Furthermore, if we define $$\label{eq:lap-int}
A=\frac{1}{(n-2)\Omega}\int_X f d\mu,$$ where $\Omega={\operatorname{Vol}}(S)$, then $u=A\rho^{2-n} +v$ with $v\in C^{k+2,\alpha}_{\beta +2}(X)$ satisfying $|A|\leq C_1 \|f\|_{C^0_\beta}$ and $\|v\|_{C^{k+2,\alpha}_{\beta +2}}\leq C_2 \|f\|_{C^{k,\alpha}_\beta}$.
We define $$\label{eq:lap-sol}
u(y)=\int_X G(y,x)f(x)d\mu(x),$$ and we first prove that $u\in C^0_{\beta +2}$. We have $|f(x)|\leq \|f\|_{C^0_{\beta}}\rho(x)^\beta$, and from (\[eq:lap-sol\]) and (\[eq:Green-bound\]) we have $$|u(y)|\leq C \|f\|_{C^0_{\beta}} \int_X d(y,x)^{2-n} \rho(x)^\beta d\mu(x).$$ Let $o\in X$ be a fixed point. We split the integral into three regions $R_1 =\{x\in X: 4d(o,x)\leq d(o,y)\}$, $R_2 = \{x\in X: \frac{1}{4}d(o,y)< d(o,x)<2d(o,y)\}$, and $R_3 =\{x\in X: d(o,x)> 2d(o,y)\}$. Estimating the integral over the three regions gives the following: $$\begin{split}
\int_{R_1} d(y,x)^{2-n} \rho(x)^\beta d\mu(x) & \leq
\begin{cases}
C'\rho(y)^{\beta+2} & \text{if } \beta\in(-n,-2) \\
C'\rho(y)^{2-n} & \text{if } \beta <-n
\end{cases},\\
\int_{R_2} d(y,x)^{2-n} \rho(x)^\beta d\mu(x) & \leq\rho(y)^{\beta+2}, \text{ and }\\
\int_{R_3} d(y,x)^{2-n} \rho(x)^\beta d\mu(x) & \leq\rho(y)^{\beta+2}
\end{split}$$ This proves that $u\in C^0_{\beta +2}(X)$ in part (i).
We now consider part (ii). If $\Delta u=f$ with $u\in C_{\beta+2}^{k+2,\alpha}(X)$ and $\beta\in (-n-\epsilon,-n)$, then $$\int_X f d\mu = \int_X \Delta u d\mu =0$$ by the arguments in Lemma \[lem:lap\]. Thus for $f\in C_{\beta}^{k,\alpha}(X)$ there exists $u\in C_{\beta+2}^{k+2,\alpha}(X)$ solving $\Delta u=f$ only if $\int_X f d\mu =0$. So suppose that $\int_X f d\mu =0$. And we replace (\[eq:lap-sol\]) with $$\label{eq:lap-sol-mod}
u(y)=\int_X [G(y,x)-G(y,o)]f(x)d\mu(x).$$ Since $\int_X f d\mu =0$, this integral is equal to (\[eq:lap-sol\]) and thus solves $\Delta u=f$. From (\[eq:Green-bound-dif\]) we have $$|u(y)|\leq C \|f\|_{C^0_{\beta}} \int_X d(x,y)^{2-n-\epsilon} d(o,x)^\epsilon \rho(x)^\beta d\mu.$$ And it is not difficult to show that $$\int_{R_1} d(x,y)^{2-n-\epsilon} d(o,x)^\epsilon \rho(x)^\beta d\mu(x) \leq
\begin{cases}
C'\rho(y)^{\beta+2} & \text{if } \beta\in(-n-\epsilon,-2) \\
C'\rho(y)^{2-n-\epsilon} & \text{if } \beta \leq -n-\epsilon
\end{cases}$$ This shows that $u\in C^0_{\beta+2}(X)$ if $\beta\in (-n-\epsilon,-n)$.
In both parts (i) and (ii) we have proved that $\|u\|_{C^0_{\beta+2}}\leq C'\|f\|_{C^0_{\beta}}$. As remarked above, we have that $u$ is locally in $C^{k+2,\alpha}(X)$. By taking an appropriate covering of $(X,g)$ using the conical structure and applying the Schauder interior estimates one can show that there is a $C>0$ so that for $\Delta u =f$ one has $$\|u\|_{C_{\beta+2}^{k+2,\alpha}} \leq C\left(\|u\|_{C^0_{\beta+2}} +\|f\|_{C^{k,\alpha}_{\beta}} \right).$$ Thus there is a $C>0$ so that $\|u\|_{C^{k+2,\alpha}_{\beta+2}}\leq C\|f\|_{C^{k,\alpha}_{\beta}}$.
We complete the proof of part (ii). Define $A$ by (\[eq:lap-int\]). Then by Lemma \[lem:lap\] we have $\int_X [f-A\Delta(\rho^{2-n})] d\mu =0$. And also by Lemma \[lem:lap\] we have $f-A\Delta(\rho^{2-n})\in C_{\beta}^{k,\alpha}(X)$. Thus by what we have already proved there is a $v\in C^{k+2,\alpha}_{\beta+2}(X)$ with $\Delta v= f-\Delta(A\rho^{2-n})$ with $$\label{eq:lap-bound}
\|v\|_{C^{k+2,\alpha}_{\beta+2}} \leq C\left(\|f\|_{C^{k,\alpha}_\beta} +|A|\|\Delta(\rho^{2-n})\|_{C^{k,\alpha}_\beta}\right).$$ Note that $$\begin{split}
|A|\leq \frac{1}{(n-2)\Omega} \int_X |f|\, d\mu & \leq \frac{\|f\|_{C^0_\beta}}{(n-2)\Omega} \int_X \rho^\beta \, d\mu \\
& \leq C_1 \|f\|_{C^0_\beta},
\end{split}$$ where $C_1 =\frac{1}{(n-2)\Omega}\int_X \rho^\beta \, d\mu$ is finite since $\beta\leq -n$. And this combined with (\[eq:lap-bound\]) completes the proof.
Kähler case
===========
Asymptotically conical Kähler manifolds will now be considered. We will begin with some definitions and preliminary results. In particular, the link $S$ in a Kähler cone is far from arbitrary. It is a *Sasaki* manifold which can be thought of as an odd dimensional analogue of a Kähler manifold. We also consider some Hodge theory which will be useful late.
Background
----------
\[defn:Sasaki\] A $2m-1$-dimensional Riemannian manifold $(S,g)$ is *Sasaki* if the metric cone $(C(S),{\bar{g}})$, $C(S)={\mathbb{R}}_{>0} \times S,\ {\bar{g}}=dr^2 +r^2 g,$ is Kähler.
This is the succinct definition of a Sasaki manifold. They were originally defined as a manifold carrying a special type of metric contact structure. For more on Sasaki manifolds see the monograph [@BG3].
It follows from the definition that the Euler vector field $r\partial_r$ acts holomorphically, i.e. $\mathcal{L}_{r\partial_r} J=0$. It is also not difficult to show that $\xi=Jr\partial_r$ is a Killing vector field which restricts to $S=\{r=1\}\subset C(S)$. Thus $\xi +ir\partial_r$ is a holomorphic vector field on $C(S)$. The restriction of $\xi$ to $S$ is the *Reeb vector field* of $(S,g)$. Sasaki manifolds can be distinguished by the action of the Reeb vector field $\xi$. If $\xi$ generates a free action of $U(1)$ then the Sasaki structure is *regular*. The Sasaki structure is *quasi-regular* if the orbits close but there are non-trivial stablizers. If the orbits do not close, then the Sasaki structure is *irregular*.
Let $\eta$ be the dual 1-form to $\xi$ with respect to $g$, that is $\eta =\frac{1}{r^2}\xi{\,\lrcorner\,}{\bar{g}}$. Then one can check $$\label{eq:contact}
\eta = -J^* \frac{dr}{r} =2d^c \log r,$$ where $d^c =\frac{i}{2}({\bar{\partial}} -\partial)$. The restriction of $\eta$ to $S$ is a contact form with Reeb vector field $\xi$.
Since $\mathcal{L}_{r\partial_r} J=0$, the Kähler form $\omega$ satisfies $$\begin{split}
2\omega & = \mathcal{L}_{r\partial_r}\omega \\
& = d(r\partial_r {\,\lrcorner\,}\omega) \\
& = d(r^2 \eta) \\
& = dd^c(r^2), \quad \text{using (\ref{eq:contact})}.
\end{split}$$ Thus the Kähler form $\omega$ on $C(S)$ has potential $\frac{1}{2}r^2$.
We are in particular interested in Ricci-flat Kähler cones. The following easily follows from the warped product structure of ${\bar{g}}$.
Let $(S,g)$ be a $2m-1$-dimensional Sasaki manifold. Then the following are equivalent.
$(S,g)$ is Sasaki-Einstein with Einstein constant $2n-2$.
$(C(S),{\bar{g}})$ is Ricci-flat Kähler.
Of course, a necessary condition for $(C(S),{\bar{g}})$ to be Ricci-flat Kähler is that $\mathbf{K}_{C(S)}^\ell$ must be trivial for some positive integer $\ell>0$. But for $C(S)$ to admit a Ricci-flat Kähler cone metric with the given Reeb vector field $\xi$ one must require a little more.
\[prop:CY-cond\] A necessary condition for $C(S)$ to admit a Ricci-flat Kähler cone metric with the same complex structure $J$ and Reeb vector field $\xi$ is that $\mathbf{K}_{C(S)}^\ell$, for some integer $\ell\geq 1$, admits a nowhere vanishing section $\Omega$ with $\mathcal{L}_\xi \Omega =im\Omega$.
If the condition in the proposition holds, then $\Omega$ satisfies $$\label{eq:CY-cond}
\frac{i^{m^2}}{2^m}\Omega\wedge{\bar{\Omega}}=\frac{1}{m!} e^h \omega^m ,$$ where $h\in C^\infty(C(S))$ is basic, meaning that $\xi h=r\partial_r h =0$. Note that the Ricci form is given by ${\operatorname{Ric}}(\omega)=dd^c h$ and is zero precisely when $h$ is constant.
*A priori* a Kähler cone $C(S)$ does not contain the vertex. But it can be proved that $C(S)\cup\{o\}$, with the vertex $o$, is an affine variety. See [@vC4] for a proof of the relevant embedding theorem.
We now define ${\operatorname{AC}}$ Kähler manifolds.
\[defn:AC-Kaehler\] Let $(C(S),g_0)$ be a Kähler cone. Then we say that a Kähler manifold $(X,g)$ is asymptotically conical of order $(\delta, k+\alpha)$, asymptotic to $(C(S), g_0)$, if there is compact subset $K\subset X$, a compact neighborhood $o\in K_0 \subset C(S)$, and a diffeomorphism $\phi: X\setminus K \rightarrow C(S)\setminus K_0$ so that $$\begin{aligned}
|\phi_* g -g_0 | \in C_{\delta} ^{k,\alpha} &\quad\text{on } C(S)\setminus K_0 \quad\text{and}\label{eq:AC-K-met}\\
|\phi_* J -J_0 | \in C_{\delta} ^{k,\alpha} &\quad\text{on } C(S)\setminus K_0 .\label{eq:AC-K-J}\end{aligned}$$ We will denote this by Kähler ${\operatorname{AC}}(\delta, k+\alpha)$.
In many cases the end of $(X,g)$ will be holomorphically a cone. In this case $\phi: X\setminus K \rightarrow C(S)\setminus K_0$ is a biholomorphism. And one can show that in this case $\phi$ extends to $\phi:X\rightarrow C(S)\cup\{o\}$, which is therefore a resolution of $C(S)$.
### Hodge theory
We review some Hodge theory that will be needed. In particular, a Hodge decomposition will be needed in a proof of a weighted version of the $\partial{\bar{\partial}}$-lemma.
We define the space of $L^2$ harmonic forms, where we assume $g$ is quasi-isometric to a ${\operatorname{AC}}$ metric, $$L^2 \mathcal{H}^p (X,g):=\{\eta\in L^2(\Lambda^p X) : \Delta\eta =0\}.$$ One can show that $L^2 \mathcal{H}^p (X,g)=\{\eta\in L^2(\Lambda^p X) : d\eta=d^* \eta =0\}$. If $\eta\in L^2 \mathcal{H}^p (X,g)$, then $\eta\in L^2 _k(\Lambda^p X)$ for all $k\geq 0$. So $L^2 \mathcal{H}^p (X,g)$ consists of smooth forms. Furthermore, one can show that it is finite dimensional, and if $\tilde{g}$ is quasi-isometric to $g$, as in (\[eq:quasi-isom\]), there is a natural isomorphism $L^2 \mathcal{H}^p (X,\tilde{g})\cong L^2 \mathcal{H}^p (X,g)$. See [@Loc] for proofs of these statements.
The $L^2$ harmonic spaces have been computed in our context. For the following, note that an ${\operatorname{AC}}$ manifold $X$ can be compactified with boundary $\partial X =S$. And $H^*(X,S)$ is isomorphic to the compactly supported cohomology $H_c^*(X)$.
\[thm:cohom\] Let $(X,g)$ be a manifold of dimension n which is ${\operatorname{AC}}$ up to quasi-isometry. Then we have the natural isomorphisms $$L^2 \mathcal{H}^p (X,g) \cong\begin{cases}
H^p(X,S), & p<n/2, \\
{\operatorname{Im}}\bigl(H^p(X,S)\rightarrow H^p(X)\bigr), & p=n/2, \\
H^p(X), & p>n/2.
\end{cases}$$
Recall the Kodaira decomposition theorem, which for arbitrary manifolds gives the orthogonal decomposition $$\label{eq:Kodaira}
L^2(\Lambda^p(X,g)) = L^2 \mathcal{H}^p(X,g) \oplus\overline{dC_0^\infty(\Lambda^{p-1})}\oplus \overline{d^* C_0^\infty(\Lambda^{p+1})},$$ where the closure in the last two summands is in $L^2$.
We need a more precise decomposition than (\[eq:Kodaira\]). Assume from now on that $(X,g)$ is ${\operatorname{AC}}(\delta,\ell+\alpha)$ with $\delta<0$ and $\ell\geq 2$. A difficulty in improving (\[eq:Kodaira\]) is that the operator $$\label{eq:Lap-forms}
\Delta: L^2_{k+2,\beta} (\Lambda^p X)\longrightarrow L^2_{k,\beta-2} (\Lambda^p X)$$ is not Fredholm for arbitrary $\beta\in{\mathbb{R}}$. The kernel of (\[eq:Lap-forms\]) is finite dimensional and the closure of the range has finite codimension. The difficulty is that the range is not always closed. It is a result of [@LocMcO] that there is a discrete set $\mathcal{D}_\Delta \subset{\mathbb{R}}$ so that (\[eq:Lap-forms\]) is Fredholm precisely when $\beta\in{\mathbb{R}}\setminus\mathcal{D}_\Delta$. But one can define a Banach space $\tilde{L}^2_{k+2,\beta} (\Lambda^p X)$ so that $$\label{eq:Lap-forms-mod}
\Delta: \tilde{L}^2_{k+2,\beta} (\Lambda^p X)\longrightarrow L^2_{k,\beta-2} (\Lambda^p X)$$ is Fredholm, and the range of (\[eq:Lap-forms-mod\]) is the closure of the range of (\[eq:Lap-forms\]).
For each $\tau\geq\beta$ with $\tau\in{\mathbb{R}}\setminus\mathcal{D}_\Delta$ let $\tilde{B}_\tau$ be the closure of $L^2_{k+2,\beta} (\Lambda^p X)$ in $$B_\tau =\{\eta \in L^2_{k+2,\tau} (\Lambda^p X) : \Delta\eta\in L^2_{k,\beta-2} (\Lambda^p X)\}$$ with respect to the norm $$\|\eta\|_{B_\tau} =\|\eta\|_{L^2_{k+2,\tau}} +\|\Delta\eta\|_{L^2_{k,\beta-2}}.$$ When $\tilde{B}_\tau$ is equipped with this norm one can show [@Loc] that $$\Delta: \tilde{B}_\tau \longrightarrow L^2_{k,\beta-2} (\Lambda^p X)$$ is Fredholm with range equal to the closure of the range of (\[eq:Lap-forms\]). One can also show that all the $\tilde{B}_\tau$ are isomorphic Banach spaces. We define $\tilde{L}^2_{k+2,\beta} (\Lambda^p X)$ to be any one of the $\tilde{B}_\tau$. In particular, we have $\tilde{L}^2_{k+2,\beta} (\Lambda^p X)=L^2_{k+2,\beta} (\Lambda^p X)$ with equivalent norms if $\beta\in{\mathbb{R}}\setminus\mathcal{D}_\Delta$. And in general $$\tilde{L}^2_{k+2,\beta} (\Lambda^p X) \subset\underset{\tau>\beta}{\bigcap}L^2_{k+2,\tau} (\Lambda^p X).$$
By our conventions we have $L^2_{0,-m}(\Lambda^p X)=L^2(\Lambda^p X)$ with equal norms. Thus consider $$\label{eq:Lap-forms-L2}
\Delta: \tilde{L}^2_{2,2-m} (\Lambda^p X)\longrightarrow L^2_{0,-m} (\Lambda^p X)=L^2(\Lambda^p X).$$ The cokernel of (\[eq:Lap-forms-L2\]) is $L^2 \mathcal{H}^p (X,g)$, so we have the following decomposition refining (\[eq:Kodaira\]).
\[thm:Hodge\] Suppose $(X,g)$ is ${\operatorname{AC}}(\delta,\ell+\alpha)$ with $\delta<0$ and $\ell\geq 2$. Then we have $$L^2(\Lambda^p X) =L^2 \mathcal{H}^p (X,g)\oplus dd^*\bigl(\tilde{L}^2_{2,2-m} (\Lambda^p X) \bigr)\oplus d^*d\bigl(\tilde{L}^2_{2,2-m} (\Lambda^p X) \bigr).$$ In particular, if $\eta\in L^2(\Lambda^p X)$, then we have the unique $L^2$ decomposition $\eta=\sigma +d\zeta +d^* \theta$, where $\zeta\in L^2_{1,\delta}(\Lambda^{p-1} X)$ for all $\delta>1-m$ and $\theta\in L^2_{1,\delta}(\Lambda^{p+1} X)$ for all $\delta>1-m$.
Of course when $(X,g)$ is Kähler the decomposition in Theorem \[thm:Hodge\] respects the decomposition into types $\Lambda^p(X) \otimes{\mathbb{C}}=\oplus_{r+s=p} \Lambda^{r,s}(X)$ as usual because $\Delta=2\Delta_{{\bar{\partial}}}$.
We now prove a weighted version of the $\partial{\bar{\partial}}$-lemma.
\[prop:dd-bar\] Let $(X,g)$ be Kähler ${\operatorname{AC}}(\delta, \ell +\alpha)$ with $\delta<0,\ \ell\geq 2$, and $H^1(S,{\mathbb{R}})=0$. Suppose $\beta\in(-2m,-m)$ and $\eta\in C_\beta^{k,\alpha}(\Lambda^{1,1}_{{\mathbb{R}}} X)$, with $0\leq k\leq\ell$, is a closed real $(1,1)$-form with $[\eta]=0$ in $H^2(X,{\mathbb{R}})$. Then there exists a unique real function $u\in C_{\beta+2}^{k+2,\alpha}(X)$ with $dd^c u=\eta$.
Recall that if $u$ is a smooth function, then $$-m\, dd^c u\wedge\omega^{m-1} =\Delta u\wedge\omega^{m}.$$ Define $f$ by $-m \eta \wedge\omega^{m-1} =f \omega^{m}$. So $f\in C_{\beta}^{k,\alpha}(X)$. By Theorem \[thm:lap-weight\] there is a $u\in C_{\beta+2}^{k+2,\alpha}(X)$ with $\Delta u=f$. Then $\gamma=\eta- dd^c u$ is an exact 2-form in $C_\beta^{k,\alpha}(\Lambda^{1,1}X)$. And since $\gamma\wedge\omega^{m-1} =-\frac{1}{m}(f-\Delta u)\omega^m =0$, one can show that $\gamma$ satisfies $$\label{eq:dd-bar-bil}
\gamma\wedge\gamma\wedge\omega^{m-2} =\frac{-1}{2m(m-1)}|\gamma|^2\omega^m .$$
Since $\beta<-m$, we have $\gamma\in L^2(\Lambda^{1,1}X)$. Therefore $\gamma=\sigma+d\zeta$ according to Theorem \[thm:Hodge\], with $\sigma\in L^2 \mathcal{H}^{1,1} (X,g)$. Since $H^1(S)=0$, the homomorphism $H^2(X,S)\rightarrow H^2(X)$ is an inclusion. It follows from Theorem \[thm:cohom\] that $L^2 \mathcal{H}^{1,1} (X,g)$ contains no exact forms. Thus $\gamma=d\zeta$, where $\zeta\in L^2_{1,\delta}(\Lambda^1 X)$ with $\delta>1-m$. Since $C_0^\infty(X)$ is dense in $L^2_{1,\delta}(\Lambda^1 X)$, we can choose a sequence $\{\zeta_j\}$ converging to $\zeta$ in $L^2_{1,\delta}(\Lambda^1 X)$. From (\[eq:dd-bar-bil\]) we have $$\begin{split}
0=\int_X d[\zeta_j \wedge\gamma\wedge\omega^{m-2} ] & =\int_X d\zeta_j \wedge\gamma\wedge\omega^{m-2} \\
& \rightarrow \frac{-1}{2m(m-1)}\int_X |\gamma|^2 \,\omega^m,
\quad\text{as }j\rightarrow\infty.
\end{split}$$ The convergence follows because $d\zeta_j \rightarrow\gamma$ in $L^2_{0,\delta-1}(\Lambda^2 X)$ and we may assume $\delta>1-m$ is chosen small enough that $\delta-1+\beta <-2m$. Therefore $\gamma=0$, and $\eta=dd^c u$.
Calabi conjecture
-----------------
On an ${\operatorname{AC}}$ Kähler manifold $(X,g,J)$ we consider the Monge-Ampére equation $$\label{eq:Monge-Amp}
(\omega+dd^c \phi)^m =e^f \omega^m.$$ If $\phi$ is a solution to (\[eq:Monge-Amp\]) and $\omega'=\omega+dd^c \phi$, then the respective Ricci forms satisfy $${\operatorname{Ric}}(\omega')={\operatorname{Ric}}(\omega)-dd^c f.$$ Equation (\[eq:Monge-Amp\]) was solved by S.-T. Yau [@Yau] for a compact Kähler manifold $(M,g,J)$ under the necessary assumption that $\int_M (1-e^f)\, d\mu_g =0$. This solved a conjecture of E. Calabi.
The Calabi conjecture for ${\operatorname{AC}}$ Kähler manifolds was solved independently by S. Bando and R. Kobayashi [@BK2] and G. Tian and S.-T. Yau [@TY2]. D. Joyce [@Joy1] gave a more exacting proof for the ALE case which gave more precise information on the solution. The proof of D. Joyce applies *mutatis mutandis* to this situation.
\[thm:Calabi\] Suppose $(X,g)$ is asymptotically conical Kähler of order $(\delta, j+\alpha)$, where $\delta<-\epsilon$, $0<\alpha<1$ and $3\leq j\leq\infty$. Let $3\leq k\leq j\leq\infty$.
If $\beta\in (-2m,-2)$, then for each $f\in C_{\beta}^{k,\alpha}(X)$ there is a unique $\phi\in C_{\beta+2}^{k+2,\alpha}(X)$ so that $\omega+dd^c \phi$ is a positive $(1,1)$-form and $(\omega+dd^c \phi)^m =e^f \omega^m$ on $X$.
If $\beta\in (-2m-\epsilon,-2m)$, then for each $f\in C_{\beta}^{k,\alpha}(X)$ there is a unique $\phi\in C_{2-2m}^{k+2,\alpha}(X)$ so that $\omega+dd^c \phi$ is a positive $(1,1)$-form and $(\omega+dd^c \phi)^m =e^f \omega^m$ on $X$. Furthermore, we have $\phi= A\rho^{2-2m} +\psi$ where $\psi\in C_{\beta+2}^{k+2,\alpha}(X)$ and $$\label{eq:const-A}
A=\frac{1}{(m-1)\Omega}\int_X (1-e^f)\, d\mu,$$ where $\Omega={\operatorname{Vol}}(S)$, $S=\{r=1\}\subset C(S)$.
Of course, by the local theory of elliptic operators, whenever $f\in C^\infty(X)$ we have $\phi\in C^\infty(X)$. The theorem is written as it is to show the precise global regularity that the proof gives.
Part (i) of Theorem \[thm:Calabi\] is already known and was essentially proved in [@TY2]. See [@Got] for a proof in the context of manifolds with a conical end. The contribution here is part (ii) which gives a sharp estimate on solutions for rapidly decaying $f\in C_{\beta}^{k,\alpha}(X),\ \beta<-2m.$
The proof of Theorem \[thm:Calabi\] goes through as in [@Joy1 §§8.6-8.7]. The proof is by the continuity method, and the essential ingredients are some *a priori* estimates on a solution $\phi$ of (\[eq:Monge-Amp\]). The Sobolev inequality (\[eq:Sobolev\]) is used to prove an *a priori* estimate on $\|\phi\|_{C^0}$. *A priori* estimates on $\|dd^c \phi\|_{C^0}$ and $\|\nabla dd^c \phi\|_{C^0}$ depending only on $\|\phi\|_{C^0}$, $\|f\|_{C^3}$, and $\|R\|_{C^1}$, where $R$ is the curvature, due to T. Aubin [@Aub1] and S.-T. Yau [@Yau] are applied as in [@TY2]. Then Theorem \[thm:lap-weight\] is applied as in [@Joy1] to show $\phi$ is in the appropriate Hölder space.
Ricci-flat metrics
------------------
Our main motivating for proving Theorem \[thm:Calabi\] is the following which is an immediate consequence.
Suppose $(X,g,J)$ is ${\operatorname{AC}}(\delta,\ell+\alpha)$ with $3 \leq\ell\leq\infty$. Let $3\leq k\leq\ell$, and suppose the Ricci form of $(X,g,J)$ satisfies $${\operatorname{Ric}}(\omega)=dd^c f, \quad f\in C^{k,\alpha}_\beta(X),\quad\text{with }\beta<-2.$$ Then there exists a $\phi\in C^{k+2,\alpha}_{\beta+2}(X)$ so that $\omega' =\omega+dd^c \phi$ is Ricci-flat, and the corresponding metric $g'$ converges to $g$ in $C^{k,\alpha}_\gamma (X)$ where $\gamma=\max(\beta, -2m)$.
Merely observe that for a solution to $(\omega+dd^c \phi)^m =e^f \omega^m$ the Ricci forms of the Kähler metrics $\omega$ and $\omega' =\omega+dd^c \phi$ satisfy $${\operatorname{Ric}}(\omega')-{\operatorname{Ric}}(\omega)=-dd^c \log\left(\frac{\omega'^m}{\omega^m} \right) =-dd^c f.$$
We have the following uniqueness result.
\[thm:unique\] Suppose $(X,g,J)$ is ${\operatorname{AC}}(\delta,\ell+\alpha),\ 2\leq\ell\leq\infty,$ and Ricci-flat. Suppose $g'$ is another Ricci-flat Kähler metric which converges to $g$ in $C^{0,\alpha}_{\gamma}(X)$ with $\gamma<-m$, i.e. $|g' -g|\in C^{0,\alpha}_{\gamma}(X)$, and the Ricci forms satisfy $[\omega']=[\omega]$. Then $g' =g$.
Set $\eta =\omega' -\omega$. Then $\eta$ is an exact form, and by Proposition \[prop:dd-bar\] there is a $\phi\in C^{4,\alpha}_{\gamma+2}(X)$ with $dd^c \phi =\eta$. Since $\phi$ solves $(\omega+dd^c\phi)^m =e^f \omega^m$, the uniqueness part of Theorem \[thm:Calabi\] gives $\phi=0$.
Ricci-flat metrics on resolutions
---------------------------------
The main motivation for proving Theorem \[thm:Calabi\] is to construct examples of asymptotically conical Ricci-flat Kähler manifolds. In this section we give a proof of the part of Theorem \[thm:main\] concerning compactly supported Kähler classes.
In order to apply Theorem \[thm:Calabi\] one must start with an ${\operatorname{AC}}$ Kähler manifold $(X,g,\omega)$ with $c_1(X)=0$. Thus we suppose there is a nowhere vanishing holomorphic $n$-form $\Omega$ on $X$. Recall that the Ricci form of $(X,g,J)$ is $$\label{eq:Ricci-form}
{\operatorname{Ric}}(\omega)=dd^c \log\left(\frac{\Omega\wedge{\bar{\Omega}}}{\omega^m}\right).$$ If $f=\log\left(\frac{\Omega\wedge{\bar{\Omega}}}{\omega^m}\right)$, then a solution $(X,g',\omega')$ to Theorem \[thm:Calabi\] has Ricci form ${\operatorname{Ric}}(\omega')={\operatorname{Ric}}(\omega) -dd^c f =0$. But order to apply Theorem \[thm:Calabi\] one must start with an ${\operatorname{AC}}$ Kähler manifold $(X,g,\omega)$ with *Ricci potential* $f\in C^k_{\beta}(X)$ with $\beta< -2$. In general, it may be difficult to find a Kähler metric on $X$ satisfying this.
The case of a quasi-projective variety $X=Y\setminus D$, where $D$ is a divisor, supporting the anti-canonical divisor $\mathbf{K}^{-1}_Y$, which admits a Kähler-Einstein metric was dealt with in [@TY2] and independently in [@BK2]. A Kähler metric $\omega$ on $X$ was perturbed to a Kähler metric $\omega_0$ whose Ricci potential $f$ satisfies $f\in C^k_{\beta}(X)$. The author considered [@vC2] and extension of this result to some cases where $D$ does not admit a Kähler-Einstein metric. One essentially needs to start with an ${\operatorname{AC}}$ Kähler metric $(X,g,\omega)$ which approximates a Ricci-flat metric at infinity to high enough order.
We consider the relatively easy case of a crepant resolution $\pi:\hat{X}\rightarrow X=C(S)\cup\{o\}$ of a Ricci-flat Kähler cone $C(S)$. We will obtain Theorem \[thm:main\] of the introduction. In the following $r$ will denote the the radius function on the cone $C(S)$.
Recall that a variety $X$ has rational singularities if for some, and it follows any, resolution $\pi:Y\rightarrow X$ $R^j \pi_* \mathcal{O}_Y =0$ for $j>0$.
\[prop:rat-sing\] Let $C(S)$ be a Kähler cone satisfying Proposition \[prop:CY-cond\]. Then $o\in X=C(S)\cup\{o\}$ is a rational singularity. In particular, if $\pi:\hat{X}\rightarrow X$ is a resolution, then $H^j(\hat{X},{\mathbb{R}})=0$ for $j\geq 1$.
We use the criterion of H. Laufer and D. Burns for the rationality of an isolated singularity $o\in X$. If $\Omega$ is an holomorphic n-form on a deleted neighborhood of $o\in X$, then $o\in X$ is rational if and only if $$\label{eq:rational}
\int_U \Omega\wedge{\bar{\Omega}}<\infty,$$ where $U$ is a small neighborhood of $o\in X$. If $\Omega$ satisfies $\mathcal{L}_\xi =im\Omega$, then (\[eq:CY-cond\]) is satisfied. And one easily see that the inequality (\[eq:rational\]) holds.
\[prop:Kahler-form\] Suppose $\omega$ is a Kähler metric on $\hat{X}$ with $[\omega]\in H^2_c(\hat{X},{\mathbb{R}})$. Then there exists a Kähler metric $\omega_0$ on $\hat{X}$ with $[\omega_0]=[\omega]$ and ${\bar{\omega}} =\pi_* \omega_0$ on $\{x\in\hat{X}: \rho(x)>R\}$ for some $R>0$, where ${\bar{\omega}}=C\frac{1}{2}dd^c(r^2),\ C>0,$ is the Kähler cone metric on $C(S)$, up to homothety.
Let $E_i,\ i=1,\ldots, d,$ be the prime divisors in the exceptional set $E=\pi^{-1}(o)\subset\hat{X}$. Since $[\omega]\in H^2_c(\hat{X},{\mathbb{R}})$, it is Poincaré dual to $\sum_{i=1}^d a_i [E_i]\in H_{2m-2}(\hat{X},{\mathbb{R}})$, for $a_i \in{\mathbb{R}}$. Thus there exists a closed compactly supported real $(1,1)$-form $\theta$ with $[\theta]=[\omega]$. Let $\eta=\omega-\theta$. Then $\eta$ is an exact real $(1,1)$-form on $\hat{X}$. There exists an $\alpha\in\Omega^1$ with $d\alpha =\eta$. We have $\alpha=\alpha^{1,0} +\alpha^{0,1}$ where $\alpha^{0,1}=\overline{\alpha^{1,0}}$. Then ${\bar{\partial}}\alpha^{0,1} =0$, and by Proposition \[prop:rat-sing\] there exists a $u\in C^\infty (X,{\mathbb{C}})$ with ${\bar{\partial}}u=\alpha^{0,1}$. Define $v=\frac{i}{2}({\bar{u}}-u)$. Then $$\begin{split}
dd^c v =i\partial{\bar{\partial}}v & =\frac{1}{2}\left(\partial{\bar{\partial}}v -\partial{\bar{\partial}}{\bar{v}}\right) \\
& =\frac{1}{2}\left(\partial{\bar{\partial}}v +{\bar{\partial}}\partial {\bar{v}}\right) \\
& =\frac{1}{2}\left(\partial\alpha^{0,1} +{\bar{\partial}}\alpha^{1,0}\right)\\
& =\eta.
\end{split}$$
We may assume the radius function $\rho$ on $\hat{X}$ is chosen so that $\rho(x)=\pi^* r(x)$, for $\rho(x)>2$, and $dd^c(\rho^2)\geq 0$. Let $\mu :{\mathbb{R}}\rightarrow [0,1]$ be a smooth function with $\mu(t)=0$ for $t>1$ and $\mu(t)=0$ for $t<0$. Define $\omega_0 =\theta +C\frac{1}{2}dd^c(\rho^2) +dd^c [\mu(\rho -R)v]$. Choose $R$ large enough that the support of $\theta$ is contained in $\{\rho <R\}\subset\hat{X}$. Then for $C>0$ chosen sufficiently large $\omega_0$ is a Kähler form with the required properties.
Now suppose $X=C(S)\cup\{o\}$ be a Ricci-flat Kähler cone. And let $\pi:\hat{X}\rightarrow X$ be a crepant resolution. Recall, this means that $\pi^* \mathbf{K}_X =\mathbf{K}_{\hat{X}}$. Thus $\mathbf{K}_{\hat{X}}$ is trivial. Let $\Omega$ be the holomorphic $n$-form on $X$ as in Proposition \[prop:CY-cond\]. Then $\pi^* \Omega$ is a nowhere vanishing holomorphic form on $\hat{X}$, which we again denote by $\Omega$. If $\omega_0$ is a Kähler form as in Proposition \[prop:Kahler-form\], then a Ricci-potential of $(\hat{X},g_0,\omega_0)$ is $f=\log\left(\frac{c \Omega\wedge{\bar{\Omega}}}{\omega_0^m}\right)$ where we choose the constant $c=i^{m^2}\frac{m!}{(2C)^m}$ so that $f=0$ outside a compact set. Part (ii) of Theorem \[thm:Calabi\] gives a $\phi\in C^\infty_{2-2m}(X)$ of the form $\phi= A\rho^{2-2m} +\psi$ where $\psi\in C^\infty_{2+\beta}$, where $\beta<-2m$. And $\omega=\omega_0 +dd^c \phi$ is the Kähler form of the Ricci-flat Kähler metric in Theorem \[thm:main\].
Suppose that $\omega'$ is another Ricci-flat Kähler form with $[\omega']=[\omega]$ and $|\omega' -\omega|\in C^{0,\alpha}_{\beta}(X)$ with $\beta<-m$. By Proposition \[prop:dd-bar\] there is a smooth $u\in C^{2,\alpha}_{\beta+2}(X)$ with $\omega' -\omega=dd^c u$. Then $u$ solves $(\omega +dd^c u)^m =\omega^m$. The uniqueness result of Theorem \[thm:Calabi\] then shows that $u=0$.
The constant $A$ in (\[eq:const-A\]) turns out to be an invariant of the Kähler class in $H^2_c (\hat{X},{\mathbb{R}})$. Recall that $\omega_0 =\theta +\frac{1}{2}Cdd^c(\rho^2) +dd^c\left[\mu(\rho-R)v\right]$ where the first and third terms have compact support. Expanding and using that the integral of a compactly supported exact form is zero gives $$\begin{split}
\int_X (1-e^f)\omega_0^m & =\int_X \omega_0^m -c\Omega\wedge{\bar{\Omega}} \\
& =\int_X \theta^m +\bigl(\frac{1}{2}Cdd^c(\rho^2)\bigr)^m -\bigl(\frac{1}{2}Cdd^c(r^2)\bigr)^m \\
& =\int_X \theta^m .
\end{split}$$ Therefore, if we consider the Kähler class $[\omega]\in H_c^*(\hat{X},{\mathbb{R}})$, then $$\label{eq:A-cohom}
A=\frac{1}{(m-1)\Omega}\int_X (1-e^f)\, d\mu =\frac{1}{(m-1)m!\Omega}[\omega]^{\cup m},$$ where $\Omega ={\operatorname{Vol}}(S)=\{r=1\}\subset C(S)$.
We also have the following result on the Kähler potential of Ricci-flat metrics of Theorem \[thm:main\].
\[thm:R-f-potent\] Let $\hat{X}$ be a crepant resolution of a Ricci-flat Kähler cone $X=C(S)\cup\{o\}$. Then in each Kähler class on $\hat{X}$, with $C>0$ as in Proposition \[prop:Kahler-form\], there is a unique Ricci-flat Kähler metric $g$ with Kähler form $\omega$ which satisfies $$\label{eq:R-f-potent}
\pi_*(\omega) =\frac{1}{2}Cdd^c (r^2) +Add^c (r^{2-2m}) +dd^c(\psi),$$ on $\{x\in X: r(x)>R\}$. Here $A$ is given by the Kähler class $[\omega]$ in (\[eq:A-cohom\]), and $\psi\in C^\infty_\gamma(X)$ with $\gamma< 2-2m$.
There remains the question of the optimal $\gamma <2-2m$ giving the decay of $\psi$ in Theorem \[thm:R-f-potent\]. This comes down to finding the largest $\epsilon>0$ in Theorem \[thm:lap-weight\]. In general, we only know $\epsilon>0$. But in Theorem \[thm:R-f-potent\] the ${\operatorname{AC}}$ Kähler manifold $(\hat{X}, g_0,\omega_0)$ has “boundary” $S$ which is Einstein. The condition for the Laplacian $$\label{eq:Lp-Fred}
\Delta: L^p_{k+2, \delta}(X)\rightarrow L^p_{k,\delta-2}(X)$$ to be Fredholm is well known [@LocMcO]. There is a family of operators on $S$ $$I(\Delta,\lambda) =\lambda^2 -(2m-2)\sqrt{-1}\lambda +\Delta_S,$$ for $\lambda\in{\mathbb{C}}$ where $\Delta_S$ is the Laplacian on $S$. Then ${\operatorname{Spec}}(I,\lambda)$ is the set of $\lambda$ for which $$I(\Delta,\lambda):L^p_{k+2}(S)\rightarrow L^p_k(S)$$ does not admit a bounded inverse. In our case $${\operatorname{Spec}}(I,\lambda)=\{0,\ (2m-2)\sqrt{-1},\ \mu_j^+ \sqrt{-1},\ \mu_j^- \sqrt{-1},\ldots| j=1,2,\ldots\}$$ where $\mu^{\pm}_j \sqrt{-1}$ are the two solutions of $x^2-(2m-2)\sqrt{-1}x+\lambda_j =0$ with $\lambda_j$ the j-th eigenvalue of $\Delta_S$. It was shown in [@LocMcO] that if ${\operatorname{Im}}{\operatorname{Spec}}(I,\lambda)$ denote the imaginary components, then (\[eq:Lp-Fred\]) is Fredholm for $-\delta\notin{\operatorname{Im}}{\operatorname{Spec}}(I,\lambda)$.
If $(S,g_S)$ is Sasaki-Einstein then by Lichnerowicz’s Theorem $\lambda_1 \geq 2m-1$ with equality only if $S$ is isometric to the sphere. And one can check that $\mu_1^+ \geq 2m-1$ with equality only if $S$ is isometric to a sphere. And for $\delta\in (-\mu_1^+ ,2-2m)$ the operator (\[eq:Lp-Fred\]) is Fredholm with index $-1$. So in Theorem \[thm:R-f-potent\] one will actually have $\psi\in C^\infty_{1-2m}(X)$. [^1]
Examples
========
We consider some examples of asymptotically conical Ricci-flat Kähler manifolds given by Theorem \[thm:main\]. These examples are ${\operatorname{AC}}(2n,\infty)$ Ricci-flat examples, and are either resolutions of toric Kähler cones or resolutions of hypersurface singularities. See [@vC3] for many more examples on resolutions of Kähler cones. Here we just give enough examples to give the reader an idea of the scope of examples.
Also in [@vC2] examples are constructed on affine varieties which are of type ${\operatorname{AC}}(2n,k)$ for large $k>0$.
Resolutions of hypersurface singularities
-----------------------------------------
We describe how examples can be constructed from resolutions of weighted homogeneous hypersurface singularities. Let $\mathbf{w}=(w_0 ,\ldots, w_m)\in ({\mathbb{Z}}_+)^{m+1}$ with $\gcd(w_0,\ldots,w_m)=1$. We have the weighted ${\mathbb{C}}^*$-action on ${\mathbb{C}}^{m+1}$ given by $(z_0,\ldots, z_m)\rightarrow (\lambda^{w_0}z_0,\ldots,\lambda^{w_m}z_m)$ for $\lambda\in{\mathbb{C}}^*$. A polynomial $f\in{\mathbb{C}}[z_0,\ldots,z_m]$ is *weighted homogeneous* of degree $d\in{\mathbb{Z}}_+$ if $$f(\lambda^{w_0}z_0,\ldots,\lambda^{w_m}z_m)=\lambda^d f(z_0,\ldots,z_m).$$
There is a *weighted Sasaki structure* on the sphere $S^{2m+1}_{\mathbf{w}}$ for which the Reeb vector field $\xi_{\mathbf{w}}$ generates the $S^1$-action induced by the above weighted action. See [@vC4] for details. The cone $C(S^{2m+1}_{\mathbf{w}})$ is biholomorphic to ${\mathbb{C}}^{m+1} \setminus\{o\}$, but with a much different metric. If $f$ is a weighted homogeneous polynomial, then the Kähler cone structure of $C(S^{2m+1}_{\mathbf{w}})$ restricts to $X_f =\{z\in{\mathbb{C}}^{m+1} : f(z)=0\}$. And similarly the Sasaki structure on $S^{2m+1}_{\mathbf{w}}$ restricts to the link $S_f := X_f \cap S^{2m+1}$. This is given in the diagram:
$$\begin{array}{ccc}
X_f & \hookrightarrow & C(S^{2m+1}_{\mathbf{w}}) \\
\cup & & \cup \\
S_f & \hookrightarrow & S^{2m+1}_{\mathbf{w}} \\
\downarrow & & \downarrow \\
Z_f & \hookrightarrow & {\mathbb{C}}P(\mathbf{w})
\end{array}$$
Here $Z_f$ is a hypersurface in the weighted projective space ${\mathbb{C}}P(\mathbf{w})$.
\[prop:hyp-CY\] The Kähler cone $C(S_f)$ admits a nowhere vanishing holomorphic $m$-form $\Omega$ satisfying $\mathcal{L}_{\xi} \Omega =im\Omega$, after possibly rescaling the Reeb vector field $\xi$, if and only if $d< |\mathbf{w}|=\sum_{j=0}^m w_j$.
This is precisely the condition that the *orbifold* canonical bundle $\mathbf{K}_{Z_f}$ on $Z_f$ is negative.
With Proposition \[prop:hyp-CY\] satisfied, we are interested in transversally deforming the Sasaki structure of $S_f$ to a Sasaki-Einstein structure. If $\eta$ is the contact structure of $S_f$ then $\omega^T =\frac{1}{2}d\eta$ is the Kähler structure transversal to the foliation generated by the Reeb field $\xi$. A *transversal deformation* is a new Sasaki structure with transversal Kähler form $$(\omega^T)' =\omega^T +dd^c \phi,$$ for some basic $\phi\in C^\infty_B (S)$. The new contact form is $\eta' =\eta +2d^c \phi$. And one can show that the Kähler structure on the cone becomes $\omega' =\frac{1}{2}dd^c (r')^2$ where $r'=e^\phi r$. Obtaining a Sasaki-Einstein structure is equivalent to solving the transversal Kähler-Einstein condition $$\label{eq:trans-KE}
{\operatorname{Ric}}((\omega^T)') =2m(\omega^T)'.$$
Condition (\[eq:CY-cond\]) implies that $${\operatorname{Ric}}(\omega^T)-2m\omega^T =dd^c h.$$ And solving (\[eq:trans-KE\]) is equivalent to solving the transversal Monge-Ampère equation $$\label{eq:trans-MA}
(\omega^T +dd^c\phi)^{m} =e^{-2m\phi +h}(\omega^T)^{m}.$$
See [@BGJ; @BG2; @BG3] for more on solving (\[eq:trans-MA\]) to find Sasaki-Einstein metrics. In particular, if $f$ is a Brieskorn-Pham polynomial, $f=\sum_{j=0}^m z_j^{a_j}$. Then [@BGJ Theorem 34] gives simple numerical conditions on the $a_j$ for (\[eq:trans-MA\]) to be solvable.
For example, consider $$f= z_0^m +z_1^m +\cdots+ z_{m-1}^m +z_m^k.$$ Then these conditions are satisfied if $k>m(m-1)$, and $X_k :=\{z\in{\mathbb{C}}: f(z)=0\}\subset{\mathbb{C}}^{m+1}$ has a Ricci-flat Kähler cone structure.
Let $\hat{{\mathbb{C}}^{m+1}}$ be the blow-up of $o\in{\mathbb{C}}^{m+1}$. If $X' \subset\hat{{\mathbb{C}}^{m+1}}$ is the birational transform and $E=X' \cap{\mathbb{C}P}^m \subset\hat{{\mathbb{C}}^{m+1}}$ is the exceptional divisor, then adjunction gives $$\mathbf{K}_{X'} =\pi^* \mathbf{K}_X + (m-\deg f) E.$$ Thus $\pi: X' \rightarrow X_k$ is crepant for $k\geq m$. It is not difficult to see that $X'$ has one singularity isomorphic to $X_{k-m}$. If $k = 0$ or $1 \mod m$, then by repeatedly blowing up $\lfloor\frac{k}{m}\rfloor$ times we get a smooth crepant resolution $\pi: \hat{X}_k \rightarrow X_k$. Therefore, if $k>n(n-1)$ and $k = 0$ or $1 \mod m$, then Theorem \[thm:main\] gives a $\lfloor\frac{k}{m}\rfloor$ family of Ricci-flat Kähler metrics on $\hat{X}_k$ converging to the cone metric as in (\[eq:conv-cpt\]).
Toric examples
--------------
One easy way to construct examples of Ricci-flat metrics on resolutions is to consider resolutions of toric Kähler cones.
A Kähler cone $(C(S),{\bar{g}}),\ \dim_{{\mathbb{C}}} C(S) =m$, is toric if it admits an effective isometric action of the torus $T=T^{m}$ which preserves the Euler vector field $r\partial_r$.
The associated Sasaki manifold $(S,g)$ is said to be toric. Let $\mathfrak{t}$ be the Lie algebra of $T^m$. It follows that the Reeb vector field $\xi\in\mathfrak{t}$.
Since $T^m$ preserves the Kähler form $\omega=\frac{1}{2}d(r^2 \eta)$ and further preserves $r^2 \eta$, there is a *moment map* $$\label{eq:moment-map}
\begin{gathered}
\mu: C(S) \longrightarrow \mathfrak{t}^* \\
\langle \mu(x),X\rangle = \frac{1}{2}r^2\eta(X_S (x)),
\end{gathered}$$ where $X_S$ denotes the vector field on $C(S)$ induced by $X\in\mathfrak{t}$. We have the moment cone defined by $$\mathcal{C} :=\mu(C(S)) \cup \{0\},$$ which from [@Ler] is a strictly convex rational polyhedral cone. Recall that this means that there are vectors $u_i,i=1,\ldots,d$ in the integral lattice ${\mathbb{Z}}_T =\ker\{\exp(2\pi i\cdot):\mathfrak{t}\rightarrow T\}$ such that $$\label{eq:moment-cone}
\mathcal{C} =\bigcap_{j=1}^{d} \{y\in\mathfrak{t}^* : \langle u_j,y\rangle\geq 0\}.$$
The elements $u_j \in{\mathbb{Z}}_T,\ j=1,\ldots,d,$ span a cone $\mathcal{C}^*$ in $\mathfrak{t}$ dual to $\mathcal{C}$. Then $\mathcal{C}^*$ and all of its faces define a fan $\Delta$ characterizing $C(S)\cup\{o\}$ as an algebraic toric variety. (cf. [@Od]) There is a nowhere vanishing holomorphic m-form satisfying Proposition \[prop:CY-cond\] precisely when there is a $\gamma\in {\operatorname{Hom}}_{{\mathbb{Z}}}({\mathbb{Z}}_T, {\mathbb{Z}})$ with $\gamma(u_j) =1,\ j=1,\ldots,d.$ This is the condition that $C(S)\cup\{o\}$ is Gorenstein.
The result of A. Futaki, H. Ono, and G. Wang on the existence of Sasaki-Einstein metrics on toric Sasaki manifolds makes toric geometry a propitious source of examples.
\[thm:FOW\] Let $C(S)\cup\{o\}$ be a Gorenstein toric Kähler cone with toric Sasaki manifold $S$. Then we can deform the Sasaki structure by varying the Reeb vector field and then performing a transverse Kähler deformation to a Sasaki-Einstein metric. The Reeb vector field and transverse Kähler deformation are unique up to isomorphism.
Define $H_\gamma =\{\gamma =1\}\subset\mathfrak{t}\cong{\mathbb{R}}^m$. The intersection $P_{\Delta} =H_\gamma \cap\mathcal{C}^*$ is an integral polytope in $H_\gamma \cong{\mathbb{R}}^{m-1}$. A toric crepant resolution $$\label{eq:toric-resol}
\pi: X_{\tilde{\Delta}}\rightarrow X_{\Delta}$$ is given by a nonsingular subdivision $\tilde{\Delta}$ of $\Delta$ with every 1-dimensional cone $\tau_i \in\tilde{\Delta}(1), i=1,\ldots,N$ generated by a primitive vector $u_i :=\tau_i \cap H_\gamma$. This is equivalent to a basic, lattice triangulation of $P_{\Delta}$. *Lattice* means that the vertices of every simplex are lattice points, and *basic* means that the vertices of every top dimensional simplex generates a basis of ${\mathbb{Z}}^{n-1}$. Note that a *maximal* triangulation of $P_{\Delta}$, meaning that the vertices of every simplex are its only lattice points, always exists. Every basic lattice triangulation is maximal, but the converse only holds in dimension 2.
We want Kähler structures on the resolution $X_{\tilde{\Delta}}$. This is given by a strictly convex support function $h\in{\operatorname{SF}}(\tilde{\Delta},{\mathbb{R}})$ on $\tilde{\Delta}$. This is a real valued function which is piecewise linear on the cones of $\tilde{\Delta}$. Convexity means that $h(x+y)\geq h(x)+h(y)$ for $x,y\in|\tilde{\Delta}|$, the support of $\tilde{\Delta}$. Let $l_\sigma$ define $h$ on the m-cone $\sigma$. Strict convexity means that $\langle l_\sigma ,x\rangle\geq h(x)$, for all $x\in|\Delta|$, with equality only if $x\in\sigma$.
The following is proved by taking a torus Hamiltonian reduction of ${\mathbb{C}}^N$. See [@BurGuiLer] and also [@vC3].
For each strictly convex support function $h\in{\operatorname{SF}}(\tilde{\Delta},{\mathbb{R}})$ there is a Kähler structure $\omega_h$ so that $(X_{\tilde{\Delta}},\omega_h)$ is a Hamiltonian Kähler manifold and the image of the moment map is the polyhedral set $$\mathcal{C}_h :=\bigcap_{j=1}^{N} \{y\in\mathfrak{t}^* : \langle u_j,y\rangle\geq\lambda_j\}.$$ If $h$ satisfies $h(u_j)=0$ for $j=1,\ldots,d$, then $[\omega_h ]\in H^2_c (X_{\tilde{\Delta}},{\mathbb{R}})$. The $u_j \in{\operatorname{Int}}P_{\Delta}, j=d+1,\ldots,N$, correspond to the prime divisors $D_j$ in $E=\pi^{-1}(o)$. For each $j=d+1,\ldots,N$, let $c_j \in H^2_c (X_{\tilde{\Delta}},{\mathbb{R}})$ be the Poincaré dual of $[D_j]$ in $H_{2n-2}(X_{\tilde{\Delta}},{\mathbb{R}})$. Then $$[\omega_h] =-2\pi\sum_{j=d+1}^N \lambda_j c_j.$$
If $[\omega_h] \in H^2_c (X_{\tilde{\Delta}},{\mathbb{R}})$, then we can apply Proposition \[prop:Kahler-form\] to construct a Kähler metric $\omega_0$ with Ricci potential $f=\log\left(\frac{c \Omega\wedge{\bar{\Omega}}}{\omega_0^m}\right)$. If $[\omega_h]$ is not compactly supported, then an initial metric is constructed by R. Goto [@Got §5] with Ricci potential $f\in C^\infty_{-4}(X)$. In both cases the initial Kähler metrics $\omega_0$ and Ricci potentials $f$ can be taken $T^m$-invariant. We get the toric version of Theorem \[thm:main\].
Let $\pi:X_{\tilde{\Delta}}\rightarrow C(S)\cup\{o\}$ be a crepant resolution of a toric Kähler cone. Then for each strictly convex $h\in{\operatorname{SF}}(\tilde{\Delta},R)$, there is a $T^m$-invariant Ricci-flat Kähler metric $g$ whose Kähler form satisfies $[\omega]=[\omega_h]$. If $[\omega_h]\in H_c^2(X_{\tilde{\Delta}},{\mathbb{R}})$, then $g$ is unique and is asymptotic to the cone metric as in (\[eq:conv-cpt\]), otherwise $g$ converges as (\[eq:conv-noncpt\]).
If $[\omega_h]\in H_c^2(X_{\tilde{\Delta}},{\mathbb{R}})$, then from (\[eq:A-cohom\]) we have $$\label{eq:A-toric}
\begin{split}
A = \frac{1}{(m-1)m!\Omega}[\omega]^{m} & =\frac{-2\pi}{(m-1)m!\Omega}\sum_{j=d+1}^N \lambda_j c_j [\omega]^{m-1}\\
& =\frac{-2\pi}{(m-1)m!\Omega}\sum_{j=d+1}^N \lambda_j\int_{E_j} [\omega_h]^{m-1} <0.
\end{split}$$ Note that all the quantities in (\[eq:A-toric\]) can be computed from $h$ in terms of volumes of various polytopes.
When $\dim_{{\mathbb{C}}} X_\Delta =3$ there always exists a toric crepant resolution $X_{\tilde{\Delta}}$. And further, if $X_\Delta$ is not the quadric cone $\{z_0^2 +z_1^2 +z_2^2 +z_3^2 =0\}\subset{\mathbb{C}}^4$, then it admits a toric crepant resolution $X_{\tilde{\Delta}}$ with a strictly convex support function $h$ so that $[\omega_h] \in H^2_c (X_{\tilde{\Delta}},{\mathbb{R}})$. See [@vC4] for more details.
[10]{}
Thierry Aubin. Métriques riemanniennes et courbure. , 4:383–424, 1970.
Shigetoshi Bando and Ryoichi Kobayashi. Ricci-flat [K]{}ähler metrics on affine algebraic manifolds. [II]{}. , 287(1):175–180, 1990.
Charles P. Boyer and Krzysztof Galicki. Sasakian geometry, hypersurface singularities, and [E]{}instein metrics. , (75):57–87, 2005.
Charles P. Boyer and Krzysztof Galicki. . Oxford Mathematical Monographs. Oxford University Press, Oxford, 2008.
Charles P. Boyer, Krzysztof Galicki, and J[á]{}nos Koll[á]{}r. Einstein metrics on spheres. , 162(1):557–580, 2005.
Dan Burns, Victor Guillemin, and Eugene Lerman. Kähler metrics on singular toric varieties. , 238(1):27–40, 2008.
Alice Chaljub-Simon and Yvonne Choquet-Bruhat. Problèmes elliptiques du second ordre sur une variété euclidienne à l’infini. , 1(1):9–25, 1979.
Koji Cho, Akito Futaki, and Hajime Ono. Uniqueness and examples of compact toric [S]{}asaki-[E]{}instein metrics. , 277(2):439–458, 2008.
E. B. Fabes and D. W. Stroock. A new proof of [M]{}oser’s parabolic [H]{}arnack inequality using the old ideas of [N]{}ash. , 96(4):327–338, 1986.
A. Futaki, H. Ono, and G. Wang. Transverse [K]{}ähler geometry of [S]{}asaki manifolds and toric [S]{}asaki-[E]{}instein manifolds. , 83(3):585–636, 2009.
Ryushi Goto. Calabi-[Y]{}au structures and [E]{}instein-[S]{}asakian structures on crepant resolutions of isolated singularities. arXiv:math.DG/0906.5191 v.2, 2009.
A. A. Grigoryan. The heat equation on noncompact [R]{}iemannian manifolds. , 182(1):55–87, 1991.
Alexander Grigor’yan and Laurent Saloff-Coste. Stability results for [H]{}arnack inequalities. , 55(3):825–890, 2005.
Tam[á]{}s Hausel, Eugenie Hunsicker, and Rafe Mazzeo. Hodge cohomology of gravitational instantons. , 122(3):485–548, 2004.
David Jerison. The [P]{}oincaré inequality for vector fields satisfying [H]{}örmander’s condition. , 53(2):503–523, 1986.
Dominic D. Joyce. . Oxford Mathematical Monographs. Oxford University Press, Oxford, 2000.
Eugene Lerman. Contact toric manifolds. , 1(4):785–828, 2003.
Robert Lockhart. Fredholm, [H]{}odge and [L]{}iouville theorems on noncompact manifolds. , 301(1):1–35, 1987.
Robert B. Lockhart and Robert C. McOwen. Elliptic differential operators on noncompact manifolds. , 12(3):409–447, 1985.
Dario Martelli and James Sparks. Baryonic branches and resolutions of [R]{}icci-flat [K]{}ähler cones. , (4):067, 44, 2008.
Dario Martelli and James Sparks. Resolutions of non-regular [R]{}icci-flat [K]{}ähler cones. , 59(8):1175–1195, 2009.
Dario Martelli and James Sparks. Symmetry-breaking vacua and baryon condensates in [A]{}d[S]{}/[CFT]{} correspondence. , 79(6):065009, 51, 2009.
Vincent Minerbe. Weighted [S]{}obolev inequalities and [R]{}icci flat manifolds. , 18(5):1696–1749, 2009.
Tadao Oda. , volume 15 of [ *Ergebnisse der Mathematik und ihrer Grenzgebiete (3) \[Results in Mathematics and Related Areas (3)\]*]{}. Springer-Verlag, Berlin, 1988. An introduction to the theory of toric varieties, Translated from the Japanese.
L. Saloff-Coste. A note on [P]{}oincaré, [S]{}obolev, and [H]{}arnack inequalities. , (2):27–38, 1992.
Laurent Saloff-Coste. Lectures on finite [M]{}arkov chains. In [*Lectures on probability theory and statistics ([S]{}aint-[F]{}lour, 1996)*]{}, volume 1665 of [*Lecture Notes in Math.*]{}, pages 301–413. Springer, Berlin, 1997.
Laurent Saloff-Coste. , volume 289 of [ *London Mathematical Society Lecture Note Series*]{}. Cambridge University Press, Cambridge, 2002.
Gang Tian and Shing-Tung Yau. Complete [K]{}ähler manifolds with zero [R]{}icci curvature. [II]{}. , 106(1):27–60, 1991.
Craig van Coevering. A construction of complete [R]{}icci-flat [K]{}ähler manifolds. arXiv:math.DG/0803.0112, 2008.
Craig van Coevering. Examples of asymptotically conical [R]{}icci-flat kähler manifolds. arXiv:math.DG/0812.4745 v.2, to appear in Math. Zeitschrift, 2008.
Craig van Coevering. Ricci-flat [K]{}ähler metrics on crepant resolutions of [K]{}ähler cones. arXiv:math.DG/0806.3728 v.3, to appear in Math. Annalen, 2008.
Shing Tung Yau. On the [R]{}icci curvature of a compact [K]{}ähler manifold and the complex [M]{}onge-[A]{}mpère equation. [I]{}. , 31(3):339–411, 1978.
[^1]: Ryushi Goto pointed this out to me.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We announce the forthcoming public release of Version 1.1 of MGGPOD, a user-friendly suite of Monte Carlo codes built around the widely used GEANT (Version 3.21) package. MGGPOD is capable of simulating [*ab initio*]{} the physical processes relevant for the production of instrumental backgrounds. These processes include the build-up and delayed decay of radioactive isotopes as well as the prompt de-excitation of excited nuclei, both of which give rise to a plethora of instrumental gamma-ray background lines in addition to continuum backgrounds. A detailed qualitative and quantitative understanding of instrumental backgrounds is crucial for most stages of high-energy astronomy missions.
Improvements implemented in Version 1.1 of the proven MGGPOD Monte Carlo suite include: additional beam geometry options, the capability of modelling polarized photons, additional output formats suitable e.g. for event reconstruction algorithms, improved neutron interaction cross sections, and improved treatment of the radioactive decay of isomeric nuclear states.
The MGGPOD package and documentation are publicly available for download from [*http://sigma-2.cesr.fr/spi/MGGPOD/*]{}.
author:
- 'G. Weidenspointner'
- 'S.J. Sturner'
- 'E.I. Novikova'
- 'M.J. Harris'
- 'A. Zoglauer'
- 'C.B. Wunderer'
- 'R.M. Kippen'
- 'P. Bloser'
- 'Ch. Zeitnitz'
title: 'MGGPOD: A Monte Carlo Suite for Gamma Ray Astronomy – Version 1.1'
---
\[2001/04/25 1.1 (PWD)\]
Introduction {#intro}
============
Intense and complex instrumental backgrounds, against which the much smaller signals from celestial sources must be discerned, are a notorious problem for low energy gamma-ray and hard X-ray astronomy. Therefore a detailed qualitative and quantitative understanding of instrumental line and continuum backgrounds is crucial for most stages of gamma-ray astronomy missions, ranging from the design, development, and performance prediction phases through calibration and response generation to data reduction. A promising approach for obtaining quantitative estimates of instrumental backgrounds is [*ab initio*]{} Monte Carlo simulation [see e.g. @Dean03].
We have developed a suite of Monte Carlo packages, named MGGPOD [@weidenspointner_mggpod], that supports this type of simulation. The original MGGPOD Monte Carlo suite (Version 1.0) and documentation have been publicly available for download from [*http://sigma-2.cesr.fr/spi/MGGPOD/*]{} since 2004. First applications include the successful modelling of the instrumental backgrounds of the TGRS Ge spectrometer onboard Wind [@weidenspointner_mggpod], the SPI Ge spectrometer onboard the INTEGRAL observatory [@weidenspointner_mggpod_aa; @weidenspointner_mggpod_seeon], and of the Reuven Ramaty High-Energy Solar Spectroscopic Imager [@wunderer_rhessi]. Since then MGGPOD has found an increasing number of applications including studies of the various instrument concepts for an Advanced Compton Telescope (ACT) [see e.g. @Boggs06a; @Boggs06b; @wunderer_act], for focal plane detectors of Laue lens gamma-ray telescopes such as MAX or the Gamma-Ray Imager (GRI) [@weidenspointner_bonifacio; @wunderer_bonifacio; @Knoedlseder_GRI], and for coded mask hard X-ray telescopes such as EXIST [@Garson05; @Garson06]. These studies required additions to and improvements of the original MGGPOD Version 1.0, which will be made available to the general community in the forthcoming public release of Version 1.1 of MGGPOD after completing rigorous testing by the ACT and GRI collaborations.
In this paper we provide an overview of the capabilities, usage, and structure of the MGGPOD package, as well as a description of the additions and improvements implemented in Version 1.1.
The MGGPOD Monte Carlo Simulation Suite {#mggpod}
=======================================
High-energy astronomy instruments are operated in an intense and complex radiation environment [see e.g. @weidenspointner_mggpod]. The MGGPOD Monte Carlo suite allows [*ab initio*]{} simulations of instrumental backgrounds – including the many gamma-ray lines – arising from interactions of the various ambient particle and photon radiation fields within the instrument and spacecraft materials. It is possible to simulate both prompt instrumental backgrounds, such as energy losses of cosmic-ray particles and their secondaries, as well as delayed instrumental backgrounds, which are due to the decay of radioactive isotopes produced in nuclear interactions. Of course, MGGPOD can also be used to study the response of gamma-ray instruments to astrophysical and calibration sources. The MGGPOD suite is therefore an ideal Monte Carlo tool for high-energy astronomy. A detailed description of the physics implemented in the original MGGPOD Monte Carlo suite (Version 1.0) can be found in @weidenspointner_mggpod. The documentation available from the MGGPOD web site provides comprehensive practical advice for users.
Capabilities and Functionalities {#cap_func}
--------------------------------
MGGPOD is a suite of five closely integrated Monte Carlo packages, namely [**MG**]{}EANT, [**G**]{}CALOR, [**P**]{}ROMPT, [**O**]{}RIHET, and [**D**]{}ECAY. The MGGPOD suite resulted from a combination of the NASA/GSFC MGEANT [@Sturner03] and the University of Southampton GGOD [@Dean03] Monte Carlo codes, which were supplemented with the newly developed PROMPT package. All these packages are based on the widely used GEANT Detector Description and Simulation Tool (Version 3.21) created and supported at CERN[^1], which is designed to simulate the passage of elementary particles through an experimental setup.
In the following, we provide a synopsis of the capabilities and functions of the five packages that constitute the MGGPOD suite:
- MGEANT [@Sturner03] is a multi-purpose simulation package that was created to increase the versatility of the GEANT simulation tool. A modular, “object oriented” approach was pursued, allowing for rapid prototyping of detector systems and easy generation of most of the radiation fields relevant to high-energy astronomy. Within the MGGPOD suite, MGEANT (i.e. GEANT) stores and transports all particles, and treats electromagnetic interactions from about 1 keV to a few TeV[^2]. MGEANT provides the option to use the GLECS and GLEPS packages to take into account the energy of bound electrons and photon polarization in Rayleigh and Compton scatterings (see Sec. \[mggpod1.1\]). The MGEANT simulation package and a user manual are available at a NASA/GSFC web site[^3].
- GCALOR [@Zeitnitz_Gabriel94] simulates hadronic interactions down to 1 MeV for nucleons and charged pions and down to thermal energies ($10^{-5}$ eV) for neutrons. Equally important, this package[^4] provides access to the energy deposits from all hadronic interactions as well as to isotope production anywhere in the simulated setup.
- PROMPT simulates prompt photon emission associated with the de-excitation of excited nuclei produced by neutron capture, inelastic neutron scattering, and spallation.
- ORIHET, originally developed for the GGOD suite [@Dean03] and improved for MGGPOD, calculates the build-up and decay of activity in any system for which the nuclide production rates are known. Hence ORIHET can be used to convert nuclide production rates, determined from simulations of cosmic-ray irradiation, to decay rates. These are required input for simulating the radioactive decays giving rise to delayed background.
- DECAY, again originally developed for GGOD and improved for our purposes, enables MGGPOD to simulate radioactive decays.
Structure
---------
The overall structure of the MGGPOD package is illustrated in Fig. \[mggpod\_flow\_chart\]. Depending on the simulated radiation field or high-energy photon source distribution one or three steps, requiring two or three input files, are needed to obtain the resulting energy deposits in the detector system under study. In general, it is advisable to simulate each component of the radiation environment separately. MGGPOD distinguishes two classes of radiation fields.
- Class I comprises radiation fields for which only prompt energy deposits are of interest, such as celestial or laboratory gamma-ray sources or cosmic-ray electrons.
- Class II comprises radiation fields for which in addition delayed energy deposits resulting from the activation of radioactive isotopes need to be considered. Examples for Class II fields are cosmic-ray protons, geomagnetically trapped protons, or neutrons leaking from the Earth’s atmosphere.
For both of these classes, the simulation of the prompt energy deposits requires two inputs: a mass model, and a model of the simulated radiation field. The mass model is a detailed computer description of the experimental setup under study. It specifies the geometrical structure of instrument and spacecraft, the atomic and/or isotopic composition of materials, and sets parameters that influence the transport of particles in different materials. Each component of the radiation environment (and analogously for gamma-ray sources) to which the instrument is exposed is characterized by three quantities: the type of the incident particles, and their spectral and angular distributions. The prompt energy deposits are written to an output event file; in case of a Class II radiation field there is an additionial output file in which all the nuclei produced in hadronic interactions are recorded.
To simulate delayed energy deposits (Class II radiation field only) two additonal steps need to be taken. These require as input the time history of the radiation field which is responsible for the activation, and the isotope production rates obtained previously during the simulation of prompt background. Based on this information, first the activity of each isotope produced in each structural element of the mass model is determined. Then these activities are used to simulate the delayed energy deposits in the instrument due to radioactive decays in each volume of the mass model. Combining prompt and delayed energy deposits from each component of the radiation environment and gamma-ray sources, it is possible to obtain the energy deposited in the system as a function of position and time. In particular, it is e.g. possible to obtain background energy spectra and rates for each detector element, and to assess the individual contributions of the various background components and mass model volumes, which is highly valuable when designing new instruments.
Additions and Improvements for MGGPOD Version 1.1 {#mggpod1.1}
=================================================
The additions to and improvements of the original MGGPOD Version 1.0 were driven by the requirements for studying various instrument concepts for an Advanced Compton Telescope (ACT), Laue lens gamma-ray telescope focal plane detectors, and coded mask hard telescopes. The original version of MGGPOD was mainly tested by modelling gamma-ray instruments that employ monolithic Ge detectors, such as the TGRS Ge spectrometer onboard Wind [@weidenspointner_mggpod], the SPI Ge spectrometer onboard the INTEGRAL observatory [@weidenspointner_mggpod_aa; @weidenspointner_mggpod_seeon], and of the Reuven Ramaty High-Energy Solar Spectroscopic Imager [@wunderer_rhessi]. Future high-energy instruments will use highly segmented or pixellated detectors. In particular detectors for Compton telescopes aim not only at recording the energy of each incoming photon, but rather at recording the complete interaction sequence. Furthermore, a large number of different detector materials is being considered for future instrumentation (e.g. Si, Xe, CZT, CdTe); hadronic interactions and nuclear de-excitations for these were not well treated in Version 1.0 of MGGPOD.
All recent upgrades of the original MGGPOD suite will soon be publicly released in Version 1.1, after completing rigorous testing by the ACT and GRI collaborations. The additions and improvements for MGGPOD Version 1.1 are as follows.
- We introduced new beam geometries. It is now possible to model radiation fields whose intensity varies with zenith angle – a requirement for simulating e.g. Earth leakage radiations in low-Earth orbit (gamma-rays, neutrons, …), or atmospheric background at balloon altitudes. Models can e.g. be generated with the ACTtools software available at [*http://public.lanl.gov/mkippen/actsim/glecs/*]{}. Also, new beam options were introduced to enable the simulation of photon beams from Laue diffraction lenses or hard X-ray mirrors.
- Additional output formats optimized for use with event reconstruction algorithms have been implemented. Among other information, it is possible to record the complete interaction information for all involved particles in passive as well as active (i.e. detector or veto shield) instrument materials. This is necessary e.g. for the creation of (probability based) response files for event reconstruction algorithms for Compton telescopes [e.g. MEGAlib, see @Zoglauer06]. It is also possible to record vertex (i.e.initial position) and momentum vector for each simulated particle. This capability to retrace the origin of each event has also been helpful for optimizing e.g. the design of active and passive shielding for high-energy instrumentation.
- One of the many exciting science goals of the next generation of high-energy instruments is to measure polarization. To include the effects of polarized incident photons in the Compton and Rayleigh scattering processes, the GLEPS package[^5] is now available in MGGPOD. GLEPS [@McConnell_Kippen04] is an extension of the GLECS GEANT3 Low-Energy Compton Scattering Package [@Kippen04], which takes into account the energy of bound electrons in photon scattering processes.
- The standard set of neutron cross sections available for GCALOR does not cover all elements/isotopes relevant for studying gamma-ray instrumentation. For selected elements/isotopes, we converted into GCALOR format the respective evaluated ENDF/B and JENDL neutron cross sections (complemented by other databases for isotopes of Zn missing in ENDF/B and JENDL). A slight modification of GCALOR was necessary to make sure that all these new cross section files can be read in properly.
- For selected elements/isotopes relevant for high-energy instrumentation (including Ge, Si, Xe, Cd, Zn, Te, Cs, I) we generated the PROMPT data files required to model de-excitations after neutron capture and inelastic neutron scattering. (De-excitations after spallations have already been modelled in Version 1.0 based on statistical considerations as described in [@weidenspointner_mggpod].)
- In Version 1.0 of MGGPOD it was assumed that nuclei in isomeric states decay exclusively through internal transition. However, this is not always true – some isomeric states can also beta-decay. The beta decay channel of isomeric states has now been implemented, improving the simulation of delayed background due to radioactive decays in common detector materials such as CZT.
- Last but not least, we introduced many checks to avoid improper input by the user or in data files (e.g. DECAY or PROMPT), which if not recognized may lead to run-time errors that are hard to diagnose. In case improper input is identified, error or warning messages are written to the log file, and the run is terminated if necessary.
Summary
=======
The MGGPOD Monte Carlo suite is an ideal tool for supporting the various stages of high-energy astronomy missions, ranging from the design, development, and performance prediction through calibration and response generation to data reduction. Version 1.0 of the MGGPOD software and documentation are publicly available for download at CESR. The package has been, and is being, successfully applied to an increasing number of past and present gamma-ray and hard X-ray missions. New applications often entail requirements for new functionalities, hence the MGGPOD suite is evolving continuously. The additions and improvements required for studying various instrument concepts for an Advanced Compton Telescope, Laue lens gamma-ray telescope focal plane detectors, and coded mask hard X-ray telescopes will be made available to the community in the forthcoming release of Version 1.1 of MGGPOD.
Boggs, S.E., et al., 2006a, Proc. of SPIE, 6266, 626624-1
Boggs, S.E., et al., 2006b, ACT Study Report, see [*astro-ph/0608532*]{} Dean, A. J., et al. 2003, Space Sci. Rev. 105, 285
Garson, A. et al., 2005, AAS meeting \#207
Garson, A. et al., 2006, Proc. of SPIE 6319, 63190D
Kippen, M., 2004, New Astr. Rev., 48, 221
Knödlseder, J., 2006, Proc. of SPIE 6266, 626623
McConnell, M., & Kippen, M., 2004, HEAD meeting \#8
Sturner, S., et al. 2003, A&A, 411, L81
Weidenspointner, G., et al. 2003a, A&A, 411, L113
Weidenspointner, G., et al. 2003b, New Astr. Rev. 48, 227
Weidenspointner, G., et al. 2004, in [*Proceedings of 5$^{th}$ INTEGRAL Workshop*]{}, ESA SP-552, 905
Weidenspointner, G., et al. 2005, ApJS, 156, 69
Weidenspointner, G., et al. 2006, Exp. Ast., 20, 377
Wunderer, C.B., et al., 2004, [*Proceedings of 5$^{th}$ INTEGRAL Workshop*]{}, ESA SP-552, 913
Wunderer, C.B., et al., 2006a, Exp. Ast., 20, 367
Wunderer, C.B., et al., 2006b, these proceedings
Zeitnitz, C., & Gabriel, T. A., 1994, NIM A, 349, 106
Zoglauer, A., et al., 2006, New Astr. Rev., 50, 629
[^1]: see [*http://wwwinfo.cern.ch/asd/geant/*]{}
[^2]: By default, the low-energy cutoff in GEANT is 10 keV. This cutoff energy can, however, be lowered, as e.g.described in the MGGPOD documentation. We have successfully lowered the cutoff energy down to 1 keV.
[^3]: see [*http://lheawww.gsfc.nasa.gov/docs/gamcosray/legr/mgeant/ mgeant.html*]{}
[^4]: see [*http://www.staff.uni-mainz.de/zeitnitz/Gcalor/gcalor.html*]{}
[^5]: see [*http://public.lanl.gov/mkippen/actsim/glecs/*]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider an asymmetric 0-$\pi$ Josephson junction consisting of $0$ and $\pi$ regions of different lengths $L_0$ and $L_\pi$. this system can be described by an effective sine-Gordon equation for the spatially averaged phase $\psi$ the effective current-phase relation of this system includes a *negative* second harmonic $\propto\sin(2\psi)$. If its amplitude is large enough, the ground state of the junction is doubly degenerate $\psi=\pm\varphi$, where $\varphi$ depends on the amplitudes of the first and second harmonics. We study the behavior of such a junction in an applied magnetic field $H$ and demonstrate that $H$ induces an additional term $\propto H \cos\psi$ in the effective current-phase relation. This results in a non-trivial ground state *tunable* by magnetic field. The dependence of the critical current on $H$ allows for revealing the ground state experimentally.'
author:
- 'E. Goldobin'
- 'D. Koelle'
- 'R. Kleiner'
- 'R. G. Mints'
bibliography:
- 'SFS.bib'
- 'SF.bib'
- 'pi.bib'
- 'LJJ.bib'
- 'ratch.bib'
- 'software.bib'
title: 'Josephson junction with magnetic-field tunable ground state'
---
\[Sec:Intro\]
Josephson junctions (JJs) with a phase shift of $\pi$ in the ground state[@Bulaevskii:pi-loop] attracted a lot of interest in the recent years[@Baselmans:1999:SNS-pi-JJ; @Ryazanov:2001:SFS-PiJJ; @Kontos:2002:SIFS-PiJJ; @Weides:2006:SIFS-HiJcPiJJ; @vanDam:2006:QuDot:SuperCurrRev; @Gumann:2007:Geometric-pi-JJ]. These JJs can be used as on-chip phase batteries for biasing various classical[@Ortlepp:2006:RSFQ-0-pi] and quantum[@Feofanov:2010:SFS:pi-qubit] circuits. This allows for removing external bias lines and reducing decoherence. Currently, it is possible to fabricate simultaneously both 0 and $\pi$ JJs using various technologies such as superconductor-ferromagnet heterostructures[@Weides:2006:SIFS-0-pi] or JJs based on d-wave superconductors[@VanHarlingen:1995:Review; @Tsuei:Review; @Smilde:ZigzagPRL].
It would be remarkable to have a JJ (a *phase battery*) providing an *arbitrary* phase shift $\varphi$, rather than just 0 or $\pi$. Long arrays of 0-$\pi$-0-$\pi$-$\ldots$ JJs with short segments were suggested as systems, where the $\varphi$ JJ can be realized[@Mints:1998:SelfGenFlux@AltJc; @Mints:2002:SplinteredVortices@GB; @Buzdin:2003:phi-LJJ]. Each of the segments are assumed to have a standard current-phase relation (CPR) $j_s=\pm j_0\sin\phi$. The phase $\phi$ in these systems can be written as a sum of two terms $\phi(x)=\psi+\xi(x)\sin\psi$, where $\psi$ is a constant spatially averaged phase and $|\xi(x)|\ll 1$ is alternating on a scale of the order of the segment’s length. The effective CPR then reads[@Mints:1998:SelfGenFlux@AltJc; @Mints:2002:SplinteredVortices@GB; @Buzdin:2003:phi-LJJ] $$j_s=j_1\sin(\psi) + j_2 \sin(2\psi)
, \label{Eq:CPR2}$$ where $j_1={\ensuremath{\left\langlej_c(x)\right\rangle_{}}\xspace}$ is the spatially averaged critical current density and $j_2<0$ is the amplitude of the effective second harmonic. The value of $j_2$ depends on parameters of the junctions. For $j_2<-j_1/2$ the ground state of the system is doubly degenerate with $\psi=\pm\varphi$, where $$\varphi = \arccos\left[ -{j_1}/(2j_2) \right]
. \label{Eq:varphi}$$ This $\varphi$ JJ[@Buzdin:2003:phi-LJJ] is the generalization of a $\pi$ JJ and can provide an *arbitrary* phase bias $0<\varphi<\pi$. These JJs have unusual physical properties[@Goldobin:CPR:2ndHarm], , non quantized Josephson vortices that were predicted[@Mints:1998:SelfGenFlux@AltJc] and observed experimentally[@Mints:2002:SplinteredVortices@GB].
The simplest system, which is suitable for the realization of a “short” $\varphi$ JJ, is an asymmetric (, $L_0 \neq L_\pi$) 0-$\pi$ JJ with relatively short segments of length $L_0,\,L_\pi\lesssim\lambda_J$, where $\lambda_J\propto 1/\sqrt{j_0}$ is the Josephson length. Such a system can be considered as one period of an infinitely long 0-$\pi$-0-$\pi$-$\ldots$ chain. The best method to reveal the $\varphi$ JJ experimentally is to include it in a superconducting loop (thus forming a SQUID) and to measure the spontaneously generated flux. Another option is to study these JJs in a magnetic field $H$. It is worth mentioning that there is an apparent contradiction here. On one hand, the usual linear phase ansatz $\phi(x)=Hx+\phi_0$ in asymmetric 0-$\pi$ JJ results in a critical current $I_c(H)$ with a cusp-like *minimum* at $H=0$, see Eq. (11) in Ref. as well as Refs. . This $I_c(H)$ should be mirror symmetric with respect to the $I_c$ axis, , $I_c(-H)=I_c(+H)$. On the other hand, the presence of the second harmonic in the CPR and the linear phase ansatz for $\psi$ leads to a Fraunhofer-like $I_{c}(H)$ with the *maximum* at $H=0$[@Goldobin:CPR:2ndHarm]. Moreover, when $|j_2|>0.5j_1$ there are two critical currents $I_{c\pm}(H)$, corresponding to depinning of the phase from different potential wells, but both of the $I_{c\pm}(H)$ dependences have the main maximum at $H=0$.[@Goldobin:CPR:2ndHarm]
In this paper we study the asymmetric 0-$\pi$ JJ in a magnetic field in detail. We demonstrate that the amplitude of the second harmonic $j_2$ in Eq. depends only on the JJ parameters, while the magnetic field $H$ induces an additional term $\propto H \cos\psi$. Further, we predict the dependence of $I_c(H)$ at small magnetic fields.
\[Sec:Theory\]
Consider an asymmetric JJ shown in Fig. \[Fig:Geometry\]. The dependence $j_c(x)$ is given by
$$\begin{aligned}
j_c(x)&=&-j_0,\quad \text{if }-L_\pi<x<0
; \label{Eq:jc(x):0}\\
j_c(x)&=&+j_0,\quad \text{if }0<x<L_0
\label{Eq:jc(x):pi}
\end{aligned}$$
\[Eq:jc(x)\]
![(Color online) Geometry of the 0-$\pi$ JJ. []{data-label="Fig:Geometry"}](sketch)
The behavior of the phase $\phi(x)$ is defined by the static sine-Gordon equation $$\phi''(x) - j_c(x)\sin[\phi(x)] =-\gamma
, \label{Eq:sG}$$ where the coordinate $x$ and lengths $L_{0,\pi}$ are normalized in the usual way to $\lambda_J$ and $\gamma=j/j_0$ is the normalized bias current density. $j_c(x)$ below is also normalized to $j_0$. Next, we write[@Mints:1998:SelfGenFlux@AltJc] $$j_c(x)={\ensuremath{\left\langlej_c\right\rangle_{}}\xspace} [1+g(x)]
, \label{Eq:g:def}$$ where $${\ensuremath{\left\langlej_c\right\rangle_{}}\xspace}
= \frac{1}{L_0+L_\pi}\int_{-L_\pi}^{+L_0} j_c(x)\,dx
= j_0\frac{L_0-L_\pi}{L_0+L_\pi}
\label{Eq:av(j_c)}$$ is the average value of the critical current density and $$g(x) =
\begin{cases}
g_\pi = \ratio{-2L_0 }{L_0-L_\pi},\text{ for } x<0;\\
g_0 = \ratio{+2L_\pi}{L_0-L_\pi},\text{ for } x>0,
\end{cases}
\label{Eq:g(x)}$$ is the deviation from the average value (${\ensuremath{\left\langleg(x)\right\rangle_{}}\xspace}=0$).
Thus, we rewrite Eq. (\[Eq:sG\]) in terms of $g(x)$ and arrive to $$\phi'' - {\ensuremath{\left\langlej_c\right\rangle_{}}\xspace}[1+g(x)]\sin\phi =-\gamma
. \label{Eq:sG(g)}$$ In what follows we treat short JJs ($L_0,L_\pi \lesssim 1$). Therefore, the solution $\phi(x)$ can be sought in the form[@Mints:1998:SelfGenFlux@AltJc] $$\phi(x)=\psi + \xi(x)\sin\psi
, \label{Eq:Ansatz}$$ where $\psi$ is a constant[^1] and $\xi(x)$ describes small variations of the phase around $\psi$, , $|\xi(x)|\ll 1$, ${\ensuremath{\left\langle\xi(x)\right\rangle_{}}\xspace}=0$. Substituting the ansatz into Eq. and expanding to the first order in $\xi(x)$ we get $$\xi'' \sin\psi - {\ensuremath{\left\langlej_c\right\rangle_{}}\xspace}[1+g(x)][1 + \xi(x)\cos\psi]\sin \psi = -\gamma
. \label{Eq:sG:approx}$$ Equation has two types of terms: the constant ones and varying ones. Note, that the term $g(x)\xi(x)$ has both the constant (average) part ${\ensuremath{\left\langleg(x)\xi(x)\right\rangle_{}}\xspace}$ and the deviation from the average $g(x)\xi(x)-{\ensuremath{\left\langleg(x)\xi(x)\right\rangle_{}}\xspace}$. Thus, from Eq. the relation for the constant terms reads $$\gamma = {\ensuremath{\left\langlej_c\right\rangle_{}}\xspace}\left[ \sin \psi + {\ensuremath{\left\langleg(x)\xi(x)\right\rangle_{}}\xspace} \sin \psi \cos \psi \right]
. \label{Eq:Slow}$$ The equation for $\xi(x)$, recalling definition , is $$\xi'' - j_c(x)\cos\psi \xi(x)
= {\ensuremath{\left\langlej_c\right\rangle_{}}\xspace}\left[ g(x)-{\ensuremath{{\ensuremath{\left\langleg\xi\right\rangle_{}}\xspace}}}\cos\psi \right]
, \label{Eq:Fast:2RHS}$$ It turns out that both terms $\propto \cos\psi$ have an extremely weak influence on the results. Therefore, for the sake of simplicity, we omit them right away to arrive to $$\xi'' = {\ensuremath{\left\langlej_c\right\rangle_{}}\xspace} g(x)
. \label{Eq:Fast:Simple}$$ Solutions of this equation include four constants to be determined from the matching conditions at $x=0$ $$\begin{aligned}
\xi_\pi(0)&=&\xi_0(0)
; \label{Eq:BC:xi(0)}\\
\xi'_\pi(0)&=&\xi'_0(0)
, \label{Eq:BC:xi'(0)}\end{aligned}$$ $$\xi_\pi(0) = \xi_0(0)
;\quad
\xi'_\pi(0) = \xi'_0(0)
, \label{Eq:BC:Zero}$$ and boundary conditions at $x=-L_\pi$ and $x=L_0$ $$\begin{aligned}
\xi'_\pi(-L_\pi) \sin\psi &=& h
; \label{Eq:BC:xi(Lpi)}\\
\xi'_0(L_0) \sin\psi &=& h
, \label{Eq:BC:xi(L0)}\end{aligned}$$ $$\xi'_\pi(-L_\pi) \sin\psi = h
; \quad
\xi'_0(L_0) \sin\psi = h
, \label{Eq:BC:Edges}$$ where $h=2H/H_{c1}$ is the normalized applied magnetic field, $H_{c1} = \Phi_0/(\pi \lambda_J \Lambda)$ is the penetration field and $\Lambda\approx 2\lambda_L$ is the effective magnetic thickness of the JJ. Thus, we arrive to the expression for $\xi(x)$ and can calculate $${\ensuremath{{\ensuremath{\left\langleg\xi\right\rangle_{}}\xspace}}}= \Gamma_0 + \Gamma_h \frac{h}{\sin\psi}
, \label{Eq:gxi}$$ where the coefficients $$\Gamma_0 = -\frac43 \frac{L_0^2 L_\pi^2}{L_0^2 - L_\pi^2}
;\quad
\Gamma_h = \frac{L_0 L_\pi}{L_0 - L_\pi}
. \label{Eq:Gammas}$$ Thus, Eq. gives the effective CPR of the JJ $$\gamma = {\ensuremath{\left\langlej_c\right\rangle_{}}\xspace}\left[
\sin \psi + \Gamma_h h \cos\psi + \frac{\Gamma_0}{2}\sin(2\psi)
\right]
. \label{Eq:CPR}$$ The result is remarkable — as in earlier works[@Mints:1998:SelfGenFlux@AltJc; @Mints:2002:SplinteredVortices@GB; @Buzdin:2003:phi-LJJ] one gets the second harmonic $\sin(2\psi)$ with the negative amplitude of $\Gamma_0/2$. In addition, the magnetic field results in a $\Gamma_h h \cos\psi$ term which additionally modifies the CPR and is tunable by magnetic field. Fig. \[Fig:CPR\](a) shows the effective CPRs $\gamma(\psi)$ of a 0-$\pi$ JJ for several different values of magnetic field $h$. At $h=0$ we might have a doubly degenerate ground state $\psi=\pm\varphi$ which, upon application of $h$, transforms into a single one.
![(Color online) (a) effective CPR $\gamma(\psi)$ and (b) effective Josephson energy $U(\psi)$ of a 0-$\pi$ JJ with $L_0=1$, $L_\pi=0.9$ for several different values of magnetic field $h$ given next to each curve. []{data-label="Fig:CPR"}](CPR)
The Josephson energy corresponding to CPR is $$U(\psi) = {\ensuremath{\left\langlej_c\right\rangle_{}}\xspace}\left[
1-\cos \psi + \Gamma_h h \sin\psi + \frac{\Gamma_0}{2} \sin^2 \psi
\right]
. \label{Eq:U(psi)}$$ The plots of $U(\psi)$ for the same set of $h$ values is shown in Fig. \[Fig:CPR\](b). Following the evolution of the curves at different $h$ one can see how the doubly degenerate ground state transforms into a single one. Moreover, one can see that at non-zero $h$ the potential energy $U(\psi)$ lacks reflection symmetry and therefore can be used to build Josephson phase ratchets[@Reimann:2002:BrownianMotors; @Haenggi:2009:ArtBrownMotors].
The effective CPR at $h=0$ allows to calculate the domain of existence of nontrivial solutions, which is defined as $|\Gamma_0|>1$ [@Goldobin:CPR:2ndHarm]. This results in $$L_\pi \geq L_0\sqrt{\ratio{3}{4L_0^2+3}}
;\quad L_0 \geq L_\pi\sqrt{\ratio{3}{4L_\pi^2+3}}
. \label{Eq:Mints:varphi-domain}$$ Since we have made certain approximations, namely used Eq. instead of Eq. , it is worth to compare the domain with the exact result[@Bulaevskii:0-pi-LJJ], which reads
$$\begin{aligned}
L_\pi &\geq& \arctan(\tanh(L_0))
; \label{Eq:Bulaevskii:0-varphi}\\
L_0 &\geq& \arctan(\tanh(L_\pi))
, \label{Eq:Bulaevskii:pi-varphi}
\end{aligned}$$
\[Eq:Bulaevskii:varphi-domain\]
for our case $j_c^0=j_c^\pi$. Both boundaries, the exact one given by Eq. and the approximate one given by Eq. are shown in Fig. \[Fig:domain\]. It is seen that our approximation works extremely well not only for small $L_0$ and $L_\pi$, but also in the limits $L_0, L_\pi \to \infty$, where it deviates from the asymptotic value by only $\sim 10\units{\%}$.
![ The domain of existence of a $\varphi$ state ($h=0$) on the $L_0,L_\pi$ plane. The exact boundary (continuous lines) and our approximation (dashed lines) agree very well. The $\varphi$ domain has regions of different colors from green (approximation $|\xi(x)\sin\psi|\ll 1$ works well) to red (approximation is invalid). The boundaries correspond to $|\xi_\mathrm{max}\sin\varphi|=0.1,\,0.2,\,0.4,\,0.8$. The dots indicate parameters used for discussion of Fig. \[Fig:Ic(H)\]. Vertical and horizontal dash-dotted lines show asymptotic behavior of exact boundary . []{data-label="Fig:domain"}](domain)
In the JJ with the CPR and energy , one may have several stable static solutions and several critical (depinning) currents corresponding to the escape of the phase from the relevant energy minimum. To find the critical currents for a given magnetic field, we look for extrema of the CPR with respect to $\psi$. We arrive to $$\cos\psi - \Gamma_h h \sin\psi - \Gamma_0 (2\cos^2\psi-1)=0
. \label{Eq:psi_c}$$ This problem can be reduced to a solution of a fourth order polynomial, , all solutions of Eq. can be found numerically. As a result we obtain up to four relevant roots and up to four corresponding critical currents.
Several examples of $\gamma_c(h)$ dependences are shown in Fig. \[Fig:Ic(H)\]. One can see that for the parameter set (a), which corresponds to a state deep in the $\varphi$ region at $h=0$, see Fig. \[Fig:domain\], one observes a characteristic rotated diamond-like shape with four critical currents in total for $|h|\lesssim0.6$. The two critical currents $\gamma_{c\pm}$ (in each direction) observed at $h=0$ coinside with $\gamma_{\pm}$ obtained in ref. . One may be able to observe the lower one experimentally in a system with low enough damping using a special sweep sequence[@Goldobin:CPR:2ndHarm]. For some values of $h$ the smaller minimum can even be absent at $\gamma=0$ \[see the arrow Fig. \[Fig:CPR\](b)\] and appear at larger $|\gamma|$ \[the vertical dotted line corresponding to $h=0.5$ in Fig. \[Fig:Ic(H)\](a)\]. The small energy minimum appears for $0.03<\gamma<0.06$, which also be seen in Fig. \[Fig:CPR\](b). Another distinct feature, which might be measurable in experiment, is the shift of the main minimum to $h=\pm0.11$ in Fig. \[Fig:Ic(H)\](a). This shift is the evidence of a non-trivial phase state in the junction. Finally, at large $|h|$ the upper branches of $\gamma_c(h)$ approach the asymptotic lines $$\gamma_c^\mathrm{as}(h) \approx \pm {\ensuremath{\left\langlej_c\right\rangle_{}}\xspace} \Gamma_h h
, \label{Eq:ass}$$ which are also shown in Fig. \[Fig:Ic(H)\]. We remind that our analysis is valid only for small values of $|h|$ and does not reproduce the global $\gamma_c(h)$ features, presented elsewhere.
When the 0-$\pi$ JJ gets more asymmetric, see Fig. \[Fig:Ic(H)\](b) and (c), the lower $\gamma_c(h)$ diamond-shaped domain becomes thinner and finally collapses. The point of collapse corresponds to the crossing of the $\varphi$-domain boundary in Fig. \[Fig:domain\]. Looking at $\gamma_c(h)$ one can see that the main cusp-like minimum shifts with asymmetry which can be used to extract the asymmetry from experimental data.
We also have performed numerical simulations of the $\gamma_c(h)$ dependence using <span style="font-variant:small-caps;">StkJJ</span> software. The above analytical results coincide with the results of direct numerical simulation with good accuracy. at $h \ll 1$
\[Sec:Conclusions\]
To summarize, we have shown that the effective CPR includes the term $\propto H\cos\psi$, which allows to tune the ground state. Thus, our system is a hybrid between $\varphi$ [@Buzdin:2003:phi-LJJ] and $\varphi_0$ [@Buzdin:2008:varphi0] JJs introduced earlier. The corresponding Josephson energy profile can be made asymmetric allowing to build Josephson phase ratchets. We have clarified how the $I_c(H)$ dependence of asymmetric 0-$\pi$ JJ looks like for low magnetic field. It does have a minimum, but the minimum is shifted to some $\pm h_\mathrm{min}$ for positive and negative bias, respectively, so that the $I_c(H)$ curve is point symmetric.
We acknowledge financial support by the German Israeli Foundation (Grant No. G-967-126.14/2007).
[^1]: In Ref. $\psi$ is a slowly changing function of $x$. Here, we have only one period (two segments) taken separately and assume $\psi=const$. The assumption $\psi=const$ makes the linear phase ansatz $\psi=hx+\psi_0$ used in Ref. invalid. Instead, $h$ enters in the boundary conditions for $\xi(x)$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'After a short introduction to the history of neutrino mixing and oscillations, we celebrate the 50th anniversary of the first analytical expression for the neutrino evolution probability obtained by Gribov and Pontecorvo in 1969 by the discussion of two new phenomena in neutrino oscillations: 1) the emergence of the neutrino spin and spin-flavour oscillations engendered by the transversal in respect to neutrino propagation matter currents, and 2) an inherent interplay of neutrino oscillations on the vacuum $\omega_{vac}=\frac{\Delta m^2}{4p}$ and magnetic $\omega_{B}=\mu B_{\perp}$ frequencies in the case of neutrino propagation in a constance magnetics field. We also predict a new phenomena of the modification of the flavour neutrino oscillations probability in moving matter that can be engendered by a non-vanishing matter transversal current ${\bf {j}}_{\perp}=n {\bf {v}}_{\perp}$.'
author:
- 'A. I. Studenikin[^1]'
title: New neutrino spin oscillations in moving matter and magnetic fields
---
\[1999/12/01 v1.4c Il Nuovo Cimento\]
Introduction {#Sec_2}
============
We give a fast look on several new aspects in the phenomena of neutrino oscillations in moving matter and a magnetic field discussed in our papers [@Pustoshny:2018jxb; @Studenikin:2004bu] and [@Popov:2019nkr; @Popov:2018seq]. By the present short notes we celebrate the 50th anniversary of the first analytic expression for the neutrino oscillation probability obtained in the paper [@Gribov:1968kq] by Gribov and Pontecorvo in 1969. Then we also predict a new phenomena of the modification of the flavour neutrino oscillations probability in moving matter that can be engendered by a non-vanishing matter transversal current $\bm{j}_{\perp}=n \bm {v}_{\perp}$.
A brief history of neutrino mixing and oscillations
===================================================
Neutrino flavour oscillations in vacuum and matter
--------------------------------------------------
The story of the neutrino mixing and oscillations started with two papers by Bruno Pontecorvo [@Pontecorvo:1957cp; @Pontecorvo:1957qd] where the above mentioned effects have been discussed for the first time. A history of neutrino mixing and oscillations can be found in [@Bilenky:2016pep]. In [@Pontecorvo:1957cp] Pontecorvo has indicated that if the neutrino charge were not conserved then the transition between a neutrino and antineutrino would become possible in vacuum. In [@Pontecorvo:1957qd] Pontecorvo has even directly introduced a phenomenon of neutrino mixing. In 1962, just after the discovery of the second flavour neutrino, the effect of neutrino mixing was discussed in [@Maki:1962mu] where the fields of the weak neutrinos $\nu_e$ and $\nu_{\mu}$ were connected with the neutrinos mass states $\nu_1$ and $\nu_2$ by the unitary mixing matrix $U$ that can be parametrized by the mixing angle $\theta$ and $$\nu_e = \nu_1 \cos \theta + \nu_2 \sin \theta, \ \ \ \nu_\mu = - \nu_1 \sin \theta + \nu_2 \cos \theta .$$
The theory of neutrino mixing and oscillations was further developed in [@Gribov:1968kq; @Bilenky:1975tb] with actual calculations of neutrino beam evolution. In [@Wolfenstein:1977ue] the effect of neutrino interaction with matter of a constant density on neutrino flavour mixing and oscillations was investigated. The existence of resonant amplification of neutrino mixing (the MSW effect) when a neutrino flux propagates through a medium with varying density was predicted in [@Mikheev:1986gs]. Generalization of the MSW resonance condition for the case of moving matter was derived in [@Grigoriev:2002zr].
The tedious studies, both experimental and theoretical, over the last 60 years has been honored by the Nobel Prize of 2015 awarded to Arthur McDonald and Takaaki Kajita for the discovery of neutrino oscillations, which shows that neutrinos have mass.
[*[Neutrino spin oscillations in magnetic fields.]{}*]{} The straightforward consequences of neutrino nonzero mass is the prediction [@Fujikawa:1980yx] that neutrinos can have nonzero magnetic moments. Studies of neutrino magnetic moments and the related phenomena attract a reasonable interest in literature. The values of neutrino magnetic moments are constrained in the terrestrial laboratory experiments and in the astrophysical considerations (see, for instance, [@Beda:2012zz; @Borexino:2017fbd] and [@Raffelt:1990pj]).
Massive neutrinos participate in electromagnetic interactions. The recent review on this topic is given in [@Giunti:2014ixa] (the upgrate can be found in [@Studenikin:2018vnp]). One of the most important phenomenon of nontrivial neutrino electromagnetic interactions is the neutrino magnetic moment precession and the corresponding spin oscillations in presence of external electromagnetic fields. The later effect has been studied in numerous papers published during several passed decades.
Within this scope the neutrino spin oscillations $\nu^{L}\Leftrightarrow \nu^{R}$ induced by the neutrino magnetic moment interaction with the transversal magnetic field ${\bf B}_{\perp}$ was first considered in [@Cisneros:1970nq]. Then spin-flavor oscillations $\nu^{L}_{e}\Leftrightarrow \nu^{R}_{\mu}$ in ${\bf B}_{\perp}$ in vacuum were discussed in [@Schechter:1981hw], the importance of the matter effect was emphasized in [@Okun:1986hi]. The effect of the resonant amplification of neutrino spin oscillations in ${\bf B}_{\perp}$ in the presence of matter was proposed in [@Akhmedov:1988uk; @Lim:1987tk]. Neutrino spin oscillations in a magnetic field with account for the effect of moving matter was studied in [@Lobanov:2001ar]. A possibility to establish conditions for the resonance in neutrino spin oscillations by the effect of matter motion discussed in [@Studenikin:2004bu]. The impact of the longitudinal magnetic field ${\bf B}_{||}$ was discussed in [@Akhmedov:1988hd]. The neutrino spin oscillations in the presence of constant twisting magnetic field were considered in [@Vidal:1990fr; @Smirnov:1991ia; @Akhmedov:1993sh; @Likhachev:1990ki; @Dvornikov:2007aj; @Dmitriev:2015ega].
In [@Egorov:1999ah] neutrino spin oscillations were considered in the presence of arbitrary constant electromagnetic fields $F_{\mu \nu}$. Neutrino spin oscillations in the presence of the field of circular and linearly polarized electromagnetic waves and superposition of an electromagnetic wave and constant magnetic field were considered in [@Lobanov:2001ar; @Dvornikov:2001ez]. The effect of the parametric resonance in neutrino oscillations in periodically varying electromagnetic fields was studied in [@Dvornikov:2004en].
More general case of neutrino spin evolution in the case when neutrino is subjected to general types of non-derative interactions with external scalar $s$, pseoudoscalar $\pi$, vector $V_{\mu}$, axial-vector $A_{\mu}$, tensor $T_{\mu\nu}$ and pseudotensor $\Pi_{\mu\nu}$ fields was considered in [@Dvornikov:2002rs]. From the general neutrino spin evolution equation, obtained in [@Dvornikov:2002rs], it follows that neither scalar $s$ nor pseudoscalar $\pi$ nor vector $V_{\mu}$ fields can induce neutrino spin evolution. On the contrary, within the general consideration of neutrino spin evolution it was shown that electromagnetic (tensor) and weak (axial-vector) interactions can contribute to the neutrino spin evolution.
We have also considered in details [@Fabbricatore:2016nec; @Studenikin:2016zdx; @Kurashvili:2017zab] neutrino mixing and oscillations in arbitrary constant magnetic field that have ${\bf B}_{\perp}$ and ${\bf B}_{||}$ nonzero components and derived an explicit expressions for the effective neutrino magnetic moments for the flavour neutrinos in terms of the corresponding magnetic moments introduced in the neutrino mass basis.
New effects in neutrino flavour, spin and spin-flavour oscillations in a constant magnetic field {#osc_B}
================================================================================================
Recently a new approach to description of neutrino spin and spin-flavor oscillations in the presence of an arbitrary constant magnetic field has been developed (see also [@Dmitriev:2015ega; @Studenikin:2016zdx]). Within the new approach exact quantum stationary states are used for classification of neutrino spin states, rather than the neutrino helicity states that have been used for this purpose within the customary approach in many published papers. Recall that the helicity states are not stationary in the presence of a magnetic field. It has been shown [@Popov:2019nkr; @Popov:2018seq], in particular, that in the presence of the transversal magnetic field $B_{\perp}$ for a given choice of parameters (the energy and magnetic moments of neutrinos and strength of the magnetic field) the amplitude of the flavour oscillations $\nu_e^L \Leftrightarrow \nu_{\mu}^L$ at the vacuum frequency $\omega_{vac}=\frac{\Delta m^2}{4p}$ is modulated by the magnetic field frequency $\omega_{B}=\mu B_{\perp}$: $$\label{fl_simp}
P^{(B_{\perp})}_{\nu_{e}^L \rightarrow \nu_{\mu}^L}(t) = \left( 1 - \sin^2(\mu B_{\perp}t) \right) \sin^2 2\theta \sin^2 \frac{\Delta m^2}{4p}t = \left(1 - P_{\nu_{e}^L \rightarrow \nu_{e}^R}^{cust}\right)
P_{\nu_{e}^L \rightarrow \nu_{\mu}^L}^{cust},$$ here $\mu$ is the effective magnetic moment of the electron neutrino and it is supposed that the following relations between diagonal and transition magnetic moments in the neutrino mass basis are valid: $\mu_1 = \mu_2, \ \ \mu_{ij}=0, \ i\neq j$. The customary expression $$\label{flavour_cust}
P_{\nu_{e}^L \rightarrow \nu_{\mu}^L}^{cust}(t)= \sin^2 2\theta \sin^2 \frac{\Delta m^2}{4p}t$$ for the neutrino flavour oscillation probability in vacuum in the presence of the transversal field $B_{\perp}$ is modified by the factor $1 - P_{\nu_{e}^L \rightarrow \nu_{e}^R}^{cust}$. Since the transition magnetic moment in the flavour basis is absent in the case $\mu_1 = \mu_2$, the process $\nu_e^L \rightarrow \nu_e^R$ is the only way for spin flip, and then $1 - P_{\nu_{e}^L \rightarrow \nu_{e}^R}^{cust}$ should be interpreted as the probability of not changing the neutrino spin polarization. And consequently, this multiplier subtracts the contribution of neutrinos $\nu^{R}_e$ with the opposite polarization providing the survival of the only contribution from the direct neutrino flavour oscillations $\nu_e^L \Leftrightarrow \nu_{\mu}^L$.
Similar results on the important influence of the transversal magnetic field on amplitudes of various types of neutrino oscillations were obtained earlier [@Kurashvili:2017zab] on the basis of the exact solution of the effective equation for neutrino evolution in the presence of a magnetic field, which accounts for four neutrino species corresponding to two different flavor states with positive and negative helicities.
Consider the probability of the neutrino spin-flavour oscillations $\nu_e^L \leftrightarrow \nu_{\mu}^R$. In the case $\mu_1 = \mu_2 = \mu$ we have [@Popov:2019nkr; @Popov:2018seq]: $$\label{spin_flavour_simplified}
P_{\nu_e^L \rightarrow \nu_{\mu}^R}(t) = \sin^2(\mu B_{\perp} t)\sin^2 2\theta \sin^2\frac{\Delta m^2}{4p}t.$$ The obtained expression (\[spin\_flavour\_simplified\]) for the probability can be expressed as a product of two probabilities derived within the customary two-neutrino-states approach $$\label{PP}
P_{\nu_e^L \rightarrow \nu_{\mu}^R} (t)=
P_{\nu_{e}^L \rightarrow \nu_{\mu}^L}^{cust} (t) P_{\nu_{e}^L \rightarrow \nu_{e}^R}^{cust}(t),$$ where the probability $P_{\nu_e^L \rightarrow \nu_{\mu}^L}^{cust}(t)$ given by (\[flavour\_cust\]) and the usual expression for the neutrino spin oscillation probability $$\label{P_mu_B}
P_{\nu_{e}^L \rightarrow \nu_{e}^R}^{cust}(t) = \sin^2(\mu B_{\perp}t),$$ are just the probabilities obtained in the customary approach. A similar neutrino spin-flavour oscillations (for the Majorana case) as a two-step neutrino conversion processes were considered in [@Akhmedov:2002mf]. Since the probability of neutrino spin-flavour oscillations was supposed to be small, this effect was calculated [@Akhmedov:2002mf] within perturbation theory.
Finally, in the case of the spin oscillations $\nu_{e}^L \rightarrow \nu_{e}^R$ in the transversal magnetic field $B_{\perp}$ within the the developed approach the effect of neutrino mixing is accounted for that leads to the modification of the customary expression $P_{\nu_{e}^L \rightarrow \nu_{e}^R}^{cust}(t)$ for the probability by the factor $1 - P_{\nu_{e}^{L} \rightarrow \nu_{\mu}^{L}}^{cust}$: $$\begin{aligned}
\label{spin_simplified}
P_{\nu_{e}^L \rightarrow \nu_{e}^R} = \left[ 1 - \sin^2 2\theta \sin^2\left(\frac{\Delta m^2}{4p}t\right) \right]\sin^2(\mu B_{\perp}t).\end{aligned}$$
The interplay between different oscillations, that follows from the obtained expressions (\[fl\_simp\]), (\[spin\_flavour\_simplified\]) and (\[spin\_simplified\]) for the oscillation probabilities, gives rise to interesting phenomena (see also [@Popov:2019nkr; @Popov:2018seq]):
1\) the amplitude modulation of the probability of flavour oscillations $\nu_e^L \rightarrow \nu_{\mu}^L$ in the transversal magnetic field with the magnetic frequency $\omega_{B}=\mu B_{\perp}$ (in the case $\mu_1 = \mu_2$) and more complicated dependence on the harmonic functions with $\omega_{B}$ for $\mu_1 \neq \mu_2$;
2\) the dependence of the spin oscillation probability $P_{\nu_{e}^L \rightarrow \nu_{e}^R}$ on the mass square difference $\Delta m^2$;
3\) the appearance of the spin-flavour oscillations in the case $\mu_1=\mu_2$ and $\mu_{12}=0$, the transition goes through the two-step processes $\nu_e^L \rightarrow \nu_{\mu}^L \rightarrow \nu_{\mu}^R$ and $\nu_e^L \rightarrow \nu_{e}^R \rightarrow \nu_{\mu}^R$.
As a result, we predict modifications of the neutrino oscillation patterns that might provide new important phenomenological consequences in case of neutrinos propagation in extreme astrophysical environments where magnetic fields are present.
New effect of neutrino spin and spin-flavour oscillations in transversal matter current {#osc_j}
=======================================================================================
Following the discussion in [@Studenikin:2004bu] consider, as an example, an electron neutrino spin precession in the case when neutrinos with the Standard Model interaction are propagating through moving and polarized matter composed of electrons (electron gas) in the presence of an electromagnetic field given by the electromagnetic-field tensor $F_{\mu \nu}=({\bf E}, {\bf B})$. To derive the neutrino spin oscillation probability in the transversal matter current we use the generalized Bargmann-Michel-Telegdi equation that describes the evolution of the three-dimensional neutrino spin vector $\bf S $, $$\label{S} {d{\bf S}
\over dt}={2 \mu \over \gamma} \Big[ {\bm S} \times ({\bm
B}_0+{\bm M}_0) \Big],$$ where the magnetic field $\bf{B}_0$ in the neutrino rest frame. The matter term ${\bf M}_0$ in Eq. (\[S\]) is composed of the transversal ${\bf M}{_{0_{\parallel}}}$ and longitudinal ${\bf M}_{0_{\perp}}$ parts in respect to the direction of the neutrino propagation , $${\bf M}_0=\bf {M}{_{0_{\parallel}}}+{\bf M}_{0_{\perp}},
\label{M_0}$$ where $$\label{M_0_parallel}
{\bf M}_{0_{\parallel}}={{{G}_{F}(1+4\sin^2 \theta _W) \over {2\sqrt{2}\mu \sqrt {1- v_{e}^{2}}
}}}n_{0}\left(1-{{\bf v}_e
{\bm\beta} \over {1- {\gamma^{-2}}}} \right)\gamma{\bm\beta},$$ $$\label{M_0_perp}
{\bf M}_{0_{\perp}}=-{{{G}_{F}(1+4\sin^2 \theta _W) \over {2\sqrt{2}\mu \sqrt {1-v_{e}^{2}}}}}
n_{0}{\bf v}_{e_{\perp}},$$ here $n_0=n_{e}\sqrt {1-v^{2}_{e}}$ is the invariant number density of the background particles in the rest reference frame of matter, $\gamma = (1-\beta^2)^{-{1 \over 2}}$, $\bm{\beta}$ is the neutrino velocity and ${\bf v}_e$ the velocity of matter.
The probability of the neutrino spin oscillations engendered by the transversal matter current (by the term ${\bf M}_{0_{\perp}}$) was obtained for the first time in [@Studenikin:2004bu] $$\label{ver2}
P_{\nu^{L}_{e} \rightarrow \nu^{R}_{e}} (x)=\sin^{2} 2\theta_\textmd{eff}
\sin^{2}{\pi x \over L_\textmd{eff}}, \ \ \sin^{2} 2\theta_\textmd{eff}={M_{0\perp}^2\over
{M_{0\parallel}^2 +M_{0\perp}^2}}, \ \ \
L_\textmd{eff}={\pi \over {\mu M_0}}\gamma.$$ Note that the possibility of neutrino spin oscillations due to interaction with matter of a rotating astrophysical object was also indicated in [@Studenikin:2004tv].
The introduced above semiclassical theory of neutrino spin oscillations engendered by the transversal matter currents has been recently appended by the direct quantum treatment of the phenomenon [@Pustoshny:2018jxb; @Studenikin:2016iwq; @Studenikin:2017pag; @Studenikin:2017mdh] Consider two flavour neutrinos with two possible helicities $\nu_{f}= (\nu_{e}^{+}, \nu_{e}^{-}, \nu_{\mu}^{+}, \nu_{\mu}^{-})^T$ in moving matter composed of neutrons. Each of the flavour neutrinos is a superposition of the neutrino mass states, $$\label{transformations}
\begin{array}{c}
%\begin{tabular}{ccc}
\nu_{e}^{\pm} =\nu_{1}^{\pm}\cos\theta+\nu_{2}^{\pm}\sin\theta,\ \ \ \ \
\nu_{\mu}^{\pm}=-\nu_{1}^{\pm}\sin\theta+\nu_{2}^{\pm}\cos\theta.
%\end{tabular}
\end{array}$$ The corresponding neutrino evolution equation is $$\label{schred_eq_fl}
i\dfrac{d}{dt}\nu_{f}=H^{f}_{v}\nu_{f},$$ where the evolution Hamiltonian reads $$\label{H_v}
H^f_v=n\tilde{G}\left( \begin{matrix}
0& (\frac{\eta}{\gamma})_{ee}v_{\perp} &0 &(\frac{\eta}{\gamma})_{e\mu}v_{\perp}\\
(\frac{\eta}{\gamma})_{ee}v_{\perp} & 2(1-v_{\parallel}) &(\frac{\eta}{\gamma})_{e\mu}v_{\perp}&
0\\
0& (\frac{\eta}{\gamma})_{e\mu}v_{\perp} &0 &(\frac{\eta}{\gamma})_{\mu \mu}v_{\perp}\\
(\frac{\eta}{\gamma})_{e\mu}v_{\perp} & 0 &(\frac{\eta}{\gamma})_{\mu\mu}v_{\perp}& 2(1-v_{\parallel})\\
\end{matrix} \right),$$ and $$\label{eta}
\Big(\frac{\eta}{\gamma}\Big)_{ee}=
\frac{\cos^2\theta}{\gamma_{11}}+\frac{\sin^2\theta}{\gamma_{22}}, \ \
\Big(\frac{\eta}{\gamma}\Big)_{\mu\mu}=
\frac{\sin^2\theta}{\gamma_{11}}+\frac{\cos^2\theta}{\gamma_{22}}, \ \
\Big(\frac{\eta}{\gamma}\Big)_{e\mu}=
\frac{\sin 2\theta}{\tilde{\gamma}_{21}},$$ where ${\widetilde{\gamma}_{2 1}}^{-1}=\frac{1}{2}\big(
\gamma_{2}^{-1}-\gamma_{1}^{-1}\big),\ \ \gamma_{\alpha}^{-1}=\frac{m_\alpha}{p^{\nu}_0}$, $p^{\nu}_0$ is neutrino energy and $\alpha = 1,2 $.
Consider the initial neutrino state $\nu_{e}^L$ moving in the background with the magnetic field ${\bm B} = {\bm B}_{\parallel}+{\bm B}_{\perp}$ and nonzero matter current ${\bm j} = {\bm j}_{\parallel}+{\bm j}_{\perp}$. One of the possible modes of neutrino transitions with the change of helicity is $\nu_{e}^L\Leftarrow (j_{\perp}, B_{\perp}) \Rightarrow \nu_{e}^R$. Here we restrict our consideration to the to the The corresponding oscillations are governed by the evolution equation [@Pustoshny:2018jxb] $$\begin{aligned}
\label{nu_L_nu_R}
i\frac{d}{dt} \begin{pmatrix}\nu^L_{e} \\ \nu^R_{e} \\ \end{pmatrix}=
\left( \begin{matrix}
(\frac{\mu}{\gamma})_{ee}{B_{||}}+{\widetilde{G}}n(1-{\bm v}{\bm \beta})\\
\mu_{ee}B_{\perp}+ (\frac{\eta}{\gamma})_{ee}{\widetilde{G}}nv_{\perp}
\end{matrix} \right.
\left. \begin{matrix} \mu_{ee}B_{\perp} + (\frac{\eta}{\gamma})_{ee}{\widetilde{G}}nv_{\perp} \\
- (\frac{\mu}{\gamma})_{ee}{B_{||}} -{\widetilde{G}}n(1-{\bm v}{\bm \beta} \end{matrix} \right)
\begin{pmatrix}\nu^L_{e} \\ \nu^R_{e} \\ \end{pmatrix}.\end{aligned}$$ Here we constraint our consideration to the binary neutrino transitions and corresponding oscillations between pairs of the neutrino states. The oscillation $\nu^L_{e} \Leftarrow (j_{\perp}, B_{\perp}) \Rightarrow \nu^R_{e}$ probability is given by $$\label{prob_oscillations}
P^{(j_{\perp},B_{\perp})}_{\nu^L_{e} \rightarrow \nu^R_{e}} (x)={E^2_\textmd{eff} \over
{E^{2}_\textmd{eff}+\Delta^{2}_\textmd{eff}}}
\sin^{2}{\pi x \over L_\textmd{eff}}, \ \ \ L_\textmd{eff}={\pi \over \sqrt{E^{2}_\textmd{eff}+\Delta^{2}_\textmd{eff}}},$$ where $$\label{E}
E_{eff}= \Big|\mu_{ee}\bm{B}_{\perp} + \Big(\frac{\eta}{\gamma}\Big)
_{ee}{\widetilde{G}}n\bm{v}_{\perp} \Big|, \ \ \
\Delta_{eff}= \Big|\Big(\frac{\mu}{\gamma}\Big)_{ee}{\bm{B}_{||}}+
{\widetilde{G}}n(1-{\bm v}{\bm \beta}){\bm {\beta}} \Big|.$$ From (\[prob\_oscillations\]) and (\[E\]) it is clearly seen that even in the absence of the transversal magnetic field the neutrino spin oscillations $\nu_{e}^L\Leftarrow (j_{\perp}) \Rightarrow \nu_{e}^R$ can be generated by the neutrino interaction with the transversal matter current $\bm{j}_{\perp}=n\bm{v}_{\perp}$.
In a quite analogous way the transversal matter current can engender the neutrino spin-flavour conversion and the corresponding oscillations $\nu_{e}^L\Leftarrow (j_{\perp}) \Rightarrow\nu_{\mu}^R$ between two different flavour states with opposite spin orientations. Note that for apperance of this effect there is no need for a magnetic field (obviously, the effect is neutrino magnetic moments independent). For the neutrino evolution $\nu_{e}^L\Leftarrow (j_{\perp}) \Rightarrow\nu_{\mu}^R$ in the case $B=0$ from (\[H\_v\]) we get $$\begin{aligned}
i\frac{d}{dt} \begin{pmatrix}\nu^L_{e} \\ \nu^R_{\mu} \\ \end{pmatrix}=
\left( \begin{matrix}
-\Delta M+{\widetilde{G}}n(1-{\bm v}{\bm \beta}) \\
(\frac{\eta}{\gamma})_{e\mu}{\widetilde{G}}nv_{\perp} \end{matrix} \right.
\left. \begin{matrix} (\frac{\eta}{\gamma})_{e\mu}{\widetilde{G}}nv_{\perp} \\
\Delta M -{\widetilde{G}}n(1-{\bm v}{\bm \beta})
\end{matrix} \right)
\begin{pmatrix}\nu^L_{e} \\ \nu^R_{\mu} \\ \end{pmatrix},\end{aligned}$$ where $\Delta M=\frac{\Delta m ^2 \cos 2\theta}{4 p^{\nu}_0 }$. The probability $P_{\nu_{e}^L \rightarrow \nu_{\mu}^R}^{(j_{\perp})}$ of the neutrino spin-flavour oscillations $\nu^L_{e} \Leftarrow (j_{\perp}) \Rightarrow \nu^R_{\mu}$ is given by the same equation (\[prob\_oscillations\]), but now $$\begin{aligned}
\label{E_1_delta_1}
E_{eff}= \Big|\Big(\frac{\eta}{\gamma}\Big)_{e\mu}{\widetilde{G}}n{v}_{\perp} \Big|, \ \
\Delta_{eff}
=\Big|\Delta M-{\widetilde{G}}n(1-{\bm v}{\bm \beta}) \Big|.\end{aligned}$$ From (\[E\_1\_delta\_1\]) it follows that the neutrino spin-flavour oscillations $\nu^L_{e} \Leftarrow (j_{\perp}) \Rightarrow \nu^R_{\mu}$ can be generated by the neutrino interaction with the transversal matter current $\bm{j}_{\perp}=n\bm{v}_{\perp}$.
New effect: Modification of flavour neutrino oscillations probability in moving matter by transversal matter current
====================================================================================================================
From the previous discussions it follows that in the presence of a magnetic field (see Section \[osc\_B\] and also [@Popov:2019nkr; @Popov:2018seq] for details) it is not possible to consider the neutrino flavour and spin oscillations as separate phenomena. On the contrary, there is an inherent communication between two. In particular, the amplitude of the neutrino flavour oscillations is modulated by the magnetic frequency $\omega_{B}=\mu B_{\perp}$. The main result of Section \[osc\_j\] (see also [@Studenikin:2004bu; @Pustoshny:2018jxb]) is the conclusion on the equal role that the transversal magnetic field ${\bm B}_{\perp}$ and the transversal matter current $\bm{j}_{\perp}$ plays in generation of the neutrino spin and spin-flavour oscillations. From these observations **we predict a new phenomena of the modification of the flavour neutrino oscillations probability in moving matter that can be engendered by a non-vanishing matter transversal current $\bm{j}_{\perp}=n \bm {v}_{\perp}$** .
The flavour neutrino oscillation probability accounting for this effect can be expressed, under quite reasonable conditions, as follows: $$\label{flav_mod_current}
P^{(j_{||}+j_{\perp})}_{\nu_{e}^L \rightarrow \nu_{\mu}^L} (t) = \left(1 - P_{\nu_{e}^L \rightarrow \nu_{e}^R}^{(j_{\perp})} - P_{\nu_{e}^L \rightarrow \nu_{\mu}^R}^{(j_{\perp})}\right)
P_{\nu_{e}^L \rightarrow \nu_{\mu}^L}^{(j_{||})},$$ where $$P_{\nu_{e}^L \rightarrow \nu_{\mu}^L}^{(j_{||})}(t) =
\sin^2 2\theta _{eff}\sin^2 \omega_{eff}t, \ \ \omega_{eff}=\frac{\Delta m^2 _{eff}}{4p_{0}^\nu},$$ is the flavour oscillation probability in moving matter [@Grigoriev:2002zr], $\theta _{eff}$ and $\Delta m^2 _{eff}$ are the corresponding quantities modified by the presence of moving matter (note that in the definition of $\theta _{eff}$ and $\Delta m^2 _{eff}$ only the longitudinal component of matter motion matters). For the probability of the neutrino spin oscillations engendered by the transversal current $\bm {j}_{\perp}$ from (\[nu\_L\_nu\_R\]) we get $$P^{j_{\perp}}_{\nu_{e}^L \rightarrow \nu_{e}^R}(t)=
\frac{\Big(\frac{\eta}{\gamma}\Big)_{ee}^2 {v}^{2}_{\perp}}
{\Big(\frac{\eta}{\gamma}\Big)_{ee}^2 {v}^{2}_{\perp}+
(1-{\bm v}{\bm \beta})^2}
\sin^2\omega^{j_{\perp}}_{ee}t.$$ For the corresponding probability of the neutrino spin-flavour oscillations due to $\bm {j}_{\perp}$ from (\[E\_1\_delta\_1\]) we get $$P^{j_{\perp}}_{\nu_{e}^L \rightarrow \nu_{\mu}^R}(t)=
\frac{\Big(\frac{\eta}{\gamma}\Big)_{e\mu}^2 {v}^{2}_{\perp}}
{\Big(\frac{\eta}{\gamma}\Big)_{e\mu}^2 {v}^{2}_{\perp}+
\Big(\frac{\Delta M}{{\widetilde{G}}n} - (1-{\bm v}
{\bm \beta})\Big)^2}
\sin^2\omega^{j_{\perp}}_{e\mu}t.$$ The discussed new effect of the modification of the flavour oscillations $\nu_{e}^L\Leftarrow (j_{||}, j_{\perp})\Rightarrow \nu_{\mu}^L$ probability is the result of an interplay of oscillations on a customary flavour oscillation frequency in moving matter $\omega_{eff}$ and two additional oscillations with changing the neutrino polarization, the neutrino spin $\nu_{e}^L\Leftarrow (j_{\perp}) \Rightarrow \nu_{e}^R$ and spin-flavour $\nu_{e}^L\Leftarrow (j_{\perp}) \Rightarrow \nu_{\mu}^R$ oscilations, that are governed by two characteristic frequencies $$\label{omega_1}
\omega^{j_{\perp}}_{ee}={\widetilde{G}}n{\sqrt{\Big(\frac{\eta}{\gamma}\Big)_{ee}^2 {v}^{2}_{\perp} +
(1-{\bm v}{\bm \beta})^2}}$$ and $$\label{omega_2}
\omega^{j_{\perp}}_{e\mu}={\widetilde{G}}n
{\sqrt{\Big(\frac{\eta}{\gamma}\Big)_{e\mu}^2 {v}^{2}_{\perp}
+
\Big(\frac{\Delta M}{{\widetilde{G}}n}-(1-{\bm v}{\bm \beta})\Big)^2}}.$$ The discussed interplay of oscillations on different frequencies can lead to important consequences in the case when neutrino fluxes are propagating in astrophysical environments peculiar to rotating compact objects or dense jets of matter.
[*Acknowledgements.*]{} The authors thanks Giorgio Bellettini, Giorgio Chiarelli, Mario Greco and Gino Isidori for the kind invitation to participate in the XXXIII Recontres de Physique de La Vallee D’Aoste on Results and Perspectives in Particle Physics and also thanks all the organizers for their hospitality in La Thuile.
[0]{}
Pustoshny P and Studenikin A 2018 [*Phys. Rev. *]{} D [**98**]{} 113009. Studenikin A 2004 [*Phys. Atom. Nucl.*]{} [**67**]{} 993. Popov A and Studenikin A 2019 [*Eur. Phys. J. *]{} C [**79**]{} 144 \[arxiv:1902.08195 \[hep-ph\]\]. Popov A and Studenikin A, arXiv:1803.05755 \[hep-ph\]. Gribov V and Pontecorvo B 1969 [*Phys. Lett. *]{} [**28B**]{} 493. Pontecorvo B 1957 [*Sov. Phys. JETP* ]{} [**6**]{} 429 \[1957 [*Zh. Eksp. Teor. Fiz. *]{} [**33**]{} 549\]. Pontecorvo B 1958 [*Sov. Phys. JETP* ]{} [**7**]{} 172 \[1957 [*Zh. Eksp. Teor. Fiz. *]{} [**34**]{} 247\]. Bilenky A 2016 [*Nucl. Phys. *]{} B [**908**]{} 2. Maki Z, Nakagawa M and Sakata S 1962 [*Prog. Theor. Phys. *]{} [**28**]{} 870. Bilenky S and Pontecorvo B 1976 [*Phys. Lett. *]{} [**61B**]{} 248. Wolfenstein L 1978 [*Phys. Rev. *]{} D [**17**]{} 2369. Mikheev S and Smirnov A 1985 [*Sov. J. Nucl. Phys. *]{} [**42**]{} 913 \[[*Yad. Fiz. *]{} [**42**]{} 1441\]. Grigoriev A, Lobanov A and Studenikin A 2002 [*Phys. Lett. *]{} B [**535**]{} 187. Fujikawa K and Shrock R 1980 [*Phys. Rev. Lett. *]{} [**45**]{} 963. Beda A, Brudanin V, Egorov V, Medvedev D [*et al.*]{} 2012 [*Adv.High Ener.Phys.*]{} [**2012**]{} 350150. Agostini M [*et al.*]{} \[Borexino Collaboration\] 2017) [*Phys. Rev. * ]{} D [**96**]{} 091103. Raffelt G 1990 [*Phys. Rev. Lett.* ]{} [**64**]{} 2856. Giunti C and Studenikin A 2015 [*Rev. Mod. Phys. *]{} [**87**]{} 531. Studenikin A 2017 [*PoS EPS-HEP2017*]{} 137. Cisneros A 1971 [*Astrophys.Space Sci.*]{} [**10**]{} 87. Schechter J and Valle J W F 1981 [*Phys. Rev. * ]{} D [**24**]{} 1883. Okun L, Voloshin M and Vysotsky M 1986 [*Sov. J. Nucl. Phys. *]{} [**44**]{} 440. Akhmedov E 1988 [*Phys. Lett. *]{} B [**213**]{} 64. Lim C S and Marciano W J 1988 [*Phys. Rev. *]{} D [**37**]{} 1368. Lobanov A and Studenikin A 2001 [*Phys. Lett. *]{} B [**515**]{} 94. Akhmedov E and Khlopov M 1988 [*Mod. Phys. Lett.* ]{} A [**3**]{} 451. Vidal J and Wudka J 1990 [*Phys. Lett. * ]{} B [**249**]{} 473. Smirnov A 1991 [*Phys. Lett. *]{} B [**260**]{} 161. Akhmedov E, Petcov S and Smirnov A 1993 [*Phys. Rev. *]{} D [**48**]{} 2167. Likhachev G and Studenikin A 1995 [*J.Exp.Theor.Phys.*]{} [**81**]{} 419. Dvornikov M 2008 [*J. Phys. *]{} G [**35**]{} 025003 Dmitriev A, Fabbricatore R and Studenikin A 2015 [*PoS CORFU* ]{} [**2014**]{} 050. Egorov A, Lobanov A and Studenikin A 2000 [*Phys. Lett.* ]{} B [**491**]{} 137. Dvornikov M and Studenikin A 2001 [*Phys.Atom.Nucl.* ]{} [**64**]{} 1624 \[[*Yad.Fiz.*]{} [**64**]{} (2001) 1705\]. Dvornikov M and Studenikin A 2004 [*Phys. Atom. Nucl. *]{} [**67**]{} 719 \[Yad. Fiz. [**67**]{} (2004) 741\]. Dvornikov M and Studenikin A 2002 [*JHEP* ]{} [**09**]{} 016. Fabbricatore R Grigoriev A and Studenikin A 2016 [*J. Phys. Conf. Ser. *]{} [**718**]{} 062058. A. Studenikin, EPJ Web Conf. [**125**]{} (2016) 04018. doi:10.1051/epjconf/201612504018 Kurashvili P, Kouzakov K, Chotorlishvili L and Studenikin A 2017 [*Phys.Rev.*]{} D [**96**]{} 103017. Akhmedov E and Pulido J 2003 [*Phys. Lett. *]{} B [**553**]{} 7. Studenikin A 2004 [*Frascati Phys. Ser.*]{} [**34**]{} 155 \[hep-ph/0407010\]. Studenikin A 2016 [*EPJ Web Conf. *]{} [**125**]{} 04018. Studenikin A 2017 [*J. Phys. Conf. Ser. *]{} [**888**]{} no.1, 012221. \[arxiv:1610.06563 \[hep-ph\]\]. Studenikin A 2017 [*PoS NOW* ]{} [**2016**]{} 070. Cirigliano V, Fuller G and Vlasenko A 2015 [*Phys. Lett.* ]{} B [**747**]{} 27. Volpe C 2015 [*Int. J. Mod. Phys.*]{} E [**24**]{} 1541009. Kartavtsev A, Raffelt G and Vogel H 2015 [*Phys. Rev.* ]{} D [**91**]{} 125020. Dobrynina A, Kartavtsev A and Raffelt G 2016 [*Phys. Rev. *]{} D [**93**]{} 125030. Tian J Y, Patwardhan A and Fuller G 2017 [*Phys. Rev. *]{} D [**95**]{} 063004.
[^1]: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We characterize the equisingularity classes of irreducible plane curve germs whose general members have a Newton nondegenerate general polar curve. In addition, we give explicit Zariski open sets of curves in such equisingularity classes whose general polars are Newton nondegenerate and describe their topology.'
author:
- |
A. Hefez (UFF), M.E. Hernandes (UEM) and\
M.F. Hernández Iglesias (UEM)[^1]
title: Plane branches with Newton nondegenerate polars
---
Authors’ e-mail addresses: [email protected], [email protected],\
[email protected]
Introduction
============
Let ${\mathbb{C}}\{x,y\}$ be the ring of convergent power series at the origin of ${\mathbb{C}}^2$. A germ of an analytic plane curve at the origin of ${\mathbb{C}}^2$ is a germ of set determined by an equation $f=0$, where $f$ belongs to the maximal ideal of the ring ${\mathbb{C}}\{x,y\}$ and has no multiple factors. Such a germ may be identified with the class $(f)$ of associated elements to the power series $f$, that is, the set of products of $f$ with the units in ${\mathbb{C}}\{x,y\}$.
The power series $f$ will be called an equation of the germ of curve $(f)$. The multiplicity of $(f)$ is the multiplicity of its equation $f$ and $(f)$ will be said singular if its multiplicity is greater or equal than $2$. When $f$ is irreducible, we call $(f)$ a branch.
We will say that two germs of analytic plane curves $(f)$ and $(f')$ are [*analytically equivalent*]{}, or equivalent for short, if there are a germ $\varphi$ of an analytic isomorphism at the origin of ${\mathbb{C}}^2$ and a unit $u$ of ${\mathbb{C}}\{x,y\}$ such that $f'=u f\circ \varphi$. We will say that $(f)$ and $(f')$ are [*topologically equivalent*]{}, or [*equisingular*]{}, if $\varphi$ is only a local homeomorphism. To verify that two germs of plane curves are equivalent is a hard task, while to verify that they are equisingular is rather easy, thanks to the classical work of Brauner and Zariski that we summarize below.
Each branch of multiplicity $n$ is analytically representable by a Weierstrass polynomial $f$ in ${\mathbb{C}}\{x\}[y]$ of degree $n$ and there is a fractional power series of the form $$y=\varphi(x^\frac 1n)= x^{\frac mn}+\sum_{i\geq 1} a_i x^\frac{m+i}{n}, \quad \text{with} \quad m>n,$$ such that $f(x,\varphi(x^\frac 1n))=0$.
If $w$ is a primitive $n$-th root of unity, the conjugates $\varphi_j(x^{\frac 1n})=\varphi(w^{j}x^{\frac 1n})$, $1\leq j \leq n$, of $\varphi(x^{\frac 1n})$, will be called the Newton-Puiseux parametrizations of $(f)$. One has that $$f=\prod_{j=1}^{n}(y- \varphi_j(x^{\frac{1}{n}})).$$
The equisingularity class of a branch is encoded by the characteristic exponents of any of its Newton-Puiseux parametrization, which, in turn, determine and are determined by the set of intersection numbers of $f$ with all series in ${\mathbb{C}}\{x,y\}$ that are not multiple of $f$, called the [*semigroup of values*]{} of the branch (cf. [@Za] for all these notions). This semigroup will be represented by $\langle v_0,\ldots, v_g\rangle$ where the $v_i$’s are its minimal set of generators. The number $g$ is called the [*genus*]{} of the semigroup and is equal to the number of characteristic exponents.
More generally, the equisingularity class of a curve is encoded by the characteristic exponents of its branches and their mutual intersection numbers, where the intersection number of two curves $(f)$ and $(h)$ is defined as $${\rm I}(f,h)=\dim_{\mathbb{C}}\frac{{\mathbb{C}}\{x,y\}}{\langle f,h \rangle},$$ where $\langle f,h \rangle$ denotes the ideal generated by $f$ and $h$ in the ring ${\mathbb{C}}\{x,y\}$ and the dimension we consider is as a ${\mathbb{C}}$-vector space.
If $f=\sum a_{i,j}x^{i}y^j\in {\mathbb{C}}\{x,y\}$, we will denote by $N(f)$ the Newton polygon of $f$ or of $(f)$, since the Newton polygon is invariant by multiplication by units. If $\ell$ is a side of $N(f)$, we define $$f_\ell=\sum_{(i,j)\in \ell} a_{i,j}x^{i}y^j.$$
We will say that a reduced $f$ is [*Newton nondegenerate*]{}, or simply nondegenerate, with respect to the coordinate system $(x,y)$, if for every side $\ell$ of $N(f)$, the polynomial $f_\ell$ has no critical point outside the lines $x=0$ and $y=0$. This is equivalent to say that the polynomial $F_\ell(z)=\frac{f_\ell(1,z)}{z^{j_0}}$, where $j_0=\min \{ j ; \ (i,j)\in \ell \}$, has no multiple roots. We call $F_\ell(z)$ the [*associated polynomial*]{} to the side $\ell$.
Notice that the notion of nondegenerate is also invariant by multiplication by units. So, to say that $f$ is nondegenerate is the same as to say that the germ of curve $(f)$ is nondegenerate.
The importance of the notion of nondegenerate curves comes from the fact that the topology of such curves is completely determined by their Newton polygons as follows from the result below (cf. [@O Propostion 4.7]).
[**Oka’s Lemma**]{} [*Let $(f)$ be a reduced plane curve germ and $(x,y)$ be a local coordinate system such that the tangent cone to $(f)$ does not contain the line $(x)$. If $(f)$ is nondegenerate, then for each side $\ell_k$ of $N(f)$ with extremal points $(i_k,j_k)$ and $(i_k',j_k')$, with $j_k' > j_k$ and $i_k > i_k'$, there correspond $d_k={\textrm}{GCD}(n_k,m_k)$ branches $\xi_{k,l}$, $1\leq l \leq d_k$, where $n_k=j_k'-j_k$ and $m_k=i_k-i_k'$, with semigroup generated by $\frac{n_k}{d_k}$ and $\frac{m_k}{d_k}$ (allowing any one of them to be $1$). If $\xi$ and $\xi'$ are two of these branches with semigroups generated by $\alpha_0,\alpha_1$ and $\beta_0,\beta_1$, respectively, then their intersection number is given by $ {\rm I}(\xi,\xi')=\min\{\alpha_0\beta_1,\beta_0\alpha_1\}.$*]{}
Oka’s Lemma shows that a nondegenerate curve may only have nonsingular branches or branches of genus one.
As a consequence, it follows that two reduced nondegenerate power series with same Newton polygons determine equisingular curves.
Let us define now our main object of study. The [*general polar curve*]{} of $f$ is the curve determined by an equation $P(f):af_x+bf_y=0$, where $(a:b)$ is a general point in $\mathbb{P}^1_{\mathbb{C}}$.
It is known that for equivalent $f$ and $f'$, their general polar curves are equisingular, while two equisingular $(f)$ and $(f')$ have not necessarily equisingular general polars (cf. [@T] and [@P]).
We will now focus on the study of equisingularity classes of plane branches whose general members have Newton nondegenerate polar curves. One of the known general results about polars is [@merle Theorem 3.1], which in particular implies that the general polar curve of a plane branch of genus greater than two always has, at least, one branch of genus greater than 1, and consequently is Newton degenerate. Hence, to study non-degeneracy conditions for $P(f)$ it is sufficient to consider $(f)$ of genus one or two. In Section 2 we will focus on the case of genus $1$, where the content of [@C1] is summarized and condensed through the use of the concept of Newton nondegenerate curves. These results will be used in Section 3 to describe the topology of the general plane branches of genus $2$ with nondegenerate general polars.
Nondegenerate polars of branches of genus $1$
=============================================
It is known (cf. [@Za] ) that an irreducible plane curve has a semigroup $\langle p,q \rangle$, with ${\rm GCD}(p,q)=1$, if and only if it has an analytic representative with an equation of the form $$f_1=y^p-x^q+\sum_{ip+jq > pq}a_{i,j}x^iy^j, \ \ a_{i,j}\in {\mathbb{C}}.$$
In what follows, we denote by $K(p,q)$ the set of all these equations. In this way, the polar curve of an element $f_1\in K(p,q)$ is given by an equation
$$P(f_1)=-qax^{q-1}+pby^{p-1}+ \sum_{ip+jq > pq}
(iaa_{i,j}y+
jba_{i,j}x)x^{i-1}y^{j-1}.$$
Let $f_1$ be a general member of $K(p,q)$. For each fixed $j$, it is easy to verify that the lowest exponent of $x$ in a term of $(f_1)_y$ having as factor the power $y^j$ is $\alpha(j) = q-\left [\frac{(j+1)q}{p}\right ]$, where $[k]$ represents the integer part of $k$. On the other hand, the lowest exponent of $x$ in a term of $(f_1)_x$ having as factor the power $y^j$ is $\gamma(j) = q-\left(\left [\frac{jq}{p}\right ] +1 \right)$. Since $\alpha(j)\leq \gamma(j)$, it follows that the powers of the monomials that determine $N(P(f_1))$ are contained in the set $$\label{pontos} A= \left \{(\alpha(j), j); \ 0 \leq
j \leq p-1 \right\}$$ and each element $(\alpha(j),j)\in A$ is associated to the following term of $P(f_1)$:
$ t_j=(j+1)ba_{\alpha(j),j+1}x^{\alpha(j)}y^j, \qquad {\textrm}{if}\,\,
\left[\frac{(j+1)q}{p}\right]\neq \left[\frac{jq}{p}\right]+1, \quad \text{or}$
$
t_j=\big( b(j+1)a_{\alpha(j),j+1}+a(q-[(jq/p])a_{\alpha(j)+1,j} \big)x^{\alpha(j)}y^j
,\,\,\, {\textrm}{if}\,\,\left[\frac{(j+1)q}{p}\right]= \left[\frac{j q}{p}\right]+1.
$
Let $\frac{q}{p}=[h_0,\ldots,h_s]$ be the continued fraction decomposition of $\frac{q}{p}$ and consider, for $i\leq s$, $\frac{q_i}{p_i}=[h_0,\ldots, h_i]$ (here necessarily ${\textrm}{GCD}(p_i,q_i)=1$),
In [@C1] it is shown that the Newton polygon $N(P(f_1))$, determined by the set $A$ in (\[pontos\]), has sides $\ell_k$, where $0\leq k \leq \frac{s-1}{2}$ (with $s$ as above). We then have that $\ell_k$ contains the points $$\label{pointsgenusone}
\begin{array}{ll}
(q-q_{2k},p_{2k}-1) + l(-q_{2k+1},p_{2k+1}); \,\, 0 \leq l \leq h_{2k+2}, & {\rm when} \ 0 \leq k < \frac{s-1}{2}, \\ \\
(q-q_{s-1},p_{s-1}-1) \ \text{and} \ (0,p-1), & {\rm when} \ k=\frac{s-1}{2},
\end{array}$$ where the last possibility only occurs when $s$ is odd.
Using (\[pointsgenusone\]) and considering the terms $t_j$, we have that the associated polynomial $F_k(z)$ to the side $\ell_k$ of the polygon $N(P(f_1))$ is the following: $$F_k(z)=\sum_{l=0}^{h_{2k+2}}\frac{t_{l
p_{2k+1}+p_{2k}-1}(1,z)}{z^{p_{2k}-1}}, \quad \ 0\leq k < \frac{s-1}{2}$$ and when $s$ is odd $$F_{\frac{s-1}{2}}(z)= \frac{t_{p_{s-1}-1}(1,z)+t_{p-1}(1,z)}{z^{p_{s-1}-1}}.$$
For a polynomial $F(z)$ we denote its discriminant and its zero set by $\Delta(F(z))$ and $Z(F(z))$, respectively.
Let us consider the Zariski closed set $Z(p,q)$ of elements in $K(p,q)$ whose coefficients are in the set $$\label{zeros}
Z \left( \prod_\lambda t_\lambda(1,1) \prod_\mu\Delta(F_\mu(z)) \right),$$ where $\lambda$ varies in the set $\{p_{2k}-1+lp_{2k+1}; \ 0\leq k < \frac{s-1}{2}, \ 0\leq l \leq h_{2k+2}\}$, union with the set $\{p_{s-1}-1, p-1\}$ if $s$ is odd; and $0\leq \mu \leq \frac{s-1}{2}$.
With the above notation we have the following proposition.
\[genusone\] If $f_1\in K(p,q)\setminus Z(p,q)$, then the polar curve $P(f_1)$ is Newton nondegenerate.
[**Proof**]{} The result is clear since the condition $f_1\not \in Z(p,q)$ implies that $N(P(f_1))$ supports exactly the points in (\[pointsgenusone\]) corresponding to monomials of the polar curve and that the associated polynomials to the side of this polygon have no multiple roots. [$\square$\
]{}
Combining Proposition and Oka’s Lemma, we get the following result.
\[topgenus1\] The polar $P(f_1)$ of an element $f_1\in
K(p,q)\setminus Z(p,q)$ has branches $\xi_{k,\lambda}$ for $0 \leq k < \frac{s-1}{2}$ and $1 \leq \lambda \leq h_{2k}$ with semigroup $\langle p_{2k+1},q_{2k+1} \rangle$. If $s$ is odd, one has additionally $d_s={\textrm}{GCD}(p_s-p_{s-1}, q_s-q_{s-1})$ branches with semigroup $\langle \frac{p_s-p_{s-1}}{d_s}, \frac{q_s-q_{s-1}}{d_s} \rangle$. Moreover, if $\xi$ and $\xi'$ are branches of $P(f_1)$ with semigroups $\langle v_0,v_1\rangle$ and $\langle w_0,w_1\rangle$, respectively, then $I(\xi,\xi')=min\{v_0w_1,w_0v_1\}$ and if $\xi$ is smooth then $I(\xi,\xi')=w_1$.
Let $f_1=y^7-x^{19}+\sum_{7i+19j > 133}a_{i,j}x^iy^j$. We have that $\frac{19}{7}=[2,1,2,2]=[h_0,h_1,h_2,h_3]$. Since $h_2=2$, corresponding to the side $\ell_0$, the polar $P(f_1)$ has $2$ branches $\xi_1$ and $\xi_2$, which are smooth because $ p_1=1$. Since $s=3$ is odd, $p_{s-1}=3$, $q_{s-1}=8$, $p_s=7$ and $q_s=19$, we have that $1={\textrm}{GCD}(4,11)={\textrm}{GCD}(p-p_{s-1},q-q_{s-1})$, then $P(f_1)$ has an extra branch $\rho$ with semigroup $\langle 4, 11 \rangle$, corresponding to the side $\ell_1$.
The side $l_0$ is determined by the terms $t_{0}=ba_{17,1}x^{17}$, $t_{1}=2ba_{14,2}x^{14}y$ and $t_{2}=3ba_{11,3}x^{11}y^2$. In this way, we have $F_0(z)=3ba_{11,3}z^2+2ba_{14,2}z+ba_{17,1}$ and $\Delta(F_0(z))=12b^3a_{11,3}(3a_{11,3}a_{17,1}-a_{14,2}^2)$. The side $l_1$ is determined by $t_{2}=3ba_{11,3}x^{11}y^{2}$ and $t_{6}=7by^6$, so $F_1(z)=b(7z^4+3a_{11,3})$ and $\Delta(F_1(z))=4(84b^2a_{11,3})^3$.
Therefore, the set (\[zeros\]) given before Proposition \[genusone\] is $$Z(a_{11,3}a_{14,2}a_{17,1}(3a_{11,3}a_{17,1}-a_{14,2}^2)).$$
So, $Z(7,19)$ is the subset of $K(7,19)$ whose coefficients belong to the above set, hence $K(7,19)\setminus Z(7,19)$ is an open set where the curves $(f_1)$ have Newton nondegenerate general polar curves. Finally one has that $I(\xi_1,\xi_2)=3$ and $I(\xi_i,\rho)=11$ for $i=1,2$.
Notice that, for all $f_1\in K(p,q)\setminus Z(p,q)$, the branches of the general polar curve have at most genus one. For $f_1 \in Z(p,q)$ this may fail as we show in the next example.
If $f_1\in K(5,12)$, then the set (\[zeros\]) described in Proposition \[genusone\] is $$Z(a_{5,3}a_{10,1}(9a_{5,3}^2-20a_{10,1})).$$
Let $f_1=y^5-x^{12}+x^5 y^3+x^8y^2+\frac{9}{20}x^{10}y \ \in Z(5,12)$. Its polar curve is given by the polynomial $$P(f_1) = 5by^4 +5ax^4y^3+(8ax^7+3bx^5)y^2+(\frac{9}{2}ax^9+ 2x^8)y-12ax^{11}+\frac{9}{20}bx^{10}.$$ By a computation using $\text{Maple}^{\text{TM}}$, we find that $P(f_1)$ has a single branch with a parametrization given by $$x=t^4 ,\,\,\,\,y=d_1t^{10}+d_2t^{11}+\cdots, \ \ \text{with} \ d_1d_2\neq 0.$$ Therefore, $P(f_1)$ is a branch with characteristic exponents $4,10,11$, hence it has genus two. In [@polar] we give the complete description of the polars of curves in $K(5,12)$.
The above example shows that the question of which equisingularity classes have all curves with nondegenerate general polars is meaningless. This is why we asked the question only for general members of equisingularity classes.
Nondegenerate polars of branches of genus $2$
=============================================
Any irreducible curve $(f)$ of genus 2 has a semigroup of values of the form $$\langle
v_0,v_1,v_2\rangle=\langle
e_1p,e_1q,e_1pq+d\rangle,$$ where $e_1={\rm GCD}(v_0,v_1)$ and ${\rm GCD}(e_1,d)=1$.
Using the construction of [@AB], one knows that any $(f)$ with semigroup as above is equivalent to one of the form $f=f_1^{e_1}+f_2$, where $f_1=y^{p}-x^q+\sum_{ip+jq > pq} a_{i,j}x^iy^j\in K(p,q)$ and $$f_2=\sum_{iv_0+jv_1 > 2v_2}a_{i,j,2}x^iy^jf_1^{e_1-2}+\cdots+
\sum_{iv_0+jv_1 > (e_1-1)v_2}a_{i,j,e_1-1}x^iy^jf_1+$$ $$+b_{i_0,j_0}x^{i_0} y^{j_0}+\sum_{iv_0+jv_1> e_1v_2}b_{i,j}x^iy^j,$$ with $i_0v_0+j_0v_1=e_1v_2, $ $i_0\geq 0$, $0\leq j_0<\frac{v_0}{e_1}$; $a_{i,j},a_{i,j,k},b_{i,j}\in
{\mathbb{C}}$ and $b_{i_0,j_0}\neq 0.$ We denote by $K(e_1p,e_1q,e_1pq+d)$ the set of the power series described above.
Notice that $P(f)=e_1f_1^{e_1-1}P(f_1)+ P(f_2)$.
With an elementary computation involving the inequalities relating the indices in the summations in the expression of $f_2$, one may conclude that $N(f)=N(f_1^{e_1})$ and since $f_1$ is irreducible, we have that $N(f_1^{e_1})=N((y^p-x^q)^{e_1})$, therefore $N(f)=N((y^p-x^q)^{e_1})$, which is precisely the segment $l$ contained in the line $pX+qY=e_1pq$ supporting only the points $(iq,(e_1-i)p)$ for $i=0,\ldots ,e_1$.
We denote by $l_{p,q}$ the segment that contains the points $(iq,(e_1-i)p-1)$ for $i=0,\ldots ,e_1-1$.
Now, since the segments $l$ and $l_{p,q}$ are parallel with slope of absolute value less than $1$ and the second coordinate of their points are shifted by one unit, it follows that the segment $l_{p,q}$ belongs to the polygon $N(P(f))$.
(7,5) (0,0)[(1,0)[7]{}]{}(0,0)[(0,1)[5]{}]{}(2,1.7)[$l_{p,q}$]{} (0,3)[(3,-2)[3]{}]{} (0,3)[$e_1p-1$]{} (0,4)[(3,-2)[6]{}]{}(0,4)[$e_1p$]{} (3,1)(3,2)[$((e_1-1)q,p)$]{} (0,0)[(3,1)]{}(-1,1)[$p-1$]{}(3,-0.3)[$(e_1-1)q$]{} (6,0) (6,-0.3)[$e_1q$]{} (2,2.7)[$l$]{}
From the expression of $f$ we have that the terms of $P(f)$ corresponding to the points on $l_{p,q} $ are obtained from $e_1f_1^{e_1-1}P(f_1)$ and the associated polynomial to $l_{p,q}$ is (modulo a nonzero constant) $F_{p,q}(z)=(z^p-1)^{e_1-1}$. Therefore, for $e_1 > 2$ we have that $P(f)$ is a degenerate curve, and since we are interested in branches with nondegenerate general polars, it will be sufficient to consider $e_1=2$, that is, the case in which $f$ has a semigroup of the form $\langle 2p,2q,2pq+d \rangle$, with odd $d$, which are coincidentally the equisingularity classes studied in [@LP] in view of their analytic classification.
So, we are reduced to study polars of curves in the sets of the form $K(2p,2q,2pq+d)$, that is, polar curves of series of the form $f=(f_1)^2+f_2$, where $$f_1\in K(p,q), \, \, \, \, f_2= b_{i_0,j_0}x^{i_0} y^{j_0}+\sum_{ip+jq> 2pq+d}b_{i,j}x^iy^j,$$ where $i_0p+j_0q=2pq+d$, $i_0\geq 0,\ 0\leq j_0 < p.$
As we have observed above, in this situation, one has that $$\label{polargen2}
P(f)=2f_1P(f_1)+P(f_2)$$ and that the Newton polygon of $f$ is supported by the line $L:pX+qY=2pq$ and it contains the points $(0,2p),\ (q,p)$ and $(2q,0)$. On the other hand, the segment $l_{p,q}$ (with $e_1=2$) with extremal points $(0,2p-1)$ and $(q,p-1)$ is contained in one side of the polygon $N(P(f))$.
Notice that the Newton polygon of $f_1(P(f_1))$ is the Minkowski sum of $N(f_1)$ and $N(P(f_1))$; that is, if $A$ is as in (\[pontos\]), then $N(f_1(P(f_1)))$ is determined by $$\{(0,p-1)+A, (q,0)+A\}.$$
Since the absolute value of the slopes of the sides of $N(P(f_1))$ are $\frac{p_{2j+1}}{q_{2j+1}}$ which are smaller than $\frac{p}{q}$ (the absolute value of the slope of $N(f_1)$), the polygon $N(f_1(P(f_1)))$ is determined by the set $\{(0,2p-1), (q,0)+A\}$.
Therefore, $N(f_1(P(f_1)))$ is given by $l_{p,q}$ and $(0,q)+N(P(f_1))$, which implies that $N(f_1P(f_1))$ has $[\frac{s+1}{2}]+1$ sides $D_k$ where $$\label{puntos}
D_k=(q,0)+l_k ,\,\,\, 0\leq k \leq \frac{s-1}{2}\ \ \ \text{and} \ \ D_{[\frac{s-1}{2}]+1}=l_{p,q},$$ where the $l_k$ are given as in Section 2 and the terms of $f_1P(f_1)$ associated to $(i,j)\in N(f_1P(f_1))$ is $$g_{j}=
\begin{cases}
-x^qt_{j}\,\,\,& \text{if}\,\,\, j\neq 2p-1\\ py^{2p-1} & \text{if} \,\, j= 2p-1.
\end{cases}$$
By a similar analysis to the case of genus one, considering the polar $P(f_2)=a(f_2)_x+b(f_2)_y$, one has that for each $j$ the smallest value of $i$ for which there exists a term in $b(f_2)_y$ containing $y^j$ as a factor is $2q-[\frac{q(j+1)-d}{p}]$ and in $a(f_2)_x$ is $2q-[\frac{qj+p-d}{p}]$. Denoting by $\beta(j)$ the minimum of these values $2q-[\frac{q(j+1)-d}{p}]$, the term of $P(f_2)$ that contains the monomial $x^{\beta(j)}y^j$ is
$
h_{j}= (j+1)bb_{\beta(j),j+1}x^{\beta(j)}y^j,\ \ {\textrm}{if}\,\,
\beta(j)\neq 2q-[\frac{jq+p-d}{p}]; \ \
{\textrm}{or}$
$h_j= \big( b(j+1)b_{\beta(j),j+1}+a(2q-[(jq+p-d)/p]+1)b_{\beta(j)+1,j}\big)x^{\beta(j)}y^j,$
if $ \beta(j) = 2q-[\frac{jq+p-d}{p}]$.
Since $\beta(j)\geq \alpha(j)+q$, we conclude that the polygons $N(P(f))$ and $N(f_1P(f_1))$ are equal and have the same points on the respective sides. So, from Formula (\[polargen2\]), it follows that the term containing $x^{\alpha(j)+q}y^j$ in $P(f)$ is $$m_{j}=
\begin{cases}
2g_{j}+h_{j},\,\,& {\textrm}{if}\,\,
\alpha(j) +q=\beta(j)\\
2g_{j} ,& {\textrm}{if}\,\,\alpha(j) +q < \beta(j).
\end{cases}$$
From the above expression of the terms $m_{j}$ and from (\[puntos\]), the associated polynomial to the side $D_k$ of $N(P(f))$ is $$\begin{array}{l}
F_k(z)=\sum_{l=0}^{h_{2k+2}}\frac{m_{p_{2k}-1+lp_{2k+1}}(1,z)}{z^{p_{2k}-1}}, \ \ {\textrm}{if} \ \ 0\leq k < \frac{s-1}{2},\\
F_{\frac{s-1}{2}}(z)= \frac{m_{p_{s-1}-1}(1,z)+m_{p-1}(1,z)}{z^{p_{s-1}-1}} \ \ \quad {\textrm}{and} \\
F_{[\frac{s-1}{2}]+1}(z)=\frac{m_{2p-1}(1,z)+m_{p-1}(1,z)}{z^{p-1}}.
\end{array}$$
Let us consider the Zariski closed set $Z(2p,2q,2pq+d)$ of elements in $K(2p,2q,2pq+d)$ whose coefficients are in the set $$Z \left( \prod_\lambda m_\lambda(1,1) \prod_\mu\Delta(F_\mu(z)) \right),
\label{zeros2}$$ where $\lambda$ varies in the set $\{p_{2k}-1+lp_{2k+1}; \ 0\leq k < \frac{s-1}{2}, \ 0\leq l \leq h_{2k+2}\}\cup\{2p-1\}$, union with the set $\{p_{s-1}-1, p-1\}$ if $s$ is odd; and $0\leq \mu \leq [\frac{s+1}{2}]+1$.
By the above analysis and Oka’s Lemma we obtain the following result.
\[genus2\] If $f \in K(2p,2q,2pq+d)\setminus Z(2p,2q,2pq+d)$, then the polar curve $P(f)$ is Newton nondegenerate. In this case, $P(f)$ has a branch with semigroup $\langle p,q\rangle$ and $\left [
\frac{s+1}{2}\right ]$ branches as described in Theorem \[topgenus1\], such that, if $(\xi)$ and $(\xi')$ are branches of $P(f)$ with semigroup $\langle v_0,v_1\rangle$ and $\langle w_0,w_1\rangle$, then $I(\xi,\xi')=min\{v_0w_1,w_0v_1\}$, if $(\xi)$ is smooth, then $I(\xi,\xi')=w_1$.
In [@polar] we showed that, for any $f\in K(4,2k,4k+d)$, with $k>1$ odd, the polar curve $P(f)$ has a smooth branch $\xi$ and a branch $\rho$ with semigroup $\langle 2,k\rangle$ such that $I(\xi,\rho)=k$ and $Z(4,2k,4k+d)$ is empty.
If $f\in K(10,24,121)$ is given by $$f=(y^5-x^{12}+\sum_{5i+12j > 60}a_{i,j}x^iy^j)^2+b_{17,3}x^{17}y^{3}
+\sum_{5i+12j > 121}b_{i,j}x^iy^j,$$ with $b_{17,3}\neq 0$, then the Newton polygon $N(P(f))$ is given by the following diagram where we wrote the term $m_j$ at the point with coordinates $(\beta(j),j)$
(9,4) (0,0)[(1,0)[9]{}]{}(0,0)[(0,1)[4]{}]{} (0,3)[(2,-1)[4]{}]{} (4,1)[(4,-1)[4]{}]{} (0,3)[$10by^9$]{} (4,1)[$-10bx^{12}y^4$]{} (8,0) (8,-0.5)[$b(-2a_{10,1}+b_{22,1})x^{22}$]{} (6,0.5)[$3b(-2a_{5,3}+b_{17,3})x^{17}y^{12}$]{} (2,1.5)[$l_{5,12}$]{} (6,0.1)[$l_0$]{}
and the associated polynomial to the side $l_0$ is $F_0(z)= b(-10z^4 + 3(b_{17,3}-2a_{5,3})z^2 + b_{22,1}-2a_{10,1})$.
In this case, we have that the set in (\[zeros2\]) is $$Z((b_{17,3}-2a_{5,3})(b_{22,1}-2a_{10,1})(9b_{17,3}^2-36a_{5,3}b_{17,3}+36a_{5,3}^2+40b_{22,1}-80a_{10,1})).$$
\[dcondition\] If $f=f_1^2+f_2\in K(2p,2q,2pq+d)$ and $d\geq q$, then the set $Z(2p,2q,2pq+d)$ depends only on $f_1$. In fact, since $f_2=\sum_{ip+jq\geq 2pq+d}b_{ij}x^iy^j$, if $d\geq q$, we have $ip+jq\geq 2pq+d\geq 2pq+q > 2pq+p$ and in this way all vertices of $N(P(f_2))$ are above $N(f_1^2)$. Hence the monomial $m_{\alpha(j),j}$ does not depend on the coefficients of $f_2$.
It is known that the topological class of the general polar curve contains analytical data of the curve, but does not allow to recover the topological class of the curve. For instance, if we consider general elements $f_k\in K(2p,2q,2pq+d_k)$, then by the previous remark $P(f_k)$ belongs to the same topological class for all $d_k\geq q$, but the topological class of $f_k$ varies according to $d_k$. In contrast, in [@C2] it was shown that the topological type of a curve may be recovered from the cluster of basis points of the general polar curve.
We summarize the above discussion in the following statement.
The general polar curve $P(f)$ of the general member of an equisingularity class of branches is Newton nondegenerate if and only if the equisingularity class corresponds to one of the following semigroups: $\langle p,q \rangle $ or $\langle 2p,2q,2pq+d\rangle $, where GCD$(p,q)=1$ and $d$ is odd. The topology of the general polar in each case is described in Theorem \[genusone\] and in Theorem \[genus2\].
[999]{} - [*Effective construction of irreducible curve singularities* ]{}; International Journal of Mathematics and Computer Science, 1, 125-149 (2006).
- [*On the singularities of polar curves*]{}; Manuscripta Math. 43, 167-190 (1983).
- [*Singularities of plane curves*]{}; London Mathematical Society, Lectures Notes Series 276 (2000).
- [*On polars of plane branches*]{}; arXiv 1505.08038.
- [*Normal forms and moduli spaces of curve singularities with semigroup $\langle 2p,2q,2pq+d \rangle$*]{}; Compositio Mathematica, 76, 247-264 (1990).
- [*Invariants polaires des courbes planes*]{}; Invent. Math. 41, n. 2, 103-–111 (1977).
- [*Nondegenerate complete intersection singularity*]{}; Actualitèes Mathématiques, Hermann (1997).
- [*Deformations equisingulières des idéaux jacobiens des courbes planes*]{}; In Proc. of Liverpool Symposium on Singularities II, SLNM, Volume 209, 218-233 (1971).
- [*Variétés polaires. II. Multiplicités polaires, sections planes, et conditions de Whitney*]{}; Algebraic Geometry (La Rábida, 1981), SLNM, Volume 961, 314-491 (1982).
- [*The moduli problem for plane branches*]{}. University lecture series AMS, Volume 39 (2006).
[^1]: The first two authors were partially supported by CNPq Grants and the third author by a fellowship from CAPES and ARAUCARIA foundations
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We realise a circular gray-field polariscope to image stress-induced birefringence in thin (sub-micron thick) silicon nitride (SiN) membranes and strings. This enables quantitative mapping of the orientation of principal stresses and stress anisotropy, complementary to, and in agreement with, finite element modeling (FEM). Furthermore, using a sample with a well known stress anisotropy, we extract a new value for the photoelastic (Brewster) coefficient of silicon nitride, $C \approx (3.4~\pm~0.1)\times~10^{-6}~\mathrm{MPa}^{-1}$. We explore possible applications of the method to analyse and quality-control stressed membranes with phononic crystal patterns.'
author:
- 'T. Capelle'
- 'Y. Tsaturyan'
- 'A. Barg'
- 'A. Schliesser'
bibliography:
- 'cleaned\_bibliography.bib'
title: Polarimetric analysis of stress anisotropy in nanomechanical silicon nitride resonators
---
Silicon nitride membranes and strings under high tensile stress have excellent mechanical and optical properties [@Zwickl2008; @Wilson2009], making them a widely used platform to study the behaviour of mechanical systems in the quantum regime [@Purdy2013; @Purdy2013a; @Zhou2013; @Underwood2015; @Wilson2015; @Peterson2016; @Nielsen2016; @Koji2016]. Recently, further enhancement of these properties through in-plane patterning has been explored. Examples include one-[@Stambaugh2014] and two-dimensional[@Chen2016a; @Norte2016; @Bernard2016] subwavelength optical grating and photonic crystal structures which can boost the reflectivity of SiN beyond $99.9\%$. In the mechanical domain, “trampoline” resonators [@Kleckner2011] combine thin tethers with a light central pad to achieve low-mass, low-frequency resonators with high quality factors [@Norte2016; @Reinhardt2016; @Weaver2016]. Patterning with phononic bandgap structures [@Ghadimi2016; @Barasheed2015; @Tsaturyan2016] can suppress radiation losses and, if combined with optimised dissipation dilution through “soft clamping”[@Tsaturyan2016], yield extreme quality factors and room-temperature $Q\cdot f$-products beyond $10^{14}\,\mathrm{Hz}$.
In all instances of patterning, the stress relaxes to a new equilibrium distribution according to the pattern boundary conditions. This changes dramatically the mechanical, and potentially, via photoelastic coupling, optical properties of the structure. In absence of a laboratory diagnostic instrument, it has so far been necessary to rely on FEM to simulate the stress redistribution. In addition, little is known about the photoelastic coupling in silicon nitride [@Borkje2012; @Grutter2015]. To address these deficiencies, we have realised a highly sensitive polarimetric setup which allows quantitative imaging of stress-induced birefringence.
Among the numerous possibilities to implement an imaging polarimeter (or polariscope)[@Ajovalasit2015], we have chosen to build a circular gray-field polariscope. Its basic idea is to illuminate the sample with circularly polarised light, and analyse the ellipticity of the transmitted beam’s polarisation in a spatially resolved manner. This approach has a decisive advantage over plane polariscopes working with linear polarisation when it comes to measuring small optical retardation $\delta$: as we will demonstrate, in the circular polariscope the optical signal is $\propto \delta$, whereas in the plane polariscope it is only $\propto \delta^2$. In our setup (Fig. \[fig:setup\] and Tab. \[tab:setup\_specs\]), we use a light emitting diode (Thorlabs M780LP1) as a light source, followed by a bandpass filter that eases the requirements on achromaticity of the subsequent optical elements. The source is imaged on a diffuser which, together with an aspheric condenser lens, provides a Köhler-like illumination of the sample. Before reaching the sample, the circular polarisation state is defined by a high-contrast polariser and a quarter-wave plate($\lambda/4$). After the sample, a motor-controlled rotating half-waveplate and a polarising beam splitter cube analyse the polarisation state. Rotating the waveplate was found to yield better results than rotating various kinds of polarisers, which tended to displace the beam and thus create image artifacts. A microscope objective (Achrovid 5x) and a plano-convex lens image part of the sample on a CCD camera (Mightex CGE B013-U) with a magnification of $10$.
{width="\linewidth"}
--------------------------------------------------- ------------------
LED central wavelength (nm) 780
Bandpass filter bandwidth (nm) 10
Working distance of the microscope objective (mm) 37
Focal length of the imaging lens (mm) 400
Pixel size ($\mu$m) 3.75$\times$3.75
Resolution of the CCD camera 1280$\times$960
Bit depth of the CCD camera ADC 12
--------------------------------------------------- ------------------
: Technical characteristics of the polarimetric setup.[]{data-label="tab:setup_specs"}
It is straightforward to compute the expected signal via Jones calculus. Each area element of the sample can be treated as a general retarder described by the Jones matrix $$\label{general retarder}
\textbf{S}_{\delta,\theta}=\textbf{R}^{-1}_{\theta}\cdot\begin{pmatrix}
e^{-i\delta/2}&0\\
0&e^{+i\delta/2}
\end{pmatrix}\cdot\textbf{R}_{\theta},$$ where $\theta$ is the azimuthal angle of the polarisation eigenstate basis with respect to a fixed laboratory reference, $\delta$ is the retardation phase, and $$\label{rotation matrix}
\textbf{R}_{\theta}=\begin{pmatrix}
\cos(\theta)& -\sin(\theta)\\
\sin(\theta)& \cos(\theta)
\end{pmatrix}$$ is the canonical rotation matrix. Specifically, in the case of stress birefringence of a SiN membrane, $\theta$ denotes the direction of the principal stress in the membrane plane, and $\delta$ is the retardation induced by stress anisotropy, $$\label{Brewster relation}
\delta=\frac{2\pi}{\lambda}\,C\,l \cdot \Delta\sigma,$$ where $\lambda$ is the wavelength of the light source, $\Delta\sigma$ the stress anisotropy, $l$ the thickness of the sample and $C$ the Brewster coefficient, which is a material constant. The other waveplates can easily be represented by similar matrices, setting $\delta = \pi/2$ for quarter waveplates, and $\delta=\pi$ for half waveplates. The analyzer at the end projects the polarisation state, as described by the matrix $$\label{polarizer}
\textbf{P}_0=\begin{pmatrix}
1 & 0\\
0 & 0
\end{pmatrix}.$$ If we introduce $\alpha$ as the angle of the output half waveplate with respect to the output polariser, the intensity at each camera pixel is given by $I_{\mathrm{gray}}\propto\vec{J}_{\mathrm{out}}^{\dagger}\cdot\vec{J}_{\mathrm{out}}$, with $$\label{}
\vec{J}_{\mathrm{out}}=\textbf{P}_{0}\cdot\textbf{S}_{\pi,\alpha}\cdot\textbf{S}_{\delta,\theta}\cdot\vec{J}_{\mathrm{in}},$$ assuming, for simplicity, a perfect input polarisation ${\vec{J}_\mathrm{in}=(1,-i)^{T}/\sqrt{2}}$.
We denote with $\tilde{I}_{\mathrm{gray}}$ the Fourier transform of the output signal with respect to time $t$, during which the half-wave plate rotates with constant angular velocity $\omega \equiv \frac{\partial \alpha}{\partial t}$. Then it is easy to show that $$\label{output signal fourier}
\frac{\tilde{I}_{\mathrm{gray}}(4\omega)}{\tilde{I}_{\mathrm{gray}}(0)}=\frac{1}{2}ie^{-2i\theta}\sin(\delta).$$ We can thus extract the entities of interest by calculating $$\label{retardation extraction}
\delta=\arcsin\left(2\left\lvert\frac{\tilde{I}_{\mathrm{gray}}(4\omega)}{\tilde{I}_{\mathrm{gray}}(0)}\right\rvert\right)$$ and $$\label{angle extraction}
\theta = -\frac{1}{2}\arg\left(\frac{\tilde{I}_{\mathrm{gray}}(4\omega)}{\tilde{I}_{\mathrm{gray}}(0)}\right)+\theta_{0}.$$ Here, $\theta_0$ is an offset angle, given by the angle of the first polariser plus $\pi/4$, which corresponds to the position of the quarter waveplate. Indeed, there is no phase shift when the axis of the quarter waveplate is aligned with the principal axis of the sample.
The data are acquired and processed as follows: images are taken at 10 angular positions for one full rotation. At each position we acquire a large number of frames (typically 150), from which a frame acquired with the shutter closed is subtracted. This is done in order to remove the dark current, which corresponds to an offset due to the thermal electrons detected by the CCD camera. The procedure is repeated without the sample to acquire background images. For each image, treated as a two-dimensional matrix, we calculate the normalized quantity $\tilde{I}_{\mathrm{gray}}(4\omega)/\tilde{I}_{\mathrm{gray}}(0)$. Finally, the background is subtracted from the measurements involving the sample. This yields a complex matrix, which is translated into the orientation $\theta$ and retardation $\delta$ following eqs. (\[retardation extraction\]) and (\[angle extraction\]).
For initial validation, we performed a measurement on a sample with a particularly simple geometry, namely a $210$ nm thick silicon nitride ribbon. We used standard nanofabrication techniques to realise this sample, starting with low-pressure chemical vapor deposition (LPCVD) of $210$ nm stoichiometric silicon nitride on a $500~\mu$m double-side polished silicon wafer. The chosen deposition parameters create a film with an isotropic tensile stress of ca. $1{.}2~\mathrm{GPa}$. Subsequently, ribbons are defined by UV photolithography and reactive ion etching. The photoresist is removed and the wafer is stripped of its native silicon oxide layer by a buffered hydrofluoric acid (BHF) dip. Finally, the ribbons are released using an anisotropic KOH etch from the wafer’s backside. Since the ribbons are clamped only from two sides, the stress in the free direction relaxes close to zero.
Figure \[ribbon\] shows the measured stress birefringence of such ribbons as obtained using our method. The results agree well with our physical intuition, as well as finite-element modelling (COMSOL Multiphysics): at the center of the ribbon the stress is large along the direction of the ribbon and nonexistent in the orthogonal direction, so the anisotropy is maximal. Closer to the clamping region the ribbon also gets stressed in the orthogonal direction, due to the proximity to the silicon support, reducing the anisotropy. The higher stress anisotropy in the ribbon’s fillets could not be recovered in our setup, possibly due to diffraction artifacts (see below). The stress direction is also reproduced correctly: homogeneously along the ribbon in its center, while close to the clamp it rotates by $\pi/4$, preserving the symmetry of the structure.
The good agreement prompts us to apply this approach to measure the Brewster coefficient, relating stress anisotropy with birefringence, of SiN, for which quantitative data is scarce [@Borkje2012; @Grutter2015]. To that end we measured a second ribbon, whose initial stress was carefully determined to be $\sigma = 1190$ MPa before releasing of the ribbon[^1]. In order to take into account the change of the boundary conditions during the release of the ribbon, we have to correct this value by a factor $1-\nu$, where $\nu$ is the Poisson ratio of the material, assumed here to be equal to $0.27$. The result is presented in Figure \[zoom\]. We obtain a retardation of $6.4\pm 0.2~\mathrm{mrad}$ by averaging a $\sim 170\times260~\mathrm{\mu m}^2$-area in the central region of the ribbon. For a quantitative comparison, we correct this value for the multiple reflections inside the film, which are not entirely negligible due to the relatively high ($n\approx 2{.}0$) refractive index of SiN. To do so, we used a transfer matrix model for the complex transmittance of this sample[^2], expanding it to first order in a small variation of refractive index. This yielded a correction parameter of $\eta = 1.26$ for the effective thickness of the sample. The Brewster coefficient of SiN can then be evaluated as $$\label{brewster coefficient}
C = \frac{\delta}{2\pi}\frac{\lambda}{\eta l}\frac{1}{\sigma(1-\nu)} \approx (3.4~\pm~0.1)\times~10^{-6}~\mathrm{MPa}^{-1},$$ where $l=210$ nm is the sample thickness. Remarkably, this is two orders of magnitude lower than the value proposed by Campillo *et al.* [@Campillo2002], but close to the value of silica ($C\approx 4\times 10^{-6}~\mathrm{MPa}^{-1}$) [@Priestley], another amorphous transparent dielectric. This mistake arises from the fact that this previous work measured the refractive index and the stress for different $\mathrm{SiH}_2\mathrm{Cl}_2\mathrm{:NH}_3$ gas flow ratios during the LPCVD process, and not the variation of the refractive index due to mechanical stress.
With such calibration at hand, we proceed to applying polarimetric stress analysis to more complicated resonator structures. Membrane resonators with phononic bandgap shield are of particular interest. In the context of silicon nitride membranes it has been shown that phononic crystal structures can suppress transmission of vibrations [@Tsaturyan2014; @Yu2014], resulting in suppression of dissipation through phonon tunneling [@Yu2014; @Wilson-Rae2008], whose avoidance had previously required delicate and often unreliable clamping techniques[@Wilson2009]. Patterning a phononic crystal structure directly onto the membrane not only suppresses phonon tunneling losses, but also enhances the dilution of internal losses dramatically, enabling an increase in the quality factor by more than an order of magnitude [@Tsaturyan2016], as compared to conventional membrane resonators embedded in silicon phononic crystal structures. Stress analysis using polarimetry in such complex geometries can be a reliable and simple tool to assess the periodicity of stress anisotropy as required for the formation of a bandgap.
To demonstrate this potential, we realised membranes (${l\approx 210~\mathrm{nm}}$) with a honeycomb pattern of ${\sim200~\mathrm{\mu m}}$-diameter holes, fabricated using the same techniques as for the ribbons. Such a patterned membrane exhibits a bandgap as previously described[@Tsaturyan2016]. Once again, we performed a measurement of the retardation, as described above, with the result shown in Figure \[patterned\_membrane\]. As expected, we observe an enhanced stress anisotropy in the tethers (i.e. the narrow regions between the circular holes), see Fig. \[patterned\_membrane\]a. In particular, a cut along a tether reveals a peak, symmetric with respect to the center of the tether, and a maximum stress anisotropy of $\Delta\sigma\sim 2.3~\mathrm{GPa}$. Excellent agreement with the predicted stress anisotropy in this region confirms the previously computed value of the Brewster coefficient.
Near the edges of the membrane, large retardation values ($\gtrsim 15~\mathrm{mrad}$) appears as dark blue rings around the holes (see Fig \[patterned\_membrane\]a). We attribute these to diffraction-related imaging artifacts. This is supported by the presence of several weak concentric rings around the holes, likely due to higher diffraction orders. In addition, the observation of these features already in single-shot images rules out excentric rotation as their cause.
In spite of such artifacts, it is straightforward to recognise defective membranes. Figure \[broken\_membrane\] shows retardation images of an undamaged membrane and one with a broken tether, as a comparison. Defects in the perforated membrane structure cause a redistribution of the stress. One could envision using this approach in assessing the overall performance of the device (e.g. the quality factor), by comparing the stress profile of the defective device with simulations of an undamaged membrane resonator.
In conclusion, applied to $210~\mathrm{nm}$-thick membrane resonators, our experimental setup achieves a $\sim 0.2~\,\mathrm{mrad}$ resolution in retardation over a $\sim(200~\mathrm{\mu m})^2$ area. Using the measured Brewster coefficient, this corresponds to a resolution in stress anisotropy of $\sim 80~\mathrm{MPa}$. Thus we expect the method to deliver relevant results also for ultrathin ($l\lesssim 50~\mathrm{nm}$) membranes of smaller dimensions. As an imaging modality, the present, very basic microscopy setup achieves moderate transverse resolution on the order of $10~\mathrm{\mu m}$, impaired also by edge artifacts (likely diffraction-caused). It is nonetheless more than sufficient to assess the overall quality of patterned membrane resonators and localise defects. We have implemented a gray field polariscope optimised for extracting stress anisotropy in membrane resonators. We have validated the technique by comparing our measurement results with FEM simulations, and performed an independent measurement of the Brewster coefficient, correcting an older value found in the litterature[@Campillo2002]. In itself, this result is of interest for optomechanical systems relying on photoelastic coupling in silicon nitride [@Borkje2012; @Grutter2015].
[^1]: This measurement was conducted on a separate sample with a $238.6$-nm thick SiN layer grown in the same conditions as the ribbon.
[^2]: the transmittance of a membrane of thickness $d$ and refractive index $n$ is given by $t = \frac{2ine^{-ikd}}{2in\cos(kdn)+(1+n^2)\sin(nkd)}$
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study how well one can recover sparse principal components of a data matrix using a sketch formed from a few of its elements. We show that for a wide class of optimization problems, if the sketch is close (in the spectral norm) to the original data matrix, then one can recover a near optimal solution to the optimization problem by using the sketch. In particular, we use this approach to obtain sparse principal components and show that for data points in dimensions, elements gives an -additive approximation to the sparse PCA problem ( is the stable rank of the data matrix). We demonstrate our algorithms extensively on image, text, biological and financial data. The results show that not only are we able to recover the sparse PCAs from the incomplete data, but by using our sparse sketch, the running time drops by a factor of five or more.'
author:
- 'Abhisek Kundu [^1]'
- 'Petros Drineas [^2]'
- 'Malik Magdon-Ismail [^3]'
title: Approximating Sparse PCA from Incomplete Data
---
[^1]: Department of Computer Science, Rensselaer Polytechnic Institute, Troy, NY, [email protected].
[^2]: Department of Computer Science, Rensselaer Polytechnic Institute, Troy, NY, [email protected].
[^3]: Department of Computer Science, Rensselaer Polytechnic Institute, Troy, NY, [email protected].
| {
"pile_set_name": "ArXiv"
} |
---
address: |
Facultat de Física, Universitat de Barcelona, Diagonal 647,\
E-08028 Barcelona, Spain
author:
- 'E. PALLANTE'
title: 'FINAL STATE INTERACTIONS IN $\vepp / \vep$ [^1] '
---
\#1\#2\#3\#4[[\#1]{} [**\#2**]{}, \#3 (\#4)]{}
Introduction
============
In a recent paper [@SHORT] it was pointed out that strong final state interaction (FSI) effects, when properly taken into account, produce a large enhancement (about a factor of two) in the Standard Model prediction of the direct CP violation parameter $\varepsilon'/\varepsilon$. FSI effects also play an important role in the observed $\Delta I =1/2$ rule [@DELTA12; @KA91].
At centre–of–mass energies around the kaon mass, the strong S–wave $\pi$–$\pi$ scattering generates a large phase-shift difference $\left(\delta^0_0 - \delta^2_0\right)(m_K^2)= 45^\circ\pm 6^\circ$ between the $I=0$ and $I=2$ partial waves [@GM]. This effect is taken into account by factoring out those phases in the usual decomposition of $K\to 2\pi$ amplitudes with definite isospin $I=0$ and $I=2$: \[eq:AIdef\] [A]{}\_I A A\_I e\^[i\^I\_0]{} . The presence of such a large phase–shift difference also signals a large dispersive FSI effect in the moduli of the isospin amplitudes, since their imaginary and real parts are related by analyticity and unitarity. The size of the FSI effect can be already roughly estimated at one loop in Chiral Perturbation Theory (ChPT), where the rescattering of the two pions in the final state produces an enhancement of about $40\%$ in the $A_0$ amplitude [@KA91; @BPP; @JHEP]. However, the fact that the one loop calculation still underestimates the observed $\delta_0^0$ phase shift indicates that a further enhancement should be produced by higher orders.
It is then clear that FSI contributions should be included and their effect resummed in the description of $K\to 2\pi$ decays. In addition, the rescattering of the two pions in the final state is a purely long–distance mechanism, which is fully decoupled from the production mechanism of the two pions in the decay process. As a consequence, FSI effects do not introduce any dependence on short–distance quantities (i.e. they do not carry any dependence on the factorization scale in the OPE description of $K\to 2\pi$ amplitudes).
Despite the importance of FSI in $K\to 2\pi$ decays has been known for more than a decade, most of the [*Standard Model predictions*]{} of $\varepsilon'/\varepsilon$ [@Buras; @Lattice] did not include those contributions and failed to reproduce the experimental measurements [@exp]. Lattice determinations of $K\to 2\pi$ amplitudes are still missing FSI effects, since they are obtained in a two–steps procedure where first the simpler $K\to \pi$ matrix element is measured on the lattice and second, the physical $K\to 2\pi$ matrix elements are obtained by using a lowest–order ChPT relation between $K\to \pi$ and $K\to 2\pi$. Neither the first nor the second step include FSI effects.
On the other hand, approaches based on effective low–energy models [@Trieste] or the $1/N_c$ expansion [@Dortmund] do include some one-loop corrections and find larger values for the $A_0$ amplitude. However, the drawback in these cases could be the possible model dependence of the matching procedure with short–distance.
The approach that has been recently proposed [@SHORT; @FUTURE_PP; @FUTURE_PPS] is the Omnès approach, which permits a resummation of the universal FSI effects to all orders in ChPT. In Section \[OMNES\] the general Omnès problem is formulated, while in Section \[SCALAR\] the Omnès problem is solved for the pion scalar form factor, a well known quantity in ChPT; this analysis clarifies the relevant properties of the Omnès solution. Finally, in Section \[KAON\] we consider the Omnès solution for the $K\to 2\pi$ amplitudes which enter the prediction of $\varepsilon'/\varepsilon$. The final result of an exact matching procedure with short–distance [@FUTURE_PPS], with the inclusion of FSI effects [@FUTURE_PP] gives \[eq:epsth\] .(’/)|\_[SM]{} = (155)10\^[-4]{} , which is in good agreement with the present experimental world average [@Ceccuchi] \[eq:exp\] .(’/)|\_[exp]{} = (19.32.4)10\^[-4]{} .
The Omnès problem {#OMNES}
=================
Let us consider a generic amplitude (or form factor) $A^I_J(s)$, with two pions in the final state which have total angular momentum and isospin given by $J$ and $I$, respectively, and invariant mass $s\equiv q^2\equiv \left(p_1 + p_2\right)^2$. The amplitude is analytic everywhere except for a cut on the real positive $s$ axis $L = [4 m_\pi^2, \infty )$.
Below the first inelastic threshold, only the $2\pi$ intermediate state contributes to the absorptive part of the amplitude and Watson’s theorem [@Watson] implies that the phase of the amplitude is equal to the phase of the $\pi\pi$ partial–wave scattering amplitude, so that \[eq:WASIM\_EL\] A\^I\_J = ( A\^I\_J )\_[2]{} &=& e\^[-i\^I\_J]{} A\^I\_J = e\^[i\^I\_J]{} A\^[I]{}\_J\
&=& A\^I\_J = A\^I\_J . Cauchy’s theorem implies instead that $A^I_J(s)$ can be written as a dispersive integral along the physical cut: \[eq:DISP\] A\^I\_J(s) = [1]{}\_[L]{} d z + . Inserting Eq. \[eq:WASIM\_EL\] in the dispersion relation \[eq:DISP\], one obtains an integral equation for $A^I_J(s)$ of the Omnès type, which has the well–known Omnès solution [@ALLOMNES] (for $n$ subtractions with subtraction point $s_0$ outside the physical cut): \[eq:NSUB\] A\^I\_J(s) = Q\^I\_[J,n]{}(s,s\_0) , where \[eq:In\] I\^I\_[J,n]{}(s,s\_0) [(s-s\_0)\^[n]{} ]{} \^\_[4m\^2\_]{} [dz(z-s\_0)\^[n]{}]{} and \[eq:NSUB\_Q\] \_[k=0]{}\^[n-1]{} [ (s-s\_0)\^[k]{} k!]{} . |\_[s=s\_0]{} , (n1) , with $Q^I_{J,0}(s,s_0)\equiv 1$. The dispersive integral $I^I_{J,n}(s,s_0)$ is uniquely determined up to a polynomial ambiguity (that does not produce any imaginary part of the amplitude), which depends on the number of subtractions and the subtraction point. The simple iterative relation for the real part of $I^I_{J,n}(s,s_0)$ \[eq:ITER\] I\^I\_[J,n]{}(s,s\_0) = I\^I\_[J,n-1]{} (s,s\_0) - (s-s\_0)\^[n-1]{}\_[ss\_0]{} [ I\^I\_[J,n-1]{}(s,s\_0) (s-s\_0)\^[n-1]{}]{} , shows that only a polynomial part of $I^I_{J,n}(s,s_0)$ does depend on the subtraction point $s_0$ and the number of subtractions $n$, while the non–polynomial part of $I^I_{J,n}(s,s_0)$, the one containing the infrared chiral logarithms, is universal (i.e. $s_0$ and $n$ independent). Thus, the Omnès solution predicts the chiral logarithmic corrections in a universal way and provides their exponentiation to all orders in the chiral expansion. The polynomial ambiguity of $I^I_{J,n}(s,s_0)$ and the subtraction function $Q^I_{J,n}(s,s_0)$ can be fixed, at a given order in the chiral expansion, by matching the Omnès formula (\[eq:NSUB\]) with the ChPT prediction of $A^I_J(s)$. It remains a polynomial ambiguity at higher orders. Notice that in the presence of a zero of the amplitude the Omnès solution can be found for the factorized amplitude $\overline{A^I_J}(s)$, such that $A^I_J(s) = (s-\zeta)^p \,\overline{A^I_J}(s)$, where $\zeta$ is a zero of order $p$.
The scalar pion form factor {#SCALAR}
===========================
The scalar pion form factor is known up to two loops in ChPT [@BGT:98] and it is a useful quantity to understand the details of the Omnès solution in the case $I=0$ and $J=0$. It is defined by the matrix element of the $SU(2)$ quark scalar density $\langle\pi^i(p^\prime )\vert \bar{u}u+\bar{d}d\vert\pi^k(p)\rangle
\,\equiv\, \delta^{ik}\, F_S^{\pi}(t)$. At low momentum transfer, ChPT provides a systematic expansion of $F_S^{\pi}(t)$ in powers of $t\equiv (p^\prime - p)^2$ and the light quark masses [@GL:84; @GL:85]: F\_S\^(t) = F\_S\^(0){ 1+ g(t) +O(p\^4)} , \[eq:SFF\_2\] where the $O(p^2)$ correction $g(t)$ contains contributions from one-loop diagrams and tree–level terms of the $O(p^4)$ ChPT lagrangian. It is given by [@GL:84; @GL:85] ($f\approx f_\pi = 92.4$ MeV): g(t) &=& [tf\^2]{}{ (1-[M\_\^22t]{}) |[J]{}\_(t) + [14]{}|[J]{}\_[KK]{}(t) + [M\_\^218 t]{} |[J]{}\_(t) +4(L\_5\^r+2L\_4\^r)() .\
&&. + [54(4)\^2]{} ( -1) - [14(4)\^2]{} } . The counterterms $L_4^r,\, L_5^r$ from the $O(p^4)$ lagrangian are kept in the numerical analysis at the standard reference value $\mu = M_\rho$, where [@EC:95] $[L_5^r +2L_4^r](M_\rho) = (0.8\pm 1.1)\times 10^{-3}$. The functions $\bar{J}_{\pi\pi}(t)$, $\bar{J}_{KK}(t)$ and $\bar{J}_{\eta\eta}(t)$ [@GL:84; @GL:85] are ultraviolet finite and, together with the logarithms, they are produced by the one-loop exchange of $\pi\pi$, $K\bar K$ and $\eta\eta$ intermediate states. Notice that at values of $t$ such that $t\ll 4 M_P^2$ |[J]{}\_[PP]{}(t) = [1(4)\^2]{} , implying that, below the $P\bar P$ threshold, $\bar{J}_{PP}(t)$ has a very smooth behaviour and it is strongly suppressed. In the case of the pion scalar form factor this means that at values of $t\ll 4 M_K^2\, , 4 M_\eta^2$ the one-loop functions $\bar{J}_{KK}(t)$ and $\bar{J}_{\eta\eta}(t)$ only give small analytic corrections, which are numerically negligible with respect to the rest of the one–loop corrections.
Let us consider now the Omnès solution for $F_S^\pi (t)$ and fix the subtraction polynomial by performing a matching with the one-loop ChPT result. This implies at the subtraction point $t_0$ F\_S\^(t) = F\_S\^(t\_0)\_0 (t,t\_0) F\_S\^(0){ 1+ g(t\_0)}\_0 (t,t\_0) . \[eq:OMNES\_SFF\] Thus, knowing the form factor at some low–energy subtraction point $t_0$, where the momentum expansion can be trusted, the Omnès factor $\Omega_0 (t,t_0)$ provides an evolution of the result to higher values of $t$, through the exponentiation of infrared effects related to FSI. The once–subtracted solution reads: \_0\^[(1)]{} (t,t\_0) = { [t-t\_0]{}\_[4M\_\^2]{}\^[|[z]{}]{} [dzz-t\_0]{} [\_0\^0(z)z-t-i]{} } \_0\^[(1)]{} (t,t\_0) e\^[i\^0\_0(t)]{} , \[eq:OMNES\_F\] where the integral has been split into its real and imaginary part, making explicit that the phase of the Omnès factor is just the original phase-shift $\delta^0_0(t)$, while $\Re_0^{(1)} (t,t_0)$ is its corresponding dispersive factor. The integral has been cut at the upper edge $\bar{z}$, which represents the first inelastic threshold ($\bar{z} = 1$ GeV${}^2$). In Table \[table:SFF\] the resulting value of $|F_S^{\pi}(M_K^2)/F_S^{\pi}(0)|$ for different subtraction points $t_0 = 0,\, M_\pi^2,\, 2 M_\pi^2, \, 3 M_\pi^2, \, 4M_\pi^2$ and $M_K^2$ and at $t=M_K^2$ is shown together with the dispersive part of the once–subtracted Omnès integral $\Re_0^{(1)}(t,t_0)$ and the one loop function $g(t_0)$. Notice that the Omnès factor is not defined for $t_0 =M_K^2$, because it lies above the threshold of the non–analyticity cut; however, by its definition in Eq. (\[eq:OMNES\_SFF\]), $\Omega (M_K^2,t_0=M_K^2) = 1$ holds.
------------- ------------------- ------------ ----------------------------------- ---------------
$t_0$ $g(t_0)$ $|F_S^{\pi}(M_K^2)/F_S^{\pi}(0)|$
$O(p^2)$ exp. $n=1$
0 0 1.23 1.45 (1.23) 1.45
$M_\pi^2$ 0.042 1.21 1.40 (1.26) 1.46
$2 M_\pi^2$ 0.091 1.17 1.34 (1.28) 1.46
$3 M_\pi^2$ 0.15 1.12 1.26 (1.29) 1.45
$4 M_\pi^2$ 0.26 1.03 1.11 (1.30) 1.40
$M_K^2$ $0.54 - 0.46\, i$ $\equiv 1$ $\equiv 1$ 1.61
------------- ------------------- ------------ ----------------------------------- ---------------
: The one loop function $g(t_0)$, the Omnès factor $\Re_0^{(1)}(t,t_0)$ and the modulus of $F_S^{\pi}(t)/F_S^{\pi}(t_0)$ are shown at $t=M_K^2$ for different values of the subtraction point $t_0\in [0,M_K^2]$. At a given $t_0$, the first value of $|F_S^{\pi}(t)/F_S^{\pi}(t_0)|$ (within brackets) is obtained with the $O(p^2)$ ChPT prediction for $\delta_0^0$, while the experimental phase-shift data has been used in the second one. The integrals have been cut at $\bar{z} = 1$ GeV${}^2$.[]{data-label="table:SFF"}
The ChPT calculation of $F_S^\pi(t_0)$ is obviously better at lower values of $t_0$ where the one-loop correction $g(t_0)$ is smaller. At $t_0 = 4 M_\pi^2$ a sizable 26% effect is already found, while at $t_0 = M_K^2$ (the point chosen in the discussion of the Omnès factor in a recent paper [@MARTI]) the correction is so large than one should worry about higher–order contributions. The Omnès exponential allows to predict $F_S^\pi(M_K^2)$ in a much more reliable way, through the evolution of safer results at lower $t_0$ values.
------------------ ------------- ------------- -------------
$\bar{z}=1$ $\bar{z}=2$ $\bar{z}=3$
$t_0=0$ 1.45 1.58 1.62
$t_0=M_\pi^2$ 1.40 1.51 1.55
$t_0= 2 M_\pi^2$ 1.34 1.44 1.47
$t_0= 3 M_\pi^2$ 1.26 1.35 1.38
$t_0= 4 M_\pi^2$ 1.11 1.86 1.21
------------------ ------------- ------------- -------------
: Dependence of $\Re_0^{(1)} (M_K^2,t_0)$ on the upper edge $\bar{z}$ (in GeV${}^2$ units) of the dispersive integral, for various choices of $t_0$. The fit to the experimental data for $\delta_0^0$ has been used[]{data-label="table:SFF_Z"}
In Table \[table:SFF\_Z\] the dependence of the Omnès integral on the upper edge $\bar{z}$ is shown. From those results one can conclude that cutting the integral at the first inelastic threshold is actually producing an underestimate of the complete FSI effect. The sensitivity of the Omnès integral to the high–energy region is reduced by considering a twice–subtracted Omnès solution [@FUTURE_PP].
Since chiral corrections are smaller at lower values of the subtraction point, values of the Omnès factor at low $t_0$ should be trusted because less sensitive to higher order corrections. Based also on the twice–subtracted result [@FUTURE_PP] one can conclude that the true value of $|F_S^{\pi}(M_K^2)/F_S^{\pi}(0)|$ is in the range 1.5 to 1.6. Taking the experimental phase-shift uncertainties into account (see Figure (\[fig:PHASESHIFT\]) for the details), the result is: |F\_S\^(M\_K\^2)/F\_S\^(0)| = 1.5 0.1 .
Final State Interactions in $\varepsilon' /\varepsilon$ {#KAON}
=======================================================
The analysis of the Omnès solution for the scalar pion form factor has clarified two main points. First, the Omnès solution given by the product of the polynomial amplitude times the Omnès factor is independent of the subtraction point as it has to be. Second, the Omnès exponential allows to predict the physical quantity in a much more reliable way, through the evolution of safer results at lower values of the subtraction point.
In analogy to the scalar pion form factor, one can derive the Omnès solution for any $K\to 2\pi$ amplitude, both CP–conserving and CP–violating (the presence of a CP–violating weak phase does not modify the Omnès derivation, being such a phase of short–distance origin [@FUTURE_PP]). At lowest order in the chiral expansion, the most general effective bosonic Lagrangian, with the same $SU(3)_L\otimes SU(3)_R$ transformation properties as the short–distance Lagrangian, contains three terms: \[eq:lg8\_g27\] \_2\^[S=1]{} &=& -[G\_F ]{} V\_[ud]{}\^ V\_[us]{}\^\* { g\_8 L\_ L\^ + g\_[27]{} ( L\_[23]{} L\^\_[11]{} + [23]{} L\_[21]{} L\^\_[13]{} ) .\
&& + 2 e\^2 f\^6 g\_[EM]{} U\^U} + . The flavour–matrix operator $L_{\mu}=-i f^2 U^\dagger D_\mu U$ represents the octet of $V-A$ currents at lowest order in derivatives, where $U = exp{i\sqrt{2}\phi/f}$ is the exponential representation of the light pseudoscalar meson field with $\phi$ the flavour octet matrix. $\cQ= {\rm diag}(\frac{2}{3},-\frac{1}{3},-\frac{1}{3})$ is the quark charge matrix, $\lambda\equiv (\lambda^6 - i \lambda^7)/2$ projects onto the $\bar s\to \bar d$ transition \[$\lambda_{ij} = \delta_{i3}\delta_{j2}$\] and $\langle A\rangle $ denotes the flavour trace of $A$.
At generic values of the squared centre–of–mass energy $s=(p_{\pi 1}+p_{\pi 2})^2$, the $I=0,2$ amplitudes generated by the lowest–order lagrangian in Eq. \[eq:lg8\_g27\] are given by A\_0 &=& -[G\_F]{} V\_[ud]{}V\^\_[us]{} f { (g\_8+[19]{} g\_[27]{}) (s-M\_\^2) -[43]{} f\^2 e\^2 g\_[EM]{}} ,\
A\_2&=& -[G\_F]{} V\_[ud]{}V\^\_[us]{} [29]{} f {5 g\_[27]{} (s-M\_\^2) - 6 f\^2 e\^2 g\_[EM]{}} . \[TREE\] In the absence of $e^2 g_{EM}$ corrections, the tree–level isospin amplitudes have a zero at $s=M_\pi^2$, because the on-shell amplitudes should vanish in the SU(3) limit [@SU3]. This is not the case for the matrix elements of the electroweak penguin operator $Q_8$, whose lowest–order realization is given by the term proportional to $e^2 g_{EM}$. Those matrix elements are constant (of order $e^2 p^0 = p^2$ in the chiral power counting) at the lowest order in ChPT.
For all the cases one can write a once–subtracted Omnès solution for the on–shell physical amplitudes [@SHORT; @FUTURE_PP] in the form \_I(M\_K\^2) = a\_I(M\_K\^2,s\_0) \_I\^[(1)]{}(M\_K\^2,s\_0) . The amplitude $a_I(M_K^2,s_0)$ depends on which weak effective operator is mediating the corresponding $K\to 2\pi$ amplitude, while the once–subtracted Omnès factor $\Omega_I^{(1)}(M_K^2,s_0)$ is universal, i.e. the same for the pion scalar form factor $F_S^\pi(s)$ and the amplitude $\cA_0(s)$ (this is in principle no more true for the twice–subtracted factor that depends on the derivative of the function itself). Moreover, taking a low subtraction point where higher–order corrections are very small [@FUTURE_PP], one can just multiply the tree–level formulae in Eq. \[TREE\] with the experimentally determined Omnès exponential. The results are [@FUTURE_PP][^2] \[eq:R0\] \_0(M\_K\^2,0) = 1.5 0.1 \_2(M\_K\^2,0) = 0.92 0.06 , where the subtraction point $s_0=0$ has been chosen. How the proper inclusion of FSI effects drastically modify the Standard Model prediction of $\varepsilon^\prime /\varepsilon$ has been already explained [@SHORT].
To obtain a complete Standard Model prediction for $\varepsilon^\prime /\varepsilon$ an exact matching procedure has been proposed [@FUTURE_PPS]. It is inspired by the large–$N_c$ expansion, but only at scales below the charm quark mass $\mu \leq m_c$ (where the logarithms that enter the Wilson coefficients are small). FSI effects, which are next–to–leading in the $1/N_c$ expansion but numerically relevant, are taken into account through the multiplicative factors $\Re_I(M_K^2,0)$ while avoiding any double counting. The general formula for $\varepsilon'/\varepsilon$ can be written as follows = [ ]{} [1A\_0]{}( [1]{} A\_2 - A\_0) , \[EPS\] where $\omega = \mbox{Re}A_2/ \mbox{Re}A_0=1/22.2$ is the inverse of the CP–conserving $\Delta I=1/2$ ratio. The CP–conserving quantities $\omega$ and $\mbox{Re}A_0$ are taken from the experiment, while the CP–violating terms $\mbox{Im} A_0$ and $\mbox{Im} A_2$ are calculated theoretically. The indirect CP violation parameter $\varepsilon$ can be either taken from the experiment or evaluated theoretically [@FUTURE_PPS]. Each contribution to $\mbox{Im} A_0$ and $\mbox{Im} A_2$ will now contain the FSI effect through the corresponding dispersive factor $\Re_I(M_K^2,0)$ with $I=0$ or 2 which was not taken into account in the previous Standard Model predictions of $\varepsilon^\prime /\varepsilon$. The result [@FUTURE_PPS] .(’/)|\_[SM]{} = (155)10\^[-4]{} , has an uncertainty which is dominated by the large–$N_c$ based matching procedure. The prediction is within one sigma respect to the present experimental world average [@Ceccuchi] .(’/)|\_[exp]{} = (19.32.4)10\^[-4]{} , and provides an enhancement of about a factor of two respect to the previous short–distance based Standard Model predictions of $\varepsilon^\prime /\varepsilon$ [@Buras; @Lattice].
Acknowledgments {#acknowledgments .unnumbered}
===============
I warmly thank my collaborators Antonio Pich and Ignazio Scimemi and the organizers of the conference. This work has been supported by the Ministerio de Educación y Cultura (Spain) and in part by the European Union TMR Network EURODAPHNE (Contract No. ERBFMX-CT98-0169).
References {#references .unnumbered}
==========
[99]{}
E. Pallante and A. Pich, .
E. Pallante and A. Pich, in preparation.
E. Pallante, A. Pich and I. Scimemi, in preparation.
T.N. Truong, ;\
N. Isgur [*et al.*]{}, ; ;\
M. Neubert and B. Stech, .
J. Kambor [*et al.*]{}, ; ; .
J. Gasser and U-G. Meissner, .
J. Bijnens, E. Pallante and J. Prades, .
E. Pallante, JHEP 01, 012 (1999).
A.J. Buras, hep-ph/9908395; S. Bosch [*et al.*]{}, hep-ph/9904408; G. Buchalla [*et al.*]{}, [*Rev. Mod. Phys.*]{} [**68**]{}, 1125 (1996); A.J. Buras [*et al.*]{}, , [**B**]{}400, 37, 75 (1993), [**B**]{}370, 69 (1992).
M. Ciuchini [*et al.*]{}, hep-ph/9910237; , [**B**]{}415, 403 (1994); .
H. Burkhardt [*et al.*]{} (NA31), ; G.D. Barr [*et al.*]{} (NA31), ; L.K. Gibbons [*et al.*]{} (E731), ; A. Alavi-Harati [*et al.*]{} (KTeV), ; V. Fanti [*et al.*]{} (NA48), hep-ex/9909022.
S. Bertolini, hep-ph/9908268; S. Bertolini [*et al.*]{}, [*Rev. Mod. Phys.*]{} 72, 65 (2000); hep-ph/0002234; [*Nucl. Phys.*]{} [**B**]{}449, 197 (1995), [**B**]{}476, 225 (1996), [**B**]{}514, 63, 93 (1998); V. Antonelli [*et al.*]{}, [*Nucl. Phys.*]{} [**B**]{}469, 143, 181 (1996); M. Fabbrichesi, hep-ph/0002235.
T. Hambye [*et al.*]{}, hep-ph/9908232; hep-ph/9906434; [*Eur. Phys. J.*]{} [**C**]{}10, 271 (1999); .
A. Ceccuchi, [*New measurement of direct CP violation in two pion decays of neutral kaons by experiment NA48 at CERN*]{}, CERN Seminar (February 29, 2000), http://na48.web.cern.ch/NA48/Welcome.html.
K.M. Watson, [*Phys. Rev.*]{} 95, 228 (1955).
R. Omnès, [*Nuovo Cimento*]{} 8, 316 (1958);\
N.I. Muskhelishvili, [*Singular Integral Equations*]{}, Noordhoof, Groningen, 1953;\
F. Guerrero and A. Pich, .
J. Bijnens, G. Colangelo and P. Talavera, JHEP 05, 014 (1998).
J. Gasser and H. Leutwyler, [*Ann. Phys.*]{} NY 158, 142 (1984).
J. Gasser and H. Leutwyler, [*Nucl. Phys.*]{} [**B**]{}250, 456; 517; 539 (1985).
G. Ecker, [*Prog. Part. Nucl. Phys.*]{} 35, 1 (1995);\
A. Pich, [*Rep. Prog. Phys.*]{} 58, 563 (1995).
A.J. Buras [*et al.*]{}, hep-ph/0002116v2.
C. Bernard [*et al.*]{}, ;\
N. Cabibbo, ;\
M. Gell-Mann, .
J.F. Donoghue [*et al.*]{}, ; A.J. Buras and J.-M. Gérard, ; H.-Y. Cheng, ; M. Lusignoli, .
U.-G. Meissner, in Proc. $5^{\rm th}$ Int. Workshop on [*Perspectives in Nuclear Physics at Intermediate Energies*]{} (Trieste, May 1991), p. 385.
M. Neubert and B. Stech, .
J.M. Frere [*et al.*]{}, .
A. Schenk, .
[^1]: Contribution to the XXXVth Rencontres de Moriond on Electroweak Interactions and Unified Theories, March 11–18, 2000.
[^2]: These values update earlier estimates of $\Re_2^{(1)}(M_K^2,M_\pi^2)$ [@ME:91; @NS:91]. The correction factors $\Re_I^{(1)}$ were also considered in an estimate of $\varepsilon'/\varepsilon$ within the $SU(2)\otimes SU(2)\otimes U(1)$ model of CP violation [@FRERE].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'It is argued considering multiple production in the deep inelastic scattering kinematics that the very high multiplicity events are extremely sensitive to the low-x partons density.'
---
=-0.4in=1.6mm=15.7cm=22.5cm
¶ ł § øØ
[**Soft partons production in deep inelastic kinematics**]{}
0.3cm [*J.Manjavidze, A.Sissakian*]{}\
[large JINR, Dubna]{}
Introduction
============
The role of soft color partons in the high energy hadron interactions is the mostly intriguing modern problem of particles physics. So, the collective phenomena and symmetry breaking in the non-Abelian gauge theories, confinement of colored charges and the infrared divergences of the pQCD are the phenomena just of the soft color particles domain.
It seems natural that the very high multiplicity (VHM) hadron interaction, where the energy of created particles is small, should be sensitive to the soft color particles densities. Indeed, the aim of this paper is to show that even in the hard by definition deep inelastic scattering (DIS) the soft color particles role becomes important in the VHM region.
The standard (mostly popular) hadron theory considers perturbative QCD (pQCD) at small distances (in the scale of $\La\simeq 0.2~Gev$) as the exact theory. This statement is confirmed by a number of experiments, e.g. the DIS data, the QCD jets observation. But one should have in mind that the pQCD has finite range of validity since the non-perturbative effects should be taken into account at distances larger then $1/\La$.
It is natural to assume building the complete theory that at large distances the non-perturbative effects exceed the perturbative ones. In result the pQCD contributions becomes $screened$ by the non-perturbative interactions.
But exist another, more natural, possibility. It should be noted here that the pQCD running coupling constant $\a_s(q^2)=1/b\ln(q^2/\La^2)$ becomes infinite at $q^2=\La^2$ and we do not know what happens with pQCD if $q^2<\La^2$. There is a suspicion that at $q^2\sim \La^2$ the properties of theory changed so drastically (being defined on new vacuum) that even the notions of pQCD is $disappeared$. This means that pQCD should be truncated from below on the ‘fundamental’ scale $\La$. It seems evident that this infrared cut-off should influence on the soft hadrons emission in the VHM domain.
So, it is important try to rise predictability of pQCD in the ‘forbidden’ area of large distances. For this purpose one should split experimentally the perturbative and non-perturbative effects at the large distances to check out quantitatively the role of soft color partons. One must realize for this purpose highly unusual condition that the non-perturbative effects must be negligible even if the distance among color charges is high.
The non-perturbative effects lead to the strong polarization of QCD vacuum and, in result, to the color charge confinement. This vacuum is unstable against creation of real quark-antiquark $(q\bar q)$ pairs if the distance among charges became large. This pairs are captured into the colorless hadrons and just emission of this ‘vacuum’ hadrons is the mostly important non-perturbative effect. Therefore, generally, the number of hadrons $n$, produced even in the hard DIS process, $n\neq n_{p}$, where $n_{p}$ is the number of $\bar{q}q$ pairs created ‘perturbatively’.
Notice, if the kinetic energy of colored partons is small, i.e. is comparable with hadron masses, creation of ‘vacuum’ hadrons should be negligible. Just this is the VHM process kinematics: because of the energy-momentum conversation law, produced (final-state) partons can not have high relative momentum and, if they was created at small distances, the production of ‘vacuum’ hadrons will be negligible (or did not play important role). Therefore, if the ‘vacuum’ channel is negligible, the pQCD contributions only should be considered .
The aim of this article is to show that $n\to\infty$ unavoidably leads to ‘low-$x$’ domain.
DIS kinematics in the VHM domain
================================
To describe the hadron production in pQCD terms the parton-hadron duality is assumed. This means that the multiplicity, momentum etc. distributions of hadron and colored partons are the same. This reduce the problem practically on the level of QED.
Let us consider now $n$ particles (gluons) creation the DIS. We would like to calculate $D_{ab}(x,q^2;n)$, where \_n D\_[ab]{}(x,q\^2;n)= D\_[ab]{}(x,q\^2). ł[1]{}As usual, let $D_{ab}(x,q^2)$ be the probability to find parton $b$ with virtuality $q^2<0$ in the parton $a$ of $\sim\la$ virtuality, $\la>>\La$ and $\a_s(\la)<<1$. We always may chose $q^2$ and $x$ so that the leading logarithm approximation (LLA) will be acceptable. One should assume also that $(1/x)>>1$ to have the phase space, into which the particles are produced, sufficiently large.
Then $D_{ab}(x,q^2)$ is described by ladder diagrams. From qualitative point of view this means approximation of Markovian process of random walk over coordinate $\ln(1/x)$ and time is $\ln\ln|q^2|$. LLA means that the ‘mobility’ $\sim \ln(1/x)/\ln\ln\le|q^2\ri|$ should be large (1/x)>>|q\^2/\^2|. ł[2]{}But, on other hand , (1/x)<<|q\^2/\^2|. ł[3]{}
The leading, able to compensate smallness of $\a_s(\la)<<1$, contributions give integration over wide range $\la^2<<k_i^2<<-q^2$, where $k_i^2>0$ is the ‘mass’ of real, i.e. time-like, gluon. If the time needed to capture the parton into the hadron is $\sim(1/\La)$ then the gluon should decay if $k_i^2>>\la^2$. This leads to creation of (mini)jets. The mean multiplicity $\bar{n}_j$ in the QCD jets is high if the gluon ‘mass’ $|k|$ is high: $\ln\bar{n}_j \simeq
\sqrt{\ln(k^2/\la^2)}$.
Rising multiplicity may (i) rise number of (mini)jets $\nu$ and/or (ii) rise the mean value mass of (mini)jets $<|k_i|>$. We will see that the mechanism (ii) would be favorable. This is consequence of the Markovian character of considered process.
But rising mean value of gluon masses $|k_i|$ decrease the range of integrability over $k_i$, i.e. violate the condition () for fixed $x$. One can remain the LLA taking $x\to0$. But this may contradict to (), i.e. in any case the LLA becomes invalid in the VHM domain and the next to leading order corrections should be taken into account.
Noting that the LLA gives main contribution, that the rising multiplicity leads to the infrared domain, where the soft gluons creation becomes dominant.
Preliminary notes
=================
First of all, neglecting the vacuum effects, we introduce definite uncertainty to the formalism. It is reasonable to define the level of strictness of our computations. Let us introduce for this purpose the generating function $T_{ab}(x,q^2;z)$: D\_[ab]{}(x,q\^2;n)=T\_[ab]{}(x,q\^2;z). ł[4]{}Having large $n$ the integral may be calculated by the saddle point method. The smallest solution $z_c$ of the equation n=zT\_[ab]{}(x,q\^2;z) ł[5]{}defines the asymptotic over $n$ behavior: D\_[ab]{}(x,q\^2;n)e\^[-nz\_c(x,q\^2;n)]{}. ł[6]{}Using the statistical interpretation of $z_c$ as the fugacity it is natural to write: z\_c(x,q\^2;n)=. ł[7]{}Notice that the solution of eq.() $z_c(x,q^2;n)$ should be the increasing function of $n$. At first glance this follows from positivity of all $D_{ab}(x,q^2;n)$. But actually this assumes that $T_{ab}(x,q^2;z)$ is a regular function of z at $z=1$. This is a natural assumption considering just the pQCD predictions.
Therefore, D\_[ab]{}(x,q\^2;n)e\^[-C\_[ab]{}(x,q\^2;n)]{}. ł[8]{}This form of $D_{ab}(x,q^2;n)$ is useful since usually $C_{ab}(x,q^2;n)$ is the slowly varying function of $n$. So, for Poisson distribution $C_{ab}(x,q^2;n)\sim\ln n$. For KNO scaling we have $C_{ab}(x,q^2;n)=const.$ over $n$.
We would like to note, that neglecting effects of vacuum polarization we introduce into the exponent so high uncertainty assuming $n\simeq
n_p$ that it is reasonable perform the calculations with the exponential accuracy. So, we would calculate -|\_[ab]{}(x,q\^2;n)== C\_[ab]{}(x,q\^2;n)(1+O(1/n)) ł[9]{}
The $n$ dependence of $C_{ab}(x,q^2;n)$ defines the asymptotic behavior of $\bar\mu_{ab}(x,q^2;n)$ and calculation of its explicit form would be our aim.
Correlation functions
=====================
Considering particles creation in the DIS processes one should distinguish correlation of particles in the (mini)jets and the correlations between (mini)jets. We will describe the jet correlations. One should introduce the $\nu$ jets creation inclusive cross section $\P_\nu^{(r)_\nu}(k_1,k_2,....,k_\nu;q^2,x)$, where $k_i$, $i=1,2,...,n$ are the jets 4-momentum. Having $\P_\nu$ we can find the correlations function $N_\nu^{(r)}(k_1,k_2,....,k_\nu;q^2,x)$, where $(r)=r_1,...,r_\nu$ and $r_i=(q,g)$ defines the sort of created color particle. It is useful introduce the generating functional F\^[ab]{}(q\^2,x;w)=\_ndØ\_n(k)\_[i=1]{}\^n w\^[r\_i]{}(k\_i) |a\^[ab]{}\_n(k\_1,k\_2,....,k\_n;q\^2,x)|\^2, ł[10]{}where $a^{ab}_n$ is the amplitude, $d\O_n(k)$ is the phase space volume and $w^{r_i}(k_i)$ are the arbitrary functions. It is evident, F\^[ab]{}(q\^2,x;w).|\_[w=1]{}=D\^[ab]{}(q\^2,x). ł[11]{}The inclusive cross sections ¶\_\^[(r)]{}(k\_1,k\_2,....,k\_;q\^2,x)=\_[1=1]{}\^F\^[ab]{}(q\^2,x;w).|\_[w=1]{}. ł[12]{}The correlation function N\_\^[(r)]{}(k\_1,k\_2,....,k\_;q\^2,x)=\_[1=1]{}\^F\^[ab]{}(q\^2,x;w).|\_[w=1]{}. ł[13]{}We will use this definitions to find the partial structure functions $D^{ab}(q^2,x;n)$.
It will be useful to introduce the Laplace transform over variable $\ln(1/x)$: F\^[ab]{}(q\^2,x;w)=\_[j<0]{}()\^j f\^[ab]{}(q\^2,j;w) ł[14]{}
The expansion parameter of our problem $\a_s\ln(-q^2/\la^2)\sim1$. By this reason one should take into account all possible cuttings of the ladder diagrams. So, calculating $D^{ab}(q^2,x)$in the LLA all possible cuttings of sceleton ladder diagrams is defined by the factor : {\^[ab]{}\_[r]{}G\_[r]{}\_[r]{}\^[ab]{}}, ł[15]{}i.e. the cutting line may get not only through the exact Green function $G_{r}(k^2_i)$ but through the exact vertex functions $\Ga^{ab}_{r}(q_i,q_{i+1},k_i)$ also ($q_i^2,q_{i+1}^2$ are negative). We have in the LLA that $$\la^2<<-q^2_i<<-q_{i+1}^2<<-q^2$$ and $$x\leq x_{i+1}\leq x_i\leq 1.$$ Following to our approximation, see previous section, we would not distinguish the way as cut line go through the Born amplitude $$a^{ab}_{r})=\le\{(\Ga^{ab}_{r})^2G_{r}\ri\}.$$ We will simply associate $w^{r}\Im a^{ab}_{r})$ to each rung of the ladder.
Considering the asymptotics over $n$ the time-like partons virtuality $k_i\simeq -q^2_i/y_i$ should be maximal. Here $y_i$ is the fraction of the longitudinal momentum of the jet. Then, slightly limiting the jets phase space, k\_i\^2=|q\_[i+1]{}|\^2(1+O((1/x)/|q\_[i+1]{}|\^2)). ł[16]{}
In result, introducing useful in the LLA variable $\tau_i=\ln(q^2_i/\La^2)$, where $\a_s(q^2)=1/\b\tau$, $\b=(11N/3)-(2n_f/3)$, we can find following set of equations: f\_[ab]{}(q\^2,j;w)=\_[c,r]{}\^r\_[ac]{}(j)w\^r() f\_[ab]{}(q\^2,x;w), ł[17]{}where \^r\_[ac]{}(j)=\_[ac]{}(j)=\^1\_0x\^jP\_[ac]{}(x) ł[18]{}and $P_{ac}(x)$ is the regular kernel of the Bethe-Salpeter equation . At $w=1$ this equation is the ordinary one for $D^{ab}(q^2,x)$.
We will search the correlation functions from eq.() in terms of Laplace transform $n_{ab}^{(r)_\nu}(k_1,k_2,....,k_\nu;q^2,j)$. Let us write: f\_[ab]{}(q\^2,j;w)=d\_[ab]{}(q\^2,j){ \_\_[i=1]{}\^( (w\^[r\_i]{}(\_i)-1))n\_[ab]{}\^[(r)\_]{}(k\_1,k\_2,....,k\_;q\^2,j) } ł[19]{}Inserting () into () and expanding over $(w-1)$ we find the sequence of coupled equation.
Omitting the cumbersome calculations, we write in the LLA that \_[ab]{}\^[(r)\_]{}(\_1,\_2,....,\_;q\^2,j)= d\_[ac\_1]{}(j,\_1)\^[r\_1]{}\_[c\_1c\_2]{}(j)d\_[c\_2c\_3]{}(j,\_2)\^[r\_]{}\_[c\_c\_[+1]{}]{}(j)d\_[c\_[+1]{}b]{}(j,\_[+1]{}). ł[20]{}One should take into account the conservation laws: \_1\_2\_[+1]{}=, \_1<\_2<…<\_[+1]{}<. ł[21]{}Computing the Laplace transform of this expression we find $\P_{ab}^{(r)_\nu}(\tau_1,\tau_2,....,\tau_\nu;q^2,x)$.
The kernel $d_{ab}(j,\tau)$ was introduced in (). Let us write it in the form: d\_[ab]{}(j,)=\_[=]{}\^[\_(j)]{}, ł[22]{}where d\^\_[qq]{}=\_s-\_[gg]{}, d\^\_[qg]{}=\_s-\_[qq]{}, d\^\_[qg]{}=\_[gq]{}, d\^\_[gq]{}=\_[qg]{} ł[23]{}and \_={\_[qq]{}+\_[gg]{}+\^[1/2]{}}. ł[24]{}If $x<<1$, then $(j-1)<<1$ are essential. In this case , \_[gg]{}\~\_[gq]{}>>\_[qg]{}\~\_[qq]{}=O(1). ł[25]{}This means the gluon jets dominance and n\^g\_[gg]{}=\_[gg]{}+O(1). ł[26]{}One can find following estimation of the two-jet correlation function: n\_[ab]{}\^[r\_1r\_2]{}(\_1,\_2;,)= O([max]{}{(\_1/)\^[\_[gg]{}]{},(\_2/)\^[\_[gg]{}]{}, (\_1/\_2)\^[\_[gg]{}]{}}}. ł[27]{}This correlation function is small since in the LLA $\tau_1<\tau_2<\tau$. This means that the jet correlations becomes high iff the mass of correlated jets are comparable. But this condition shrinks the range of integration over $\tau$ and by this reason one may neglect the ‘short-range’ correlations among jets. Therefore, as follows from (), f\_[ab]{}(q\^2,j;w)=d\_[gg]{}(,j){ \_[gg]{}\^\_[\_0]{} w\^[g]{}(’)} ł[28]{}We will use this expression to find the multiplicity distribution in the DIS domain.
Generating function in the DIS kinematics
=========================================
To describe particles production one should replace $w^{r}\Im
a^{ab}_{r})$ on $w^{r}_n\Im a^{ab}_{r})$, where $w^{r}_n$ is the $probability$ of $n$ particles production, \_nw\^[r]{}\_n=1. ł[29]{}Having $\nu$ jets one should take into account the conservation condition $n_1+n_2+...+n_\nu=n$. By this reason the generating functions formalism is useful. In result one can find that if we take () w\^g=w\^g(,z), w\^g(,z).|\_[z=1]{}=1, ł[30]{}then $f_{ab}(q^2,j;w)$ defined by () is the generating functional of the multiplicity distribution in the ‘$j$ representation’. In this expression $w^g(\tau,z)$ is the generating function of the multiplicity distribution in the jet of mass $|k|=\la
e^{\tau/2}$.
In result, see (), F\^[ab]{}(q\^2,x;w)\_[j<0]{}([1]{}/[x]{})\^j e\^[\_[gg]{}ø(,z)]{} ł[31]{}where ø(,z)=\^\_[\_0]{}w\^[g]{}(’,z). ł[32]{}Noting the normalization condition (), ø(,z=1)=. ł[33]{}
The integral () may be calculated by steepest descent method. It is not hard to see that jj\_c=1+{4Nø(,z)/(1/x)}\^[1/2]{} ł[34]{}is essential. Notice that $j-1<<1$ should be essential we find, instead of the constraint (), that ø(,z)<<(1/x). ł[35]{}In the frame of this constraint, F\^[ab]{}(q\^2,x;w){4}. ł[36]{}
Generally speaking, exist such values of $z$ that $j_c-1\sim1$. This is possible if $\o(\tau,z)$ is a regular function of $z$ at $z=1$. Then $z_c$ should be the increasing function of $n$ and consequently $\o(\tau,z_c)$ would be the increasing function of $n$. Therefore, one may expect that in the VHM domain $j_c-1\sim1$.
Then $j\simeq 1+\o(\tau,z)/\ln(1/x)$ would be essential in the integral (). This leads to following estimation: $$F^{ab}(q^2,x;w)\propto e^{-\o(\tau,z)}.$$ But this is impossible since $F^{ab}(q^2,x;w)$ should be the increasing function of $z$. This shows that the estimation () has a finite range of validity.
Solution of this problem with unitarity is evident. One should take into account correlations among jets considering the expansion (). Indeed, smallness of $n_{ab}^{(r)_\nu}$ may be compensated by large values of $\prod_i^\nu w^{r_i}(\tau_i,z)$ in the VHM domain.
Conclusion
==========
We can conclude that our LLA is applicable in the VHM domain till ø(,z)<<(1/x)<<=(-q\^2/). ł[37]{}The mean multiplicity of gluons created in the DIS kinematics |[n]{}\_g(,x)=F\^[ab]{}(q\^2,x;w).|\_[z=1]{} =ø\_1()>>ø\_1(), ł[38]{}where ø\_1()=\^\_[\_0]{}|[n]{}\_j() ł[39]{}and the mean gluon multiplicity in the jet $\bar{n}_j(\tau)$ has following estimation : |[n]{}\_j() ł[40]{}Inserting () into (), $$\o_1(\tau)=\bar{n}_j(\tau)/\sqrt{\tau}.$$ Therefore, noting (), |[n]{}\_g(,x)|[n]{}\_j() <<|[n]{}\_j(). ł[41]{}This means that the considered ‘t-channel’ ladder is important in the narrow domain of multiplicities n\~|[n]{}\_g<<|[n]{}\_j. ł[42]{}
So, in the VHM domain $n>>\bar{n}_g$ one should consider\
(i) The ladder diagrams with small number of rungs;\
(ii) To take into account the malty-jet correlations assuming that increasing multiplicity leads to the increasing number of rungs in the ladder diagram.\
To choose one of this possibilities one should consider the structure of $\o(\tau,z)$ much more carefully. This will be done in the consequent paper.
We can conclude that the VHM domain multiplicities production unavoidably destroy the ladder LLA. To conserve this leading approximation one should choose $x\to0$ and, in result, to get to the multi-ladder diagrams, since in this case $\a_s\ln (-q^2/\la^2)\sim
1$ and $\a_s\ln(1/x)\sim 1$. Such theory was considered in .
0.3cm [**Acknowledgments**]{} 0.2cm
We are grateful to V.G.Kadyshevski for interest to discussed in the paper questions. We would like to note with gratitude that the conversations with E.Kuraev was always interesting and important.
[99]{}
S.Mandelstam, [*Phys. Rep., **67***]{}, 109 (1980)
C.Callan, R.Dashen and D.Gross, [*Phys. Rev., **D19***]{}, 1826 (1979)
A.Casher, H.Neuberger and S.Nussinov, [*Phys. Rev., **D20***]{}, 179 (1979)
J.Manjavidze and A.Sissakian, [*JINR Rap. Comm., *****]{}
Yu.L.Dokshitser, D.I.Dyakonov and S.I.Troyan, [*Phys. Rep., **58***]{}, 211 (1980)
W.Furmansky, R.Petroncio and S.Pokorski, [*Nucl. Phys., **B155***]{}, 253 (1979)
L.V.Gribov, E.M.Levin and M.G.Ryskin, [*Phys. Rep., **100***]{}, 1 (1983)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The Bohm motion for a particle moving on the line in a quantum state that is a superposition of $n+1$ energy eigenstates is quasiperiodic with $n$ frequencies.'
address: 'Department of Mathematics, Rutgers University, Piscataway, NJ 08854-8019'
author:
- 'Sheldon Goldstein[^1]'
date: 'January 5, 1999'
title: Absence of Chaos in Bohmian Dynamics
---
In a recent paper [@bonfim], O. F. de Alcantara Bonfim, J. Florencio, and F. C. Sá Barreto claim to have found numerical evidence of chaos in the motion of a Bohmian quantum particle in a double square-well potential, for a wave function that is a superposition of five energy eigenstates. But according to the result proven here, chaos for this motion is impossible. We prove in fact that for a particle on the line in a superposition of $n+1$ energy eigenstates, the Bohm motion $x(t)$ is always quasiperiodic, with (at most) $n$ frequencies. This means that there is a function $F(y_1, \dots, y_n)$ of period 2$\pi$ in each of its variables and $n$ frequencies $\omega_1, \dots, \omega_n$ such that $x(t)= F(\omega_1 t,
\dots, \omega_n t).$
The Bohm motion for a quantum particle of mass $m$ with wave function $\psi=\psi(x,t)$, a solution to Schrödinger’s equation, is defined by $$\label{be}
dx/dt= (\hbar/m) \mbox{Im}\nabla \psi/\psi.$$ The right hand side of (\[be\]) depends upon $\psi$ only through its associated ray. In particular, if the wave function $$\label{sup}
\psi(x,t) = \Sigma_{i=0}^n a_ie^{-iE_i t/\hbar}\phi_i(x)$$ is a superposition of $n+1$ energy eigenstates $\phi_i$, then the right hand side of (\[be\]) is, in its dependence upon $t$, quasiperiodic with $n$ frequencies, as is $|\psi|$.
The quasiperiodicity in time of the vector field defining a dynamical system in general does not imply any corresponding property of the motion, since an autonomous system (one defined by a time independent vector field) can be chaotic. (In fact, it is autonomous systems that are normally studied in chaos theory.) However, for the Bohm motion on the line, the position of the particle is anchored in the (normalized) wave function, in such a way that its motion $x(t)$ inherits the quasiperiodicity of $|\psi|$:
A crucial feature of the motion (\[be\]) is the equivariance of $|\psi|^2$, i.e., the fact that probabilities for configurations given by $|\psi(x,t)|^2$ are consistent with the dynamics (\[be\]). This is a completely general feature of the Bohmian dynamics, valid in any dimension for any wave function satisfying Schrödinger’s equation. For a single particle moving on the line, it has the following important consequence: $$\label{nd}
\int_{-\infty}^{x(t)}|\psi(x',t)|^2\,dx'=\int_{-\infty}^{x(0)}|\psi(x',0)|^2\,dx',$$ which follows from equivariance since in one-dimension the dynamics is order-preserving, and in particular the evolution from time 0 to time $t$ carries the interval $(-\infty, x(0))$ to $(-\infty, x(t))$.
Given $\psi(x,0)$ and $x(0)$, equation (\[nd\]) determines $x(t)$ as a functional of $|\psi(x,t)|^2$, and thus $x(t)$, like $|\psi(x,t)|^2$, is quasiperiodic with $n$ frequencies. In fact $x(t)= F(\omega_1 t, \dots,
\omega_n t)$ with $\omega_i=(E_i-E_0)/\hbar$ for $i=1,\dots,n$ and $F(y_1,
\dots, y_n)=G(\int_{-\infty}^{x(0)}|\psi(x',0)|^2\,dx')$ where $G$ is the inverse of the function $H(x)=\int_{-\infty}^{x}|\psi(x')|^2\,dx'$ with $$\psi(x)\equiv\psi_{y_1,\dots,y_n}(x)= a_0\phi_0(x)+\Sigma_{i=1}^n a_ie^{-iy_i}\phi_i(x).$$
For the one-dimensional motion $x(t)$ the Lyapunov exponent $\lambda$ is given by $$\lambda=\lim_{t\to\infty}t^{-1}\ln {dx(t)\over dx(0)}.$$ It presumably follows from the quasiperiodicity of $x(t)$ alone that $dx(t)/dx(0)$ is similarly quasiperiodic. In any case, we have by equivariance that $|\psi(x(t),t)|^2dx(t)=|\psi(x(0),0)|^2dx(0)$, so that $$\label{dx}
dx(t)/dx(0)={|\psi(x(0),0)|^2\over |\psi_{y_1,\dots,y_n}(F(y_1,\dots,y_n))|^2}$$ with $y_i=\omega_i t$. Hence $dx(t)/dx(0)$ is quasiperiodic with $n$ frequencies and thus $\lambda=0$.
Remarks: (i) In one-dimension we always have that $dx(t)/dx(0)=|\psi(x(0),0)|^2/|\psi(x(t),t)|^2$. Thus the vanishing of the Lyapunov exponent $\lambda$ is more general than described here, and should be valid for any wave function, on the circle as well as the line. After all, for bound states the ratio on the right is not likely to grow or decrease in any systematic way at all, while for states with continuous spectrum the behavior will be at most power law; in no case will there be exponential growth or decay. (ii) Another aspect of chaos, the weak convergence of densities to the “equilibrium” distribution (for Bohmian mechanics given by $|\psi(x(t),t)|^2$) will, as a simple consequence of the order preserving character of such motions, almost always fail for any one-dimensional flow, Bohmian or otherwise. The sole exception can occur only when the asymptotic “equilibrium” distribution is concentrated on a single (perhaps moving) point, something that is impossible for Bohmian mechanics.
I am grateful to Michael Kiessling for helpful suggestions. This work was supported in part by NSF Grant No. DMS-9504556.
O. F. de Alcantara Bonfim, J. Florencio, and F. C. Sá Barreto, Phys. Rev. E [**58**]{}, 6851 (1998).
[^1]: Electronic address: [email protected]
| {
"pile_set_name": "ArXiv"
} |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.